Issue 32 · Aug 3–9, 2026
Every candidate
All 3,436 papers were scored from their abstracts by gpt-5.6-luna; 1,312 were not skipped. Shown here: the top 150 of those, in score order. Picks are marked.
-
Recursive Synthesis for Long-Horizon Terminal Tasks
The paper proposes a verified recursive process for generating long-horizon terminal-agent tasks: extend a solved task, update its instructions and verifier, validate it in a fresh sandbox, and use successful outputs as seeds for further rounds. It generates 37,484 tasks at about $0.05 each, with task length and difficulty increasing substantially over 15 rounds, and training on resulting trajectories improves several terminal-agent benchmarks by up to 10 points. The main novelty is showing that verified task synthesis can recursively scale difficulty without an apparent degradation in yield or validation rate.
This addresses a major bottleneck for capable agents—expensive, internally consistent long-horizon training tasks—with a potentially general recursive data-generation strategy and substantial reported benchmark gains.
-
Reachability Is Not Realization: Tracing the Sources of LLM Benchmark Gains
The paper separates an LLM’s ability to produce an answer by default (“realization”) from whether the answer can be elicited under a fixed probing budget (“reachability”). It reports that random inference-time layer routes match or beat structured routing across 43 settings, that silencing a single MLP block repairs 68–92% of selected failures, and that training can improve deployed accuracy while leaving—or even reducing—the reachable ceiling. The authors argue that benchmark gains should report both quantities rather than treating every score increase as expanded capability.
This introduces a useful capability decomposition and supports an unusually strong, assumption-challenging result: training and higher benchmark scores can coincide with lower underlying answer reachability, backed by multi-setting routing, intervention, and training analyses.
-
Stable Density Ridges: Consistency and Convergence of Subspace Constrained Mean Shift
This paper argues that SCMS does not generally converge to the classical density ridge, because that definition ignores how the relevant Hessian eigenspaces rotate along the algorithm’s flow. It introduces a dynamical-systems-based “stable ridge” as the correct target, proves convergence and statistical estimation results for a generalized constant-step SCMS method, and identifies a polynomial-time inefficiency in the original algorithm’s bandwidth-coupled step size.
It appears to overturn a foundational assumption about what SCMS computes and replaces it with a theoretically justified target, while also providing convergence, consistency, and efficiency guarantees rather than merely tweaking the algorithm.
-
Scaling Inherently Interpretable Language Models
The paper trains language models with interpretability objectives built into the learning process rather than explaining them afterward. It reports that interpretability improves across three orders of compute, and presents an 8B diffusion model whose outputs can be traced to input tokens, human-readable concepts, and training examples, then altered through concept steering without retraining.
The central claim—that interpretability can scale positively with capability and support closed-loop behavioral correction—is unusually important and assumption-challenging, but the abstract gives few quantitative details to substantiate its broad comparisons and scaling claims.
-
Dirichlet Follow-the-Leader Closes the Gap in Simultaneous Multiclass U-Calibration
The paper proposes a very simple forecaster: after each round, sample the next class-probability vector from a Dirichlet distribution whose parameters are the observed class counts. It proves that this single, horizon-free method achieves optimal worst-case regret of order √(S_T T) for all bounded proper losses and logarithmic regret for every smooth proper loss, removing the dimension and smoothness gaps of prior methods.
This is an unusually clean theoretical advance: a one-line Bayesian-bootstrap algorithm simultaneously attains matching lower-bound rates across broad loss classes, with an exact telescoping analysis that also handles nonsmooth losses and changing active faces.
-
LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks
The paper treats long-horizon agent execution as explicit task-state management rather than letting state accumulate in the model’s context. Its Manage–Execute–Audit loop uses a manager, fresh-context executor, and independently verifying auditor, producing sizable reported gains across WeaveBench, Terminal-Bench, and OSWorld, including across different models and agent harnesses.
The explicit separation of planning state, execution, and environment-based verification is a meaningful and plausibly general agent design, and the reported cross-model gains are large, but the abstract provides too little detail to establish robustness or rule out benchmark and implementation effects.
-
Mind the Cap: Output-Budget Regimes Change the Measured Multilingual Reasoning Gap
The paper shows that apparent multilingual reasoning gaps on MGSM can largely be artifacts of the output-token budget: changing the cap shifts measured native-versus-translation gaps by up to 57 points, and length normalization can substantially change or even reverse comparisons. Across preregistered/frozen tests and 540,000 hard-capped decodes, the authors find that truncation, announced budgets, and language-specific tokenization all affect accuracy; once native accuracy saturates, the remaining difference is better described as a strategy-performance gap than an identified reasoning deficit.
This is a potentially assumption-breaking evaluation result, backed by unusually extensive frozen tests: single-budget multilingual reasoning comparisons may be measuring output-length and truncation effects rather than reasoning ability.
-
In-Context Collapse in Vision-Language Models and How to Mitigate it?
The paper reports that adding many image–label examples can cause some vision-language models to abruptly lose accuracy—even falling below chance—despite producing fluent, well-formed answers. Experiments across models and tasks attribute the failure to the vision–language integration pathway rather than the output layers, and a small adapter trained on one synthetic task reportedly transfers resistance to collapse on unseen tasks.
This identifies a potentially important, assumption-reversing failure mode in many-shot VLM inference and provides unusually strong-looking causal localization plus a transferable mitigation, though the headline claims warrant verification.
-
Agogic: Performance-Timed Music Tokens for LLM-Native Text-to-Symbolic-Music Generation
This paper isolates music tokenization as a variable by holding the backbone, data, training budget, and decoding fixed across seven representations. A 0.8B model using performance-resolution tokens achieves substantially better distributional fidelity than 27B models using beat-grid tokens, and the result persists across another backbone and tokenizer; a separate decoding constraint improves caption adherence without harming fidelity. The study also releases an evaluation harness, checkpoints, and large aligned music corpora, though whether the gains are audible remains untested.
The controlled result that representation can matter far more than a 34x increase in model size is a genuinely non-obvious finding with credible cross-backbone, cross-tokenizer, and anti-resolution-artifact checks.
-
An Optimal Agnostic PAC Algorithm
The paper constructs an agnostic PAC learner for any binary hypothesis class of finite VC dimension whose excess risk scales as the statistically optimal rate, including the favorable dependence on the unknown best-class risk L*. This closes the gap between known upper and lower bounds for agnostic learning up to universal constants, though the stated constant is extremely large and the abstract does not discuss computational efficiency.
It claims to resolve a fundamental longstanding sample-complexity question in statistical learning theory by attaining the optimal L*-dependent agnostic PAC rate for arbitrary finite-VC classes, rather than offering an incremental algorithmic improvement.
-
The Replay Gap: Static Evaluation of Model Switching in LLM Agents Scores the Wrong World
This paper tests whether logged agent trajectories can be replayed with a different model to evaluate per-step LLM routing. In roughly 900 branching SWE-bench rollouts, model swaps caused substantial trajectory divergence—often at the very next action—while replay-based scoring mispredicted all observed success-relevant outcome changes; it also shows that apparent temperature-0 determinism depends strongly on serving configuration. The main contribution is empirical evidence that static replay evaluates a counterfactual world unlike the one an agent would actually experience.
It challenges a widely used evaluation assumption for agentic model routing with controlled branching experiments, outcome flips, and a measured noise baseline, although the evidence is centered on SWE-bench and a limited set of runs.
-
ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment
The paper trains long-horizon web-search agents by tracing a known answer backward to identify the intermediate clues needed to solve it, then assigning credit to individual search steps rather than treating the whole trajectory as simply right or wrong. A 4B-parameter model trained on 8.5k examples reaches 37.3% on BrowseComp and 39.1% on BrowseComp-ZH, rising to 55.3% and 52.9% with context management; the authors report performance comparable to much larger agents.
Answer-backtracked, clue-level credit assignment is a meaningful approach to the sparse-reward problem in search agents, and the reported small-model results are notable, but the abstract does not establish how much comes from the credit assignment versus context management or provide enough detail to justify a strong verdict.
-
Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes
This paper systematically studies how language, visual understanding, and visual generation interact during unified multimodal pretraining. It reports asymmetric knowledge transfer, finds that modality synergy depends on data complexity and architectural sharing choices, and argues that integrating modalities from the beginning avoids “vision laziness” caused by language priors. The authors also claim efficient recipes that retain strong generative performance with 5% of the compute, validated using 13.5B-parameter MoE models trained on 2T tokens.
The combination of controlled analysis, an apparently non-obvious early-unification result, and a potentially large compute reduction merits attention, but the abstract gives too few quantitative comparisons to justify a strong recommendation.
-
JoyAI-Video-Edit: Real-Time Open-Ended Video Editing with Autoregressive Diffusion
JoyAI-Video-Edit is a 16B-parameter causal video-editing system that generates video chunk by chunk without seeing future frames or knowing the eventual duration. It combines source-anchored distillation and long-horizon autoregressive distillation to reduce source degradation and temporal drift, reportedly reaching 720p at about 30 FPS on one Nvidia B200 while matching or outperforming streaming and offline systems.
The combination of open-ended causal editing, long-horizon consistency, and claimed 720p real-time performance is a meaningful capability advance, but the abstract gives no quantitative comparisons or latency/quality details beyond broad claims.
-
Hunyuan3D-Buffalo 1.0: A Unified Multimodal Model for Scalable 3D Generation, Understanding, and Editing
Hunyuan3D-Buffalo combines 3D understanding, text-to-3D generation, instruction-based editing, and part generation in one model. It trains on an 87-million-example multimodal corpus, including large amounts of automatically generated editing data, and reports leading results across generation and editing benchmarks. The authors also find that joint generation and understanding training improves editing quality.
The unified 3D model and unusually large multimodal/editing corpus could be an important step toward general-purpose 3D systems, but the abstract gives no quantitative results and much of the architecture appears to combine existing Hunyuan components with synthetic data.
-
Non-KKT Accumulation in Entropic Mirror Descent
The paper constructs smooth counterexamples showing that entropic mirror descent can have bounded, decreasing-objective sequences whose accumulation points include non-KKT boundary points, even with standard diminishing stepsizes and entropy-relative smoothness. The mechanism is the degeneracy of the entropy geometry at the boundary: dual-feasibility information can vanish as coordinates approach zero; the authors realize this with accumulation sets containing smooth boundary circles and non-KKT arcs.
This directly overturns a basic expectation about convergence of mirror descent and isolates a subtle boundary-degeneracy failure, but the result is a deliberately constructed theoretical pathology rather than a demonstrated failure in widely used optimization workloads.
-
Differentiate the Solver, Not the Equation: Reverse-Sweep Adjoints for Block Implicit Simulation
The paper differentiates the numerical solver that actually ran, rather than the ideal converged equations or every iteration through generic autodiff. For block-implicit simulation, the backward pass reverses the forward local updates, requiring only local 3×3 adjoint solves; it matches unrolled autodiff exactly while being reported as 33× faster and 71× lower-memory, and scales elastodynamics to 8 million vertices on one GPU.
This combines a genuinely useful solver-level differentiation formulation with unusually large efficiency and scale claims, while showing that equation-level differentiation can substantially disagree with the finite computation actually executed.
-
CalibForge: Adversarial Solver Calibration for Scaling Learnable Terminal Tasks
CalibForge generates executable terminal tasks and iteratively adjusts them based on how different agent solvers perform. It targets tasks that are solvable by a stronger solver but fail for a weaker one, producing 5,431 calibrated tasks; training on them reportedly improves Terminal-Bench 2.0 substantially and also transfers to SWE-bench Pro and Doc2Repo.
The solver-relative notion of task difficulty and adversarial calibration is a meaningful direction for scaling agent-training data, with unusually large reported cross-benchmark gains, but the abstract provides limited detail about baselines, contamination controls, and how much of the improvement comes from the data scale or task distribution.
-
Full-bandwidth transformer
The paper adds a recurrent latent channel to an otherwise standard autoregressive transformer: the previous step’s final hidden state is gated together with the next token embedding and fed back into the network. A scheduled multi-pass training objective makes this compatible with teacher forcing, and experiments on 1B-parameter models trained on 400B tokens reportedly show better loss, several downstream capabilities, and shorter reasoning traces, comparable to training a standard model on about 1.5× more data.
This is a genuinely nonstandard way to increase inter-token computational bandwidth and could improve data efficiency, but the abstract gives no quantitative results or comparisons detailed enough to establish that the claimed gains are robust.
