Issue 36 · Aug 31 – Sep 6, 2026
Every candidate
All 2,798 papers were scored from their abstracts by gpt-5.6-luna; 1,159 were not skipped. Shown here: the top 150 of those, in score order. Picks are marked.
-
Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning
The paper finds that KV-cache eviction for long-chain reasoning can be done by keeping the prompt and randomly evicting other tokens, rather than ranking tokens with an importance score. Across four models and six reasoning tasks, this matches the best prior eviction methods while improving vLLM throughput by 32–43%; the authors argue that reasoning traces are redundant both within the text and across attention heads, whereas the prompt is the cache component that must be protected.
It challenges the core assumption behind KV-cache compression—that sophisticated token-importance scores are necessary—and reports a substantial practical speedup with multi-model, multi-task evidence.
-
Post-Training Language Models for Gold-Medal Performance in Coding Competitions
The paper builds competition-specialized language models using curated problems, synthetic reasoning traces, supervised fine-tuning, reinforcement learning, and iterative test-time solution refinement. On IOI 2025, the smaller model exceeded the gold threshold with refinement, while a prospective IOI 2026 evaluation reports a 550B-parameter system scoring 535.4/600—above both the gold threshold and the top human contestant under comparable constraints.
The prospective result of an AI system surpassing the highest-scoring human at IOI is a major capability milestone, although the abstract alone leaves important details about reproducibility and evaluation conditions unverified.
-
Constant regret in general games via higher-order optimism
The paper proposes HOOD, an uncoupled learning algorithm for repeated play in arbitrary finite-player games. It claims that when every player uses it, each player’s regret remains bounded by O(N^3 log^2 K), independent of the number of rounds, using high-order prediction, discounting, and a lifted entropic regularizer to suppress oscillations. This would substantially improve on the usual regret bounds, though the abstract provides no experimental or detailed theorem evidence beyond the stated result.
A uniform constant-regret guarantee in arbitrary games would be a major theoretical advance and overturn the expectation that regret must grow with the horizon; the claim is highly consequential, although its credibility must be checked in the full proof.
-
Restricted Eigenvalues Beyond Gaussian Width: Threshold Occupancy under Heavy Tails
This paper gives a constructive counterexample to the proposed claim that Gaussian-width sample complexity guarantees restricted eigenvalues under only a uniform small-ball condition. It identifies simultaneous threshold occupancy as the missing obstruction: even a constant-width polyhedral cone can have zero empirical RE for many samples under an isotropic, finite-moment, smooth heavy-tailed design, while Gaussian measurements succeed with only logarithmically many samples. It also derives a sharp VC-dimension-dependent worst-case sample complexity and a dimension-based fallback bound.
It appears to resolve a longstanding open question with an explicit, robust counterexample that overturns a widely expected Gaussian-width law and explains the mechanism behind the failure.
-
Constant Individual Regret in General Games
The paper presents ECHO-OFTRL, a deterministic and fully uncoupled learning rule for finite multiplayer normal-form games. It claims that, with full-information feedback, every player’s cumulative individual regret remains bounded by a quantity depending only polynomially on the number of players and logarithmically on the largest action set—not on the time horizon. The key idea is a cascade of exponential moving averages that supplies higher-order optimism, inspired by filter design.
A horizon-independent regret guarantee for every finite general game would remove a longstanding polylogarithmic dependence and uses a genuinely unusual optimism mechanism, though the abstract provides no constants or detailed theorem conditions.
-
Extremely Sparse Supervision Incentivizes Reasoning Ability
The paper studies on-policy distillation and finds that training on only one or two selected tokens per reasoning trajectory—about 0.05% of generated tokens—can improve reasoning as much as, or better than, supervising every token. The effect appears across multiple Qwen3 teacher–student scales and is also reported for coding, Llama models, and PPO-based RL with verifiable rewards. If robust, this suggests that post-training may need feedback on only a few critical reasoning points rather than dense token-level correction.
The claimed 2,000-fold reduction in supervised tokens while retaining reasoning gains is a genuinely counterintuitive training result, with replication across model families and learning setups, though the abstract lacks quantitative details needed to fully validate its magnitude.
-
The Head Complexity of Boolean Functions in Single-Layer Attention
This paper characterizes how many attention heads a single-layer attention-only model needs to compute Boolean functions. It proves that exactly k heads can compute k-bit parity but not (k+1)-bit parity, even with unlimited embedding dimension and numerical precision, and gives analogous lower bounds for induction-head tasks. It also shows that almost all n-bit Boolean functions require nearly 2^n heads, while 2^n heads always suffice, largely pinning down the model’s computational capacity.
The unconditional separation of head count from dimension and precision is a genuinely informative result about what single-layer attention can compute, with broad parity and counting arguments rather than a benchmark-specific improvement.
-
SMELT: Scaling Laws for Compute-Matched MoE Looped Transformers
This paper studies whether repeatedly reusing Transformer layers helps when looping is compared fairly at matched compute, parameter count, and KV-cache size. Its SMELT design loops the middle layers of a sparse MoE Transformer and reports 6.8–18% lower training compute on the compute-optimal frontier across models up to 54B parameters, with especially strong gains on code and long-context/in-context learning; analysis links this to reduced attention sinks.
The budget-matched scaling-law study and mechanistic explanation make a credible case that depth reuse provides real efficiency gains rather than merely more computation, but the abstract lacks enough detailed comparisons and independent validation for a strong verdict.
-
Evidence Integration in Large Language Models
This paper develops a distributional theory of how LLMs incorporate external candidate answers after beginning to form their own answer, and tests it across 10 million trials, 12 models, and eight domains. It reports that models preferentially accept evidence resembling their own likely errors, that the same evidence can help weaker models but hurt stronger ones, and that answer integration can proceed even when the model internally represents the evidence as invalid. Causal and representation analyses suggest candidate integration occurs late and is partly dissociated from verbalized verification.
The receiver-relative account of evidence integration, especially the finding that verifiable-invalid candidates can still control answers and that verification representations have little causal influence, challenges simple trust-based explanations of RAG and is supported by unusually broad experiments and causal analyses.
-
Does On-Policy Distillation Really Distill? From Noisy Teacher to Self-Improvement
The paper argues that on-policy distillation gains do not mainly come from accurately copying a teacher. Instead, training appears to suppress low-probability (“tail”) tokens, and a simple fixed negative learning signal can match teacher-based supervision. Based on this analysis, the authors propose OPSA, a teacher-free, entropy-adaptive method that reportedly gives large gains on mathematical reasoning benchmarks, including a 35.41-point Avg@32 improvement on AIME24 for Qwen3-1.7B.
The potentially important contribution is the claim that OPD is largely teacher-free tail-token suppression, paired with a simple self-adaptation method, but the abstract does not provide enough experimental detail to establish that the large benchmark gains generalize beyond the reported setup.
-
Unlocking Lossless Speedups in LLMs via Discrete Diffusion
The paper adds lightweight diffusion modules to standard autoregressive LLMs so they can generate multiple tokens in parallel while preserving the original autoregressive distribution. Its proposed Ψ-Spec samplers reportedly deliver up to 3× higher throughput without a separate draft model, and an 8B model beats larger diffusion-based models on tool use, coding, and long-context reasoning benchmarks.
The combination of distribution-preserving parallel diffusion decoding and no separate speculative model is a genuinely interesting efficiency direction, but the abstract provides too few experimental details to establish that the claimed losslessness and broad throughput advantage hold generally.
-
Lucida: Parse, Generate, and Place for Composable Real-to-Sim Scene Modeling
Lucida builds editable real-to-simulation indoor scenes from video by maintaining a scene graph with multi-view evidence for each object, generating complete assets, and placing them through a closed-loop VLM policy that manipulates object gizmos until alignment is sufficient. It reports large improvements over prior systems in detection, pose estimation, and scene reconstruction, including scene F-Score rising from 0.794 to 0.924.
The staged redistribution of uncertainty and the closed-loop GUI-based placement policy are interesting, with unusually large reported gains, but the abstract gives limited detail about benchmark difficulty, ablations, and how much of the improvement comes from the pipeline versus stronger components.
-
Driving on Memory
The authors test whether autonomous-driving models actually need the current camera view to score well by replacing it with retrieved memories from previous drives at the same location. On NAVSIM, memory alone nearly matches or exceeds leading end-to-end systems, while performance drops much more on Bench2Drive and RealEngine, suggesting NAVSIM can be solved largely from static map-like context rather than reacting to current traffic.
This is a striking, experimentally supported challenge to the interpretation of a widely used driving benchmark: high NAVSIM scores may not demonstrate perception of the evaluated scene or responsiveness to dynamic traffic.
-
Cheap Verifiers, Large Blind Spots: Measuring the Reliability Cost of Cost-Saving Cascades
The paper studies cascades that use a cheap student model and a stronger verifier, then fine-tune the student on cases rejected by the verifier. Across real LLMs, it finds that verifier blind spots grow as the student improves, corrective fine-tuning can destabilize or collapse the student, and verifier-based dashboards can report roughly constant low error while true delivered error rises substantially. It also proposes a theory explaining how all in-loop metrics can improve without improving actual quality.
This presents a potentially assumption-breaking failure mode for cost-saving and self-improving inference systems, backed by real-model experiments and a mechanistic theory rather than just a benchmark result.
-
SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models
SolarWM presents an open end-to-end infrastructure for training interactive video world models. It standardizes 1.43 million clips from 10 datasets, supports multiple native video backbones, and combines bidirectional adaptation, autoregressive initialization, and distillation to produce causal models that reportedly support real-time rollouts lasting minutes to hours despite training on 5-second clips. The abstract emphasizes reproducibility and extensibility, but gives no quantitative capability or comparison results.
The unified data contract and backbone-agnostic training framework, together with the claim of very long real-time rollouts from short training sequences, could be important for world-model research, but the abstract provides too little empirical evidence to justify a strong recommendation.
-
Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills
The paper treats practical ML know-how in code repositories as reusable “skills” rather than expecting an agent to rediscover it during each task. Its DisCo system distills repository knowledge into compact, verified skills—both general-purpose and task-specific—and reports large gains over an otherwise identical agent on MLE-bench, PaperBench, FrontierCS, and PassNet, including a 134.3% improvement on MLE-bench.
The repository-to-skill distillation framing and broad claimed gains are substantial, but the abstract gives limited detail about skill verification, evaluation controls, and whether the results generalize beyond one agent/backbone and benchmark setup.
-
Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments
Terminal-Universe reconstructs executable coding environments from recorded terminal-agent trajectories rather than treating each trajectory as a fixed demonstration. It restores partial workspaces, fills in missing files and dependencies, and generates original, cross-project, and multi-round tasks; 37.3k environments reportedly improve Qwen3.5-27B by 11.9 points on Terminal-Bench 2.1 and 13.8 points on EvoCode-Bench v2 MT@4.
The trajectory-to-reusable-environment idea and the reported gains are substantial, but the abstract does not establish how much comes from environment reconstruction versus scale, filtering, or standard synthetic-task generation.
-
ZimaBlue: Evolving Generalizable World Action Models through Scalable Video Pre-training
ZimaBlue trains robot-control models from large amounts of action-free egocentric human and robot video, then connects those representations to robot trajectories and specializes them for a target platform. Its Slow-Fast architecture reportedly enables 30 Hz control, while scaling to more than 120,000 hours of video raises real-robot zero-shot success from 36.1% to 77.8%, especially on unseen tasks.
The combination of large-scale video pretraining, unified video-action grounding, and real-time dual-speed control could be a meaningful route to general-purpose manipulation, but the abstract lacks enough benchmark, baseline, and task detail to justify a strong verdict.
-
Scal3R: Learning Efficient Multi-Relative Pose Query for Scalable Online 3D Reconstruction
Scal3R addresses long-video 3D reconstruction by separating stable per-frame depth from failing global pose estimation. It adds lightweight learned tokens to a frozen backbone to query poses relative to multiple past keyframes, then uses online pose-graph optimization and loop closure to control drift. The method reportedly cuts average KITTI trajectory error by over 60% versus an online baseline and performs strongly across several reconstruction datasets.
The multi-reference pose-query formulation targets a real long-horizon failure mode and reports a substantial error reduction across multiple datasets, but the abstract does not establish how much comes from the learned component versus standard keyframe pose-graph optimization.
