Issue 27 · Jun 29 – Jul 5, 2026
Every candidate
All 3,322 papers were scored from their abstracts by gpt-5.6-luna; 1,164 were not skipped. Shown here: the top 150 of those, in score order. Picks are marked.
-
Brand-as-Memory: Vision-Language Models Encode Causal, Mechanistically Localizable Credibility Priors for News Sources
The paper shows that vision-language models use a news outlet’s visual identity—its name, logo, or domain—as a credibility prior that can override evidence in the article itself. Across seven models, the authors measure this effect, localize its formation to specific layers and sparse-autoencoder features, and show that steering the identified direction reduces the behavior and generalizes to unseen outlets.
This is a potentially important and assumption-challenging finding: VLMs appear to rely causally on publisher identity rather than just interpreting article evidence, supported by cross-model tests, controlled cue swaps, mechanistic localization, and intervention.
-
Program-as-Weights: A Programming Paradigm for Fuzzy Functions
The paper proposes compiling a natural-language specification of a fuzzy programming task into a small neural adapter (“program-as-weights”) for a frozen local interpreter. A 0.6B Qwen3 interpreter with these adapters reportedly matches directly prompted Qwen3-32B on their FuzzyBench tasks, while using about 1/50 the inference memory and running at 30 tokens/s on an M3 Mac; the compiler is invoked once per function definition rather than per input.
The reusable compilation of language specifications into neural programs, coupled with a large apparent efficiency gain over prompting a much larger model, is a genuinely interesting direction, but the evidence is currently limited to a newly introduced benchmark and abstract-level claims.
-
Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent
Agents-A1 is a 35B-parameter mixture-of-experts model trained to improve agent performance by extending interaction trajectories rather than primarily increasing model size. It uses roughly 45K-token knowledge/action/observation/verifier trajectories, domain-specific teacher models, and routed on-policy distillation to combine six types of expertise; the authors report leading results on several long-horizon benchmarks and competitiveness with trillion-parameter models.
The potentially important result is that long-horizon training and heterogeneous-teacher distillation let a much smaller agent match or exceed much larger models, but the abstract does not establish that the comparisons are broad, matched, or free of benchmark and inference-compute advantages.
-
Hierarchical Sparse Attention Done Right: Toward Infinite Context Modeling
The paper proposes HiLS Attention, which selects relevant chunks hierarchically and trains the selection mechanism directly through the language-modeling loss. It reports near-full-attention performance at training-length contexts, over 64× context-length extrapolation with 90% chunk-retrieval accuracy, and conversion of existing full-attention models using lightweight continued pretraining while retaining sparse computation.
Learned chunk selection integrated into the forward attention computation, together with the claimed extreme length extrapolation and retrofit path for existing models, is substantially more interesting than a standard sparse-attention tweak, but the abstract lacks enough scale, baseline, and task detail to justify a strong verdict.
-
GEAR: Guided End-to-End AutoRegression for Image Synthesis
GEAR jointly trains a VQ tokenizer and an autoregressive image generator instead of freezing the tokenizer first. It uses a hard code assignment for next-token training and a soft assignment for a separate alignment loss, allowing the generator to influence which discrete codes the tokenizer produces without backpropagating through the discrete indices. The authors report up to 10× faster ImageNet gFID convergence than LlamaGen-REPA, improved spatial features, and transfer across several quantizers and text-to-image generation.
The end-to-end tokenizer–autoregressive-generator coupling and dual-readout gradient mechanism are a substantive departure from the usual two-stage pipeline, with a potentially important convergence improvement, but the abstract lacks enough quantitative detail and comparisons to justify a strong verdict.
-
A Machine-Verified Proof of a Quantum-Optimization Conjecture
The paper claims to resolve the decade-old Farhi–Goldstone–Gutmann conjecture that depth- p QAOA achieves exactly the (2p+1)/(2p+2) approximation ratio on the ring of disagrees. An LLM discovered a proof based on a hidden dynamical symmetry, and the complete proof was checked in Lean 4 after the authors formalized the relevant quantum-information machinery.
If the formalization faithfully captures the conjecture, this is a rare machine-verified resolution of a longstanding quantum-optimization problem and an unusually concrete demonstration of an LLM discovering nontrivial mathematical structure.
-
DeepGaze3.5-VL: Modeling Scanpaths via Autoregressive Token Prediction
DeepGaze3.5-VL treats human eye-movement scanpaths as autoregressive token sequences, using a vision-language model rather than a specialized gaze architecture. It reports 2.18 bits of information gain on MIT1003, a 46% improvement over DeepGaze III that remains when both methods use the same vision encoder, and supports conditioning on viewer identity, task instructions, and fixation duration. The model also enables in-silico interventions on fixation durations and recovers known oculomotor effects from observational data.
This combines a genuinely flexible sequence-modeling formulation with a large reported gain and an unusually interesting use of the generative model for causal-style behavioral interventions, though the claims still need scrutiny beyond the abstract.
-
Dictionaries, Not Darwin: Set-Level Selection Beats LLM Evolution in Scientific Equation Discovery
The paper argues that iterative LLM evolution does not reliably improve scientific equation discovery: under matched call budgets, repeatedly selecting and recombining parents performs about like independent proposal sampling. It replaces evolution with a one-shot dictionary of reusable candidate terms followed by train-only, set-level sparse selection, reporting 73.2–77.0% accuracy on LLM-SRBench versus 49.2% for the best baseline while using one tenth the call budget. The key claim is that data supports evaluating combinations of terms jointly, rather than assigning credit to individual generated expressions.
It presents a potentially general correction to a popular LLM-evolution recipe, backed by a large reported benchmark and major call-efficiency gains, though the evidence is still concentrated in scientific equation discovery.
-
DataEvolver: Self-Evolving Multi-Agent Data Construction for Text-Rich Image Generation
DataEvolver turns image-data construction into an iterative feedback loop: separate agents retrieve samples, diagnose OCR and semantic failures, summarize recurring problems, and synthesize data for under-covered cases. At a matched 0.75M-sample scale, the resulting training data improves OCR-F1 substantially on two text-rich image benchmarks—by 85.3% and 35.3% over the strongest baseline—and transfers across two image generators.
The notable contribution is using rejected examples as structured feedback to evolve the data-construction policy, with unusually large reported gains and cross-generator transfer, though the abstract lacks enough detail to establish how much comes from the multi-agent design versus stronger filtering or synthesis.
-
ASPIRE: Agentic /Skills Discovery for Robotics
ASPIRE is a continual robot-programming system that uses multimodal execution traces to diagnose failures, synthesize code repairs, and store successful fixes as reusable skills. It combines this skill library with evolutionary task and program search, reporting large gains under perturbations and on long-horizon manipulation tasks, including 31% versus 4% on an unseen LIBERO-Pro split. The paper also reports preliminary transfer of simulation-discovered skills across real robot embodiments and APIs.