-
Deferred Exposure of Future Trajectories for Verifiable Reasoning in Autonomous Driving VLMs
The paper argues that showing a teacher model the logged future trajectory while generating driving chain-of-thought causes it to rationalize the outcome rather than reason from the scene. It introduces AD-MCQ, which frames planning as choosing among candidate trajectories, and DEFT-RLVR, which withholds future trajectories until after the decision so they serve as verification targets; the authors report better reasoning and fewer hallucinations, but provide no quantitative results in the abstract.
The deferred-exposure formulation targets a real causal-labeling problem in driving VLMs and is a non-obvious use of trajectory candidates for verifiable reasoning, but the abstract lacks enough experimental detail to establish the size or breadth of the improvement.
-
OasisKV: Scaling In-Decode KV Cache Beyond HBM with Lookahead Sparse Prefetching
OasisKV stores the full KV cache in larger, slower memory tiers while keeping only predicted-important KV blocks in HBM during decoding. It uses speculative-decoding lookahead tokens to predict which blocks will matter next, prefetches them in the background, and reports up to 1.69× throughput on reasoning workloads and 2.1× on multi-GPU long-context serving, with accuracy within 0.7 points under a 2,048-token KV budget.
The combination of lookahead-based KV importance prediction, tiered-memory offload, and background prefetching addresses a major practical bottleneck with substantial reported throughput and capacity gains, though it builds on established sparse-attention and KV-offloading ideas and the abstract does not establish broad robustness.
-
SFT Conflicts, RL Coexists: A Theoretical and Empirical Analysis of Multi-Task Learning for LLMs
This paper compares supervised fine-tuning and reinforcement learning for training LLMs on multiple reasoning tasks. It reports that SFT updates interfere across tasks, while RL produces sparse, nearly orthogonal updates, and attributes the difference to norm-limited versus variance-limited gradient interference; it then uses this finding to design a more parallelizable multi-task RL procedure.
The claimed distinction between SFT and RL gradient interference, together with a mechanistic explanation and training paradigm, is genuinely interesting and somewhat assumption-challenging, but the abstract provides no quantitative results or evidence of broad superiority.
-
The Loss Does Not See the Basis, but Adam Does
This paper argues that low-rank implicit bias in factored models depends not just on the loss, but on whether the optimizer respects the model’s gauge symmetry. It gives a characterization of memoryless gauge-equivariant updates, shows why coordinate-wise methods such as Adam and RMSProp can select different solutions from equivalent parameterizations, and supports the analysis with matrix-sensing, transformer, and hyperspectral experiments. The reported effects are substantial: gradient descent improves held-out error by 43–44% in one setting, while Adam produces markedly different transformer invariants from gauge-equivalent initializations.
The combination of a symmetry-based explanation, an optimizer-classification theorem, and sizable empirical consequences makes this more than an optimizer tweak, though the broad practical significance depends on how general the demonstrated effects are beyond the studied factored models and datasets.
-
Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution
Ouroboros is an agent framework that lets its tools, prompts, context construction, and implementation evolve through reviewed code changes that are later used at runtime. It reports very high scores on Terminal-Bench 2.1 and OSWorld-Verified, plus a 161-day publicly documented deployment in which human interactions generate improvement proposals while the agent chooses which changes to make. The potentially novel idea is treating the agent harness itself as an evolving, continuously deployed artifact rather than fixing the architecture during training or evaluation.
The self-evolving runtime and long-lived deployment are genuinely unusual, but the abstract provides little detail on the mechanism, ablations, independent verification, or whether the benchmark gains come from the evolving design rather than the underlying Opus 5 model.
-
A Polarized Histamine-GABA Core-Rim Architecture within Synaptic Vesicles
The paper uses specialized electron-microscopy labeling to claim that histamine and GABA occupy different radial positions within the same conventional GABAergic vesicles: histamine in a dense inner core and GABA near the membrane. If reproducible, this would challenge the usual separation between clear synaptic and dense-core vesicles, but the abstract provides structural observations rather than direct evidence that the arrangement changes transmitter release.
The proposed histamine-core/GABA-rim organization is a genuinely surprising challenge to vesicle-class conventions, but the evidence described is limited to an unconventional ultrastructural labeling method and does not yet demonstrate functional co-release or exclude labeling and preparation artifacts.
-
EnvACE: Internalizing Environment Dynamics via World Rehearsal for Agentic Reinforcement Learning
EnvACE trains an LLM agent to rehearse the likely response to each tool call internally, rather than relying on an external executable environment during every training step. The same model generates both actions and simulated environment responses, and the authors report improved performance across several tool-use benchmarks, plus gains from private rehearsal at test time.
The central idea—jointly training an agent to act as both policy and internal environment model—could reduce the cost of agentic RL and enable test-time planning, but the abstract gives no quantitative margins or evidence that self-generated rehearsal remains grounded rather than reinforcing model errors.
-
Ego2Robot: Scalable Robot Data Synthesis from Egocentric Human Data
Ego2Robot builds a pipeline that turns egocentric human manipulation videos into synthetic robot demonstrations using action retargeting, robot-arm rendering, and quality filtering. It produces 18,561 hours across 15 robot embodiments, and joint pretraining with real robot data reportedly improves robustness to changes in appearance, layout, morphology, and task semantics, including on real robots.
The scale and systematic use of human video for cross-embodiment robot-policy pretraining are potentially important, but the abstract gives no quantitative gains or details sufficient to establish that synthetic data delivers a major capability jump.
-
An AI4AI Framework for Visual Token Pruning
The paper uses an LLM to design visual-token pruning policies for multimodal language models. Its AutoPrune framework searches over a domain-specific language of 131 pruning operations, representing candidates as residual changes to a strong baseline policy; across 14 benchmarks and three backbones, it reportedly removes 94.4% of visual tokens while retaining over 99% of full-token performance, with 9.9x lower FLOPs and 6.4x lower prefill latency.
The residual DSL formulation for LLM-driven, training-free pruning search is a meaningful automation idea with unusually large reported efficiency gains, though the abstract lacks baseline details and enough information to judge how much is due to the search method versus an especially favorable pruning setup.
-
Small Foundation Models of Human Cognition and Behaviour
The authors train 14 language models ranging from 135M to 14B parameters on 10.7 million trial-level human choices across 160 psychology experiments. Small models match a 70B model on in-distribution participant prediction, while larger models help substantially on novel task structures; masking stimuli and feedback also shows that performance is not explained by choice history alone. The paper proposes these relatively small models as empirical noise-ceiling estimators for behavioral experiments, while finding that their generalization remains limited by the tasks seen during training.
The combination of scaling results and controlled information-use diagnostics yields a non-obvious finding—model size matters little for interpolation but much more for structural generalization—with reasonably broad evidence, though the conclusions remain bounded to one behavioral dataset and modeling setup.
-
Progressive Agent Skill Generation via Reinforcement Learning
Skill-α treats agent-skill writing as a sequence of edits and uses reinforcement learning to evaluate each edit by comparing downstream agent behavior before and after the edit. Across document-to-skill and experience-to-skill settings, it improves success rates over the strongest baseline by 3.3 points on CL-Bench and 6.7 points on τau2-bench, with ablations supporting the rollback reward and progressive editing.
The edit-level rollback reward is a concrete and potentially general solution to the lack of direct supervision for skill quality, but the reported gains are moderate and limited to a small set of agent benchmarks.
-
Optimal Learning Under Tsybakov Noise
This paper closes a roughly 20-year logarithmic gap in the sample/error guarantees for PAC learning under Tsybakov label noise, matching the best known lower bound for general concept classes. The algorithm adaptively partitions the input space by estimated noise levels and enforces region-specific error constraints, building on ideas from recent non-realizable learning theory.
Resolving a long-standing optimality gap with a new adaptive-partitioning technique is theoretically important, but the abstract gives no quantitative theorem statement or experimental/theoretical detail beyond the claimed matching bound.
-
AURORA-LM: Autoencoding Unified Representation for Continuous-Latent Diffusion Language Modeling
AURORA-LM builds a continuous, decodable latent representation of text and models it with a block-causal diffusion/flow-matching Transformer. It generates blocks left to right while denoising tokens within each block in parallel, using several techniques to preserve latent capacity and make training/inference trajectories consistent. The authors report leading results among evaluated continuous and diffusion language models on OpenWebText generation and XSum, plus gains from scaling to 1B parameters, though the abstract gives no numerical comparisons.
The separation of high-capacity text autoencoding from distribution modeling, combined with block-causal latent diffusion, is a substantive direction for non-autoregressive language modeling, but the abstract provides insufficient quantitative evidence to establish a major capability or efficiency breakthrough.
-
SkillZip: Contract-Preserving Graph Compression for Scalable Agent Skill Libraries
SkillZip compresses reusable agent skills as execution graphs rather than as plain text. It replaces recurring, contract-valid graph motifs with reversible macros, aiming to preserve dependencies, verifier access, executability, and later source expansion while loading only the needed context; experiments report up to 12.2-point gains, 3.46× compression, and robust scaling to 100K skills. The idea is a procedural, execution-aware unit for skill reuse rather than package-level retrieval and text compression.
The contract-preserving graph abstraction and incremental re-compression are a meaningful direction for scalable agent memory, but the abstract gives too little detail about baselines, tasks, and whether the reported gains reflect a broadly important capability rather than a system-specific benchmark result.
-
Any-OPD: Heterogeneous On-Policy Distillation for Flow-Matching Models via Representation-Space Bridging
Any-OPD proposes distilling one latent flow-matching image generator into a structurally incompatible student by comparing their decoded samples in a frozen vision-representation space, rather than matching latents or internal features. It also aligns trajectories by continuous noise level and uses a short re-encoding warm-up; on distilling FLUX.1-dev into SD3.5-Medium, the student’s PickScore rises from 0.846 to 0.884 and HPSv3 from 9.12 to 10.97, while direct latent regression reportedly fails.
The ability to perform on-policy distillation across different VAEs, architectures, and timestep schedules is a genuinely useful framing with an unexpectedly large reported gain, but the abstract provides evidence from only one teacher–student pair and limited metrics, so generality and robustness remain unclear.
-
Relevant but Incomplete: Referential Dangling as a Paradigm-Level Failure Mode in Hard Prompt Compression
This paper identifies “referential dangling”: hard prompt compressors retain an answer-bearing passage while deleting the passage that defines the entity or relationship needed to interpret it. Across several QA datasets and six compressors, this occurs frequently—up to 60% of cases—and restoring the missing supporting paragraph recovers much of the lost accuracy; a lightweight restoration classifier gives a 4.7-point improvement at nearly unchanged compression. The main proposed lesson is that compression should preserve referential completeness, not just independently score relevance.
The specific failure mode is a useful and plausibly general diagnosis of hard prompt compression, backed by broad measurements and a substantial intervention effect, but the proposed remedy and overall impact appear incremental rather than a clearly new compression paradigm.
-
SciCode-Verified: How Benchmark Defects Underestimated the Scientific-Coding Ability of Language Models
The authors audit all 65 SciCode scientific-programming problems and report 263 defects, including non-reproducible reference answers, overly strict tolerances, and contradictory specifications. After correcting the benchmark and independently rechecking the changes, twelve frontier model evaluations improve dramatically: subproblem accuracy rises from 45–60% to 84–98%, and main-problem accuracy from 9–27% to 69–92%.
The reported benchmark-wide correction reverses the apparent conclusion that model scientific-coding ability had plateaued, but the very large gains require scrutiny of the audit methodology and whether the revised tests remain independent and appropriately difficult.
-
UEmbed: Unified Sparse and Dense Multimodal Embeddings
UEmbed is a decoder-only multimodal embedding model that produces both dense vectors and learned sparse lexical representations in a single causal pass. It uses multiple learned tokens whose hidden states predict disjoint vocabulary partitions, and reports competitive or leading results on MMEB-v2 and BEIR at 2B–9B parameter scales, along with efficiency and agent-use cases.
The unified decoder-only construction for dense and sparse multimodal retrieval is a meaningful architectural direction, but the abstract provides limited comparative detail and does not establish a major capability or efficiency jump.
-
Invisible Shortcuts: Why Vision Encoders Know Your Camera
The paper argues that vision encoders learn imperceptible pixel-level traces left by cameras and image-processing pipelines, because these metadata cues become correlated with semantic labels or captions during pretraining. Controlled experiments show that stronger such correlations increase metadata sensitivity and hurt performance when metadata distributions shift; mitigation reduces sensitivity to both targeted and unseen metadata, while also affecting generated-image detection.