-
The Missing Temporal Link: Temporal Context Routing for Script-Driven Audio-Video Generation
The paper adds Temporal Context Routing, a mechanism that projects time annotations from a structured script onto the shared temporal axis used by joint audio-video generation. On 200 test scripts, it reportedly reduces shot-timing error from 1.11 seconds to 0.042 seconds and improves dialogue timing accuracy from 28.3% to 84.1%, without reducing audiovisual synchronization or visual quality.
The explicit alignment of script timing with both generated modalities is a meaningful and plausibly general direction, with unusually large reported gains, but the abstract provides limited detail about the baseline, scripts, and evaluation robustness, so it falls short of a strong recommendation.
-
MINERVA: How Small Can a Manipulation Policy Be and Still Solve LIBERO?
MINERVA shows that LIBERO can be solved with a deliberately small 0.54M-parameter visuomotor policy: it reaches 95.1% success, close to much larger VLA models, while running 113–1,400× faster on a laptop CPU. Broad ablations suggest that action-chunk length and vision capacity matter, while flow matching does not; a task-ID permutation test further indicates that standard LIBERO performance relies heavily on memorized task mappings, and robustness collapses under photometric shifts.
The combination of a roughly 7,700× parameter reduction, large practical speedup, and evidence that LIBERO rewards task memorization rather than broad instruction-conditioned manipulation challenges assumptions behind current VLA benchmark results.
-
TacPAC: Tactile Prediction and Real-Time Action Correction in World-Action Models for Contact-Rich Manipulation
TacPAC uses predicted tactile contact states from a planned action as a reference for interpreting tactile feedback during execution. A lightweight tactile module then corrects the remaining actions in the chunk instead of replanning from scratch, achieving a 20.7× lower correction cost. On five real-robot contact-rich manipulation tasks, it improves the vision-only model’s average success from 22% to 64%.
The paper presents a genuinely useful timing-aware integration of tactile prediction and feedback, with a large real-robot improvement and substantial computational savings across diverse manipulation tasks.
-
Microglia drive demyelination via multiple sclerosis antibodies and BTK signaling
The study shows that MS patient-derived antibodies can cause microglia to engulf and destroy intact myelin in living mouse cortex, rather than merely clearing pre-existing debris. The effect depends on antibody specificity, Fc/complement signaling, and BTK; similar microglial changes appear in human MS lesions, and BTK inhibition prevents the observed myelin loss.
It provides an unusually direct, multimodal case that microglia actively drive antibody-mediated demyelination—challenging the debris-scavenging interpretation—and links the mechanism to an actionable BTK intervention with both in vivo imaging and human-lesion support.
-
BeaconKV: Key-Value Cache Compression Guided by Beacon Queries for Efficient Large Reasoning Model Inference
BeaconKV compresses the KV cache during long chain-of-thought inference by storing representative “beacon” queries for clusters of queries that may later revisit distant reasoning context. Across four open-source reasoning models and several benchmarks, it reportedly reduces memory by up to 5.8× and improves throughput by over 4.3× while nearly preserving accuracy, addressing a failure mode of methods that use only recent queries to predict future attention.
The identification of clustered long-range revisit queries and its use for training-free KV compression is a meaningful, potentially broadly useful idea, but the abstract gives limited detail on accuracy tradeoffs and robustness beyond headline best-case numbers.
-
Rethinking On-Policy Distillation of Large Language Models II: One Training Example
The paper studies how much data on-policy distillation actually needs by training language models from rollouts of only one query. A single query reaches 71.5% of the states visited by full-data training and recovers much of its improvement; 16 semantically diverse queries reach 98.9% coverage and match full-data performance across several settings. The authors argue that OPD is mainly limited by the slow absorption of supervision, not by insufficient training data, and show similar behavior with multiple teachers and generic or off-domain prompts.
The data-minimal result and state-coverage framing are genuinely non-obvious and could change how OPD experiments are designed, but the abstract does not establish a major capability or efficiency breakthrough beyond a focused empirical finding.
-
LatentPress: Context Compression Beyond Text and Vision
LatentPress compresses conversations and documents into continuous memory tokens that a frozen language model can consume directly, avoiding text summaries or image/OCR reconstruction. With only a small adapter trained, it reports 4–16× compression, faster writing and reading, and near- or slightly better-than-raw-context QA at moderate compression, including some cross-domain transfer.
The direct machine-readable soft-memory interface and reported efficiency/cross-domain transfer are interesting, but latent context compression has substantial prior work and the abstract gives limited evidence beyond a few benchmarks and adapter settings.
-
ZipTok3D: High-Fidelity 3D Tokenization with Compact Token Prefixes
ZipTok3D trains a 3D tokenizer so that the first few latent tokens contain the most important geometry, using random prefix truncation during training and a shared iterative decoder. It reports reconstruction comparable to a 32-token baseline with only one token on ShapeNet and four on TRELLIS, substantially reducing sequence length, though the abstract does not establish downstream generation quality.
The prefix-organized representation and extreme reported compression are genuinely notable, but the evidence is limited to reconstruction claims on two benchmarks and does not yet show a comparable gain for full 3D generation or broader data.
-
Language Models Can Control Their Own Attention
The paper lets a language model emit control tokens declaring whether it needs global context, a focused region, or only recent output, allowing the inference engine to skip most KV-cache reads. Across 15 long-context tasks, this reduced attended tokens by 52.0% and 31.1% on two off-the-shelf models, with accuracy drops of 1.27 and 2.75 percentage points respectively; larger models narrowed the loss.
This is a genuinely interesting intrinsic alternative to proxy-based sparse attention with substantial reported reductions in context reads, but the abstract does not establish end-to-end latency or cost gains, and the zero-shot protocol still incurs notable accuracy losses.
-
WorldSculpt: Generating Compositional Worlds from Grounded Videos
WorldSculpt reconstructs densely cluttered scenes as collections of separate object meshes in a shared coordinate frame. Its key approach adapts a single-object 3D generative model with multi-view conditioning, yet trains only on canonical single-object data; the authors report that this generalizes to scenes containing hundreds of heavily occluded objects. They also introduce a photorealistic, object-annotated benchmark and show conversions from existing 3D Gaussian-splat worlds into compositional meshes.
The scene-level generalization from single-object training to highly cluttered, compositional worlds is a genuinely interesting direction, but the abstract provides no quantitative results or details sufficient to judge whether the claimed scalability is reliable.
-
Hi-Q: Hierarchical Evidence-guided Query Refinement for Multi-Hop Question Answering
Hi-Q treats multi-hop QA as discovering the right retrieval granularity rather than following a fixed decomposition or corpus graph. It recursively tests whether a query unit is supported by retrieved evidence and splits only unresolved units, reporting large gains over IRCoT and graph-based PropRAG across three benchmarks, especially with full-corpus distractors.
The evidence-conditioned, adaptive query-tree formulation and sizable full-corpus gains are genuinely interesting, but the core components build on established iterative retrieval and query decomposition ideas, and the abstract does not establish robustness beyond benchmark results.
-
DreamX-Creator: Democratizing Native Audio-Video Generation at 2K Resolution
DreamX-Creator is a 7B model that generates video and synchronized audio jointly rather than adding audio afterward. It combines late cross-modal attention, staged multimodal training and feedback, and a distilled one-step 2K refinement process; the abstract claims performance competitive with open-source systems but gives no quantitative results.
Native joint audio-video generation at 2K with a compact model is a relevant and potentially useful direction, but the abstract provides too little comparative evidence to establish a major advance.
-
On the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability
The paper presents Qwen3.8-Flash-Next, a 125B-parameter sparse MoE language model that activates only 6B parameters per token. It combines mostly Gated DeltaNet layers with occasional global or sparse attention, gated residual branches, host-memory n-gram embeddings, and Muon optimization; compared with a 397B predecessor, it reportedly uses one-third the activated parameters and tokens and about one-ninth the training FLOPs while achieving similar or better results on 14 benchmarks. The paper also argues that architecture, efficiency, and training stability must be optimized jointly, with several design choices changing the best learning-rate and batch-size regime.
The claimed near-parity with a much larger predecessor at roughly one-ninth the training compute, together with a fairly unusual combination of recurrent-style token mixing, sparse attention, gated residuals, and host-resident embeddings, merits attention, though the abstract lacks detailed benchmark numbers and enough evidence to justify a strong verdict.
-
Aspire: Can Models Self-Evolve from Vague Goals?
ASPIRE tests whether language-model agents can improve from only a vague capability goal, without being given the downstream tasks or evaluation metric. Agents must interpret the goal, select data and update methods, create their own training and validation signals, and decide when to stop; across six goals and 520 hidden expert-authored items, weight-level improvements were sparse and unstable, while harness evolution stayed below a strong engineered baseline. The main finding is that current agents can execute self-improvement loops but commonly optimize mismatched data or narrow self-evaluations, so local progress fails to transfer and can later be erased.
The vague-goal, hidden-evaluation formulation is a meaningful and relatively underexplored test of genuine self-directed learning, supported by broad experiments, but the current results mainly document failure modes rather than demonstrating a major new capability.
-
HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?
HarnessDev evaluates whether LLMs can build and iteratively improve the external software infrastructure—or “harness”—that runs agents, rather than merely producing task answers. Across six models, four domains, and 2,207 downstream instances, generated harnesses lag human-designed ones on coding and research but can match or exceed them for writing and ML experimentation; iterative improvement is unstable and transfers poorly across tasks and runtime models.
The paper introduces a useful and relatively underexplored evaluation target—agents developing their own execution scaffolding—with broad experiments that reveal strong domain and model-transfer limitations, though it is primarily a benchmark and diagnostic study rather than a demonstrated capability breakthrough.
-
Facet-0: A Robotic Foundation Model for Contact-Rich Precise Manipulation
Facet-0 is a multimodal robotic foundation model for precise assembly that predicts not only actions but the wrist forces those actions will cause. It combines force-history conditioning, flow-matching action generation, and an action-wrench critic with contact-focused RL and lightweight on-robot adaptation; on five sub-millimeter computer-assembly tasks it reports 82% success versus 15% for its strongest baseline, with 0.5 mm placement accuracy and 50 ms latency.
The explicit action-to-contact-outcome modeling and contact-selective post-training target a major weakness of vision-only robot policies, and the reported fivefold success improvement on real, force-synchronized assembly tasks is large enough to merit close inspection, though the abstract does not establish baseline quality or broader generality.
-
Why Gated DeltaNet Survives 4-Bit Quantization: NVFP4 W4A4 for the Recurrent Half of a Hybrid 27B LLM
The paper quantizes every linear layer, including the recurrent Gated DeltaNet layers and their gates, to NVFP4 W4A4 in a 27B hybrid LLM. It reports near-BF16 quality across language, reasoning, coding, and long-context retrieval tests, while reducing the checkpoint to 17.5 GiB and improving prefill speed; mechanistic experiments suggest the recurrence does not accumulate quantization noise because state updates overwrite and forget errors. It also identifies a serving issue caused by mismatched calibration scales when modules are fused into GEMMs.
This is a useful and somewhat counterintuitive result showing that the supposedly fragile recurrent component can be quantized as aggressively as the rest of a large hybrid LLM, backed by broad evaluations and a proposed mechanism, though it remains a recipe and analysis centered on one model family.
-
Scaffolding Foundation Models into Physical-World Agents Pushes the Frontier of Long-Horizon Navigation
NavMCP combines a vision-language model for persistent task-level reasoning with a navigation foundation model for reliable closed-loop movement. It adds intent, observation, and memory interfaces so the system can decompose long-horizon embodied questions into navigation calls without retraining either model; it reports a 14.9-point gain on HM-EQA and 78.3% success on a Unitree Go2, with larger gains on longer tasks.
The compelling contribution is a demonstrated way to turn bounded navigation models into persistent long-horizon agents, including real-robot results, although the core idea is scaffolding and the abstract does not establish broad robustness beyond the reported tasks.
-
ChessQueries: Toward Better Chess Board Recognition
The paper uses a ViT encoder with a DETR-style decoder to infer the piece arrangement from chess-board images. It reports a dramatic improvement on ChessReD, from 15.3% to 99.2% board-level accuracy, near-perfect performance on the existing benchmarks, and strong out-of-distribution results; it also introduces a harder dataset from broadcast tournament footage.
The architecture combines established components, but the reported jump from very poor benchmark performance to near-saturation, plus evidence on broadcast-style and other out-of-distribution images, is substantial enough to merit close inspection.