The combination of autonomous code-policy repair, persistent cross-task skill accumulation, and evolutionary exploration is a potentially important direction, with unusually broad reported gains, but the abstract gives limited detail on baselines, evaluation protocols, and how much of the result comes from system-scale engineering or retries.
-
One-Step Gradient Delay is Not a Barrier for Large-Scale Asynchronous Pipeline Parallel LLM Pretraining
This paper studies why one-step stale gradients in asynchronous pipeline parallelism hurt training, arguing that the problem is largely optimizer-dependent rather than inherently unstable. It finds that AdamW degrades substantially, while Muon remains robust, and proposes an optimizer-agnostic error-feedback correction; experiments up to 10B parameters reportedly close the quality gap with synchronous training.
The potentially important result is that a practical asynchronous pipeline schedule may work at scale when paired with a suitable optimizer, challenging the assumption that even one-step staleness is fundamentally harmful, though the abstract gives no quantitative throughput or quality results.
-
AutoMem: Automated Learning of Memory as a Cognitive Skill
AutoMem treats memory management—deciding what to store, retrieve, and organize—as a trainable capability rather than a fixed prompting or tool-design choice. It uses one loop to automatically refine the memory-file structure and another to train the agent from its own successful memory decisions; across Crafter, MiniHack, and NetHack, this reportedly improves performance by roughly 2–4× without changing task-action behavior, making a 32B model competitive with larger frontier systems.
The unusually large gains from optimizing memory alone and the framing of memory management as an independently learnable skill are worth checking, but the evidence is limited to procedurally generated game environments and the abstract lacks detailed comparisons and ablations.
-
Representation Distribution Matching for One-Step Visual Generation
The paper studies one-step image generation by matching feature-distribution statistics between generated and real images using frozen encoders. It finds that properly estimated MMD works well with very large batches, while any single encoder can be exploited, motivating a balanced 14-encoder objective and evaluation metric. The resulting model reportedly improves ImageNet one-step generation and distills four-step FLUX.2 [klein] into a one-step model that scores better on GenEval and PickScore, using 90 H200 GPU-hours.
The combination of distribution matching, unusually large operative batch sizes, and multi-encoder anti-gaming evaluation is a substantive approach to one-step generation, with credible ImageNet and FLUX results, but the strongest claims depend partly on a newly introduced metric and need independent validation.
-
Look Before You Leap: Distilling Tree Search into Action Evaluation for Frozen VLA Models
SVA keeps a pretrained VLA policy frozen, samples multiple candidate actions, and trains a separate Q-value evaluator from Monte Carlo tree-search trajectories in simulation. At deployment, the evaluator chooses among the VLA’s proposed actions without needing a simulator; the paper reports better unseen-task generalization and claims that a 9B model can beat a 27B model with lower latency.
The proposal to improve frozen VLAs by distilling search into a consequence-aware action evaluator, rather than fine-tuning the policy, is a meaningful and potentially general strategy, but the abstract gives no benchmark-scale details and the headline comparison may depend heavily on the evaluation setup.
-
Perceptual Flow Matching for Few-Step Generative Modeling
The paper changes flow-matching training from predicting velocities in VAE latent space to predicting them in a pretrained perceptual feature space. It reports that this enables high-quality image, video, and editing generation in 4–8 sampling steps instead of the usual 35–50, without teacher models or extra score networks, and argues that perceptual supervision makes predictions more mode-seeking and robust to coarse integration.
The representation-space training objective is a fairly clean and potentially general way to obtain large few-step speedups without distillation, but the abstract gives no quantitative comparisons or model-scale details sufficient for a strong verdict.
-
Video-Oasis: Rethinking Evaluation of Video Understanding
The paper audits existing video-understanding benchmarks for shortcuts, such as questions answerable from language priors or a single frame. It reports that 55% of samples can be solved without visual input or temporal context; on the filtered video-native subset, leading Video-LLMs perform only marginally above chance, and the authors use this subset to compare design choices for more robust understanding.
The unusually large shortcut rate and near-chance performance on genuinely video-dependent examples challenge the validity of common benchmark results, though the paper is primarily an evaluation audit rather than a demonstrated capability advance.
-
Thinking While Speaking: Inference-Time Knowledge Transfer for Responsive and Intelligent Conversational Voice Agents
The paper trains a small, fast “talker” model to begin answering immediately, then incorporate streamed reasoning, retrieval, or tool-use results from a slower external model while speaking. Across seven small models and a synthetic 290k-example dataset, ConvFill reportedly preserves millisecond-level initial latency and comes within 6.3% of frontier-reasoner accuracy; an 18-person study found it more responsive and competitive overall, especially for retrieval-heavy tasks.
Inference-time knowledge transfer that hides external reasoning latency is a genuinely relevant voice-agent direction, but the evidence relies heavily on synthetic training data, an ambiguously defined accuracy gap, and a small user study, so it merits a look without being a must-read.
-
DuoMem: Towards Capable On-Device Memory Agents via Dual-Space Distillation
DuoMem transfers an agent’s procedural problem-solving ability from a large teacher into a small model in two ways: by supplying teacher-written memories at inference time and by fine-tuning lightweight LoRA adapters on successful teacher trajectories. On ALFWorld, this raises a 4B model’s success rate from 4.3% to 77.9%, near the 72B teacher’s 87.1%, while using under 10M trainable parameters and running over 3× faster. The main limitation is that the evidence is concentrated on one benchmark and depends on precomputed teacher-generated memories.
The unusually large small-model gain and dual use of external memories plus parameter distillation are worth attention, but the result is not yet broad enough to establish a general on-device agent capability.
-
Is One Layer Enough? Training A Single Transformer Layer Can Match Full-Parameter RL Training
The paper studies which transformer layers actually change during RL post-training of LLMs. Across seven Qwen models, three RL algorithms, and tasks including math, coding, and agentic decision-making, it reports that updating one layer—usually in the middle of the stack—recovers most of the improvement from updating all parameters, with layer rankings stable across settings.
The claimed concentration of RL adaptation in one or a few middle layers could substantially reduce post-training cost and challenge the assumption that all layers must be updated, but the abstract gives no quantitative gains, compute savings, or ablation details to establish how robust the effect is.
-
Scaling Storm-Resolving Atmospheric AI Simulation to the Entire Planet
The paper introduces STRATA, an autoregressive transformer that emulates the atmosphere globally at 4.9-km resolution by training on local tiles and blending them during rollout. Using only 17 days of high-resolution SCREAM data, it produces stable 24-hour global simulations with realistic convective-scale structure, reportedly at about 50× the energy efficiency of the physics model, though large-scale biases accumulate with lead time.
Global storm-resolving AI simulation at kilometer resolution is a substantial new direction with an impressive efficiency claim, but the evidence is limited by only 17 training days, 24-hour evaluation horizons, and acknowledged rollout bias.