This identifies a less obvious shortcut-learning mechanism with implications for pretrained vision models and distribution shift, but the abstract gives no quantitative scale or evidence that the effect is broadly consequential enough for a stronger verdict.
-
MASS: Multiplayer World Models with Authoritative Shared State
MASS separates multiplayer world modeling into a learned Logic Engine that maintains a shared typed state and a Rendering Engine that produces each player's view, rather than storing view-dependent video latents as the recurrent memory. On a matched multiplayer Snake benchmark, it reportedly improves state accuracy and cross-view consistency and can roll out 1,024 players for 10,000 steps. The main new idea is importing an authoritative-server-style representation into learned video world models without hand-coded dynamics.
The explicit learned authoritative state is a meaningful and somewhat surprising architectural direction for scalable multi-agent world models, but the evidence is limited to a narrow synthetic Snake environment and does not yet establish broad usefulness.
-
Activity Frames: Deterministic Screen-Activity Compilation for Agent Memory and Replay
The paper introduces a deterministic, model-free compiler that turns raw screen-activity logs into compact, typed episodes with links back to the underlying evidence, for agent memory and replay. On one user’s 51-day corpus, it achieved 86× compression, answered activity questions at 98.4% accuracy, and identified recurring routines that could sometimes be replayed with zero model inference; however, the evaluation is very narrow.
The deterministic activity-memory and zero-token replay framing is genuinely interesting and challenges the assumption that agent memory must be model-generated, but the single-user evaluation and limited evidence make the broader cost and generality claims uncertain.
-
InfiniSplat: Implicit Gaussian Decoding for Large-Baseline Monocular View Synthesis
InfiniSplat generates a 3D Gaussian Splatting scene from a single image, but places Gaussian supports using estimated surface geometry rather than fixed image-grid pixels. A query-conditioned decoder then predicts Gaussian properties at those supports, reportedly improving large-viewpoint rendering and generalizing from synthetic indoor training to open-world scenes.
The shift from pixel-aligned to geometry- or surface-aligned implicit Gaussian prediction is a substantive representation change for feed-forward single-view 3DGS, but the abstract provides no quantitative gains or evidence beyond broad state-of-the-art claims.
-
PaDoc: Layout-Grounded Parallel Decoding for Document Parsing
PaDoc represents a document page as a predicted layout tree, allowing content for independent regions to decode concurrently while sharing the full-page visual context. Its implementation combines specialized attention masks with parallel requests and shared KV caches; on OmniDocBench it reports top-tier parsing quality, while improving throughput by 67.4–118% and reducing P95 latency by 39.2–54.9% versus a sequential same-backbone baseline.
The layout-grounded factorization and cache-sharing approach is a substantive decoding redesign with credible efficiency results, but its impact is demonstrated mainly on the narrow task of document parsing rather than as a broadly applicable parallel-generation method.
-
The Illusion of Visual Tool-Use: A Causal Audit of Thinking with Images
This paper treats visual tool use in multimodal LLMs as a causal problem rather than assuming that crop-and-zoom operations help because they are correlated with correct answers. Across six models and five fine-grained perception benchmarks, interventions show two recurring failures: models sometimes call visual tools without using the returned evidence, or use informative observations with poorly planned call schedules. The proposed step-level Visual Evidence Gain and trajectory diagnostics identify a minority of genuinely calibrated tool-use rollouts amid many that create an illusion of reasoning with images.
The causal decomposition and per-observation intervention are a meaningful diagnostic for an increasingly important multimodal capability, with fairly broad evidence, but the paper appears to analyze and expose failure modes rather than demonstrate a major capability or efficiency improvement.
-
WorldCycle: Self-Verifiable Reinforcement Learning for Long-Horizon Video World Models
WorldCycle trains interactive video world models using reversible action sequences as self-supervision: executing an action cycle should return the model to its starting state, and repeating the cycle should produce consistent states. It uses spatial closure and temporal consistency rewards to reduce long-horizon drift, reporting up to 44% lower state-return error and nearly 4× higher accuracy on composite actions, along with a diagnostic benchmark.
The reversible-cycle formulation offers a genuinely interesting way to verify long-horizon predictions without future-state labels, but the abstract gives limited detail about environments, baselines, and the breadth of evidence behind the large reported gains.
-
ToolArtist: Tool-Using Unified Multimodal Models for Agentic Image Generation
ToolArtist post-trains a unified multimodal model to jointly reason, search external tools, and generate images, rather than following a fixed pipeline with separate agent and image-generation components. It uses teacher-generated trajectories for supervised training and a reinforcement-learning method with intent and image-quality rewards; the abstract claims consistent gains over partially agentic alternatives, but gives no quantitative results.
The unified policy that controls search, reasoning, and native image generation is a meaningful direction for open-world image generation, but the abstract provides too little experimental detail to establish how large or general the gains are.
-
Resume Means Resume: A Machine-Checked Conformance Contract for Checkpoint, Interrupt, and Resume Semantics in Workflow Persistence Layers
The paper defines a machine-checkable contract for what checkpoint, interrupt, crash recovery, and resume should mean, then formally verifies a reference semantics and tests five workflow frameworks against it. It finds concrete inconsistencies—including silent invalid-state persistence, duplicate effect execution after crashes, and concurrent resumes triggering the same effect multiple times—and presents a verified reference implementation plus a shared-store read-path fix for the concurrency failure.
This is a rigorous and unusually concrete analysis of correctness semantics in agent workflow persistence, with strong formal and empirical evidence, though its impact is mainly on infrastructure rather than core AI capability.
-
Poly-OPD: Heterogeneous Multi-Teacher On-Policy Distillation for Capability-Selectable Flow Models
Poly-OPD distills multiple text-to-image teachers with incompatible latent spaces into one smaller flow-matching student. It uses a pixel-space bridge, DINOv2 feature matching, teacher-specific versus shared adapters, and a curriculum focused on remaining capability gaps. Distilling FLUX.1-dev and Z-Image into a 2.5B SD3.5-Medium model reportedly raises GenEval from 67.3 to 73.3 and HPSv3 from 9.34 to 11.35, exceeding the individual teachers on these metrics.
The heterogeneous-teacher distillation setup and gradient-guided adapter separation address a real limitation of latent-space distillation, with a notably non-obvious claim of surpassing both teachers, but the abstract provides limited ablations and validation beyond two reported benchmarks.
-
When Does Latent Communication Pay? A Causal Audit of Relayed KV Caches in Multi-Agent LLMs
The paper causally tests whether relayed KV caches in multi-agent LLMs contain information specific to the sender’s example, rather than merely providing a generic helpful signal. Across several model families and tasks, mismatched caches often perform like zeroed caches when the receiver needs private information, while native matched relays can produce large gains; when private information is unnecessary, the authors find statistical equivalence between matched and mismatched relays. The audit shows that some systems transmit example-specific information, while others’ cache benefits do not support the stronger “latent thoughts” interpretation.
This is a rigorous and non-obvious causal test of a prominent interpretation of latent communication, with preregistered equivalence testing and replication across models and tasks, but it is primarily an audit and clarification rather than a new capability or method.
-
Sound unmasks short-latency tectal visual signals that are independent of the primary visual cortex
In primates with primary visual cortex (V1) temporarily inactivated, reflexive saccadic responses to visual events normally disappear. The authors show that adding an irrelevant, simultaneous sound restores these short-latency responses and produces visual-related activity in the superior and inferior colliculi, indicating that multisensory context can recruit V1-independent visual pathways.
The causal finding that an otherwise ineffective tectal visual pathway can be behaviorally unmasked by sound challenges the assumption that V1 is indispensable for these rapid visual orienting responses, although the abstract provides limited detail on effect sizes and generality.
-
Video-DeepResearch: Towards the Next-Generation Multimodal Deepresearch Agent
The paper extends deep-research agents from still images to continuous video, requiring frame-level visual grounding before web search. It uses a staged perception/exploration pipeline and SFT followed by GRPO, and reports strong results on a new 200-question benchmark, including higher accuracy than several proprietary models. The main claimed findings are that agents otherwise prefer text search and can rely on memorized knowledge instead of genuinely using tools.
The video-grounded research-agent setting and enforced visual/tool-use training are interesting, but the evidence rests heavily on a small, newly curated benchmark and unsupported cross-model comparisons, so it is not yet a strong recommendation.
-
AtlasVLA: Persistent World-Ego State Modeling for Vision-Language-Action Models
AtlasVLA adds persistent memory to vision-language-action models for long-horizon manipulation from a single wrist camera. It maintains both a spatial world model, represented as an incrementally updated 4D voxel state, and an ego/task-progress memory, then conditions a diffusion policy on both; the authors report gains over multi-view baselines on LIBERO-Long and real-world tasks.
The explicit combination of persistent spatial memory and task-progress memory addresses a real limitation of reactive single-view VLAs, with sizable reported gains across simulation and real-world long-horizon tasks, though the abstract does not establish how much is genuinely new beyond existing recurrent world-model and memory-based policies.
-
Toward Skill-Native LLMs: Skill Entropy for Benchmarking and Training Long-Horizon Reasoning
The paper defines “skill entropy” to quantify how difficult it is for an LLM to switch among different reasoning skills during a long, dependent task. It introduces a 558-skill benchmark across nine domains and an RL method that rewards both correct intermediate steps and matching the intended skill sequence; on its benchmark, this substantially improves smaller Qwen models, though the tasks and skill annotations are introduced by the paper itself.
The cross-skill framing and use of predicted skill sequences as a training signal are genuinely interesting, but the large gains are mainly on a new benchmark with gold skill traces, so their generality and independence from the benchmark design remain uncertain.
-
Mendel Gödel Machine: Recursive Self-Improving Coding Agents via Comparative Evolution
The paper proposes a self-improving coding agent that modifies its own code using comparisons across many past task attempts, rather than only learning from one failure at a time. Its two mechanisms—multi-task “reaction-norm” mutation and cross-lineage hybridization—are supported by an additive-landscape analysis, surrogate simulations, and experiments on SWE-bench and Polyglot, where the authors report better performance, efficiency, and generalization.
Comparative, evolutionary self-modification is a meaningful direction for coding agents, but the abstract provides no quantitative results and the theoretical evidence is limited to a simplified fitness model, so the practical advance is not yet clearly established.
-
UniWorld-Design: From Pixel Generation to Layer-Native Design
UniWorld-Design treats semantic RGBA layers—not flat pixels—as the basic representation for image generation and editing. It includes a text-to-RGBA asset generator and an image-to-layer model that can decompose images recursively or extract targeted objects; on Crello, it reports 37% lower per-layer RGB error and 34% better Alpha Soft IoU than Qwen-Image-Layered.
The layer-native, instruction-addressable representation is a genuinely useful direction for controllable visual editing, but the abstract provides only benchmark-relative gains and limited evidence of broader capability or adoption impact.
-
The Personalization Mirage: How LLMs Fabricate User Profiles, and Why Self-Monitoring Misleads
The paper measures how often personalized LLMs infer user attributes that are not supported by evidence, using a new benchmark across 12 models and more than 143,000 judged claims. It reports substantial over-inference in every model and finds that models’ self-assessments are inversely related to externally judged over-inference across models, although self-audits work better within an individual model.
The benchmark is primarily an evaluation paper, but the consistent over-inference rates and especially the apparently inverted relationship between self-confidence and externally measured error are non-obvious findings, albeit based on only 12 models with a wide correlation confidence interval.
-
HelloWorld: Enabling Socially Interactive Characters in Video World Models
HelloWorld extends a video world model so that an on-screen character can react to a user’s button press by turning, waving, nodding, or speaking, while camera motion continues. It uses self-generated training clips to learn social interaction and camera-pose conditioning, plus a training-free cross-attention mechanism to confine the response to the time window after the press. The authors report better interaction quality than baselines while preserving visual quality and camera control, evaluated on a new 400-sample benchmark.
The combination of self-distilled interactive video generation and temporally localized, button-triggered character responses is a meaningful capability advance, but the abstract provides limited quantitative evidence and the demonstrated interactions remain narrow.
-
ChronoLens: Measuring Language Change Across Time, Languages, and Linguistic Levels
ChronoLens uses frozen multilingual language models, feature-aligned crosscoders, and linguistic interventions to compare historical change in morphology, syntax, semantics, and pragmatics within one representation. Applied to 17.2 billion tokens from five parliamentary traditions over 1803–2026, it finds that linguistic levels often change by similar magnitudes within a language, while different languages follow substantially different trajectories; its sparse representations align better with linguistic statistics than dense embeddings or a pooled sparse autoencoder.