-
StudentSim: Training LLM-based Student Simulators
StudentSim trains a pooled LLM student model and then specializes it using sparse data from each individual, aiming to reproduce both their responses and how they change after tutor guidance. Across chess, English writing, and math, it reports higher behavioral fidelity and guidance responsiveness than general LLM role-play and prior student models, and uses the simulator as a reward model to train a more personalized chess tutor in a small human evaluation.
The individualized simulator-plus-tutor-RL setup is a meaningful direction with cross-domain results, but the evaluation is relatively small and the abstract does not establish that simulator fidelity reliably translates into better tutoring beyond the chess proof of concept.
-
Bilevel Coordinated Reflection: A Game-Theoretic Approach to Multi-Agent LLM Systems
The paper formalizes orchestrator–worker LLM systems as a bilevel coordination game and models reflection as movement through semantic memory states. It proves bounds on coordination and reflection, shows that transcript-only gates cannot uniformly improve reflection without environment-grounded evaluation, and proposes SRMA, which accepts memory updates only when grounded risk decreases. The empirical result is modest: 72.2% on 500 SWE-bench instances versus 70.8% for a public mini-SWE-agent reference.
The formal separation between transcript-only and environment-grounded reflection, together with tight convergence analyses, is a potentially useful conceptual contribution, but the practical evidence is weak and the benchmark gain is small.
-
Environment Evolution for Terminal Agents
The paper proposes evolving terminal-agent environments off-policy: instead of repeatedly generating tasks from current model failures, it gradually increases task difficulty along three directions and schedules those tasks during RL training. The generated environments become harder for several frontier models, and training Qwen models with them improves Terminal-Bench 2.1 performance by 14.4 and 18.0 percentage points.
The off-policy environment-evolution framing and sizable reported gains are notable, but the abstract gives limited evidence about baselines, generalization beyond one benchmark, and whether the gains come from the evolution method rather than additional task or training scale.
-
Compile by Training: Turning Natural-Language Specifications into Local Neural Functions
The paper proposes compiling a natural-language description into a small reusable neural function: teacher models generate examples, then an adapter is trained for a compact interpreter so future inputs run locally without the teachers. On a hard subset of FuzzyBench-Hard, it reports 83.6% semantic accuracy where a faster Program-as-Weights compiler got no exact matches, at the cost of about a minute of compilation; demos include website helpers, a 3D avatar, and an English–Claudish translator.
The compile-once, run-locally framing is a plausible and useful direction for turning natural-language specifications into deployable software-like neural functions, but the abstract gives limited comparative and experimental evidence beyond one benchmark subset and demonstrations.
-
Hakken: Predicting future discoveries to fill the gaps in today's knowledge
Hakken predicts previously undocumented scientific relationships from temporal knowledge graphs combined with LLM-derived semantic information, and provides explanations for its hypotheses. Applied to biomedicine, it generated 1.5 million aging-related predictions; biologists reviewed batches, and wet-lab work reportedly confirmed two new interactions involving TP53–BAMBI and RAF1–TNF.
The end-to-end goal of generating experimentally testable scientific discoveries is genuinely distinctive, and two reported confirmations are notable, but the abstract gives too little detail about validation quality, novelty relative to existing biomedical knowledge, and the causal role of the model to justify a strong verdict.
-
FlashRender: Few-Step Generative Rendering via Camera-Controlled Video MeanFlow
FlashRender generates a source video re-rendered along a specified camera path using only a few denoising steps. It combines representation alignment to encode the geometric transformation, MeanFlow training to reduce discretization error, and on-policy distillation to improve fixed-step rollouts; the authors report matching multi-step quality and geometric consistency at 25x lower sampling cost, with better control on unseen camera trajectories.
The combination of step-consistent camera conditioning and flow-based distillation addresses a meaningful bottleneck in controllable video generation, but the abstract provides no quantitative details or comparisons sufficient to establish that the claimed 25x gain is broadly reliable.
-
Harness-of-Harness: Multi-Day Autonomous Software Development with Continual Improvement
The paper proposes Harness-of-Harness, a framework that repeatedly plans, implements, tests, evaluates, and incrementally improves software produced by coding agents. Across three agent/model combinations and three benchmarks, it reports 52.25% average relative gains over standalone harnesses, plus a 70+ iteration deployment that produced a playable first-person-shooter game with visuals, audio, and a storyline.
The multi-loop harness-level approach and multi-day autonomous software construction are substantially more ambitious than ordinary coding-agent tweaks, but the abstract lacks absolute benchmark results, comparison details, and enough evidence to establish how much of the reported gain is robust.
-
Principia: Relational Physics Tests for Video Models
Principia evaluates video generators using calibration-independent relationships between paired objects rather than absolute motion measurements. Across eight Newtonian phenomena and thousands of generations from six leading models, all perform poorly on the relational physics score (below 0.42) despite high VBench scores, while vision-language models are also weak at identifying violations.
The relational, calibration-independent formulation exposes a substantial gap between conventional video quality scores and physical consistency, though it is primarily a benchmark paper and the abstract does not establish whether the score predicts real-world usefulness.
-
DRACO: Fine-Grained Credit Assignment with Dynamic Rubrics for Long-Horizon Agent Training
DRACO trains long-horizon agents without task verifiers by dynamically generating multi-criteria rubrics, scoring completed trajectories, and redistributing rubric-level feedback to the steps that contributed to each criterion. On AppWorld and Tau-Bench, it reports sizable gains over base models, sparse ground-truth reward training, and other rubric-based approaches, including some out-of-domain transfer.
The step-level credit assignment from dynamically evolving rubric judgments is a plausible and nontrivial alternative to sparse trajectory rewards, with encouraging cross-domain results, but the abstract does not establish how robust the gains are or whether rubric generation and attribution work beyond these agent benchmarks.
-
Exploring Collaboration between a language and a non-language agent
This paper studies how LLMs can collaborate with non-language agents such as chess engines, comparing textual summaries with direct injection of the agent’s continuous state into the LLM as learned latent state tokens. It reports that verbal communication creates a growing performance gap, while latent state internalization lets a 14B model match or exceed task specialists and larger frontier models across six collaborative chess tasks, including out-of-distribution tests.
The proposed latent interface for integrating non-language agents is a genuinely interesting direction and the claimed scaling and OOD results would be important, but the abstract gives no quantitative results or detail sufficient to trust the unusually strong comparisons.
-
It Takes Two to Match: Co-Evolving Generative Retriever with Reinforcement Learning
CoGR trains separate LLM-based generators for queries and items to produce keywords that can be matched directly with a conventional inverted index. After supervised alignment, the two generators alternately optimize retrieval F1 using reinforcement learning, with the item side rewarded according to its counterfactual impact on query retrieval. The method reportedly improves F1 by 10.9% on an internal marketplace dataset and 36.1% on WANDS over the strongest baseline, though the abstract provides limited detail about scale and experimental controls.
The co-evolution of query- and item-side generative representations while retaining standard inverted-index infrastructure is a substantive retrieval idea, and the reported gains are large enough to merit checking, but the evidence is not detailed enough to establish broad significance.
-
RoboTok: An Internet-Scale Data Engine for Human Demonstration Retrieval and Dexterous Manipulation Learning
RoboTok retrieves human manipulation videos from the web using a learned latent representation of 3D hand-motion trajectories, rather than relying mainly on visual appearance. The authors report better retrieval and improved dexterous robot-policy success, suggesting that human videos could provide a continually expanding source of manipulation supervision, though the abstract gives no quantitative results or details about the robot-learning pipeline.
The hand-trajectory-aware, internet-scale retrieval framing is a meaningful direction for reducing robot-data collection costs, but the abstract provides insufficient quantitative evidence to justify a stronger recommendation.
-
CRISP: Cliff-awaRe Input-adaptive Sparse Prefilling with Structural-Mass-Motivated Routing
CRISP is a dynamic sparse-attention method for long-context prefilling that routes attention heads using structural information directly from a proxy attention map, avoiding the overhead of Jensen–Shannon-divergence routing. It also introduces a sink-aware threshold to avoid selecting large amounts of low-value background mass at long sequence lengths. Across three long-context benchmarks and two model families, it reports up to 5.3× attention speedup at 512k tokens and up to 28 percentage points better retrieval performance than prior sparse methods, sometimes matching or exceeding dense attention.
The combination of proxy-free structural routing and a theoretically motivated treatment of the post-softmax noise floor addresses real failure modes in long-context sparse attention, with substantial reported speed and retrieval gains, though the abstract does not establish how broadly these results transfer beyond the tested models and benchmarks.
-
A Borel Concept Class of VC Dimension One with a Non-PAC Consistent Learner in ZFC
The paper constructs, in ordinary ZFC set theory without assuming the Continuum Hypothesis, a Borel concept class with VC dimension one and a proper consistent learner that nevertheless fails PAC learning. For a suitable distribution and target, the learner has risk one at every sample size on a set of samples with outer probability one, showing that measurability of individual concepts alone is insufficient.
This is a sharp, assumption-breaking learning-theory counterexample that removes the Continuum Hypothesis from a known construction, though its practical impact is limited and the abstract provides no details of the construction or proof.
-
How Output Format Confounds Data Quality and Capability in Instruction Tuning
This paper argues that the surface format used for instruction-tuning data and model outputs can substantially distort both data-quality metrics and measured capabilities. Across 12 tasks, four equivalent interfaces, and three model families, it finds that common spectral diagnostics miss semantic corruption, while interface-relative update directions reveal it; abilities can also appear or disappear depending on the format and generation budget. The main contribution is treating data quality and capability as interface-conditioned rather than intrinsic properties.
The interface-dependent capability and gradient-geometry effects are unusually strong and challenge standard evaluation assumptions, with reasonably broad controlled evidence, but the abstract’s striking claims need close inspection of methodology and reproducibility before meriting a strong recommendation.
-
Exact Limits of Random Projections for Preserving Geometry: Distance Recovery, Nearest-Neighbor Rankings, and Covariance Shape in Gaussian Models
This paper analyzes what information about high-dimensional geometry survives a rank-m Gaussian projection, rather than relying only on the usual JL distance guarantee. For isotropic Gaussian data, it derives the optimal recovery spectrum and shows that only about an m/d fraction of single-distance variance survives, Kendall rank correlation scales as sqrt(m/d), nearest-neighbor agreement can become random, and covariance-shape information scales as (m/d)^2. The key message is that a projection can satisfy JL while retaining little useful information for comparing or inferring geometry.
The sharp distinction between relative distance preservation and preservation of rankings or geometric structure is genuinely non-obvious and potentially important, but the strongest results are asymptotic and restricted mainly to Gaussian models, so this merits a closer look rather than an automatic must-read.
-
WebWorld: The Browser as a World Model for Self-Improving Web Code
WebWorld uses the browser itself as an executable verifier for self-improving web-code models. A VLM proposes repairs, but only browser-checked interactions that make progress without breaking previously verified behavior are admitted as training data; the authors report gains of 5.3 points on HTMLBench-400 and 14.9 on MiniAppBench-Val, with the effect largely disappearing without certification.
The browser-backed, regression-aware admission loop is a meaningful alternative to self-judgment in code-generation training, but the abstract gives limited detail about the benchmarks, certificate reliability, and whether gains generalize beyond web artifacts.
-
$τ^τ$-Bench: An Environment for End-To-End, Realistic Agent Construction
τ^τ-Bench evaluates coding agents on building complete customer-service agents from realistic business records, requirements, APIs, inherited code, and cost/model constraints, with performance measured through held-out simulated users. Across 53 tasks in four domains, the best reported setup passes only 23.9% of simulations versus 82.2% for an expert-authored reference, with failures involving shallow data understanding, weak client communication, and insufficient architectural experimentation.
This is a materially more realistic framing of agent-building ability than standard coding benchmarks, and the large gap between strong coding agents and expert solutions provides a useful, non-obvious diagnosis, though it is still primarily a benchmark paper rather than a capability advance.
-
PaperGym: Rubric-Centered Evolution for Research-Plan Generation
PaperGym builds training environments for generating research plans by deriving the research question from a paper’s background and the evaluation rubric from its methods and experiments, reducing the chance that models can simply paraphrase the prompt. It combines rubric-conditioned self-training with rubric-based GRPO reinforcement learning and reports gains across several model sizes and benchmarks, including a 58.1% preference rate over a competing dataset-based recipe. The authors release a 20,000-example corpus, training pipeline, and evaluation benchmarks.
The separation of question generation from method/experiment-derived rubrics is a meaningful way to make research-plan rewards less vulnerable to leakage, but the abstract provides limited detail about benchmark validity and relies heavily on self-generated data and reported comparisons.