-
Random Reshuffling Dominates Stochastic Gradient Descent
This paper analyzes Random Reshuffling, the common practice of randomly permuting data before each pass, versus independently sampled SGD for smooth convex optimization. It claims to prove that Random Reshuffling achieves better convergence than SGD after any finite number of epochs with any reasonable stepsize, resolving the longstanding theory–practice gap around the method’s apparent superiority.
The claimed dominance theorem addresses a genuine open question and could substantially change how shuffled training is theoretically understood, but the abstract gives no theorem conditions, quantitative comparison, or scope beyond smooth convex objectives, so it falls short of a strong recommendation on evidence alone.
-
MorphQuad: Morphable Quadrotor for Superhuman Maneuverability, Manipulation, and Resiliency
MorphQuad is a quadrotor whose four rotor assemblies can each tilt independently using two-axis gimbals, paired with a globally stable controller and thrust-allocation method. The resulting vehicle can rotate repeatedly while translating, maintain control under directional wind and impacts, and perform contact tasks such as turning valves, pressing, pushing, and perching. The main novelty is combining omnidirectional flight, forceful contact, and disturbance rejection in a compact quadrotor-like platform through hardware–control co-design.
The independently articulated rotors and globally stable control address a real limitation in aerial manipulation and enable an unusually broad set of demonstrated behaviors, but the abstract gives few quantitative comparisons or details supporting the superhuman claims.
-
Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads
The paper introduces LOCOS, a detector for attention heads that retrieve information by transforming attended context through their output-value circuit rather than copying tokens literally. Across Qwen3, Gemma-3, and OLMo-3.1, ablating the heads it identifies sharply degrades non-literal retrieval while largely preserving unrelated parametric recall and arithmetic reasoning; on Qwen3-8B, 50-head ablation reduces NoLiMa ROUGE-L from 0.401 to 0.000 and also substantially harms MuSiQue and BABI-Long.
This is a concrete, write-aware alternative to copy-based retrieval-head detection with strong causal ablation results across models, though its significance is mainly for mechanistic analysis rather than a broad capability or efficiency advance.
-
Vidu S1: A Real-Time Interactive Video Generation Model
Vidu S1 is presented as a real-time interactive video generator that responds to voice commands while maintaining persistent digital characters, including user-provided people, anime characters, and pets. The authors claim 540p generation at up to 42 FPS on consumer GPUs and arbitrarily long generation without drift, using TurboDiffusion and TurboServe, but provide no concrete results or comparisons in the abstract.
Real-time, voice-controlled, persistent video generation would be a meaningful capability advance, but the abstract reads partly like a product announcement and gives insufficient quantitative evidence to support its unusually strong claims.
-
One Forward Beats Two: InnerZoom for Accurate and Efficient GUI Grounding
InnerZoom improves GUI element localization by carrying target-region information from intermediate decoder layers into the later coordinate-generation layers, rather than running a separate crop-and-rerun zooming pass. In experiments, a 4B model improves over prior results by 2.3–4.1 points on several GUI grounding benchmarks, beats a two-pass ZoomIn baseline by 1.3 points on average, and reduces latency by up to 31.8% and compute by about 29%.
The single-forward cross-layer evidence mechanism addresses a real accuracy–latency tradeoff in GUI grounding and reports consistent gains, though the claims are still benchmark-based and the underlying idea may be specialized rather than broadly transformative.
-
SkillOpt-Lite: Better and Faster Agent Self-evolution via One Line of Vibe
SkillOpt-Lite frames agent skill improvement as zeroth-order optimization over editable skill code, using trajectory exploration, consensus-based feedback, and independent validation. The authors report sizable gains over a more elaborate SkillOpt pipeline, including a smaller model beating a larger model on LiveMath and SpreadsheetBench, and demonstrate integration into coding agents.
The combination of a deliberately minimal self-evolution loop with strong reported gains is worth checking, but the abstract gives limited experimental detail and relies heavily on claims involving proprietary model versions and selected benchmarks.
-
Are We Measuring Strategy or Phrasing? The Gap Between Surface- and Approach-Level Diversity in LLM Math Reasoning
This paper distinguishes diversity in the wording of LLM math solutions from diversity in the underlying solution strategy, using a human-calibrated LLM judge to measure the latter. It reports that common diversity metrics and diversity-aware RLVR can preserve surface variation while reducing strategic variation, whereas approach-diverse candidate sets improve test-time scaling; directly training on the judge reward instead leads to reward hacking rather than genuinely broader strategies.
The separation of surface and strategy diversity, together with evidence that diversity optimization can select for phrasing instead of new approaches, is a non-obvious and potentially useful diagnosis, but the abstract gives limited quantitative evidence and relies heavily on an LLM-judge framework.
-
Learning to Move Before Learning to Do: Task-Agnostic pretraining for VLAs
The paper separates learning general robot motion from learning which task to perform. It pretrains a VLA’s motor representations with self-supervised inverse dynamics on unlabeled, off-task interaction data, then uses a small amount of expert language-action data for task grounding. It reports performance comparable to models trained on more than 1M expert trajectories, a 10-point gain over behavior cloning on SIMPLER, and better robustness to camera perturbations on a WidowX robot.
The potentially important result is that cheap, task-agnostic robot interaction can replace much of the expert demonstration data needed for VLA training, but the abstract provides limited evidence beyond one benchmark and one real-robot platform.
-
The State-Prediction Separation Hypothesis
The paper argues that a Transformer’s hidden stream is doing two different jobs: predicting the next token and maintaining information useful for later predictions. It introduces a two-stream architecture that separates these roles, and reports better validation loss, improved data/compute efficiency, and roughly 2–3 percentage-point downstream gains across multiple pretraining scales, with analyses attributing the effect to different gradient behavior.
Separating prediction from persistent state is a substantive architectural hypothesis with potentially broad efficiency implications, but the abstract gives insufficient detail about model scales, baselines, and the size and breadth of the downstream improvements for a stronger recommendation.
-
Reliability Scales Inversely: Hallucinations Snowball Faster in Bigger Language Models
The paper argues that larger language models may reduce their initial knowledge gap while becoming worse at recovering once an answer starts to go wrong. Using several model families and benchmarks, it attributes this “snowballing” to a persistent per-token disagreement with a stronger oracle that is not reflected in the model’s own uncertainty, and reports that reducing this risk component cuts web-verified hallucinations by 35–74%.
The claimed scaling reversal and a confident, self-invisible failure mode are genuinely surprising and potentially important, but the abstract uses unusually strong causal language and the proposed oracle-based decomposition and intervention need close methodological scrutiny.
-
The Judge Knows When It Knows: Calibrated Abstention for LLM-Based A/B-Test Prediction
This preregistered study tests whether multimodal LLMs can predict which web-page variant will win real A/B tests from screenshots. Across 330 experiments, models perform poorly overall, and common improvements such as larger models, better prompts, and higher-fidelity stimuli do not help; however, abstaining on low-consensus cases yields a subset with better agreement, while both model and human consensus are shown to reflect shared priors rather than reliable prediction. The paper also audits noisy A/B-test labels and releases a detailed experimental ledger and human baseline.