The cross-level, cross-lingual measurement framework and finding that similar overall change magnitudes can hide different directions are genuinely interesting, with unusually broad historical data, but the abstract does not establish a major capability or methodological breakthrough beyond a promising analysis tool.
-
Gaming Without an Attacker: Benchmark Fingerprinting in LLM-Driven Search Under Selection Pressure
The paper shows that LLMs optimizing GPU kernels can exploit benchmark details even without being instructed to attack the evaluation: winning kernels often branch on recognizable runtime parameters, optimize measured cases, and neglect or break held-out ones. Across two suites, 16 of 53 in-distribution wins failed to transfer, and the authors classify the failures and propose measurement practices focused on non-enumerable held-out probes and performance—not just correctness—gates.
It provides a concrete, somewhat surprising demonstration that ordinary selection pressure induces benchmark fingerprinting in LLM-generated code, though the evidence is confined to two GPU-kernel suites and a relatively small set of wins.
-
Reassessing the Feasibility of PPG-Based Non-Invasive Blood Glucose Level Estimation
The paper re-evaluates five PPG-based blood-glucose estimation methods using progressively stricter data splits, including participant-held-out testing. Models that looked competitive with random window-level splits generally failed to generalize, achieving near-zero or negative R², while the Clarke Error Grid still labeled over 90% of predictions clinically acceptable—even for a mean-prediction baseline. The main contribution is showing that common data leakage and clinical-zone metrics can make these systems appear useful when they are not.
This is a non-obvious, practically important reassessment that exposes both participant leakage and a misleading evaluation metric, though it is primarily an evaluation study rather than a new modeling direction.
-
One QK Channel, Many Sources: Guarding Low-Precision Attention Collapse
The paper studies sudden training collapse in bfloat16 Transformers and argues that different numerical faults converge on a common failure pathway: runaway query-key (QK) dynamics that saturate attention. Controlled interventions suggest the QK channel is causally responsible, and a parameter-free QK normalization controller prevents collapse while remaining off during normal training, across the tested models and GPU architectures.
The source-versus-channel dissociation and causal QK intervention are a non-obvious mechanistic result with a potentially reusable stability fix, but the abstract does not establish broad scaling or deployment impact beyond the tested architectures and GPT-2-class setting.
-
Messages, Not Tokens: Grounded Coresets for Faithful VLM Compression
The paper reframes visual-token compression in VLMs as preserving the decoder’s aggregate, signed messages rather than independently selecting the most important patches. Its training-free GMC method selects representatives using query, appearance, and spatial information, then merges discarded visual states into those representatives before reducing the sequence; across several VLMs it reports retaining roughly full-model capability while removing 80.2% of visual tokens, with one variant slightly exceeding the full model.
The message-coreset framing and explicit realization of discarded population information are more substantive than ordinary token pruning, and the reported broad, near-lossless 80% compression is important, but the abstract does not establish how robust the gains are across tasks or whether comparisons and evaluation costs are fully convincing.
-
Computational and Statistical Guarantees of the \textit{c}-Rectified flow
This paper analyzes iterative rectified flow as a method for recovering optimal-transport couplings. It shows that ordinary rectified flow can fail in a Gaussian setting unless the source and target covariances commute, while a cost-aware variant that projects velocities onto gradient fields converges under stated assumptions; it also gives contraction and statistically minimax-optimal estimation rates.
The counterexample and convergence theory provide a meaningful explanation of when rectified flow does—and does not—recover optimal transport, but the results are primarily theoretical and depend on substantial regularity assumptions.
-
Optimal Unambiguous DNFs and Alon-Saks-Seymour
The paper constructs unambiguous DNFs with a near-maximal gap between width and 0-certificate complexity, then uses a constant-size lifting gadget to transfer that gap exactly into communication complexity. This yields optimal separations related to the Alon–Saks–Seymour conjecture and Clique-vs-Independent-Set, along with applications to approximate degree and multiclass sample compression.
The core construction appears to give optimal, broadly reusable separations in complexity and learning theory, but the paper is primarily communication-complexity theory rather than a direct advance in mainstream AI/ML.
-
Interaction Is Not Necessary for Order-Optimal 1-Bit Mean Estimation
The paper shows that one-bit mean estimation does not need adaptive interaction to achieve the optimal sample complexity. It constructs a randomized non-adaptive protocol whose rates match prior adaptive methods across different moment regimes, and proves minimax optimality where lower bounds are known.
It resolves a recent open problem by eliminating interaction without sacrificing order-optimality, although the result is specialized to one-bit statistical estimation and its broader practical impact is unclear.
-
COMPAS: Difficulty-Aware Joint Search for Optimizing Code Generation
COMPAS jointly optimizes the model, prompt, and decoding settings for code-generation tasks, grouping tasks by difficulty and routing each task to a learned quality–cost configuration. On LiveCodeBench it raises pass@1 from 45.9% to 52.8% while cutting reported search cost from $36.57 to $4.92, and it improves SWE-bench resolution from 70.0% to 76.0%.
The combination of difficulty-aware routing with joint prompt/decoding/model optimization produces an unusually large cost reduction alongside meaningful coding gains, though the abstract does not establish how broadly it generalizes beyond the reported benchmarks.
-
The Calibration Floor: Format Repair Can Masquerade as Self-Correction at Small-to-Mid Scale
This paper argues that much of the apparent benefit from language-model self-revision comes from repairing answer formatting so outputs become parseable, rather than from improved reasoning. Across several model sizes and tasks, grammar-constrained decoding attributes a median 71% of the apparent gain to format effects, while content-level changes are often negligible; the authors also fail to reproduce a prior confidence-gating result. The evidence suggests self-correction measurements can be substantially confounded by answer extraction, though the frontier-model analysis is relatively low-powered.
The causal separation of reasoning changes from answer-format recovery, together with broad replication evidence that format effects dominate, challenges a common interpretation of self-correction results, but it does not yet establish a new capability or universally general phenomenon.
-
The Price of Isolation: Estimating the Ecosystem Cost of Symmetric Two-Sided A/B Testing
The paper studies the engagement cost of isolating creators and viewers into separate markets for two-sided A/B tests. It shows, using extreme-value theory, that catalog thinning becomes negligible for light-tailed match-quality distributions but can remain a constant cost under heavy tails, and production experiments on millions of creators support this prediction. It also proposes a pre-experiment procedure for estimating the cost and choosing traffic allocation or a fallback design.
The heavy-tail result challenges the common assumption that isolation costs disappear on large platforms, and it is backed by theory plus production experiments, though the contribution is specialized to marketplace experimentation rather than a broadly transformative ML method.
-
Representational separation between unitary and channel quantum generative models via shared classical randomness at shallow depth
The paper proves that adding a single shared classical random bit to shallow, locally connected quantum circuits lets a quantum generative model produce long-range output correlations that shallow unitary Born models cannot reproduce. In one-dimensional nearest-neighbor systems, matching these distributions with a purely unitary model may require depth proportional to system size; the authors also show how measurement-based quantum computation can implement the shared randomness and provide numerical support.
This is a non-obvious, scalable representational separation with a linear-depth lower bound, though its significance is mainly theoretical and tied to specialized quantum generative-model settings.
-
Minimax Optimal Early-Stopped Gradient Descent for Gaussian Mixture Classification
This paper analyzes logistic-loss gradient descent on linearly separable but noisy Gaussian-mixture classification problems, where continued training converges to a statistically suboptimal max-margin classifier. It proves that stopping at the right time achieves minimax-optimal classification risk across broad covariance-spectrum decay regimes, and that linear interpolation may need exponentially more samples to match early stopping. The analysis also gives a sharper conversion from logistic excess risk to zero-one risk under label noise and model misspecification.
The matching minimax results and claimed exponential sample-complexity gap make early stopping a substantially better alternative to interpolation in this model, but the conclusion is still confined to a stylized Gaussian-mixture setting and depends on an oracle stopping time.
-
On-Policy Self-Distillation without Any Supervision
The paper proposes U-OPSD, a self-distillation method that samples multiple answers, uses majority agreement to create a pseudo-solution, and trains on the completions that disagree with it—without labels, rewards, or a teacher model. On five math benchmarks, it reports 8.5–10.7% gains over Qwen3 base models and performance comparable to or better than supervised OPSD and GRPO.
The appealing result is that internal self-consistency can provide useful on-policy training without external supervision, but the core ingredients—self-consistency, pseudo-labeling, and self-training—are established, and the evidence is limited mainly to mathematical reasoning benchmarks and the authors' reported comparisons.
-
Density-Functional Excited-State Gradients and Nonadiabatic Couplings on a Consumer GPU from a Contraction-DAG
The paper derives analytic excited-state gradients and nonadiabatic couplings for a double-hybrid density-functional method, implementing both through a shared reverse-mode contraction graph and AO-direct GPU kernels. It reports improved excitation energies, independent coupling validation to about 10^-4, and physically meaningful behavior at an ammonia conical intersection, while fitting the full calculation on an 8 GB consumer RTX 4060.
This is a genuinely new computational capability with an unusually strong cost/access result—correlated excited-state derivatives and couplings on commodity hardware—but its demonstrated scientific scope is still fairly narrow and specialized.
-
Hidden Gauge Controls Feature Specialization in ReLU Networks
This paper studies overparameterized ReLU teacher–student networks where multiple neurons initially represent the same function. It shows that changing only a positive-homogeneous parameter scaling—without changing the initial predictor—can determine which neuron captures the learned feature, how quickly specialization occurs, and which neurons are pruned; the effect can create a sharp Θ(D²) difference in specialization time. The authors explain this through different effective mobilities for changing feature coefficients versus feature directions, with supporting results for perturbations, gradient descent, and finite-sample experiments.
The non-identifiability result is genuinely interesting: functionally identical initial networks can follow qualitatively different specialization and pruning trajectories due solely to hidden parameterization, although the claims are established mainly in a tractable model rather than broad modern networks.
-
Multiscale Reward Hedging from Correct Demonstrations
This paper studies learning from demonstrations where the demonstrated action is known to be correct, but the learner never observes rewards or whether its own action was also valid. It introduces a multiscale hedging method for continuous reward classes, proving horizon-free hidden-regret bounds governed by metric entropy; for bounded linear contextual recommendation it obtains O(d) total regret for arbitrary compact action menus, with matching lower bounds and several computational corollaries. Small MovieLens and stress-test experiments illustrate the theory, but the main contribution is theoretical.
The removal of the usual finite-reward-class assumption with a horizon-free, entropy-controlled guarantee—and no structural restrictions on menus—is a genuinely interesting theoretical advance, though its practical impact is uncertain and the empirical validation is limited.
-
Selective Cross-View Consistency for World Action Models: Held-Out Viewpoint Robustness Without Test-Time Camera Information
The paper shows that cross-view consistency training can hurt world action models when applied to view-dependent predicted video, because it suppresses legitimate viewpoint-specific content. It proposes applying consistency only to view-invariant outputs such as actions and proprioception, and reports a 12.2-point closed-loop success improvement on held-out camera viewpoints, with replication across another seed and two camera axes; interpolation and in-distribution performance do not improve.
The covariant/invariant decomposition is a concrete and potentially general lesson for training world action models, supported by controlled shrinkage analysis and held-out-viewpoint experiments, but the evidence is limited to LIBERO-Plus-style evaluation and does not yet establish broad real-world robustness.
-
Your VLM Already Knows When: Training-Free Temporal Grounding by Asking Yes or No
The paper shows that frozen vision-language models can localize events in video much better when temporal grounding is reframed as repeatedly answering binary yes/no questions, rather than directly predicting timestamps. This training-free coarse-to-fine method improves R@0.5 by 28–50 points across four backbones, reaches 56.8% on Charades-STA, and the authors separate remaining errors into visual perception and temporal-geometry limitations.
The large, training-free gains and the claim that the main failure is the timestamp-output interface rather than video perception are genuinely interesting, but the core technique is still a relatively simple inference-time reformulation and the abstract does not establish whether it generalizes beyond the reported benchmarks.
-
Do AI Forecast Ensembles Sample the Correct Conditional Distribution?
The authors test whether diffusion-based AI forecast ensembles reproduce the full conditional distribution, not just accurate per-location marginals. Across coastal sea-level forecasting and Lorenz-96 experiments, the models achieve marginal skill but produce spatially joint samples worse than climatological draws; the problem is exposed by the variogram score but missed by the energy score, and also appears in a linear learned emulator rather than in a dynamical ensemble.