-
UniMate: One Unified Model to Animate Diverse Skeletons
UniMate is a text-conditioned motion-generation model intended to animate many different rigged skeletons without category-specific templates, per-skeleton retraining, or test-time optimization. It combines graph- and spectrum-based topology encodings with a diffusion transformer and trains on a 13,006-sequence dataset covering animals and articulated objects; the abstract claims zero-shot transfer across skeleton types plus motion editing and in-betweening.
A genuinely broad, topology-agnostic animation direction with potentially useful zero-shot transfer, but the abstract gives no quantitative results or details sufficient to validate the claimed generality and efficiency gains.
-
Puffin-World: Scaling a Unified Multimodal Model with Native 3D World States
Puffin-World is a unified multimodal world model that jointly represents physics, geometry, and appearance, using an omni-camera representation to generate and reconstruct future 3D views under camera motion. It also introduces a large training corpus of vision-language-camera data and trajectories, and demonstrates closed-loop tasks such as mimicry and self-calibrated exploration. The central new idea is to make physical state, depth, and images native components of one generative model rather than relying on separate offline modules.
The native joint modeling of physics, geometry, and appearance could be a meaningful direction for interactive world models, but the abstract gives no quantitative comparisons or concrete capability results to establish a major advance.
-
ExecRetrieval: Measuring the Functional-Correctness Gap in Code-Embedding Retrieval
ExecRetrieval evaluates whether code embeddings retrieve functionally correct implementations over near-identical but execution-verified buggy variants. Across 939 Python tasks and 23 embedding configurations, the best systems achieve perfect top-10 functional recall but only 0.331 top-1 recall; rank-1 errors overwhelmingly select a paired buggy variant. The benchmark and released artifacts make a concrete case that standard code-retrieval performance can hide a substantial functional-correctness gap.
This is more than a routine benchmark because controlled executable counterfactuals expose a non-obvious failure mode—strong retrieval systems often rank a nearly identical incorrect implementation above the correct one—but it does not yet demonstrate a remedy or a broad capability advance.
-
Corporate-Family Resolution Is Not a String-Matching Problem: A Public Benchmark Stratified by Name Visibility
This paper argues that corporate-family resolution is primarily a candidate-retrieval problem rather than a pairwise matching problem. On a large, publicly released benchmark, existing blocking methods retrieve only 6.8% of links between companies with no shared name tokens, and ranking improvements cannot recover the remaining 93.2%; independent SEC evidence supports that these links are genuine.
The unusually strong, independently validated finding that candidate generation—not matching accuracy—is the dominant bottleneck gives this otherwise domain-specific benchmark a useful and non-obvious methodological lesson.
-
Interface-Induced Trajectory Censoring
This paper shows that tool-use evaluations can report zero tool calls even when the model is generating valid calls, because the serving adapter, chat template, and parser can silently censor the trajectory. Controlled swaps on BFCL v4 and tau-bench produce dramatic changes with identical model weights and inputs, while additional experiments across Qwen and Llama models show the failure also affects training loops and can depend on scale and configuration. The authors release a 98-line preflight check, although repairing the interface restores calls without producing a statistically significant task-success improvement in their evaluation.
The carefully isolated interface-by-parser interaction and its large, reproducible impact on reported tool use are important and likely practically useful, but the work is primarily a diagnostic of evaluation infrastructure rather than a new capability or broadly general algorithmic advance.
-
A cognitive representation in primary visual cortex modulated by vision
In rats navigating in darkness, many V1 neurons represented progress through a shared sequence of turns across physically different paths, suggesting an internally generated, path-invariant code. When vision was available, visual input modulated this activity rather than replacing it, and these neurons were coordinated with hippocampal activity during sharp-wave ripples.
The result substantially challenges the usual view of V1 as primarily stimulus-driven, but the abstract provides limited quantitative and cross-task evidence for such a broad claim.
-
Rapid spatial cognition in mice, with and without neocortex and hippocampus
Mice learned a reconfigurable 3D maze very rapidly: they developed useful navigation biases before receiving rewards, reached efficient routes within roughly 20 rewards, and retained routes for weeks to months. Mice lacking the hippocampus and most of the neocortex were slower during initial exploration but showed largely preserved learning, retention, and generalization, while a neuromorphic model suggested that reward-independent environmental structure could support this behavior.
The striking, assumption-challenging result is that rapid spatial learning and generalization remain largely intact without hippocampus and neocortex, but the abstract provides limited detail on lesion validation, sample sizes, and the breadth of behavioral tests.
-
WHALE: A Simple Recipe for Joint Harness-Weight Optimization
WHALE jointly improves an agent’s model weights and its executable harness—the code that handles context and control flow—by alternating weight fine-tuning with harness search. Across search QA, math, and chess using Qwen3.5 agents, it reports 4.15–24.38 percentage-point gains over optimizing either component alone, with small interleaved updates also reducing rollout cost versus stagewise optimization.
The explicit treatment of agent harness code as a co-optimized counterpart to model weights is a meaningful direction, but the evidence is limited to three domains and a relatively simple alternating recipe, so the broad significance is not yet clear.
-
A Common Measure of Communication for Speech Brain-Computer Interfaces
The paper introduces open-vocabulary mutual information (OVMI), an information-theoretic metric intended to compare speech BCI decoders across different vocabularies, datasets, and recording setups. It argues that standard accuracy and WER can exaggerate communication ability when evaluated only on supported words, and reports up to a 16.3% relative accuracy improvement from choosing vocabularies to maximize OVMI across three speech domains.
The open-vocabulary evaluation framing addresses a real comparability problem and exposes a non-obvious weakness in standard speech-BCI metrics, but the abstract provides limited evidence that OVMI will become a broadly accepted measure or materially improve decoding beyond vocabulary selection.
-
NeoMME: A Single-Tower Multimodal-Native Multilingual Foundation Encoder for Efficient Fine-Tuning and Inference
NeoMME is a 260M/800M bidirectional multimodal encoder trained from scratch on multilingual text and image patches, rather than adapting a large generative vision-language model. It combines dense and late-interaction retrieval heads, reports competitive ViDoRe v3 results, roughly 2× page-encoding throughput versus ColModernVBERT, and 255× embedding compression with limited retrieval loss.
The single-tower, multimodal-native encoder and diffusion-style pretraining offer a credible efficiency-oriented alternative to repurposed VLMs, but the evidence is primarily one retrieval benchmark and the abstract does not establish broad superiority or general-purpose capability.
-
EarlyEval: Cheaper Agent Evaluation via Early Outcome Prediction
EarlyEval predicts whether an LLM agent will ultimately succeed or fail from its intermediate behavior, then stops the run once a calibrated LightGBM classifier is confident. On SWE-bench Verified, TerminalBench, and Toolathlon, it removes 13–26% of agent steps and up to 44.1% of input tokens and 29.4% of output tokens, while changing solve rates by only 1–2 percentage points.
The within-task early-stopping formulation is a useful complement to benchmark distillation, with credible multi-benchmark cost savings, but the lightweight classifier approach is fairly intuitive and the gains are meaningful rather than transformative.
-
Percolation Dynamics in Optimization : Variance Cascades and Discrete Scale Invariance
The paper models stochastic gradient flow as a percolation process in which symmetry-related subnetworks merge in synchronized discrete events. It argues that these structural transitions produce macroscopic variance spikes and a cascade-like scaling pattern, and claims the same trapping mechanism extends to Adam and AdamW under heavy-tailed gradient noise.
The percolation and discrete-scale-invariance framing could offer a genuinely new explanation of SGD’s implicit structural simplification, but the abstract gives no quantitative results or indication of how broadly the theory matches real neural-network training.
-
Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall
The paper finds that standard forward-KL distillation behaves differently depending on when it is applied: during mid-training it improves reasoning but slows factual knowledge acquisition. It attributes this to teachers being more confident on procedural data while students learn low-entropy factual knowledge earlier, and proposes Switch Distillation, which uses teacher entropy to choose between distillation and ordinary cross-entropy; the method reportedly improves reasoning and knowledge performance while largely preserving factual recall, including after post-training.
The stage-dependent failure mode and entropy-based routing are non-obvious and potentially useful for large-scale LM training, but the abstract gives relative gains without enough detail about absolute results, model scales, and evaluation breadth to justify a stronger verdict.
-
From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix
The authors build a self-hosted LLM for a diverse corporate workload by analyzing production failures and training separate GRPO experts for instruction following, function calling, and task-specific behavior, then merging them with two-stage SLERP. They report that the resulting model beats a roughly 7× larger baseline on internal evaluations and serves 50% of platform traffic—about 116 million requests per month—at lower cost.
The production-scale consolidation and explicit treatment of conflicting post-training objectives are useful and potentially significant, but the core techniques are familiar and the evidence is mostly from undisclosed internal benchmarks and limited headline comparisons.
-
Don't Drop Dropout: Optimizing Layer Sparsity for Efficient LLM Training and Inference
This paper revisits layer dropout during large-language-model pretraining, tuning its layer-selection distribution, schedule, and optimizer settings across models from 271M to 8.2B parameters. It reports up to 25% lower training compute for comparable validation loss, and shows that the resulting models support early exit, layer skipping, and self-speculative decoding for up to 1.5x faster inference with little accuracy loss. The main contribution is a large-scale empirical recipe showing that a previously neglected technique remains useful when properly configured.
The underlying idea is established stochastic depth, but the unusually broad scaling study and substantial training/inference efficiency gains make this worth knowing about, though the gains depend on careful tuning and are not a fundamentally new method.
-
Confounding Masquerading as Improvement: A Systematic Evaluation of Offline Reinforcement Learning for Stroke Antithrombotic Treatment in a 129,000-Patient Registry
This paper shows that apparent gains from offline RL for stroke antithrombotic treatment largely disappear when the terminal reward is separated from baseline disease severity and prognosis. Across five RL families and 14 reward designs in a large stroke registry, FQE’s estimated improvement fell from +0.0069 to +0.0025 and became statistically non-significant after reward deconfounding, with several analyses converging on no clinically meaningful aggregate benefit. The main contribution is an empirical warning and evaluation checklist for confounding in clinical offline RL.
The large, partially crossed analysis identifies a specific failure mode—prognostic information embedded in rewards masquerading as treatment-policy improvement—and tests it with multiple algorithms and outcome analyses, though the evidence remains observational and focused on one clinical task.
-
Where Do Multilingual Vision-Language Encoders Fail on Low-Resource Languages?
This paper investigates why multilingual vision-language encoders perform much worse on low-resource languages. It finds that removing easily detectable linear language information from the output representation does not improve retrieval; instead, the harmful divergence appears in the EOS representation’s trajectory through intermediate layers. Replacing or steering that representation toward a parallel English/content centroid produces large gains for Swahili and other low-resource languages while preserving high-resource performance.
The paper offers a non-obvious causal diagnosis that challenges the usual output-space language-bias explanation, with unusually large intervention gains and replication, but the evidence is still limited to a small number of encoders and a 1,000-image retrieval subset.
-
Hard-ReLU Gradient Descent Selects an Event-Free Sensitivity Limit
This paper shows that for hard-ReLU networks, gradient descent can converge to gradient flow in its parameter trajectory while exact automatic differentiation through the discrete updates converges to the wrong sensitivity. The missing terms are event-time “saltation” corrections at activation crossings; the authors give convex constructions with arbitrarily large sensitivity gaps, a coordinate reversal, and a 17-parameter experiment where the discrepancy persists as the step size shrinks.
It identifies a non-commutation between the gradient-flow limit and differentiation through nonsmooth training dynamics, with explicit counterexamples and empirical confirmation, but the result is specialized to hard-ReLU systems and its broader practical impact is not yet established.
-
One note in three: a verified census of three deployed AI scribes, and the instrument that counted it
This paper audits three commercial AI clinical scribes on 565 notes from 142 consultations, using adversarial model review plus blinded clinician checks. It finds a verified error in about 31% of notes (24.8% after excluding issues a patient record might have prevented), especially involving allergies, medications, invented identity or dates, and telephone-call histories misrepresented as examinations. A major additional result is that the measured error rate changes dramatically with the review instructions, model family, and definition of failure, and the authors release the evidence and reproducible audit pipeline.
The combination of real-world commercial-scribe auditing and a demonstrated, quantified dependence of error rates on the evaluation instrument is genuinely useful, but the study covers only three products and 142 consultations, so it is not yet a definitive general result.