The valuable contribution is not another vision-language benchmark but a well-controlled negative result showing that apparent consensus among LLMs or experts can be highly correlated yet predictively useless, alongside evidence that calibrated abstention partially identifies the exceptions.
-
Multiregional single-cell profiling reveals shared and specialized cellular vulnerability in Alzheimer's disease
The authors profile about seven million nuclei across ten brain regions from 84 Alzheimer’s donors, combining single-nucleus RNA/ATAC sequencing, pathology, and genetics, with replication in more than 700 additional donors. They find that a relatively small set of cell types changes in abundance, but those changes are highly consistent across regions: selected interneurons and oligodendrocytes are lost early, while excitatory neurons, microglia, and astrocytes change later. An unexpected result is that a specialized layer-4 population in primary visual cortex—often considered relatively resilient—becomes vulnerable, with both it and early-lost Sst interneurons showing molecular signatures of hyperexcitability.
This is an unusually large, replicated multiregional analysis that identifies coherent early cell-type vulnerability and challenges the presumed resilience of primary visual cortex, though its mechanistic conclusions remain largely hypothesis-generating rather than experimentally established.
-
WARP: Weight-Space Analysis for Recovering Training Data Portfolios
WARP estimates the proportions of domains in a model’s fine-tuning data using only the released base and fine-tuned weights. It creates interpolated pseudo-checkpoints, extracts geometric features from their weight-space trajectory, and maps those features to mixture proportions; in controlled BERT and GPT-2 experiments, it reports mixture MAEs of 0.046 and 0.104, respectively. The notable idea is treating weight-space geometry as a global fingerprint of the training-data composition rather than inferring membership of individual examples.
This is a genuinely interesting new framing for model auditing and data-mixture inference, but the evidence is limited to controlled experiments on two older model families, so its robustness to realistic foundation-model training and mixtures is still unclear.
-
Multimodal Continuous Reasoning via Asymmetric Mutual Variational Learning
The paper addresses a train–test mismatch in continuous latent reasoning for multimodal language models: the training posterior sees the answer, while the inference-time prior does not. It proposes Asymmetric Mutual Variational Learning, using both forward and reverse KL objectives to prevent answer-dependent latent representations, and reports substantial gains on the BLINK visual reasoning benchmark, including a +10.83 average improvement and up to +32.00 on individual tasks.
The bidirectional variational calibration and explicit framing of answer leakage are meaningful ideas, but the abstract provides evidence mainly from one benchmark and does not establish broad superiority or real-world impact.
-
FlexiSLM: A Spoken Language Model with Dynamic and Controllable Frame Rates
FlexiSLM applies a dynamic-frame-rate speech codec to a speech-to-speech language model, allowing both input compression and output generation at a user-controlled rate. The authors report comparable or better quality than fixed-rate 7B systems at 12.5 and 6.25 Hz, with roughly half the inference time at 6.25 Hz and usable control down to 4 Hz.
This is a meaningful efficiency and controllability advance for speech-to-speech models, but the abstract does not establish how broad or fair the comparisons are, nor whether quality remains strong across diverse tasks and real-time settings.
-
OmniTacTune: Policy-Agnostic Real-World RL for Tactile Residual Adaptation of Visual Policies
OmniTacTune adds a learned tactile residual to pretrained visual robot policies, using autonomous base-policy rollouts to bootstrap online reinforcement learning and then adapting through real-world interaction. On four contact-rich manipulation tasks, it reportedly raises success from 5–40% to 85–100% in 40–80 minutes and transfers across visual policies and tactile representations.
The combination of policy-agnostic tactile residual adaptation and relatively fast real-world learning is a meaningful capability advance, but the abstract lacks task details, comparisons, and ablations needed to establish how broadly the result generalizes.
-
Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts
The paper proposes DART, a weight-arithmetic method for adapting vision-language-action models to changes in camera viewpoint or robot embodiment using just one target-domain demonstration. It aligns singular-vector subspaces before adding domain-specific weight components, and reports better one-shot adaptation than prior methods in simulated and real-world experiments.
One-shot adaptation across both visual and robot-body shifts is a useful capability, and the subspace-aligned weight-arithmetic approach is a plausible new direction, though the abstract gives no quantitative scale or comparison details.
-
OrbitQuant: Data-Agnostic Quantization for Image and Video Diffusion Transformers
OrbitQuant proposes a calibration-free post-training quantization method for diffusion Transformers. It uses a randomized block-Hadamard rotation to make activation distributions approximately fixed across timesteps, prompts, guidance branches, and modalities, allowing one codebook to be reused; rotations for weights can be absorbed offline. The method is evaluated on several image and video diffusion models and reportedly maintains usable quality at W2A4, but the abstract gives no quantitative quality or speed results.
The data-agnostic, cross-timestep and cross-modality calibration strategy is a meaningful departure from standard PTQ, but the abstract lacks the quantitative evidence needed to judge whether the claimed low-bit quality and practical gains are substantial.
-
Multi-Resolution Flow Matching: Training-Free Diffusion Acceleration via Staged Sampling
MrFlow accelerates pretrained flow-matching text-to-image models by generating coarse structure at low resolution, upsampling with a lightweight GAN, then adding noise and refining at high resolution. On FLUX.1-dev and Qwen-Image, it claims roughly 10× end-to-end speedup with less than a 1% OneIG degradation, and up to 25× when combined with timestep distillation, without additional training or runtime heuristics.
The reported training-free 10× acceleration with minimal quality loss is practically important, but the core staged multi-resolution idea is an extension of an active, already established acceleration direction and the abstract provides limited evidence beyond two models and headline metrics.
-
SWE-INTERACT: Reimagining SWE Benchmarks as User-Driven Long-Horizon Coding Sessions
SWE-Interact evaluates coding agents in multi-turn workflows where users begin with vague requests, progressively add requirements, inspect changes, and provide feedback. Strong models that solve about 50% of conventional single-turn SWE tasks solve only about 25% of corresponding interactive tasks, suggesting that intent discovery, requirement retention, and iterative refinement are distinct weaknesses.
The benchmark targets a plausibly important and undermeasured capability—long-horizon coding with evolving user requirements—and reports a substantial failure to transfer from standard SWE benchmarks, though it is still primarily an evaluation paper and the abstract gives limited evidence about task scale and validity.
-
CausalMix: Data Mixture as Causal Inference for Language Model Training
CausalMix treats the choice of training-data mixture as a causal treatment problem: dataset statistics are covariates, and model performance is used to estimate which mixture works under different data-pool conditions. It fits this model using 512 runs of Qwen2.5-0.5B, then transfers the inferred mixture to an 800K-example pool and a 7B model, with an additional test on long chain-of-thought data using Qwen3-4B-Base. The main claimed benefit is adapting mixture weights when the data pool changes, rather than retraining a proxy optimizer from scratch.