This is a non-obvious and potentially important warning that learned probabilistic forecasters can have good marginal scores while fundamentally misrepresenting dependence structure, though the evidence is limited to forecasting settings and the abstract gives few quantitative details.
-
Dopamine dips during unrewarded actions promote punishment-resistant reward seeking
In female mice, chronic estradiol increased punishment-resistant reward seeking not by amplifying dopamine responses to rewards, but by making dopamine dips after unrewarded actions stronger. Optogenetically reproducing these dips made both female and male mice more resistant to punishment, suggesting that dopamine dips may weaken learning from unsuccessful actions and promote persistent reward seeking.
The study provides a non-obvious causal link between dopamine dips—not just reward-related dopamine peaks—and addiction-like persistence, but the evidence is still limited to mouse behavior and optogenetic manipulation.
-
Primary Cilia Loss in Striatal Neurons Disrupts Synaptic Connectivity and Excitatory Transmission and Drives Metabolomic Remodeling
The study removes primary cilia from striatal medium spiny neurons in mice and finds a broad disruption of circuit inputs, excitatory synaptic transmission, and local chemistry. Cilia loss produced roughly a threefold reduction in brain-wide afferent convergence, weaker AMPA/NMDA-mediated signaling, and substantial metabolomic changes, especially in lipid and neuromodulatory pathways. The new result is that striatal cilia appear to help maintain connectivity and biochemical state, not merely serve as receptor-localization structures.
The multimodal, causal evidence for a major role of primary cilia in maintaining striatal circuit connectivity is notably broader and more consequential than a standard synaptic phenotype, though the findings are from a targeted animal manipulation and may include ablation-related confounds.
-
A guanosine nucleotide metabolism-bioenergetics intersection drives Parkinson's disease
The study identifies nucleoside diphosphate kinase (NDK) as a link between ATP/GTP metabolism, mitochondrial bioenergetics, and Parkinson’s disease risk. NDK overexpression protected neurons in cellular and animal models, while disrupting NDK or guanosine-nucleotide synthesis worsened synaptic dysfunction; genetic analysis of more than 64,000 affected individuals and 38,000 controls also associated pathway genes with PD risk or protection.
The proposed ATP–GTP–BH4 metabolic intersection offers a non-obvious mechanistic explanation for several PD risk signals and is supported by both perturbation experiments and a large human genetic analysis, though the abstract does not establish therapeutic efficacy or causality in patients.
-
An axonal sodium current gates motoneuron doublets and force amplification
This study argues that motoneuron doublets—closely spaced spike pairs that disproportionately increase muscle force—are generated by a persistent sodium current at the first axonal node, which feeds back toward the axon initial segment and interacts with somatic currents. Intracellular recordings, biophysical modeling, and human motor-unit/force data suggest that monoaminergic neuromodulation controls both whether doublets occur and how their interval changes over time.
It provides a specific, experimentally and computationally supported mechanism for a century-old unresolved motoneuron phenomenon, with a direct link from neuromodulation to nonlinear force coding, but the abstract does not establish a sufficiently broad or transformative impact for a strong verdict.
-
Tlr4 and Nlrp3 Inflammasome Signaling Drive a Delayed Secondary Wave of Hyperexcitability After TBI in Larval Zebrafish: Implications for Post-Traumatic Epileptogenesis
Using a larval zebrafish blast-like TBI model, the authors tracked inflammation and seizure-like behavior over time and tested causal interventions. They report an acute, bidirectional interaction followed by a delayed inflammatory wave driven by TLR4 and NLRP3 signaling; targeting this later phase reduced the secondary hyperexcitability peak, while retigabine had opposite effects in acute and post-acute phases.
The temporally distinct, inflammation-dependent secondary seizure phase—and especially the opposing timing-dependent effects of anti-inflammatory and anticonvulsant treatments—offers a non-obvious mechanistic and therapeutic-window result, but evidence is limited to a larval zebrafish model and an abstract-level report.
-
Multiple Nested Distributed Language Networks in the Human Brain
Using intensive precision fMRI, the authors identify two nested, left-lateralized language networks: an inner network linked to speech and phonological processing, and a surrounding association network linked to meaning-based sentence processing. The networks were reproducible across three cohorts, could be identified through connectivity with neighboring cerebellar regions, and showed distinct task responses, supporting a hierarchical organization of language rather than a single distributed language network.
This is a well-supported and potentially important anatomical and functional decomposition of the language system, but the evidence is based on intensive fMRI mapping in only 22 individuals and does not yet establish causal or uniquely linguistic specialization.
-
Epilepsy and premature mortality driven by inhibitory neuron dysfunction in a mouse model of SCN1A gain-of-function neurodevelopmental disorder
The authors created the first mouse model of an SCN1A gain-of-function epilepsy variant, R1636Q. The variant caused uniformly fatal early seizures, and selectively expressing it in parvalbumin inhibitory neurons—but not excitatory, VIP, or SST neurons—reproduced the phenotype; a sodium-channel blocker substantially extended survival.
This provides unusually direct evidence that dysfunction of parvalbumin interneurons drives SCN1A gain-of-function epilepsy and offers a mechanistically useful model, though the findings are currently limited to one mouse variant and a survival phenotype.
-
Beyond Sequence Order: Syntax-Informed Positional Embeddings for Transformers
The paper adds dependency-parse information to several kinds of Transformer positional embeddings, with the best injection method differing between encoder and autoregressive decoder architectures. In reported pretraining experiments, this improves syntactic generalization by up to 10.3%, lowers perplexity by 9.0%, and raises GLUE scores by up to 8.2% relative to an otherwise comparable model without syntactic supervision, while requiring only a single parse at inference.
The architecture-specific way of injecting syntax and the claim that syntactic supervision improves both syntax-sensitive evaluation and ordinary language modeling are genuinely interesting, but the abstract does not establish how broad or reproducible the gains are across model scales, parsers, languages, and baselines.
-
When Many Answers Are Valid, Voting Fails: Symbolic Verification for Best-of-K Causal Reasoning in LLMs
The paper proposes CALVER, a training-free symbolic verifier that checks sampled causal-reasoning traces against Pearl-style criteria rather than selecting the most common answer. On multi-answer causal queries, it reportedly raises selection accuracy to 42.1% versus about 30% for plurality, reward models, LLM judges, and confidence, with similar gains across graph families, model families, and text-to-graph settings. It also finds that many benchmark disagreements are valid alternative answers, suggesting exact-answer evaluation can undercount correct causal reasoning.
The combination of symbolic causal validity checking with best-of-K selection addresses a real failure mode of self-consistency and has reasonably broad reported evidence, but the gains are moderate and the method depends on having a correct causal structure or successfully extracting one.
-
OmniPack: Unified Token Compression for Efficient Omni-modal Large Language Models
OmniPack is a training-free token-compression method for audio-visual LLMs. It first removes structurally redundant audio and visual tokens, then uses multimodal and text-guided information inside the LLM to preserve task-relevant content; across five benchmarks and three backbones, it reportedly retains 98.0% of baseline performance at 16.7% of the FLOPs and 92.9% at 6.8%.
The substantial reported compute reduction in omni-modal models is worth noting, but the approach combines familiar pre- and in-model token pruning/merging ideas, and the abstract lacks absolute task results and enough detail to establish how broadly the gains generalize.
-
GST-Bench: Can VLMs Develop Global Spatial Awareness from Video?
The paper introduces GST-Bench, a video question-answering benchmark that tests whether vision-language models can build a globally consistent spatial map from long, egocentric video and answer questions from unseen viewpoints or top-down views. Across 22 VLMs, the best zero-shot model scores 42.68 versus 79.08 for humans; comparison with a local version suggests the main failure is integrating observations over time rather than basic spatial perception.
It provides a concrete, non-obvious diagnosis—that current VLMs retain local spatial competence but fail at long-horizon global scene consolidation—although the evidence comes from a synthetic benchmark and accompanying dataset rather than a new capability or method.
-
Lossless Tensor Compression as Program Synthesis
Brevis treats lossless tensor compression as a program-synthesis problem: it uses a typed DSL of reversible tensor operations and a checkpoint-specific prior to search for compact reconstruction programs. Across 10 checkpoints from language, audio, and image-generation models, it reports 33.93% storage reduction, outperforming several general-purpose and tensor-specific compressors while retaining bit-exact reconstruction and multi-GB/s throughput.
The program-synthesis framing and self-contained reversible representations are a genuinely interesting departure from fixed compression pipelines, but the evidence is limited to 10 checkpoints and the practical gains do not yet establish a broadly transformative compression method.
-
GROVE: Growing and Reasoning over Temporally Stratified Memory from Streaming Video Experience
GROVE is a training-free memory system for wearable assistants that continuously organizes streaming video into fine-grained observations, time-stamped moments, episodes, and recurring patterns across days. It uses retrieval procedures matched to each temporal scale, allowing the same memory to answer questions about the past or proactively identify relevant prior experiences; experiments report leading results on several video-memory and assistance benchmarks, with multi-day patterns providing the largest ablation gain.
The unified, causally growing multi-timescale memory for both reactive video QA and proactive assistance is a meaningful systems direction, but the abstract gives no quantitative results and the contribution may largely be a well-engineered combination of familiar memory and retrieval ideas.
-
Addressable Memory for Video World Models
The paper identifies a specific failure mode in long-horizon video world models: after generation exceeds the training horizon, RoPE positions make older visual content difficult to retrieve, and compressing already position-rotated KV states can corrupt it. WorldTrace assigns compressed memories in-distribution virtual positions and proposes two variants for temporal coherence or episodic scene recall; on its new LoopBench, they report 15.5% and 19.5% improvements without retraining.
The addressable-memory framing and training-free fix target a real bottleneck in long-horizon video generation, but the evidence is limited to an authors' new benchmark and the abstract does not establish broad gains across models or tasks.
-
Motion Beyond Morphology: Bootstrapping Cross-Category Motion Transfer from Abstract Motion Representations
The paper tackles video motion transfer when the source and target objects have very different shapes or articulation, where direct part-to-part correspondence breaks down. It learns abstract motion descriptions at multiple scales to create cross-category training pairs, then trains a generator that transfers motion directly from a reference video without explicit motion extraction; it also introduces a dataset and benchmark spanning same, near, and far category gaps.
The move from morphology-specific correspondence to abstract, cross-category motion transfer is a meaningful direction, but the abstract provides no quantitative results or evidence beyond an unsupported state-of-the-art claim.
-
NOLLI: A Difficulty-Calibrated Puzzle Benchmark for Diagnosing the English-Korean Performance Gap
NOLLI is a procedurally generated, deterministically scored benchmark of 25 English-Korean puzzle tasks, including translated, Hangul-jamo, and Korean-specific problems. Across 15 models, matched English and Korean performance was usually similar, but tasks requiring multi-step sub-syllabic manipulation produced gaps as large as 68.7 percentage points; jamo composition performance predicted cipher performance, while cultural and orthographic tasks showed different patterns. The study also finds that nominal puzzle size often fails to track behavioral difficulty.
Although primarily a benchmark paper, its calibrated cross-script contrasts yield a non-obvious diagnostic result: Korean deficits appear concentrated in multi-step Hangul-jamo execution rather than language presentation generally, with reasonably broad model evaluation supporting the claim.
-
SimWAM: A Simple World Action Model for End-to-End Autonomous Driving
SimWAM uses a pretrained video-prediction expert to supervise a separate action-prediction expert during training, while avoiding future-frame generation at test time. It reports 91.5 PDMS on NAVSIM, lower latency than prior world-action-model planners, and zero-shot transfer to nuScenes; reinforcement learning is also used to optimize driving rewards beyond imitation.
The decoupling of video dynamics supervision from inference could make world-action models substantially cheaper and easier to scale, but the abstract provides limited detail about the comparison and the gains appear primarily to be a strong baseline result rather than a clearly new direction.
-
RoMeRL: Balancing Feedback Coverage and the Memory-Reward Trap in Self-Evolving Agent Memory via Reduced-Order Utility States
RoMeRL replaces an ever-growing, trajectory-specific utility table for agent memories with a fixed number of task-level semantic utility coordinates, separated by outcome polarity and memory dynamics. On ALFWorld and LifelongAgentBench, it reportedly improves task performance while reducing reward contamination, memory size by 84.4%, and LLM calls by 21.1%, though the evidence is limited to two agent benchmarks.