-
Wrong Prediction, Right Answer: Recovering Evidence from Collapsed LLM Sequence Scores
The paper argues that some apparent reasoning failures come from the model’s output scoring rather than from missing logical competence. It shows that hidden-state probes can recover correct answers when native sequence scoring fails, and that a two-parameter, label-free correction fitted on as few as 25 unlabeled examples improves accuracy by 9–34 points across several models and reasoning tasks. The main novelty is separating latent decision quality from the final token-level readout bottleneck.
The cross-model, label-free recovery of substantial reasoning accuracy would be important if robust, but the evidence is still limited to controlled tasks and the abstract does not establish how broadly the diagnostic generalizes beyond this evaluation setup.
-
Non-Prehensile Throwing: A Reinforcement Learning Perspective
The paper uses reinforcement learning to plan joint-space motions for grasp-free throwing, allowing sliding and rolling contacts instead of relying on simplified analytical contact models. In simulation it reports 99% success and, on a UR5e, 97% real-world success while throwing heavy or large objects up to 3.5 m away or 1.8 m high without task-specific real-world tuning.
The combination of model-free contact-mode handling, uncertainty-aware sim-to-real transfer, and strong zero-shot throwing results is notably beyond standard pick-and-place, but the contribution is still focused on a specialized manipulation task and the abstract gives limited comparative evidence.
-
VOIM: Training-Free Open-Vocabulary 3D Instance Mapping for RGB-D and Monocular SLAM
VOIM builds open-vocabulary 3D instance maps by accumulating uncertain voxel-level evidence across views before deciding object identities, rather than labeling instances at their first detection. This training-free strategy improves substantially over OVO-SLAM on ScanNet++ (44.07 vs. 32.37 mIoU under matched inputs) and transfers unchanged from RGB-D to monocular RGB, though labeling is not real-time and performance is limited by vocabulary coverage and SLAM drift at larger scales.
The deferred, voxel-grounded decision process appears to provide a large and fairly robust gain for open-vocabulary 3D mapping, including a difficult monocular setting, but the contribution remains a specialized mapping system with benchmark-based evidence and clear operational limitations.
-
MIDR: Enrichment-Augmented Indexing for Multimodal Document Retrieval
MIDR uses a multimodal LLM during document ingestion to extract verified text about page content, tables, figures, and layout, then serves retrieval through conventional BM25F and optional dense search instead of visual late-interaction models. On ViDoRe V3 it substantially improves over BM25, remains competitive with ColQwen2.5, handles English queries over French documents, and reportedly uses about 9× less index memory with 2× lower query latency.
The index-time-versus-query-time multimodal retrieval tradeoff is a useful and non-obvious systems direction, supported by sizable cross-domain accuracy and deployment-efficiency gains, though the evidence is limited to benchmark evaluation and one framework.
-
Coverage, Not Targeting: A Structural Regime in Multi-Turn Agent Credit Assignment
The paper argues that in long-horizon tool-use tasks, terminal verifiers reveal too little information to identify which turns deserve credit. Across controlled experiments, uniformly distributed dense rewards outperform both sparse terminal rewards and equally concentrated rewards targeted at presumed progress turns, with performance improving as reward coverage approaches the full causal chain. It proposes verifier information density, V_d, as a way to predict when targeting is appropriate, and reports consistent results across tau^2-bench, BFCL V3, model families, and many seeds.
The coverage-versus-targeting framing and matched-concentration controls address a real credit-assignment assumption with unusually thorough evidence, but the result is still demonstrated mainly on two tool-use benchmarks rather than established as a broadly general regime.
-
Trace as State: Reasoning Traces as Conditional States for Long-Context Transformers
The paper treats an earlier reasoning trace as a compact task-state summary, then places that trace before the original long context in a second transformer pass. Compared with putting the same trace after the context, this ordering wins in 26 of 27 model/task/metric combinations, with exact-match gains such as 43.0% to 81.8% on one model and 83.2% to 100.0% on another.
The controlled comparison and unusually large gains make the ordering effect worth examining, though the approach depends on externally collected traces and the abstract does not establish how broadly it transfers beyond the tested long-context tasks.
-
Exemplar: Classical Priors Complement Frozen Features for Few-Shot Microscopy Segmentation at Native Resolution
Exemplar combines frozen DINOv3 features with a fixed bank of classical, native-resolution image filters in a lightweight head trained only from a few support masks. Across 11 biomedical segmentation datasets, the combination substantially outperforms either component alone and reportedly beats five competing few-shot methods in 54 of 55 comparisons, while fitting much faster than a from-scratch nnU-Net. The main limitation is that nnU-Net catches up and surpasses it at eight annotated masks, especially on centreline metrics.
The strong and somewhat surprising result is that hand-crafted native-resolution filters complement frozen foundation-model features across many microscopy datasets, yielding a practical few-shot method with large fitting-time advantages; however, it remains a specialized segmentation study and the abstract does not establish broader generality beyond this panel.
-
LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes
LLaDA-Image combines a 6B diffusion-transformer image generator, pretrained largely on image-only data, with a frozen diffusion-language-model-based vision-language module for conditioning and editing. The authors report strong photorealism and instruction following, top open-source scores on Qwen-Image-Bench, and a distilled 2–4-step Turbo version, while releasing weights, code, and training recipes.
The open end-to-end recipe, image-first training strategy, and very fast distilled variant are worth noting, but the abstract provides limited comparative detail and the claimed benchmark leadership appears to be an incremental advance rather than a clearly new generation paradigm.
-
What Attention Recalls and Recurrence Controls in Hybrid Language Models
The paper uses cache-level interventions to separate the roles of attention KV caches and recurrent state in hybrid language models. Across Qwen3.5 and Falcon-H1, exact retrieval depends almost entirely on attention, while language choice and persona depend mainly on the recurrent state; swapping the two channels causally reproduces this division. The recurrent state can also generalize to semantically or morphologically related words rather than only copying context items.
The causal cache-swapping analysis gives a clear and non-obvious functional decomposition of hybrid models, but the evidence is limited to two model families and does not yet establish broad architectural generality.
-
Training-Free Halving of Activated Experts in Fine-Grained Mixture-of-Experts Models
The paper identifies that reducing the number of active MoE experts at inference changes both expert selection and the router’s output gain because of top-k renormalization. It avoids this by selecting fewer experts while normalizing against the probability mass of a larger reference set, cutting routed-expert compute roughly in half with only a 0.35–0.55 point MMLU loss on two large Qwen models. The effect is also shown to depend on the choice of reference mass and to differ between perplexity and labeled accuracy.
A simple training-free intervention appears to preserve most MoE quality while halving inference expert compute, and the replication on a 397B model makes it worth examining, though the evidence is still limited to the reported models and benchmarks.
-
A reactivated thalamocortical plasticity window promotes learning and is reshaped by experience
In adult mice, cutting the infraorbital nerve briefly reopened a critical-period-like state in barrel cortex, increasing thalamocortical gain and improving whisker-detection learning when training occurred during that window. The timing of experience also reshaped the window: early training advanced potentiation, training during it prolonged potentiation and sensory memory, while later training could not restore it; blocking local GluN2B receptors eliminated the cortical and behavioral effects.
This is a well-supported, causally tested demonstration that adult sensory plasticity is a temporally gated and experience-reciprocal circuit state, but its significance beyond the mouse sensory-loss model remains uncertain.
-
Mice and humans use ambiguous local motion to stabilize gaze
This study tested whether gaze-stabilizing circuits resolve the ambiguity of local motion signals before acting on them. In mice, optic-tract neurons and eye movements were dominated by locally measured edge motion, even producing horizontal movements when vertically moving objects changed edge orientation; humans showed a similar bias, especially when central vision was masked. The results support a conserved, efficient heuristic rather than an accurate reconstruction of global object motion for this reflex.
The cross-species physiological and behavioral evidence challenges the assumption that ambiguous retinal motion must be disambiguated before driving behavior, though the underlying local-motion bias builds on established work on the aperture problem and reflexive vision.
-
A canonical excitatory-inhibitory motif embedded in the connectome underlies rhythmic locomotion in Drosophila larvae
This study combines larval fly connectomics, physiological recordings, modeling, and perturbations to argue that a repeated excitatory A18f–inhibitory A02j circuit generates forward locomotion. A model using reconstructed synapse counts produces traveling motor waves with little tuning, and experiments support the predicted roles and timing of the two neuron types, including wave initiation by activating posterior A18f neurons.
It provides unusually direct evidence that a textbook excitatory-inhibitory oscillator is physically instantiated in a complete neural connectome and causally drives behavior, though the result is still confined to one small locomotor circuit and the abstract does not establish broad generality.
-
LightNav-0: Eliciting VLM Spatial Intelligence for Generalist Embodied Navigation
LightNav-0 uses a compact pretrained vision-language model as a general navigation controller, representing spatial intent with pointing tokens and embodiment-specific trajectories with a residual action tokenizer. It combines instruction following, object navigation, and visual tracking, and claims strong simulation results plus zero-shot transfer across robot bodies and real-world scenes. The main new idea is a unified interface intended to replace task- and embodiment-specific navigation heads.
The unified spatial-intent/action-token interface and claimed cross-embodiment transfer are plausibly important, but the abstract provides no quantitative results or details sufficient to distinguish a major advance from a well-engineered integration of existing VLM, tokenizer, and RL techniques.
-
ShallowStream: Index Shallow then Answer Deep for Streaming Video Understanding
ShallowStream reduces the cost of streaming video understanding by running only shallow MLLM layers continuously, using their KV cache to build a lightweight frame index. At query time, it uses shallow-layer attention to retrieve diverse relevant frames before running deeper processing, reportedly matching strong streaming baselines while reducing per-frame prefill latency by up to 52.1× and 10-second end-to-end latency by up to 11.9×.
The explicit use of shallow model depth for always-on indexing is a meaningful and potentially general efficiency direction, with unusually large claimed speedups, but the abstract provides no benchmark details and labels the work as in progress, so the strength of the result is not yet established.
-
Learning Where Outcomes Change:Credit-Addressable Reasoning for Multimodal Geometry
The paper proposes aligning the units used to represent multimodal reasoning with the units used for reinforcement-learning credit assignment. Its Code-CoT represents diagram relations as line-addressable executable code and its CE-GRPO assigns localized advantages at typed reasoning events; across nine geometry benchmarks it reports 76.04 average accuracy, 8.09 points above Qwen3-VL-8B and 3.43 above trajectory-level GRPO, with larger gains on longer reasoning chains.
The representation–optimization co-design and event-level credit assignment are a meaningful step beyond standard trajectory-level RL, supported by broad benchmark results, but the core idea remains related to established process-level supervision and the evidence is confined to geometry reasoning.
-
Autoregressive Mosaics: Probing 2D Spatial Reasoning in Text-Only Language Models
The paper tests whether text-and-code-only language models genuinely represent 2D layouts or merely translate verbal geometry into drawing code. Across eight models, translation ability does not predict open-ended composition; raw SVG produces better layouts than procedural code, and activation probes suggest models revise geometric state during generation rather than follow a fixed initial plan.
It offers a useful separation between spatial understanding and code translation, with the unexpected finding that output medium strongly affects apparent spatial reasoning, though the abstract gives no quantitative results or evidence of broader downstream impact.
-
Adapting Without Gradients: Affine Statistics Transport and What Its Certificate Can Tell You
The paper proposes CASTER, a gradient-free test-time adaptation method for frozen models. It stores compact class statistics, estimates a class-shared affine transformation from target-batch moments, and transports the stored distributions without backpropagation or a feature bank. Across seven datasets and four backbones it beats k-NN on frozen features in 27/28 settings, but can fail badly on sparse-class batches; a residual-to-margin certificate detects many destructive cases and improves the average effect from -3.35 to +1.69 points when used for gating.
The combination of frozen-model affine adaptation and an explicit, mechanism-specific transportability certificate is a credible new direction for accelerator-constrained TTA, with broad experiments, but the gains are moderate and the method remains unreliable in important settings.
-
A.X K2 Technical Report
A.X K2 is a 688B-parameter mixture-of-experts model trained on 8.5T tokens, with more agentic and software-engineering data than its predecessor. It combines Sparse Gated Attention for efficient long-context processing, Gated Norm for more stable low-bit serving, and a unified thinking/non-thinking mode; the authors report 94.6 on RULER through 256K context and large gains over A.X K1 despite less training data.