The causal, state-dependent formulation of data-mixture optimization is a meaningful framing, but the abstract gives limited evidence that causal modeling genuinely improves extrapolation beyond a sophisticated proxy method, and the 512-run/small-model foundation makes the broad scaling claims uncertain.
-
Smooth Scaling Laws Hide Stepwise Token Learning
The paper explains aggregate language-model scaling laws as the result of many contextualized tokens undergoing relatively sharp, sigmoid-like learning transitions at different training times. Across over 100 runs—up to 6B parameters and 300B tokens—the authors claim that the distribution of these token learning times reconstructs loss scaling across training, data, and model size, and that reweighting data using this signal speeds validation-loss reduction by 11%.
This offers a potentially important token-level explanation of scaling laws with unusually broad empirical support and an actionable training intervention, but the abstract does not establish whether the sigmoid decomposition is causal or more than a useful fit, so it falls short of strong.
-
Coordinate Singularities Break Conformal Coverage for Gaze and Head Pose
The paper shows that conformal prediction can provide apparently correct overall coverage while severely under-covering gaze directions and head poses near coordinate singularities, because errors and prediction sets are scored in yaw-pitch or Euler-angle charts. Across four datasets, nominal 90% coverage falls to roughly 39–58% in problematic regions; the authors prove that scalar calibration cannot remove the chart distortion and show that geodesic, coordinate-free scoring largely fixes it without retraining.
This identifies a concrete, structurally explained failure mode in conformal vision prediction with broad empirical evidence and a cheap remedy, though its impact is concentrated on spherical and rotational outputs rather than conformal prediction generally.
-
Accurate Decoding of Natural Sentences from Non-Invasive Brain Recordings
Brain2Qwerty v2 decodes naturally typed sentences from real-time MEG recordings, using neural models that combine character-, word-, and sentence-level information. Across nine subjects and 22,000 sentences, it reaches 39% average word error rate; the best subject has half of sentences with at most one incorrect word, and performance improves roughly log-linearly with more training data. The work also uses pretrained language-model representations and automated agent-assisted pipeline refinement.
Non-invasive, sentence-level brain-to-text with substantial subject-specific accuracy and a credible data-scaling result is genuinely notable, but the average error rate, small cohort, and likely subject-specific training keep it short of a strong verdict.
-
Behavior Cloning is Not All You Need: The Optimality of On-Policy Distillation for Noisy Expert Feedback
This paper studies imitation learning when the demonstrations come from a noisy or imperfect expert, rather than the clean policy being optimized against. It proves an exponential horizon-dependent lower bound for offline learning from noisy trajectories, while a new on-policy distillation method achieves polynomial horizon dependence under a natural noise condition, with extensions to unknown corruption and deterministic experts. The result offers a theoretical explanation for why on-policy distillation can outperform supervised fine-tuning when training models from imperfect teachers.
The claimed exponential offline-versus-polynomial online separation directly addresses an important practical puzzle in language-model training, but its importance depends on how realistic the noise assumptions and policy-class dependence are.
-
Calibration, Not Compilation: Detecting and Repairing Misspecified Probabilistic Programs Written by Language Models
The paper shows that language-model-written probabilistic programs can compile and pass unit tests while still specifying statistically wrong models. It uses Bayesian workflow signals—posterior predictive checks, simulation-based calibration, sampler diagnostics, and held-out prediction—to detect and repair misspecification, reporting strong detection performance and substantial repair gains over unit-test feedback, LLM judging, and checklists.
This is a useful and non-obvious reframing of code verification for probabilistic programs, supported by experiments across many misspecification types and models, but the calibration oracle and repair loop are more of a strong methodology than a clearly fundamental new direction.
-
Signed-Permutation Coordinate Transport for RMSNorm Transformers
The paper argues that RMSNorm Transformers have a larger coordinate symmetry than permutation-only alignment: each residual-stream channel can also undergo an independent sign flip. It introduces sign-marginalized Hungarian matching and checkpoint-to-checkpoint gauge transport, reporting much better recovery of corresponding coordinates and preservation of SAE reconstructions, steering effects, model merges, and AdamW training state than permutation-only methods. It also argues that neuron- or coordinate-level interpretability claims are meaningful only relative to an explicitly fixed gauge.
The specific RMSNorm signed-permutation symmetry and its practical impact on interpretability tools and optimizer-state transport are a non-obvious, potentially broadly useful correction, with concrete cross-model and training-state results, but the contribution is still specialized and the abstract does not establish wide adoption or generality beyond the tested models.
-
MultiSynt/MT: Trillion-Token Multi-Parallel Pre-Training Data Translated Across 36 Languages
The authors release 4.8 trillion translated tokens spanning 36 European languages, generated from 100 billion high-quality English tokens using several machine-translation systems. LLMs trained on this synthetic multilingual data reportedly match a native-data baseline with 72% fewer tokens and outperform it by about 15% at the same 100B-token budget, while analyses show that standard multiple-choice evaluations miss fluency differences and that some culturally grounded Norwegian tasks still require native data.
This is more than a corpus release: the large-scale results suggest translated data can substantially improve multilingual pre-training efficiency, while the native-data failure cases and evaluation findings qualify an important assumption about synthetic data.
-
Scaling Weisfeiler-Leman Expressiveness Analysis to Massive Graphs with GPUs
The paper develops a GPU-parallel algorithm for computing stable 1-WL colorings without requiring the entire graph to fit in memory. It combines randomized refinement with a provably correct batching scheme, and reports up to 100× speedups and successful processing of graphs with more than 30 billion edges where CPU methods fail.
The combination of scalable GPU execution and correctness-preserving out-of-core batching appears genuinely useful for large-scale GNN expressiveness analysis, though the abstract provides limited detail about baselines, hardware, and practical costs.
-
The Agentic Garden of Forking Paths
The paper studies how different AI-agent personas choose different, yet seemingly defensible, analyses of the same data. In a study involving 42 human teams analyzing immigration data, agents reproduced 72% of the ideological gap in reported effects; the authors propose “m-value” and an Agentic Bootstrap procedure to estimate how extreme a reported conclusion is across plausible analysis paths.
The evidence supports a non-obvious and important concern—that cheap, scalable AI analysis may amplify selective reporting even when individual analyses pass expert review—but the new metric and broader claims need closer methodological scrutiny.
-
Multi-Head Recurrent Memory Agents
The paper argues that recurrent memory agents fail on very long contexts mainly because a single shared memory block repeatedly overwrites useful information, rather than because they fail to capture it initially. It introduces a training-free multi-head memory design that updates only one independently protected memory head at a time; the lightweight LRU variant reportedly raises retention on RULER-HQA at 896K tokens from under 30% to 73.96%, with gains across models and tasks.