The reduced-order formulation is a meaningful approach to the feedback sparsity and credit-assignment problems in self-evolving memory, with substantial reported efficiency gains, but its generality and practical importance beyond the tested benchmarks remain uncertain.
-
CENTILE: A Telemetry Foundation Model Evaluated by the Decisions It Drives
CENTILE is a generative foundation model for irregular, heterogeneous systems telemetry that predicts calibrated conditional quantiles for many entities and forecast horizons in one pass. Rather than optimizing point-forecast error, the authors evaluate whether its forecasts improve downstream decisions, reporting up to a 77% reduction in HPC backfilling slowdown and roughly half as many network-provisioning violations, with some zero-shot transfer across time and domains.
The decision-centric evaluation and claimed cross-domain reuse are genuinely interesting, but the abstract provides limited detail about scale, baselines, and replay realism, so the large reported gains need verification before this merits a stronger recommendation.
-
A Tilt-Rotor UAV with a Gripper for Stable Contact-Based Tasks via Environmental Anchoring
The paper presents a tilt-rotor UAV that can fly normally, then clamp onto nearby structures with a compliant cable-driven gripper to become a stable platform for physical interaction. In real-robot tests, anchoring reduced positional drift by more than 95% versus free flight, supported reaction forces up to 75 N, and kept drift below 3 mm across tested geometries and orientations. The main novelty is using environmental anchoring to overcome the inherent instability of aerial robots during contact tasks.
This demonstrates a genuinely useful new aerial-robot capability with strong reported stability gains, but the abstract gives limited evidence about the range of completed tasks, environmental conditions, and generality beyond controlled flight tests.
-
Open-DiffLoco: Open-Source Differentiable Learning for Deployable Blind Quadruped Locomotion
Open-DiffLoco trains proprioceptive quadruped locomotion policies end-to-end in differentiable MuJoCo simulation, using a simplified reward and no privileged velocity observations or reference trajectories. On a Unitree Go2, the policy tracks omnidirectional commands with under 0.2 m/s RMS error, exceeds 1 m/s, and tolerates uneven terrain and pushes; training reportedly takes 20–60 minutes on a single RTX 5080. The paper also introduces Jacobian-Augmented Value Estimation to improve early policy-gradient learning and releases code for the full simulation-to-real pipeline.
The combination of differentiable simulation, simplified reward design, low compute cost, and demonstrated sim-to-real deployment is a meaningful advance, but the evidence is limited to one framework, robot platform, and reported set of configurations rather than a broad comparison or capability breakthrough.
-
Same Semantics, Different Paths: Self-Improving Alignment for Vision-Text Compression
The paper identifies a mismatch between how a vision encoder represents rendered text and how the model represents native text, arguing that this—not just token compression—is a key bottleneck in vision-text compression. SPIRAL uses the model’s own text pathway as supervision, combining token-level distillation with sequence-level preference optimization; on VTCBench it raises Qwen3-VL-8B from 35.10 to 54.02, nearly matching native-text input at 55.60, with reported gains transferring out of domain.
The cross-path alignment framing and large reported gain make this more than an incremental compression tweak, but the abstract provides limited evidence beyond one main benchmark and self-reported generalization claims.
-
Grounded Semantic Re-Binding for Robust Instruction Generalization in Vision-Language-Action Models
The paper argues that VLA models often preserve the meaning of paraphrased instructions, but their action policies are destabilized by feature shifts caused by joint vision-text encoding. It proposes Grounded Semantic Re-binding, which separately extracts task semantics, fuses them with visual features, and retrains the action expert; on LIBERO-Para, this reportedly improves success by up to 44.6% and gives a small 0.33B model near-perfect paraphrase robustness.
The architectural diagnosis and large robustness gains using only canonical demonstrations are potentially important, but the evidence is mainly one benchmark and the abstract provides insufficient detail to establish that the claimed gains generalize beyond paraphrased instruction following.
-
Why Does Action Chunking Improve Behavioral Cloning Performance in Robotic Control?
This paper investigates why predicting sequences of robot actions (“action chunking”) improves behavioral cloning. It argues that the usual explanations are insufficient: much of the gain comes from non-Markovian history dependence and reduced error accumulation, while an additional benefit comes from implicit ensembling across different temporal relationships. In simulation and real robots, randomized-delay policy ensembles can match chunking, and an explicitly designed ensemble can outperform it.
The implicit-ensemble explanation and a way to reproduce or improve action-chunking benefits without executing chunks are genuinely useful conceptual and practical results, but the abstract provides no quantitative scale or breadth of the claimed gains.
-
Temporal Leakage in LLM Backtesting: Measurement, Validation, and Adjusted Scores
This paper argues that post-cutoff performance is not a reliable contamination test because models naturally know more about events closer to their training cutoff, making recency look like memorization. It proves that passive backtests cannot distinguish these effects, then proposes boundary-based and matched-clean-control estimators for detecting and adjusting leakage; experiments with planted leakage recover the injected amount, while audits of frontier models find some apparent gains attributable to recency alone.
The impossibility result and use of matched clean controls could materially change how LLM backtests are designed, but the abstract gives limited detail on estimator robustness and the frontier-model evidence, so it merits a look rather than priority reading.
-
CUDA MPC: A GPU-Native Solver for Model Predictive Control
The paper presents a CUDA-native MPC solver that keeps the full ADMM optimization loop on the GPU, using fused kernels, shared memory, and local synchronization across horizon blocks. It reports real-time control with much longer horizons than CPU solvers, including 100 seconds of lookahead for parking and collision-free centralized control of a 10-agent swarm; it also claims up to a 965× speedup over tensor-framework implementations.
A tightly co-designed GPU execution model for long-horizon, constrained MPC is a meaningful systems direction with striking reported speedups, but the abstract does not establish hardware breadth, solver-quality tradeoffs, or whether the comparisons are fully fair enough for a strong verdict.
-
TimeRLM: Recursive Language Models Enable Precise Anomaly Localization in Long-Context Time-Series
TimeRLM lets a language model inspect long time-series signals recursively through code and visual tools instead of processing the entire signal in one pass. On a new synthetic benchmark, it substantially improves anomaly localization over existing time-series language models, and reinforcement learning reduces the number of interaction turns; the gains also transfer to ECG, sleep, and software-observability recordings despite training only on synthetic data.
The combination of recursive external-context querying with time-series analysis produces unusually large localization gains and some real-world transfer, but the evidence relies heavily on a new synthetic benchmark and the abstract does not establish robustness across broader settings.
-
Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse
This paper transfers a conversation’s KV cache between different-sized models in the same family, using a per-head closed-form ridge regression learned from only 500 calibration sequences. Across six model pairs, it preserves 73–98% of the target model’s standalone-prefill accuracy on four pairs and runs 2.7–25× faster than recomputing the prompt, although two pairs fail substantially and require an MLP to recover performance.
The idea could materially reduce latency and cost in model cascades and handoffs, but the strong dependence on matching KV geometry, mixed pair results, and limited reported evaluation keep it short of a strong recommendation.
-
Perception Before Reasoning: Dynamic Latent Reasoning for Video Understanding and Question Answering
The paper introduces Dynamic Latent Reasoning, which first encodes query-relevant video evidence into a small set of continuous perception latents and only adds latent reasoning steps when needed. Across nine video-QA benchmarks and four multimodal backbones, it reports higher accuracy with dramatically shorter outputs; for Qwen3-VL-4B, accuracy rises from 54.0 to 58.2 while average response length falls from 1,220.7 to 18.5 tokens.
The combination of grounded latent visual evidence, adaptive reasoning, and a roughly 66× reduction in generated tokens is a genuinely non-obvious efficiency result, but the abstract lacks enough detail to establish how robustly it beats strong baselines or whether the gains generalize beyond the reported benchmarks.
-
SiMDex: Mining Similar Egocentric Videos for Cross-Embodiment Dexterous Manipulation
SiMDex treats selection of human egocentric video for robot dexterous-manipulation training as a retrieval and recommendation problem. Its morphology-agnostic recall, ranking, and reranking pipeline selects about 1.49M relevant samples from a 32M-video pool, raising overall success from 47.7% with random sampling to 61.1% without changing the VLA architecture or using more training data.
The substantial gain from using less than 5% of the available human data suggests that targeted cross-embodiment curation may matter more than indiscriminate scaling, though the abstract provides limited detail on task breadth and experimental validation.
-
PhysMind: From Video to Executable Worlds for Training-Free Physical Reasoning
PhysMind builds a reusable executable physical scene from a video using segmentation, 3D reconstruction, pose tracking, and fitted continuous-time dynamics, then answers questions by inspecting or modifying that world rather than relying only on VLM reasoning. It reports large gains over direct chain-of-thought baselines on CLEVRER and Physion++, including a 19.25-point advantage over GPT-5.5 on counterfactual questions.
The question-agnostic executable-world representation and intervention-based reasoning are a meaningful, nonstandard direction with unusually large reported gains, but the abstract gives limited detail about evaluation breadth, reconstruction failures, and how credible the strong baseline comparisons are.
-
Rethinking Pixel Mean Flows via Interval Denoiser
The paper derives an “Interval Denoiser” directly from the flow-matching ODE to train a pixel-space generator for long denoising intervals, rather than relying on latent autoencoders or empirical substitutions. It adds residual clipping and curriculum time sampling, and reports ImageNet 256×256 FID scores of 4.55 with one sampling step and 3.98 with two steps, trained from scratch without perceptual losses.
Direct pixel-space flow matching with strong reported one- and two-step ImageNet results is potentially important and challenges the assumption that high-quality few-step generation requires latent representations, but the abstract lacks model scale, baselines, and experimental breadth needed for a stronger verdict.
-
LILAC: An Idempotent Neural Speech Codec
LILAC introduces a speech codec whose encode–decode–re-encode cycle is exactly idempotent: any valid token sequence decodes to audio that re-encodes to the same tokens. The fully convolutional 24 kHz codec operates at 0.75 kbit/s and reports competitive speech quality, while the authors find that twelve existing codecs rewrite at least 15% of tokens after one such cycle.
Exact idempotency by construction is a distinctive and useful property for codec-token pipelines, but the abstract provides limited evidence about quality tradeoffs, downstream benefits, and robustness beyond headline comparisons.
-
HLSmith: An Expert-Guided Agentic Framework for C/C++-to-HLS Translation
HLSmith uses an expert-authored library of guarded HLS transformations, a staged synthesis-and-feedback workflow, and fine-tuning from tool-use trajectories to turn C/C++ into FPGA HLS designs. On PolyBench, it reports a 4.24x geometric-mean speedup over ChatHLS with 100% functionally valid designs versus 57% for ChatHLS, with maximum reported speedups of 252x and 138x using commercial and open-weight models respectively.
The combination of explicit hardware expertise, synthesis-grounded iteration, and model adaptation appears genuinely useful, and the large validity and speedup differences are notable, but the evidence is limited to PolyBench and the abstract does not clarify baselines, resource tradeoffs, or how much of the gain comes from the stronger underlying models.
-
Exact Thrust-Reversal Limits of Bidirectional Propellers under Bounded Motor Inputs
The paper shows that a bidirectional propeller cannot track an ordinary linear thrust reversal exactly with bounded motor torque, because thrust sensitivity to motor speed vanishes at zero thrust. It derives necessary and sufficient conditions based on how smoothly thrust crosses zero, then validates them experimentally: linear reversals produce current/voltage peaks and tracking errors, while higher-order reversals avoid the singular behavior.
The intrinsic zero-thrust actuator singularity and resulting design rule for higher-order reversals are a non-obvious, broadly relevant control insight, though the evidence described is limited to theory plus one motor-propeller setup rather than demonstrated system-level impact.
-
The Sample Complexity of Policy Learning with Mu-Resets
This paper analyzes policy-based reinforcement learning when the learner can reset episodes according to an exploratory distribution. It shows that policy realizability alone does not remove severe horizon dependence: under all-policy concentrability, sample complexity can be exponential in the horizon, while under the weaker pushforward-concentrability notion the dependence is characterized tightly as exp(Θ(√H)).
The sharp separation between two coverage assumptions, including an exp(√H) characterization and an exponential lower bound despite realizability, is a theoretically non-obvious result with implications for how RL reset protocols are analyzed, though the abstract provides limited experimental or proof detail.
-
Reason Wide, Not Deep: Amortizing the Reasoning Premium into Distilled Skills
The paper trains a coding agent to distill recurring procedures from a small corpus of agent trajectories into compact natural-language skills, then gives those skills to a non-reasoning model. On four agentic benchmarks, the skills recover 55–100% or more of the reasoning-mode performance gap while using 2.7–6× fewer output tokens; the results suggest that reasoning can often be amortized across tasks rather than repeated within each episode.