The native sparse long-context design and reported token/serving efficiency gains are genuinely interesting, but the abstract gives limited comparative detail and the headline claims come from a technical report with unclear independent validation.
-
Three Steps at a Time: Learning Representations from Action Sequences in Contrastive RL
The paper extends contrastive reinforcement learning from predicting single actions to predicting short action chunks. Across 18 offline and 11 online environments, this reportedly produces large gains, and the authors argue the main benefit is not simply better multi-step return propagation: action chunks contain more information about the agent’s goal, leading to better critic representations.
The broad gains and the proposed explanation—that action chunks improve goal information in contrastive critics—are genuinely interesting, but the abstract lacks enough detail about baselines, effect consistency, and experimental controls for a strong recommendation.
-
Can Video World Models Track Unobserved World States?
The paper tests whether video world models preserve hidden world state rather than merely generate plausible-looking frames, using an action-conditioned shell game with longer swap sequences than seen during training. Standard Transformers, Mamba, and constrained linear attention fit the training horizon but collapse to chance on longer sequences; models that maintain and revise an explicit recurrent state—linear attention with negative eigenvalues or nonlinear test-time training (TTT)—extrapolate better. The results suggest that visual prediction loss alone does not force a model to represent unobserved state, and that stateful architectures may be necessary.
The hidden-state extrapolation test and architectural diagnosis are genuinely interesting, especially the finding that plausible video generation can coexist with chance-level state tracking, but the evidence appears centered on a specialized shell-game setup and is not yet enough for a strong verdict.
-
A Universal Context-Reuse Layer for Cross-Model KV Sharing
The paper introduces a translation layer that converts one LLM’s KV cache into a form usable by a different model, even across model families, sizes, tokenizers, and attention designs. It reports substantial target-side prefill savings—up to 67% in one setting and latency falling from 899 ms to 138 ms in another—while retaining near-native quality, though the results are preliminary.
Cross-model KV reuse is a genuinely interesting systems abstraction with potentially large serving benefits, but the abstract provides limited methodological detail and only a few experimental configurations, so the evidence is not yet strong enough for a top-priority recommendation.
-
QTEA: Ternary LLMs with Sparse Residual Salient Weight and By-Column Optimization
QTEA is a weight-only post-training quantization method that stores most LLM weights as ternary values and uses structured sparse residuals in selected columns to recover accuracy. It combines this with column-wise scale refinement and error attenuation, reporting large gains over a ternary baseline at roughly 1.7 bits per weight, plus a lookup-table kernel claiming substantially faster generation than FP16.
The combination of ternary quantization, hardware-friendly sparse residuals, and column-wise optimization is promising, and the reported accuracy and speed gains are unusually large, but the abstract does not establish how strong the baselines or hardware comparisons are.
-
Vision Is Not Overhead: One-Pass Block Drafting for Lossless Speculative Decoding in Vision-Language Models
GLANCE is a speculative-decoding drafter for vision-language models that uses the target model’s fused image-text state and a diffusion-style head to propose an entire token block in one pass. It claims exact greedy-output preservation and reports up to 2.93× decoding speedup, with especially long accepted blocks on visually grounded, copy-heavy tasks; it also proposes a target-entropy-based explanation for when block drafting works.
The combination of one-pass block drafting, direct reuse of vision-language representations, and lossless verification is a genuinely interesting direction with meaningful reported speedups, but the abstract does not provide enough model, workload, or baseline detail to justify a strong verdict.
-
Independent Reinforcement Learning in Discounted Markov Games
This paper studies what independently learning agents can achieve in general-sum discounted Markov games without sharing information or coordinating their algorithms. It proves a conditional computational-hardness result for finding approximate coarse correlated equilibria, while also giving the first fully uncoupled algorithm with sub-exponential convergence, including versions for full and partial feedback; the method uses layered optimistic mirror descent with increasing step sizes.
The combination of a broad impossibility result and a first general-purpose sub-exponential uncoupled-learning algorithm is genuinely notable, but the abstract gives no concrete convergence rates or experimental/theoretical detail to support a stronger recommendation.
-
ViTAL-X: Video-Text Alignment with Cross-Modal Temporal Edits
The paper argues that video-text models often recognize static content but miss temporal structure such as order, direction, and motion. It introduces XTE-Bench to expose this weakness and Cross-Modal Temporal Edits, a self-supervised way to create hard temporal negatives by jointly transforming videos and captions; a 0.4B-parameter ViTAL-X model reportedly beats much larger models across six temporal benchmarks.
The combination of a diagnostic for temporal blindness and annotation-free, cross-modal temporal edits could be a useful direction, with a potentially large efficiency gain, but the abstract gives no numerical results or detail sufficient to validate the strong scaling and data-efficiency claims.
-
Panda Diplomacy: Foundation Model Pre-training across Particle Imaging Detectors for High Energy and Nuclear Physics
The paper proposes a point-cloud self-distillation recipe that can pre-train one broadly reusable model across three very different particle-detector modalities: liquid-argon TPCs, collider TPCs, and water-Cherenkov detectors. With only 1,000 labeled events for adaptation, it reportedly matches or beats detector-specific foundation models trained with far more labels, including up to 70× and 1,000× reductions in labeled data, while its representations encode physically meaningful properties such as causality and track curvature.
The cross-modality sensor pre-training result and very large label-efficiency claims are genuinely interesting, but the abstract gives limited detail about dataset scale, comparisons, and whether the gains transfer beyond these specialized physics benchmarks.
-
On Global Regulatability of Robot Manipulators by Classical PID
The paper proves a sharp distinction for classical PID control of robot manipulators. For one-degree-of-freedom systems, it derives gain conditions guaranteeing global asymptotic regulation, while for some multi-degree-of-freedom manipulators it proves that no fixed classical PID gains can achieve global regulation. The main contribution is a theoretical characterization of both the power and limitation of standard PID control under common structural assumptions.
It claims to resolve a long-standing theoretical question with an unexpected 1-DOF versus multi-DOF impossibility boundary, but the abstract gives no proof details or scope beyond the stated structural assumptions, so it merits a closer look rather than a strong recommendation.
-
Making Prospective Memory SLM-Shaped: Typed Intention Stores for Small-Model Agents
The paper treats prospective memory in agents as structured state tracking rather than free-form reasoning. Its Prospective Intention Store keeps lifecycle logic in code and restricts the model to typed operations, raising PM-Bench Set-F1 from 65.1% for a prior large-model scaffold to 82.9%, and from 4.2% to 66.2% for Gemma-E2B without training. The main novelty is showing that a lightweight external state machine can substantially improve deferred-intention handling in small models.
The large, training-free improvement and small-model result are notably stronger than a routine memory tweak, but the evidence is confined to one specialized benchmark and needs replication beyond PM-Bench.
-
Zeta-Lite: A Concurrent, Branchable In-Browser SQL Database for Agentic Memory
Zeta-Lite is a 2.87 MB WebAssembly build of the Zeta database for browser applications. Its asynchronous MVCC design supports overlapping snapshot-isolated transactions on one thread, plus copy-on-write database fork, merge, and rebase; it also exposes broad PostgreSQL features and OPFS persistence. The authors report 268k–315k point reads per second across browsers and sustained mixed read/write performance over millions of operations, positioning speculative database branches as memory for AI agents.
Branchable, concurrently usable SQL state in a small browser runtime is a genuinely distinctive capability with plausible importance for local agent memory, but the abstract provides limited comparative and workload evidence for the broad performance and feature claims.
-
The Dynamics of Continuous Mixture Collapse in Language Models
This paper analyzes why language models tend to destroy continuous mixtures of possible latent reasoning states instead of preserving them. It identifies architectural distortion, training-amplified geometric changes, and— even under perfect linear transport—softmax readout plus autoregressive feedback that either collapses mixtures or makes them indistinguishable. The theory predicts a contraction-versus-amplification threshold that is observed in pretrained-model rollouts, and the paper argues that preserving many-way mixtures generally requires increasingly high-dimensional, context-dependent corrections.
The dynamical-systems account of mixture collapse and its predicted threshold could materially affect how continuous latent reasoning methods are designed, but the abstract provides limited quantitative detail and does not establish impact across large or practically important systems.
-
A Power Law in Logarithm's Clothing: On the Scalability of Graph-Based Vector Search
The paper argues that graph-based approximate nearest-neighbor search does not generally scale polylogarithmically at realistic dataset sizes. Instead, search cost initially follows a sublinear power law, with a later transition to subpolynomial growth once dataset size exceeds the data’s intrinsic dimensionality; the authors support this with experiments, a beam-search theory, and proofs for simplified graph constructions.
The proposed scale-dependent power-law regime challenges a common assumption about vector-index scalability and could materially affect capacity planning and recall/cost trade-offs, but the abstract gives no quantitative results and the strongest proofs appear limited to idealized constructions.
-
The Diagnosis a Reporter Leaves Unspoken: Surfacing Frozen Tumor Features for Brain-Tumor MRI Reporting
The paper finds that a brain-MRI report generator appears unable to diagnose tumors, even though its frozen segmentation features contain strong tumor-type information. NeuroFusion extracts that signal with classifier heads and feeds it into a constrained draft-and-review reporter, substantially improving diagnosis recall and report quality while reducing latency from about 457 to 80 seconds per case; a small neurologist pilot also preferred it and found no critical errors.
The interesting contribution is the demonstrated gap between latent diagnostic information and what a generative medical model verbalizes, plus a practical mechanism for surfacing it without hard overriding; however, the evidence is based on limited cohorts and a nine-case pilot, so it falls short of strong.
-
Following a Unique Path: A Fast Certifier Applied to Outlier-Robust Pose Registration
The paper introduces CP-Cert, a central-path-based method for certifying candidate solutions when standard local-solve-then-certify methods fail because the SDP relaxation is degenerate. Applied to robust pose registration and point-cloud data association, it reportedly runs up to three orders of magnitude faster than direct SDP solvers and is demonstrated in a real-world outlier-robust pose-estimation pipeline.
The combination of a new certification strategy for degenerate relaxations and a potentially very large speedup is compelling, but the abstract gives limited detail behind the best-case runtime claim and does not establish broad superiority beyond the tested registration problems.
-
Introducing SINFONIA: Symplectic, slimplectic and Magnusian (Neural) Flows for Orbital Numerical Integration and Acceleration
The paper learns finite-time evolution maps for dissipative gravitational-wave inspirals using three structure-preserving neural-flow designs, including symplectic/slimplectic and Magnus-style constructions. On a 2.5PN neutron-star inspiral, the learned maps reportedly remain accurate over 10^2–10^5 repeated orbital-scale steps and use a secular energy–angular-momentum error channel to support both long-horizon integration and inference of an unmodeled drag force.
The combination of learned finite-time maps, geometric structure preservation, and secular-error control is genuinely interesting, but the evidence is currently limited to a single modeled inspiral and the abstract does not establish robustness beyond this proof of concept.
-
Scalable Context Orchestration for Serving LLMs Over Voice
The paper introduces llmovoice, middleware that explicitly represents speaking style and environmental conditions—such as speaking rate, noise, and packet loss—in a bounded context supplied to an LLM. It reports substantially better voice interaction behavior, including 52.4% lower speaking-rate mismatch, a reduction in false interruptions from 46.0% to 0.9% under packet loss, and up to 79.2% lower model cost while preserving most answer quality in long sessions.
The explicit orchestration of paralinguistic and transmission context is a useful systems direction, and the reported robustness and cost improvements are unusually large, but the abstract gives too little detail about workloads, baselines, and evaluation breadth for a strong verdict.
-
GraFT: A Training-Free Framework for Spatial Reasoning in Multimodal Large Language Models via 3D Scene Graphs
GraFT augments a frozen multimodal language model with a compact 3D scene graph rather than retraining it. It combines symbolic geometry tools, bird’s-eye-view renderings for global layout, and selected egocentric views for appearance grounding; on ScanQA and VSI-Bench, the authors report substantial gains, including up to 65% on VSI-Bench and a 27% CIDEr increase on ScanQA.
The training-free combination of explicit 3D structure, symbolic computation, and complementary viewpoints is a credible and potentially reusable approach, but the abstract provides only relative benchmark claims and no evidence yet of robustness beyond these evaluations.