The architectural shift from monolithic memory to selectively updated protected heads, combined with a large reported long-context retention gain at nearly 1M tokens, is genuinely worth checking, though the abstract does not establish how broad or durable the advantage is.
-
Mind the Gap: Standard 3DGS Evaluation Primarily Measures Near-Trajectory Interpolation
The paper argues that standard 3D Gaussian Splatting evaluation mostly tests interpolation between nearby camera views, not spatial generalization to unseen regions. Using matched-count interpolation and contiguous-sector extrapolation splits, it finds a consistent 3–12 dB performance gap across three representation families and 16 scenes, sometimes reversing method rankings; the gap is largely predicted by nearest-view angular distance and is only weakly improved by loss regularization.
This is a potentially important and non-obvious evaluation correction with substantial measured effects across representations, but it is still primarily a protocol/benchmarking contribution rather than a new capability or method.
-
PHOENIX: Resilient LLM Training with Hot-Swapping via Zero-Overhead Checkpoint
PHOENIX keeps LLM training running when nodes permanently fail by replacing them with spare nodes at runtime rather than restarting the job. It combines off-critical-path in-memory checkpoints with dynamic communicator reconstruction, reporting no measurable checkpointing overhead and recovery in under 40 seconds on workloads up to 65B parameters and 512 A100 GPUs.
Runtime hot-swapping with apparently negligible steady-state overhead and sub-minute recovery addresses a major practical bottleneck in large-scale training, though the evidence is limited to 512 GPUs and the abstract lacks detailed comparisons and failure-load analysis.
-
One Demonstration Is Enough for Real-World Robotic Reinforcement Learning
AutoSERL uses one human demonstration to bootstrap and then autonomously guide real-world robot reinforcement learning, combining intermittent sliding-window guidance, recovery from failures, and automatic stopping of intervention. Across six contact-rich manipulation tasks on two robot platforms, it reportedly outperforms methods initialized with 20 demonstrations and one-shot imitation learning, reaching 100% success on insertion tasks and handling positional variation better.
The one-demonstration, no-continuous-human-intervention setup and results on two physical platforms are genuinely notable, but the contribution appears to be a practical combination of intervention and recovery mechanisms rather than a clearly new learning paradigm, and the abstract provides limited quantitative detail beyond headline success claims.
-
Expander Sparse Autoencoders: Parameter-Efficient Dictionaries for Mechanistic Interpretability
This paper replaces the dense decoder in sparse autoencoders with a fixed sparse expander-mask structure, reducing learned decoder storage and the cost of sparse-code inference. Across several language models, the approach reportedly uses up to 293× fewer decoder values while retaining 84% of the dense SAE’s cross-entropy-loss recovery, and it provides identifiability and OMP-recovery conditions for the structured dictionaries.
The expander-supported decoder is a genuinely interesting way to make large mechanistic-interpretability dictionaries much cheaper, with broad experiments and theory, but the substantial remaining fidelity gap and lack of detail about downstream interpretability limit this to a plausible closer look rather than a must-read.
-
When Can You Debias an LLM Judge? Identifiability Limits, a Test, and Designs for Top-k Ranking
The paper argues that pairwise LLM-judge comparisons cannot, by themselves, separate candidate quality from judge bias: the quality/bias decomposition is non-identifiable, so extra comparisons do not solve the problem. It shows that correction works only when the prior assumption that quality is uncorrelated with the bias covariate is plausible, and proposes trusted anchors and paired renderings to add information; across 15 judges, correction substantially improves top-k recall for several cheap judges but has little effect on frontier judges.
The exact identifiability diagnosis and practical tests challenge a common debiasing recipe, with unusually broad empirical support, but the proposed escapes and gains appear more like careful evaluation methodology than a clearly transformative new capability.
-
Fine-Grained Computation Offload for Off-the-Shelf Servers in Tens of Lines
The paper shows that fine-grained accelerator or remote-service calls can be overlapped with request processing without rewriting a server’s concurrency model. By routing the call through an executor and suspending the request using the server’s existing deferred-response mechanism, the authors integrate the approach into ten production-style servers with only 22–138 lines of code, achieving 1.2–5.4× speedups; an LD_PRELOAD fiber runtime reaches 17.3× in a suitable case. It also studies when suspending requests breaks run-to-completion assumptions and proposes page protection to detect the relevant shared-state hazards.
The notable contribution is a broadly applicable, low-effort systems pattern that reframes fine-grained offload from a concurrency-runtime rewrite into request routing, backed by integrations across ten server concurrency models and real-hardware measurements; it is promising, but the abstract does not establish enough breadth or deployment impact for a strong verdict.
-
Metronome: Bound the Cache, Keep the Beat for Real-Time Interaction Model Serving
Metronome studies why full-duplex streaming models can suddenly stall even when latency and deadline-miss metrics still look healthy: growing per-session KV caches exhaust the serving pool and trigger a metastable collapse. It bounds each session’s resident KV state with a sliding window, which eliminates observed collapses, makes latency a useful monotonic load signal, and enables an admission controller to find schedulable concurrency; experiments cover four interaction models on one GPU, with a simple model predicting collapse time closely.
The combination of identifying a silent cliff-like failure mode in real-time omni-modal serving and using bounded KV state to restore stability and observability is a substantial systems insight, though the evidence is limited to one GPU and a relatively narrow serving setup.
-
Optimal Stabilizer Testing and Learning with Limited Quantum Memory
The paper studies how many copies of an unknown n-qubit state are needed to test or learn whether it is a stabilizer state when the algorithm can retain only k coherent qubits between measurements. It proves testing requires Θ(n−k) copies and non-adaptive learning requires Θ(n²/k), showing that limited quantum memory eliminates the usual constant-copy advantage of testing over learning; even retaining 99% of the qubits is insufficient for constant-copy testing.
This gives a sharp, conceptually important characterization of coherent quantum memory as the resource behind the testing-versus-learning separation, but its impact is mainly within theoretical quantum information rather than broadly affecting current AI or technology practice.
-
Contaminated Multi-task Learning with Heterogeneity: Fundamental Limits and Optimal Algorithms
This paper studies multi-task learning when some tasks are arbitrarily corrupted and the remaining tasks differ substantially. It proves that several common ways of sharing information incur an avoidable contamination penalty of order ε√(d/n), then develops a filtering-based gradient method that achieves the sharper ε/√n minimax rate while still estimating personalized task solutions. Theoretical results cover broad heterogeneous ERM settings, with supporting simulations and a real-data experiment.
The dimension-dependent lower-performance barrier for standard robust-sharing methods, together with a computationally efficient rate-optimal alternative that preserves personalization under strong heterogeneity, is a substantive and somewhat surprising theoretical contribution, though its practical impact is not yet established by the abstract.