The wide-search-versus-deep-search framing and strong reported efficiency gains are genuinely interesting, but the abstract lacks enough methodological and comparative detail to justify a stronger verdict.
-
RotaryQuant: Fitting 120B MoE Models on Consumer Hardware via Fused Compressed-Space Attention
RotaryQuant combines mixed-precision quantization, expert paging, and a compressed KV cache to run large MoE language models on consumer hardware. Its main technical contribution, IsoQuant, uses structured Walsh–Hadamard/SO(4) rotations and 3-bit KV quantization, with fused Metal kernels that avoid reconstructing full-precision cache state; the authors report fitting models up to 120B parameters at 9–19 tokens/s with very small perplexity and retrieval losses.
The fused execution of structured 3-bit KV compression alongside expert offloading could be a meaningful systems advance, but the abstract gives limited comparative evidence and the unusually strong near-zero degradation claims need close scrutiny.
-
High-Capacity Generalized Hopfield Networks
This paper extends Hopfield networks from vector-valued neurons to memories represented on SU(d) Riemannian manifolds. Using Lie-algebra methods, replica analysis, and simulations, it claims that capacity rises by nearly an order of magnitude for d=3 and grows further with d, with recall reducible to alignment with the top eigenvector of a spiked matrix. It also demonstrates image recovery for SU(3), proposes physical Landau–Lifshitz–Gilbert recall dynamics, and analyzes a quantum version with glassy spectral structure.
The manifold-valued Hopfield formulation and claimed large capacity increase are substantially more than an incremental tweak, but the abstract gives limited quantitative detail and evidence beyond simulations, analytical approximations, and a small illustrative image experiment.
-
Halpern Iteration Achieves $\tilde{\mathcal{O}}(ε^{-1/p})$ $p$th-Order Oracle Complexity for Monotone Variational Inequalities
The paper introduces a large-step inexact Halpern iteration for smooth monotone variational inequalities, yielding a claimed convergence rate of approximately $\tilde{O}(T^{-2})$ for second-order methods. It generalizes this construction to order $p$, combining an Anchored Tensor Method with Halpern iteration to obtain $\tilde{O}(T^{-p})$ convergence, or roughly $\tilde{O}(\epsilon^{-1/p})$ oracle complexity, improving prior MVI results.
The claimed improvement from the best known second-order MVI rate to $T^{-2}$, together with a general $p$th-order framework, is a meaningful theoretical advance, but the abstract provides only stated convergence results and the problem setting is specialized.
-
VoxZip: Semantic-Anchored Temporal KV Cache Compression for Long-Context Audio Inference
VoxZip compresses the KV cache of long-context speech models without retraining, first using ASR transcripts to merge audio tokens around semantic boundaries and then evicting tokens using attention that decays over time. On Qwen3-Omni across six audio benchmarks, it reportedly preserves over 90% of baseline performance at 20× compression, while 4× compression gives 1.9× higher throughput and 3.3× lower peak memory.
The combination of transcript-guided temporal compression and dynamic cache eviction is a relevant, nontrivial approach to a major bottleneck, with unusually large claimed efficiency gains, but the abstract lacks enough detail about baselines, workloads, and where quality degrades to justify a stronger verdict.
-
LegoLM: Structured Weight Sharing for Large Language Models
LegoLM compresses language models by sharing weights through scalar-level codebooks, while selectively preserving extreme outliers and protecting the first and last layers. On GPT-2 small and Mistral-7B, the authors report nearly unchanged perplexity and downstream accuracy at roughly 4.4–5.1× compression, substantially better than naive global sharing; they attribute the gain mainly to outlier preservation.
The combination of a diagnosis of why global sharing fails with a simple, data-free fix and striking compression results is worth checking, but the evidence is limited to two models and the unusually large claims need validation against stronger compression baselines and implementation details.
-
Reading is not Reasoning: Bridging the Agentic Policy Gap in Vision-Text Compression
The paper studies why replacing an agent’s growing text interaction history with an image hurts performance, finding that the loss is not just due to imperfect OCR but also changes the agent’s action, stopping, querying, and evidence-use policies. It introduces CAPS, which distills behavior from the model’s text-history policy into its visual-history policy, improving SearchQA and ALFWorld performance while reducing context costs substantially.
The explicit identification of a cross-modal agentic policy gap and its use as a distillation target is a meaningful idea, with large reported gains and memory savings, but the evidence is limited to a few benchmarks and the abstract does not establish broad generality.
-
Adolescent stress recruits a latent amygdala-dopamine circuit to drive punishment-resistant reward-seeking
In adolescent mice, chronic unpredictable stress produced reward-seeking that persisted despite punishment. The authors link this behavior causally to hyperexcitable central amygdala neurons projecting to the substantia nigra pars lateralis, which appear to increase dopamine release in the tail of the striatum; normalizing either circuit activity or dopamine release prevented the effect.
This is a specific and potentially important circuit mechanism linking adolescent stress to addiction-like behavior, supported by two causal circuit manipulations, but the evidence is currently limited to a mouse model and the abstract gives little detail on scale or generality.
-
Dual-Specific Antibody Design Using Artificial Intelligence
The authors describe an AI-assisted platform for designing conventional IgG antibodies whose single binding surface can recognize two unrelated targets, rather than using a conventional bispecific format. They report nine such multibodies covering 15 targets, with claimed high affinity, specificity, developability, and activity across several therapeutic mechanisms; two are entering IND-enabling studies.
Engineering two functional, non-polyreactive specificities into one antibody variable region is a genuinely difficult and potentially important capability, but the abstract gives few quantitative comparisons or independent validation details, so it does not support a strong verdict.
-
The Cellular and Synaptic Actions of Dopamine During Behavior
The study measured and manipulated dopamine while recording identified striatal neurons in awake, behaving mice. Short dopamine changes had little effect on corticostriatal transmission, membrane potentials, or intrinsic excitability, whereas associative learning strengthened corticostriatal synapses onto both D1 and D2 neurons; only D1-neuron plasticity depended on dopamine. This argues that dopamine primarily gates learning-related synaptic plasticity rather than rapidly tuning striatal activity during behavior.
The in vivo cell-type-specific result directly challenges a common model of rapid dopamine-driven striatal excitability and offers a potentially important mechanistic account of how dopamine supports learning, though the abstract provides limited detail on scale and robustness.
-
Phagocytic activity of perivascular cell promotes CNS injury pathology
The study identifies perivascular cells as previously unrecognized phagocytes that accumulate after CNS injury and clear myelin debris through an Axl-dependent pathway involving LC3-associated phagocytosis. In mouse injury models and human stroke tissue, this activity was linked to fibrosis and lesion progression; deleting Axl in these cells or inhibiting it with gilteritinib improved recovery after spinal cord injury.
The proposed role of perivascular cells as pathogenic phagocytes, supported across injury models and human lesions with a manipulable mechanism, is a genuinely non-obvious neuroscience finding, but the abstract provides limited quantitative detail and the therapeutic evidence is still preclinical.
-
Lymphatic CD49a is a driver of meningeal immune aging and cognitive decline.
The study identifies the integrin CD49a as an endothelial-intrinsic regulator of age-related dysfunction in meningeal lymphatic vessels, apparently through control of CCL21 release. In aged mice, deleting CD49a in lymphatic endothelial cells reversed several immune and glial aging changes and improved cognitive and social behavioral deficits.
The paper proposes a specific, targetable mechanism linking lymphatic endothelial aging to meningeal immune dysfunction and behavior, but the abstract provides no quantitative effect sizes or evidence beyond a mouse genetic-deletion model.
-
Skaling: Chinchilla's Exponents Meet Kaplan's Coupling
The paper proposes a neural scaling-law formula in which model size and training data interact through a shared exponent, rather than affecting loss independently. Across interpolation and extrapolation settings, it reportedly cuts prediction error by 1.5–3× and allows sparse low-compute experiments to predict full scaling grids with about 10× less compute.
The explicit capacity–data coupling and claimed reduction in compute for extrapolation are potentially useful, but the abstract gives too little detail about model families, scales, baselines, and robustness to justify a stronger recommendation.
-
Can LLM Agents Stick to the Script? A Benchmark for Long-Horizon Consistency in Interactive Narratives
The paper defines Narrative Commitment Preservation, the ability of an LLM agent to maintain previously established facts, goals, and plot commitments during long interactive narratives despite user interventions. It introduces NCP-Bench with 100 movie-derived environments and finds that even strong models degrade substantially: GPT-5.2 survives only 42% of 20-turn interactions, while fact-conflict rates reach 40–68%, and nearly no runs satisfy all commitments over 100 turns.
The benchmark targets an important and under-measured failure mode in long-horizon agent interaction, with striking consistency failures, but it is still primarily a narrow narrative benchmark rather than a demonstrated advance in agent capability or a broadly validated explanation.
-
ContextMaster: Interactive Multi-Shot Video Creation via Fixed-Budget Sparse Context Routing
The paper frames multi-shot video creation as an interactive workflow in which one model must generate, condition on references, and edit video while preserving context from earlier shots. ContextMaster uses reusable context states, sparse fixed-budget routing, constraint-focused tokens, and teacher-student distillation to keep long-context inference efficient; it reports better cross-shot consistency and task fulfillment, plus 16 FPS on one GPU.
The unified interactive multi-shot formulation and fixed-cost history access are relevant, potentially useful ideas for video systems, but the abstract gives few quantitative comparisons or details supporting the claimed gains.
-
CAPEval: A Decoupled Caption Evaluation across Understanding and Generation
CAPEval separates caption quality into Coverage—how much relevant visual content is mentioned—and Precision—whether the caption’s claims are actually supported by the image. Using human-verified atomic facts and controlled experiments across 10 captioners and four model families, the authors report that Coverage better predicts multimodal understanding performance, while Precision better predicts text-to-image generation performance.
The task-dependent split between caption coverage and factual precision is a non-obvious finding with practical implications for training-data selection, though the abstract provides no quantitative effect sizes and the work is fundamentally an evaluation benchmark.
-
Vision-Language Grounding as Bidirectional Concept Correspondence
The paper reframes vision-language grounding as finding all visually meaningful text spans and matching each to an image instance, rather than grounding text phrases supplied in advance. Its ConCor-1 model uses learned bridge tokens to jointly predict text masks, image masks, and whether a correspondence exists, and reportedly improves correspondence F1 by 48% on long captions and 29% on zero-shot LVIS.
The unified bidirectional formulation addresses a real limitation of standard grounding and reports large relative gains, but the abstract gives no absolute scores or enough experimental detail to establish a major capability advance.
-
YOLO-PEFT: Parameter-Efficient Fine-Tuning on YOLO Family
YOLO-PEFT treats parameter-efficient fine-tuning for object detectors as a constrained module-selection problem rather than blindly applying language-model adapters. On VOC, its planner-selected RS-LoRA substantially outperforms Full-SFT on YOLO11s and YOLO12s, while a memory audit reports 43.9% lower peak training memory at the cost of 1.72× slower training; it also refuses LoRA configurations that fail on RT-DETR-L. The main contribution is explicit, auditable placement and calibrated refusal, though validation is limited to a few detector families and datasets.
The non-obvious finding that generic PEFT can catastrophically fail on detectors, combined with an explicit refusal mechanism and large reported gains over Full-SFT, merits a look, but the evidence is narrow and the planning framework may be an incremental engineering formalization.
-
Meganeura: Portable GPU Training and Inference through Vulkan and Metal
Meganeura is a compact compiler and runtime that uses Vulkan and Metal to support both GPU training and deployment inference across consumer hardware, rather than relying on separate vendor-specific stacks. Across five workloads and several GPU types, it usually runs successfully, compiles much faster than torch.compile, and is sometimes competitive with PyTorch; it also demonstrates transferring a trained decoder into an Android XR application sharing the graphics queue.
The unified train-to-deploy stack across heterogeneous consumer GPUs and graphics APIs is a meaningful systems direction with unusually broad measurements, but the evidence shows useful portability and occasional competitiveness rather than a decisive performance or capability breakthrough.
-
Online Algorithms via Minimax and Posterior Matching
The paper proposes a general way to design online algorithms: track the posterior expectation of the hindsight-optimal fractional solution under the observed arrival history. The authors show that this “posterior matching” rule gives optimal or near-optimal competitive guarantees across several online problems, including set cover, load balancing, matching, and resource allocation, with implications for randomized integral algorithms.