-
VestigeKV: The NoPE-MLA KV Cache Carries Its Own Eviction Signal in a Vestigial Branch
VestigeKV uses a 64-dimensional decoupled branch in Kimi Linear’s NoPE MLA cache as a query-independent estimate of token salience. It keeps only the most salient rows in the fast attended tier and moves the rest, without deleting or quantizing them, to an exact GPU archive; the reported retrieval remains 1.00 at 8× compression and 0.92 at 32× over 8k–65k contexts. The signal appears specific to NoPE models: the same method performs poorly with RoPE MLA.
The potentially important finding is that a supposedly vestigial NoPE branch carries query-independent cache-eviction information, enabling unusually aggressive exact KV-cache tiering, but the evidence is limited to one measured model and needs replication across NoPE architectures and workloads.
-
Conformity Breaks Conformal Prediction
The paper shows that conformal prediction guarantees can fail when an LLM answers after seeing peers that confidently give the wrong answer, even though the question distribution is unchanged. Across open-weight models and multiple-choice tasks, nominal 90% coverage drops to 74%, and targeted attacks reduce coverage on selected low-confidence cases from 87% to 47%, sometimes causing the system to act rather than escalate.
The score-mechanism-shift framing identifies a non-obvious failure mode for conformal guarantees in multi-agent LLM systems, with substantial reported coverage losses, but the evidence is limited to the abstract's model and QA-task experiments.
-
Distilled Continuous Diffusion Language Models Can Write Code in Few Steps---or One
PlaidQ turns a pretrained autoregressive model into a continuous diffusion denoiser over token embeddings, then distills its long sampling process into 16 steps or even one. A 16-step student reportedly reaches 31.78 pass@10 on HumanEval and 40.49 on MBPP+, exceeding the same model sampled for 512 steps; the one-step model achieves 7.07 pass@1 on HumanEval, showing functional but substantially weaker code generation.
The combination of continuous diffusion, bidirectional denoising, and trajectory distillation offers a genuinely interesting route to much faster non-autoregressive code generation, but the abstract does not establish broad superiority beyond a 0.7B code-focused model and the one-step result is still weak.
-
LookThere! Sparse Vision by Reinforced Selection
LookThere jointly trains a lightweight selector and a deeper vision model with reinforcement learning, so the system learns which image tokens are worth processing for each task rather than relying on attention or hand-designed diversity heuristics. It reports maintaining accuracy while processing as little as 0.2% of the input across classification, segmentation, zero-shot recognition, and counting, especially in high-resolution images.
Extreme, task-specific token sparsity across several vision tasks could be a meaningful efficiency advance, but the abstract gives no quantitative comparisons or details sufficient to establish that the claimed Pareto improvement is robust rather than an incremental adaptive-computation variant.
-
Compression Beyond the Uncompressed: A Two-Stage Training Recipe for Soft Context Compression in RAG
The paper proposes DEX-Comp, a two-stage method for compressing retrieved documents into much shorter soft-token sequences. It first distills only correct answers from an uncompressed RAG system, then uses reinforcement learning on queries where that teacher fails, and reports 16× context compression with 4×–24× faster inference while matching or exceeding the uncompressed baseline on five QA benchmarks.
The notable idea is training specifically on the uncompressed system’s failures to discover behaviors better suited to compressed contexts, paired with a substantial efficiency claim, but the abstract lacks enough quantitative detail to establish how robust or broadly meaningful the gains are.
-
Democratizing three-dimensional surface phenotyping: an open structured-light platform reveals and removes the projection bias in biological imaging
The authors build an open, sub-$4,000 structured-light system for reconstructing biological surfaces in 3D, including hardware, software, and analysis workflows. Across leaves, mouse anatomy, and insect specimens, they report that 2D projection underestimates surface area by 11–41% and can substantially distort measured geometry, such as limb angles, depending on viewing direction.
The combination of an unusually accessible end-to-end 3D phenotyping platform and evidence that routine projection measurements can introduce biology-scale errors is genuinely useful, though the evidence is limited to a few example applications and the platform’s broader robustness is not established.
-
KAT3 Shuttling Between Neuronal Identity and Activity-Dependent Plasticity Programs Drives Large-Scale Chromatin Remodeling
This study finds that neuronal activation causes CBP/p300 (KAT3) cofactors to move from enhancers maintaining neuronal identity to enhancers controlling activity-dependent plasticity genes. Proneural factors such as NeuroD2 and activity-induced AP-1 factors appear to regulate this redistribution, producing reversible changes in histone marks, chromatin accessibility, and 3D genome organization; FOS overexpression alone can reproduce repression of identity genes.
The proposed competition for shared CBP/p300 cofactors offers a specific, potentially general mechanism linking neuronal activity to coordinated large-scale chromatin remodeling, though the abstract provides mainly molecular evidence without enough quantitative or functional detail for a strong verdict.
-
PURELIGHT: a quantitative photon-counting framework unifying intensity and lifetime imaging at video rate across detector technologies
PURELIGHT is a hardware/software approach for correcting photon pile-up in time-correlated single-photon counting, allowing intensity and fluorescence-lifetime measurements to remain quantitatively accurate at much higher count rates. It works across several detector types and reportedly enables video-rate, subcellular lifetime imaging in awake mice, while retaining over three times as many photons as competing corrections and supporting new ratiometric and multiplexing modes.
The combination of detector-agnostic pile-up correction with video-rate lifetime imaging in awake animals is a substantial microscopy capability, but the abstract provides limited quantitative validation beyond the headline photon-retention and demonstration claims.
-
Spatial-filtering nanoscopy for 40-nm label-free Raman imaging
The paper introduces spatial-filtering nanoscopy, combining a microsphere lens with confocal Raman microscopy to improve resolution by narrowing excitation and suppressing background rather than amplifying weak Raman signals. It reports roughly 40 nm lateral resolution and label-free chemical imaging of nanostructures, cells, and tissue, including subcellular features.
A potentially important and conceptually distinct route to nanoscale, label-free chemical imaging, but the extraordinary 40 nm claim needs close inspection of calibration, artifacts, contrast, and reproducibility before warranting a strong recommendation.
-
Sensory cortical disinhibition is required for long-term fear memory in humans
This study combines EEG-fMRI, neural decoding, and targeted stimulation to test whether early visual cortex contributes to lasting fear memories. It reports that conditioning-related visual-cortex disinhibition predicts memory at 24 hours and 7 days, while stimulation intended to inhibit V1/V2 disrupts later consolidation and retention without affecting initial learning. The notable claim is that early sensory cortex is not merely a passive analyzer but a causal contributor to long-term fear-memory formation.
The causal manipulation of early visual cortex in humans, paired with longitudinal memory decoding, is a non-obvious and potentially important result, but the abstract gives no sample sizes or quantitative effects and the specificity of tACS-based inhibition remains uncertain.
-
AGC kinase homology requires and enables co-targeting for CNS regeneration
The study argues that CNS axon regeneration requires simultaneous inhibition of several closely related AGC-family kinases, rather than targeting one kinase alone. Using phenotype-guided optimization, the authors developed TMP-316, which co-targets these kinases and produced sustained motor recovery after a single intrathecal dose in a rat spinal-cord injury model, with supporting neurite-growth results in rodent and human neurons.
The potentially generalizable idea that shared evolutionary architecture can make deliberate polypharmacology both necessary and achievable is notable, but the abstract provides limited quantitative evidence and only a narrow set of injury models.
-
Memory encoding begins with spontaneous neural reactivation
Using MEG from 67 people learning short melodies, the authors decoded stimulus-related neural patterns during encoding and found that similar patterns continued spontaneously during the following rest period. Reactivation was associated with later recognition, especially after a 10-day delay, and appeared to depend more on hippocampal and broader associative-memory networks than auditory regions.
This provides a relatively direct challenge to the usual view that replay is mainly a post-learning sleep/rest or consolidation process, but the abstract gives limited methodological detail and no effect sizes, so it merits a look rather than a strong recommendation.
-
Natural prey capture reveals flexible reconfiguration of canonical motor-cortical dynamics for online control
The authors recorded marmoset sensorimotor cortex during natural live-prey capture, contrasting fast ballistic captures with pursuits that require continuous correction. Both behaviors remained in a shared low-dimensional dynamical subspace, but corrective pursuit movements shifted into a slower, feedback-dominated regime with weaker premovement predictability and more tangled trajectories. The work suggests canonical motor-cortical dynamics are not replaced during natural behavior but are flexibly reconfigured for online control.
Naturalistic neural recordings reveal a specific, non-obvious reconfiguration of canonical motor dynamics during feedback-rich behavior, but the evidence is from one behavioral paradigm and the abstract provides limited quantitative detail.
-
Cortical isolation separates rhythmic synchrony from network integration in the human neocortex.
The authors compare laminar field and single-unit activity from the same human cortical regions during wakefulness, NREM sleep, and after acute surgical isolation. Isolated cortex retained gamma-band coordination and laminar organization, but lost population coupling, long-range integration, and dynamical dimensionality, suggesting that local circuits can generate coherent gamma while brain-wide inputs support richer, higher-dimensional activity.
This is a relatively direct human test separating intrinsic cortical rhythm generation from network-dependent integration, with an unexpected dissociation between stronger synchrony and less complex computation, but the evidence is based on only eight patients and an unusual ex vivo comparison.
-
Select, Compress, Reinvest: A Controlled Study of Visual-Token Allocation in Long-Video MLLMs
This controlled study isolates how long-video multimodal models spend a fixed visual-token budget across frame selection, spatial compression, and adding more frames. Query-based selection mattered most: eight selected frames outperformed sixteen uniform frames by 6.9 points on an hour-long LongVideoBench setting, while simple Orthogonal Matching Pursuit matched specialized selectors; halving spatial detail was nearly free when the savings were reinvested in more frames. The study also finds substantial evaluation-harness discrepancies and identifies a bug in a published baseline.
The non-obvious result that an old generic sparse-approximation method matches specialized selectors, together with a careful controlled budget analysis, is useful for long-video MLLMs, though it remains an empirical optimization study rather than a major new capability or method.
-
Iris: Climbing to the Search Frontier
Iris trains large ReAct search agents using automatically constructed multi-hop questions, filtered tool-use trajectories, and alternating SFT with RL on live web search. It also emphasizes request-level rollout recovery and inference-time context management, reporting strong results on BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE without sub-agents or test-time verification. The main contribution appears to be an integrated data-generation and training recipe rather than a fundamentally new agent architecture.
The automated hard-question construction and SFT-RL iteration could be useful for training search agents, and the reported results are substantial, but the abstract does not isolate which component drives the gains or provide enough comparative detail to establish a major advance.
-
When Models Edit Too Much: On the Fidelity of Minimal Code Edits
This paper studies whether code-repair models change more of a program than necessary, using AST-level corruptions of 400 BigCodeBench solutions where the minimal fix is known. It finds that frontier models often over-edit, and that a preservation instruction reduces excess edits and complexity; reinforcement learning appears to generalize better than supervised fine-tuning to unseen corruption patterns.
The explicit treatment of edit fidelity as a separate code-repair objective, plus evidence that RL improves out-of-distribution minimality, is a useful and non-obvious direction, though the evaluation is limited to synthetic corruptions and a moderate-sized benchmark.
-
StreamScout: Learning When to Look Deeper for Streaming Video Understanding
StreamScout answers questions about an ongoing, unbounded video stream by starting with a cheap textual timeline and escalating only when needed: recent-frame inspection, uniform historical sampling, and query-focused retrieval. Its learned stop-or-escalate policies improve accuracy while reducing computation; on OVO-Bench, one variant reportedly gains 14.65 points for Qwen3-VL-8B, uses 59% fewer tokens than uniform sampling, and answers in 1.04 seconds on average.
The query-adaptive depth of memory access is a meaningful and potentially useful direction for real-time streaming video, supported by results across three backbones and benchmarks, but it remains an inference-policy framework rather than a clearly fundamental new video architecture or capability.
-
Unmasking Face Embeddings: Reading, Rendering and Naming with Foundation Models
The paper aligns existing face-recognition embeddings with foundation-model embeddings using only a learned linear map from paired examples. It claims that the aligned representation can support natural-language description and search, reconstruct a face with an unchanged diffusion decoder, and infer a person’s name without an enrolled gallery. The central finding is that identity embeddings retain substantially more interpretable and visually recoverable information than their usual biometric-matching interface suggests.
The proposed interoperability and the claimed gallery-free naming/reconstruction capabilities are genuinely notable, but the abstract provides no quantitative results, scale, or comparisons to establish how reliable these capabilities are.