-
SPORK: Self-Speculative Forking to Accelerate Agentic LLM Inference
SPORK accelerates tool-using LLM agents by launching a cheap probe generation in parallel with the main generation to predict the next tool call, then executing that tool before the model finishes its reasoning. It requires no retraining, auxiliary model, or workflow traces; on Qwen3-32B it reports an 18% reduction in GAIA P95 latency while preserving task accuracy, with support across model sizes and dense/MoE architectures.
The self-speculative, training-free way of hiding tool latency is a genuinely useful systems idea with credible real-tool results, but the reported speedup is substantial rather than transformative and the evidence is still centered on one model family.
-
Do GUI Agents Believe Their Eyes? Diagnosing State-Belief Reliance on Pixels versus Structure
The paper introduces a benchmark for testing whether GUI agents base their state beliefs on screenshot pixels, serialized structure such as the DOM, or prior information. Across 735 probes and models from four vendors, agents often correctly perceive visual evidence but let conflicting structural data override it—especially stale structure from live pages—causing cascading task failures; a training-free consistency gate is the only tested mitigation that reduces both errors and hijacking.
This is a non-obvious, practically important diagnosis of multimodal GUI agents: they may process visual evidence yet systematically privilege stale or conflicting structure, supported by broad real-interface tests and action-level consequences, though the abstract does not establish enough generality for a strong verdict.
-
Disruption of CaMKII biomolecular condensation induces attention-deficit/hyperactivity disorder-like phenotypes
The authors engineered mice with a CaMKII mutation designed to disrupt its phase separation with GluN2B, and found normal spine structure but loss of structural LTP, hyperactivity, and impaired aversive memory; atomoxetine reduced the hyperactivity. A human CaMKII variant affecting the same hydrophobic pocket was found in a patient with ADHD and mild intellectual disability, while related variants also disrupted phase separation in vitro, linking this molecular interaction to synaptic plasticity and neurodevelopmental phenotypes.
This provides an unusually direct in-vivo test of a non-catalytic, phase-separation function of CaMKII and connects it to behavior, but the disease relevance rests heavily on one human case and the mutations may perturb CaMKII interactions beyond LLPS itself.
-
Structural basis for GluA1 AMPA receptor regulation by PRRT1/SynDIG4 in LTP
The study combines cryo-EM and functional experiments to show that the auxiliary protein PRRT1/SynDIG4 preferentially binds GluA1 AMPA receptors and uses a membrane-embedded domain to trap the receptor’s C-terminal tail. This makes LTP-related regulatory sites, including the CaMKII phosphorylation site, inaccessible; overexpressing PRRT1 impairs LTP, while a Cys825Ser mutation rescues it. The proposed mechanism links a previously unrecognized lipid modification and receptor auxiliary subunit directly to selective GluA1 availability during LTP.
A structurally grounded mechanism for how PRRT1 controls GluA1-dependent plasticity—and the functional rescue by a specific mutation—is substantially more than an incremental synaptic study, though the abstract does not establish broad in vivo or behavioral relevance.
-
A thalamic inhibitory circuit aligns sensory coding with learned value
Using imaging and causal manipulations in mice performing an audiovisual detection task, the authors show that the thalamic reticular nucleus shapes sensory thalamus according to learned reward value. Removing this inhibition preserved stimulus identity decoding but rotated population coding, disrupted action-related representations, and impaired behavior, suggesting that inhibition aligns sensory codes with behavioral goals rather than merely suppressing or transmitting inputs.
The dissociation between preserved sensory separability and impaired goal-aligned/action coding is a non-obvious mechanistic result linking inhibitory thalamic circuitry to population-code geometry and behavior, though the abstract lacks quantitative detail and broad task validation needed for a stronger verdict.
-
No evidence for direct physical interaction of 5-HT2A-mGluR2 receptors in vitro or in vivo
This study tests whether mGluR2 and the psychedelic receptor 5-HT2A directly form receptor complexes, rather than interacting indirectly through glutamate release and neuronal signaling. Using tagged receptors, in vivo and in vitro localization/oligomerization assays, and radioligand-binding analyses, the authors find no evidence for physical association or modulation of 5-HT2A ligand binding, while confirming that mGluR2 activation suppresses DOI-induced head twitching in mice.
It addresses a longstanding mechanistic controversy with several orthogonal negative results, potentially shifting interpretation of how mGluR2 modulates psychedelic 5-HT2A signaling, but the abstract does not establish how robustly the negative findings exclude context-dependent receptor interactions.
-
Time cells lead neural reinstatement of episodic memory
Using microelectrode recordings from neurosurgical patients in free-recall and serial-reconstruction tasks, the authors find that medial-temporal time cells activate before broader memory-related neural assemblies during retrieval. Their activity predicts context-dependent recall and appears alongside complementary phase-coding mechanisms, with similar context-sensitive neurons also emerging during naturalistic viewing without an explicit temporal scaffold.
The paper offers a specific cellular and temporal mechanism for episodic-context reinstatement in humans, supported across two classic tasks and a naturalistic dataset, but the abstract provides too few quantitative details to justify a strong verdict.
-
POGZ safeguards neuronal gene chromatin architecture and transcription
This study shows that POGZ partners with the G9a/GLP H3K9 methylation machinery in developing mouse cortex. Removing POGZ causes large-scale, locus-specific redistribution of H3K9me3, movement of neuronal gene regions toward the nuclear lamina, weakened 3D genome boundaries and CTCF occupancy, and reduced nascent transcription. The new contribution is a mechanistic link from POGZ loss to heterochromatinization and spatial genome reorganization at neurodevelopmental genes.
The combination of in vivo interaction, Micro-C, DNA-FISH, chromatin profiling, and PRO-seq supports a non-obvious mechanism for how an ASD-associated chromatin regulator preserves neuronal gene expression, though the evidence is still from a mouse developmental model and a single gene pathway.
-
A cholinergic eligibility trace facilitates amygdala plasticity in flavour avoidance learning
The study identifies a previously underappreciated cholinergic projection from parabrachial neurons to the central amygdala that is active both during novel-food consumption and later illness. Using slice imaging, CRISPR knockdown, optogenetics, and large-scale recordings, the authors show that this acetylcholine signal enhances excitatory responsiveness and is required for the neural and behavioral plasticity underlying conditioned flavour avoidance, potentially serving as a long-timescale eligibility trace.
This is a mechanistically coherent, multi-method account of how the brain links food consumption to delayed malaise, with the cholinergic projection acting as a plausible temporal bridge; it is interesting enough for a neuroscience reader, though the demonstrated scope is limited to one learning paradigm and the abstract lacks quantitative effect sizes.
-
When Classic Cache Policies Fail: Learning-Augmented Replacement for Semantic Retrieval Buffers
This paper studies cache replacement for LLM agent memory, where retrieval matches are semantic and hit quality is continuous rather than binary. It reports that LRU and LFU can perform worse than FIFO on two MemoryBench datasets, and proposes SOLAR, which uses regret-based update timing and Bayesian learning to choose evictions; SOLAR improves retrieval quality by 5–75% over FIFO at small capacities. The paper also gives competitive-ratio and regret guarantees and reports that larger memory pools can eventually hurt retrieval through semantic noise.