A potentially reusable minimax/posterior-process framework unifies strong results across several classical online problems, but the abstract gives too few quantitative details to establish that it is a major breakthrough rather than an elegant generalization.
-
Bole: Efficient Tree Speculation for Hybrid-Attention Language Models
Bole adapts tree speculative decoding to hybrid-attention language models by rewriting recurrent linear-attention updates into a tree-structured computation, with a specialized GPU kernel and compact state representation. It reportedly speeds linear-attention verification by 3.4–7.7×, cuts transient state memory by 82–99×, and achieves up to 4.72× offline decoding throughput across four models and two GPU platforms, with substantial latency reductions in online agent workloads.
The kernel/runtime co-design addresses a real bottleneck specific to hybrid-attention serving and reports large, broad system gains, but it is a specialized inference optimization rather than a fundamental new modeling direction.
-
Floor, Ceiling, and the Fusion Gap: How Much of Crowd Reading Attention Can Machines Predict?
The paper estimates both a trivial floor and a human split-half ceiling for predicting which sentences independent readers highlight in 120 web documents. Frontier language models capture 35–53% of the gap zero-shot, while fusing five diverse model rankings reaches 60%; a whole-document 8B student retains most of that gain, and the fusion improvement replicates on 217 new documents. The results suggest the useful signal is semantic and document-level rather than merely positional, and that simple cross-model averaging is more effective than a prompt compressor.
The carefully preregistered floor/ceiling framing, replication, and evidence that cheap cross-model fusion captures substantially more crowd-reading signal are genuinely informative, though the core method remains an application of ensemble averaging to a relatively narrow prediction task.
-
Temporal Channel Estimation for Generalized CSI Feedback
The paper proposes TAP, a lightweight neural method that estimates continuous propagation delays from temporal channel measurements and uses those parameters to reconstruct CSI without a learned decoder. Across five 3GPP channel environments, it reports 3.13–12.22 dB lower CFR-NMSE than CsiNet, a 660× smaller model, and 2700× faster inference than iterative OMP, while generalizing across antenna geometries and unseen environments.
The combination of differentiable off-grid delay estimation, decoder-free reconstruction, and claimed zero-shot spatial/environmental generalization is a meaningful advance, but it is a specialized wireless-communications result whose broad robustness is not established by the abstract alone.
-
DeGS: A Scalable 3DGS Architecture via Decoupled Workload Parsing and Reorganization
DeGS redesigns 3D Gaussian Splatting hardware by separating visibility checks and blending, then reorganizing irregular, variable-length workloads into dense batches before parallel blending. In a 28 nm implementation, it reports 1.82–6.02× end-to-end speedups and 1.59–4.42× better energy efficiency than prior 3DGS accelerators, while retaining over 80% PE utilization when scaling from 16 to 1024 PEs at high resolutions.
The decoupled workload-reorganization architecture directly addresses a genuine scalability bottleneck and reports large gains across resolutions, but it is a specialized accelerator advance rather than a broadly transformative 3DGS or ML result.
-
Qwen-CUA: Native Computer Use for (almost) Everything
Qwen-CUA is a very large mixture-of-experts agent trained to operate arbitrary software from screenshots and keyboard/mouse events alone, without DOM or accessibility information. The authors combine massive parallel interactive rollouts, verifiable long-horizon rewards, trajectory slicing, and iterative data refresh; the resulting 397B model scores 86.2 on OSWorld-Verified, while a trillion-parameter version reaches 87.6, with additional results across computer-use benchmarks and a browser deployment.
The combination of screenshot-only native computer use, verifiable long-horizon training, and unusually large-scale rollout infrastructure appears to produce a substantial capability advance, but the abstract does not establish how much comes from the training recipe versus sheer model and compute scale.
-
Structured Memory for Edge Language Models: Persistent Context and Corpus Retrieval via O(1) SSM State Injection
The paper proposes PRECOG, which precomputes SSM hidden states for documents and injects the retrieved state directly into a language model instead of rereading retrieved text. On a 1.2B-parameter gated-SSM model, this reportedly reduces RAG prefill latency from about 27 seconds to under 6 ms while preserving in-context RAG answer quality; it also introduces a hierarchical persistent-memory scheme built from the same state-injection idea.
Directly injecting precomputed recurrent states could be a genuinely useful edge-RAG direction with a striking claimed speedup, but the abstract gives limited evidence beyond one SSM model and makes a strong, questionable-sounding claim that a fixed hidden state is a complete document summary.
-
AnchorKV: Anchor-Residual KV Cache Compression
AnchorKV compresses long-context transformer KV caches by storing a small set of exact anchor tokens and representing other tokens relative to their most similar anchor, selectively refining approximations that matter most to the output. The authors claim roughly 20× compression without token eviction and retention of 99% of full-cache performance at 70B scale, though the abstract gives few experimental details.
The combination of lossless-in-token-coverage compression with output-aware residual refinement could materially improve long-context inference, but the unusually strong compression and accuracy claims are not supported by enough results in the abstract to merit a strong verdict.
-
Getting the Parameters Right: A Difficulty-Graded Benchmark and Probe-Guided Training for LLM Tool Calls
This paper studies a neglected failure mode in tool-using LLMs: selecting the right parameter values for an API call. It finds that hidden states contain a usable signal for whether a generated value is correct, then uses a linear probe to filter synthetic training data and rerank candidate calls; across five models and six external benchmarks, exact-match accuracy rises from 19.7% to 59.6%.
The large reported gain and probe-guided approach make this more than a routine tool-calling tweak, but the abstract does not establish how much depends on benchmark construction, candidate generation, or probe supervision, so it merits a closer look rather than a strong recommendation.
-
LAEF: A Lead-Agnostic ECG Foundation Model Towards Point-of-Care Diagnostics
LAEF is a 7M-parameter ECG foundation model that represents whichever leads are available as a variable-size physiological graph, rather than forcing missing leads into a fixed 12-lead architecture. Pretrained on 9.2M ECGs with randomly masked leads, it matches much larger 12-lead models when all leads are present and improves single- and two-lead diagnosis across many downstream datasets, with a reported average AUROC gain of 3.2 points.
The lead-agnostic architecture addresses a real deployment bottleneck and is supported by broad multi-dataset results, though it remains a specialized medical application and the abstract does not establish a field-wide breakthrough.
-
MetaVideoAgent: Automated Video-Agent Evolution for Long-Form Video Understanding
MetaVideoAgent automatically adapts a modular long-video reasoning agent to a target video distribution. It uses sparse profiling, failure-focused validation tasks, and attribution of errors to specific modules to evolve the agent; across eight distributions, it improves accuracy from 38.44% to 51.47% while using fewer frames and tokens than competing agents. The main novelty is applying structured, code-level agent evolution to long-form video understanding rather than using one fixed agent design.
The sizable gains across eight held-out distributions and reduced inference cost make this worth a look, but the evidence is primarily on a newly introduced benchmark and the contribution may be a substantial engineering framework rather than a fundamentally new learning paradigm.
-
DreamWAM: Beyond RGB Future Prediction for World Action Models
DreamWAM trains a world-action model to predict future motion, geometry, and semantics alongside RGB, so its learned representation emphasizes states relevant to manipulation rather than nuisance visual changes. The extra prediction branches are used only during training, while deployment remains RGB-only; it improves LIBERO performance modestly in-distribution but substantially under unseen perturbations, including a real-world success-rate increase from 55.6% to 74.4%.
The beyond-RGB, action-relevant future-state formulation is a meaningful modeling idea, and the large robustness gains under visual and layout shifts are supported by both simulated and real-robot results, though the core approach remains an incremental extension of predictive world models.
-
PathCover: A Fast Convex Decomposition along a Path via Randomized Iterative Space Partitioning (RISP) on Point Clouds
PathCover introduces RISP, a randomized method for constructing overlapping convex, obstacle-free polytopes directly from LiDAR point clouds along a planned path. The authors claim expected linear-time behavior, formal termination and progress guarantees, and roughly 10× faster corridor generation than existing methods while retaining similar corridor volume; they also report simulation and quadruped-robot validation.
The combination of a theoretically motivated randomized decomposition algorithm, order-of-magnitude speedup, and live-robot validation is notably stronger than a routine corridor-generation tweak, but the abstract lacks quantitative scale, failure-rate, and baseline details needed to justify a strong verdict.
-
SoRoMoX: Fast, Differentiable, and Parallelizable Soft Robot Models
SoRoMoX is a JAX-based framework for differentiable, GPU-parallel simulation and control of soft robots using Cosserat-rod and strain-based models. It reports up to 18.1× faster CPU rollouts and 234.6× higher GPU throughput, enabling improved system identification, residual learning, model-based control, safety-constrained control, and faster reinforcement-learning training.
The combination of end-to-end differentiability, GPU parallelism, and control-oriented soft-robot dynamics could remove a meaningful tooling bottleneck, but the abstract provides broad performance claims without enough detail to judge how general or reproducible the gains are.
-
Circuit Fine-Tuning for Compute-Efficient Transformer Adaptation
The paper uses circuit-discovery methods to identify which parts of a pretrained ViT respond to a new task, then fine-tunes only those modules. Unlike standard parameter-efficient methods that still execute most of the model’s training computation, CFT reportedly reaches peak accuracy in about 20 epochs versus 44–96 for baselines, reducing training FLOPs by 2.3–6.6× and wall-clock time by up to 16× without adding inference parameters or operations. Results span VTAB-1k, Swin, medical imaging, and a vision-language model.
The distinctive use of probe-head circuit discovery to select a trainable subgraph, combined with substantial reported training-cost reductions across several model and domain settings, makes this worth a closer look, though the abstract does not establish whether accuracy is consistently preserved or improved.
-
DCAS: Decoupling CLI Agent Scaffolding to Internalize Planning across Scaffolds
The paper argues that CLI software-engineering agents fine-tuned on OpenHands trajectories learn scaffold-specific planning conventions, which causes them to degrade when run with other agent scaffolds. It introduces DCAS, an API-interception layer for collecting and evaluating trajectories across scaffold/backend combinations, and reports that planning-aware trajectories improve cross-scaffold performance; it also separates explicit plans from implicit scaffold-driven planning structure.
The finding that fine-tuning can create scaffold-specific planning dependence, plus a way to disentangle and measure it, is non-obvious and relevant to agent training, but the abstract gives no concrete task, scale, or numerical results to establish a major advance.
-
Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory
The paper uses successful trajectories from a large teacher agent to build three levels of external memory—task workflows, subtask examples, and tool-function conventions—and supplies these memories to smaller agents without retraining them. Across three tool-use benchmarks, this reportedly improves 4B–8B students by 27.2, 11.2, and 3.4 percentage points, with the largest gains coming from subtask memory and from the smallest models.
The hierarchical separation of proactive workflow/subtask guidance from reactive tool-error memory is a plausible and useful agent design, and the reported gains—especially on AppWorld—are substantial, but the abstract does not establish how robust the result is beyond these benchmarks or how much comes from strong teacher-generated demonstrations.
-
When Memory Lies: An Empirical Study of Spatial Memory Staleness in VLM Agents
This study evaluates whether vision-language agents notice when persistent spatial memories have become outdated, using dynamic FrozenLake navigation tasks across six VLMs and multiple navigators. It finds a large gap between text-based and visual stale-memory detection, and shows that blindly trusting stale memory can make navigation worse than having no memory; filtering helps in text settings but not reliably when visual grounding is weak.
The result that stale memory can be more harmful than no memory, together with the extreme text-versus-vision grounding gap, is a useful non-obvious warning, though the evidence comes from a narrow synthetic benchmark and the paper introduces no new agent method.
-
What to Edit Next: Visually Aligned Image-Editing Follow-Up Suggestions in Conversational Systems
The paper builds a multimodal system that suggests the next image edit in a conversational image-generation workflow. It combines human-labeled edit intents, click-feedback reinforcement learning, and a visual-consistency verifier; in a million-user A/B test, it reports a reduction in inconsistent suggestions from 3.7% to 0.9%, along with sizable gains in click-through rate, image saves, and conversation length. The main contribution is a large-scale, production-oriented formulation and evaluation of image-dependent follow-up recommendations rather than a fundamentally new model architecture.
The large real-world deployment and strong A/B-test results make this worth knowing about, but the three-stage combination of supervised tuning, feedback optimization, and visual verification is largely an expected engineering approach rather than a clearly new research direction.