-
HBQ: Hierarchical Scaling Block Quantization with Hardware-Efficiency-Aware Design for Accurate LLM Inference
This paper explores the hardware/accuracy tradeoff in block quantization for LLM inference and proposes Hierarchical Block Quantization (HBQ): large quantization blocks for efficient hardware, combined with a second-level significand-based scale to recover accuracy. In a 28nm ASIC implementation, the authors report W4A5 accuracy comparable to W4A16, 1.6–3.3× lower system energy, and 1.5–3× speedups over prior block-quantization methods, including quantization of weights, activations, and KV cache.
The combination of large-block quantization with hierarchical significand scaling and a fabricated end-to-end accelerator is a substantive hardware–algorithm co-design result, though the unusually strong gains depend on ASIC-specific comparisons and details not available in the abstract.
-
Revisiting Cross-View Completion: Self-Supervised Pre-Training via Reconstruction Error Comparison
Gekko augments cross-view self-supervised learning with a masked-autoencoder comparison: the relative reconstruction improvement estimates which image regions are actually co-visible, providing a binocular signal even where views do not overlap. It improves several 3D-vision tasks over CroCo under matched conditions, including a reported 6x gain at the strictest pose threshold and 22% lower ETH3D endpoint error, and can be trained from raw videos without 3D preprocessing.
The co-visibility proxy and raw-video training route are substantive additions with credible multi-task results, but the claims remain within an established cross-view-pretraining paradigm rather than clearly establishing a field-changing capability.
-
Emergence of Fibrations, Compression, and Symmetry Breaking in Artificial Neural Networks
The paper argues that training neural networks causes repeated local structures—graph-theoretic fibrations and coverings—to emerge as stable outcomes of stochastic gradient descent. It reports finding these symmetries across MLPs, CNNs, RNNs, and transformers, using them to compress models to 17% of their original size without performance loss and to improve continual learning by deliberately breaking the symmetry.
The symmetry-based explanation and its proposed links to compression and plasticity are unusually broad and potentially important, but the abstract provides few experimental details and makes ambitious claims such as state-of-the-art continual learning without enough evidence for a stronger verdict.
-
FOCUS: Foot Observation Confidence for Robust Humanoid Proprioceptive Odometry
FOCUS replaces binary foot-contact decisions in humanoid proprioceptive odometry with a learned continuous confidence score for each foot. Trained from automatically generated simulation signals, the model adjusts FK observation blending and EKF covariance so that partial support, toe dragging, and slip are handled smoothly; it reports large ATE reductions in simulation and on real walking and dynamic-motion sequences.
The continuous, label-free reliability estimation is a meaningful improvement over hard contact switching, and the reported gains transfer to real humanoid motion, but it remains a focused estimator enhancement rather than a clearly new robotics capability or broadly validated breakthrough.
-
Sensing Which Modality Matters: Evidence-Gated Regularization for Robust VLA Policies
The paper identifies “modality entanglement” in vision-language-action policies: models trained on limited demonstrations can rely on spurious correlations between sensors, making them brittle when sensors are occluded or distracted. Evidence-Gated Regularization trains the policy to ignore low-evidence modalities while making high-evidence modalities individually sufficient, with no inference-time cost. It reports large relative gains under corrupted or missing sensors in simulation and on two real-robot platforms, although absolute success remains modest in several settings.
The state-conditional combination of modality invariance and single-sensor sufficiency, supported by both simulation and two physical embodiments, is a substantive robustness idea, but the low baselines and limited reported evaluations make it worth a closer look rather than an obvious must-read.
-
HaptiNet: Networked Haptic Robots Enable Physical Co-presence in Geographically-Unconstrained Rehabilitation
HaptiNet is a distributed robotic rehabilitation system that lets people in different locations exchange forces and coordinate movements through networked haptic robots. An imitation-learning delay compensator supported laboratory, cross-city, and clinical tests with 284 healthy participants and 111 patients; haptic cooperation reportedly improved task performance, engagement, synchrony, and training gains over solo or visual-only conditions across links totaling about 4,000 km.
The combination of low-latency networked force interaction, large-scale human evaluation, and intercity testing in neurological patients is a meaningful real-world robotics capability, but the abstract does not establish how robust the clinical benefits are or how much depends on the specific tasks and comparison design.
-
CUA-Universe: A Scalable and Dynamic Environment for Hybrid GUI+CLI Agents
CUA-Universe builds reproducible desktop environments where agents can combine GUI actions with CLI commands over shared application state. It provides tools to adapt 16 applications, synthesize hybrid tasks, and collect efficient trajectories; training a 9B model reportedly improves success while substantially reducing steps and tokens on its own benchmark and OSWorld variants.
The scalable GUI+CLI environment and trajectory-generation pipeline address a real limitation in computer-use agents, with sizable reported efficiency and success gains, but the evidence is primarily benchmark-based and the contribution may be more infrastructure/data engineering than a broadly new agent capability.
-
Compression Sequencing enables ultra-sensitive and scalable scRNA-seq
The paper introduces a sequencing workflow that applies an approximately logarithmic, abundance-preserving transformation to molecular counts before sequencing, reducing oversampling of abundant transcripts and enriching rare ones. In scRNA-seq tests, the authors claim 2–5× more detected UMIs, up to 200× lower sequencing cost, and comparable cell-type and differential-expression results across 500–2,000 genes, including AML samples. The central novelty is using abundance compression to trade redundant reads for broader molecular coverage.
The proposed logarithmic compression could be a genuinely important change to sequencing economics and rare-transcript sensitivity, but the abstract provides no experimental detail or quantitative baselines sufficient to validate its unusually large 200× cost-reduction claims.
-
Sparse Readout Prism: Explaining Logit-Lens Scores in Features Instead of Tokens
The paper argues that logit-lens interpretations depend partly on the corpus used to fit the lens, not just on the model’s hidden states. It introduces Sparse Readout Prism, which decomposes the unembedding matrix into sparse features so token logits and logit differences can be explained in terms of readout features rather than token identities; the decomposition reconstructs substantially more logit-difference structure than six geometric baselines, and feature importance is validated by ablation.
This is a genuinely useful new interpretability framing and offers a corpus-independent control for logit-lens analysis, but its demonstrated impact is currently methodological and narrow rather than a major advance in model understanding.
-
Evaluating Multimodal LLMs as Generalist Vision-Language-Action Agents for Drone Control: Commanding, Approaching, Tracking and Searching
The paper evaluates multimodal language models as zero-shot drone controllers across approaching, tracking, searching, and multi-drone command tasks, with the model choosing actions from a prompt-defined protocol. It reports that smaller open models can navigate to targets more reliably than frontier models, but often fail by declaring success too early, failing to terminate, or copying coordinates across different drone views. The main contribution is the DroneCATS-Agent architecture and benchmark for exposing these protocol-following failures rather than only measuring visual navigation.
The finding that high-end multimodal models fail more in action-protocol discipline than in basic spatial navigation is non-obvious and relevant to deployable embodied agents, but the abstract gives no quantitative results or evidence of real-world drone deployment.
-
Ignorance or Incompetence? Constructing Knowledge-Gated, Verifiable Tasks for LLM Agents
The paper proposes a protocol for building agent tasks where success genuinely depends on a supplied private artefact, such as conventions, reference tables, or utility rules, rather than on memorized knowledge or ambiguous instructions. Across 15 calibration tasks, one agent configuration passed 68% with the artefact and 0% without it, with executable solvers, rubrics, provenance checks, and leak audits used to verify the setup; seven tasks survived a stricter empirical screen. The work validates task construction rather than showing improved agent training or general reasoning.
The explicit separation of ignorance from incompetence and the construction-time verification machinery are useful, nontrivial benchmark methodology, but the evidence is small, configuration-relative, and does not demonstrate a downstream capability gain.
-
DICS: Exploring Data Intrinsic Consistency for Visual Instruction Selection
The paper proposes selecting visual instruction-tuning examples using “data intrinsic consistency”: agreement between the image and instruction, plus coherence between the instruction and response. Its selection method reportedly matches or exceeds full-data fine-tuning with 25% of LLaVA-1.5’s 665K examples, and reaches 94.52% of InternVL3-8B-Instruct performance using under 25% of the reported training data.
The potentially substantial reduction in visual instruction-tuning data is interesting, but the core idea is a plausible extension of existing data-quality and data-selection methods, and the abstract lacks enough experimental detail to establish how robust the gains are.
-
Contrast-Free Autonomous Navigation of Untethered Endovascular Microrobots Using Single-Plane Fluoroscopy
The paper presents VISTA, which combines a preconstructed 3D vascular digital twin with single-plane fluoroscopy to infer a microrobot’s progress along vessel centerlines and autonomously generate magnetic-control commands. In vascular phantoms and an inferior-vena-cava rat experiment, it reportedly reduced navigation time by up to 62%, corrective commands by 98%, and radiation exposure by 57% versus human-in-the-loop control, without contrast injection during navigation.
Using a topology-constrained digital twin to recover actionable 3D navigation states from a single 2D fluoroscopic view is a meaningful and clinically relevant direction, but the evidence is still limited to phantoms and one small-animal setting, with unclear robustness to anatomical variation and occlusion.
-
Exact Recovery Thresholds for Weighted Data Selection in Vector-Valued Linear Regression
This paper gives exact thresholds for how many weighted examples are needed to preserve the full-data least-squares loss in vector-valued linear regression: (m+1)d examples are necessary and sufficient in the worst case. It also characterizes several nearby regimes, obtains bounds for the smallest unresolved case, and disproves a recent claim that 2d examples always suffice when there are two output dimensions. The work introduces several geometric and sparsification tools, but its impact is mainly on a narrow theoretical data-selection problem.
It resolves a named open problem with exact worst-case thresholds and overturns a plausible 2d -point sufficiency claim, though the result is specialized and has no demonstrated practical consequence beyond regression coreset theory.
-
Tail-Replay: Escaping the Curse of Linear Attention in Prefix Caching for Hybrid LLMs
Tail-Replay enables token-level prefix caching for hybrid LLMs that combine full attention with recurrent linear-attention layers. Instead of storing recurrent-state checkpoints, it reconstructs the state by replaying only a recent suffix of the cached prefix, preserving 92.8–99.9% of full-prefill quality with a 5–10% replay budget and achieving 9.1–14.3× TTFT speedups at 32K prefixes.
The lossy-compression view of recurrent attention leads to a useful new caching strategy with potentially large serving gains, though the evidence is limited to three Gated DeltaNet hybrids and benchmark-based quality measurements.
-
No Equivariant Architecture Covers All Equivariant Attention
This paper characterizes when multi-head self-attention can be exactly equivariant to a symmetry group, showing that the group action must effectively permute clusters of attention heads while constraining the QK and OV parameters. It further argues that the valid parameter settings split into many algebraically distinct components, so any one polynomially parameterized equivariant architecture can represent only part of the full equivariant function class; for a D4 example with eight heads, it proves at least Ω(C^64) components.
The potentially important result is a structural impossibility for fixed equivariant-attention parameterizations, backed by an explicit algebraic characterization and a striking component-count lower bound, though its practical consequences for model design are not demonstrated here.
-
Graph Evidence Is Not Enough: Diagnosing Native Decoder Use in Graph-Augmented LLMs
The paper tests whether graph-augmented LLMs can actually use graph structure supplied in their prompts, rather than merely receiving it. On a bounded shortest-hop task, matched readable, shuffled, and absent-graph conditions show that evidence inclusion does not guarantee topology use; a query-aware ordering and structure-preserving interface improves exact match substantially across four datasets.
The intervention-based diagnosis and large reported gains make this more than a routine graph-LLM tweak, but the result is demonstrated on a narrow synthetic-like topology task, so its broader impact on graph reasoning remains uncertain.
-
Compressed Single-Tone Frequency Estimation With Unknown Complex Gain: Singular Rates and Global Identifiability
This paper analyzes when a fixed linear sketch of a noisy single complex sinusoid can still support frequency estimation when the signal gain is unknown. It distinguishes local Fisher-information preservation from global identifiability, derives sharp estimation rates near frequencies where the sketch response vanishes, and characterizes when exact local aliases occur. It also solves several worst-case compression-design problems, showing that rank-two sketches can be intrinsically aliased while sufficiently higher-rank sketches can be globally identifying without sacrificing worst-case information.
The paper offers a mathematically substantive and somewhat surprising separation between local information preservation, singular estimation rates, and global identifiability, with exact design results rather than a routine estimator tweak, though its impact appears confined to specialized compressed spectral estimation.