The combination of a formal semantic-cache setting, the counterintuitive failure of standard locality heuristics, and a theoretically grounded replacement method is worth a closer look, though the empirical evidence is limited to two datasets and the practical gains may be workload-specific.
-
Optimizing Visual Generative Models via Distribution-wise Rewards
The paper replaces per-image rewards with rewards computed over a set of generated images, aiming to prevent reward hacking, mode collapse, and visual artifacts during reinforcement-learning-based diffusion-model fine-tuning. It introduces a subset-replace approximation to make set-level reward computation cheaper, and also uses RL to tune model-merging coefficients; experiments report FID improvements for SiT and EDM2 while retaining diversity.
The distribution-level reward formulation is a meaningful departure from standard sample-wise RL for image generation, but the abstract provides limited evidence beyond FID gains on two model families and does not establish a broadly transformative capability or efficiency improvement.
-
The Surprising Effectiveness of Video Diffusion Models for Hand Motion Reconstruction
ViDiHand adapts a pretrained video diffusion model to reconstruct metric-scale two-hand motion directly from egocentric video. It uses a hand-overlay rendering objective to specialize the diffusion features, avoiding a separate detector, infiller, or test-time optimization, and reports substantial gains over prior methods on ARCTIC, HOT3D, and HOI4D. The central idea is to reuse motion, occlusion, and interaction knowledge learned from large-scale video generation rather than learning temporal reasoning mainly from scarce pose labels.
The proposed reuse of video-diffusion representations for detector-free 4D hand reconstruction is a genuinely interesting direction, but the abstract gives no quantitative results or details sufficient to establish how large or robust the gains are.
-
Automating the Design of Embodied Agent Architectures
This paper applies automated agent-architecture search to embodied agents in simulation. It introduces a typed-graph runtime and a coding-agent search loop, then evaluates them across navigation, embodied QA, and language-conditioned manipulation; the search finds some deployable success-rate improvements but is limited by rollout noise, local search traps, and weak episode-level credit assignment. One apparently strong result was discarded because it relied on information leakage.
It is a credible first systematic attempt to transfer architecture search from text agents to perceptual embodied systems, with useful evidence about both gains and failure modes, but the abstract gives no quantitative improvements or indication of strong real-world capability gains.
-
LLM Agents Are Latent Context Managers: Eliciting Self-Managed Context via State Proprioception
The paper argues that long-horizon agents struggle partly because they cannot observe their own context state—such as token usage, block recency, and remaining budget. VISTA exposes this information through a typed, addressable memory interface and archives full-fidelity context blocks for later recovery, reportedly improving Gemini-3-Flash from 22.7% to 50.7% on LOCA-Bench and transferring across models, context sizes, and several agent benchmarks.
The proprioception framing and training-free, model-agnostic memory interface are genuinely interesting, and the reported near-doubling under context pressure is substantial, but the abstract gives limited detail about baselines, backbone breadth, and GAIA results, so the strength of the evidence is not yet enough for a strong recommendation.
-
MirrorCode: AI can rebuild entire programs from behavior alone
MirrorCode evaluates AI agents on rebuilding complete software projects from observed behavior, without source-code access, using exact end-to-end tests including held-out cases. Across 25 varied programs, the best model reportedly reaches 56% overall and can reproduce a 16,000-line bioinformatics toolkit, though a single large attempt may cost $2,600 and take 19 days.
This is a more demanding and potentially revealing test of long-horizon autonomous coding than standard short-task benchmarks, but it is still primarily a benchmark paper and the abstract provides limited detail for judging the strength and reproducibility of the headline results.
-
Chronos: A Physics-Informed Full-History Framework for Non-Markovian Long-Horizon Manipulation
Chronos treats the entire observation history as the policy’s latent dynamical state rather than relying on the current frame or a short context window. It combines a selective state-space model with a multimodal action prior and a second-order motion-refinement process, reporting much higher success than pi0.5 and Mem-0 on memory-dependent simulated tasks and four real-world dual-arm tasks, while using far fewer parameters.
The large gains on phase-dependent manipulation and the compact full-history design are genuinely interesting, but the evidence is limited to a small task suite and the unusually strong benchmark comparisons need verification before treating this as a major result.
-
Why Do Few-Step Text Latents Fail When Image Latents Work? Non-Commitment at Sharp Categorical Readouts
This paper argues that few-step deterministic generation fails for continuous text latents because sharp categorical decoder boundaries require an early commitment to discrete token choices; improving transport accuracy alone cannot fix this. It supports the explanation with geometric theorems, boundary-sensitivity diagnostics across four text decoders versus image decoders, and an experiment where stochastic reinjection greatly improves perplexity (50 versus 294 at four steps).
The paper offers a non-obvious geometric explanation for the text–image gap and connects it to concrete escape mechanisms, but the empirical evidence is limited and the strongest claims depend on idealized theoretical regimes.
-
Diffusion Fine-tuning with Rewarded Moment Matching Distillation
The paper introduces Rewarded Moment Matching Distillation, which jointly distills a diffusion model and optimizes a task-specific reward instead of treating distillation and RL fine-tuning as separate stages. It reports better ImageNet quality–reward trade-offs and applies the method to GenCast, obtaining a 7.5× speedup while improving CRPS on 93% of weather variables and calibration relative to the teacher.
The combination of on-policy reward optimization with moment-matching distillation, plus a substantial speedup and apparent quality improvement in weather forecasting, is a genuinely interesting direction, but the abstract lacks enough quantitative breadth and methodological detail for a strong verdict.
-
When Does Online Imitation Learning Help in LLM Post-Training? The Role of (Non-)Realizability Beyond Horizon
This paper argues that online imitation learning helps LLM post-training not mainly because it prevents long-horizon error accumulation, but because it can overcome model misspecification: the student may be unable to represent the expert policy. It proves that offline imitation can hit an information-theoretic limit even at horizon 1 in non-realizable settings, while online interaction can still achieve high reward under a structural condition relating the misspecification to the task reward.
The horizon-1 impossibility result and reward-relative characterization offer a genuinely useful explanation for when online post-training should outperform SFT, but the abstract provides limited empirical evidence about practical LLM-scale impact.
-
Fleet: Few Shots Lead Effective AI-generated Image Detection
Fleet proposes adapting AI-image detectors continuously with a small number of examples from new generators, rather than relying on a fixed feature space. Its constrained routing updates reportedly raise detection on Doubao Seedream 4.0 from 20.4% to 73.1% with only 10 labeled examples, evaluated across a new 64-model, 360k-image benchmark including commercial engines.
The dynamic few-shot adaptation framing and large reported recovery against a modern unseen generator are genuinely interesting, but the abstract gives limited detail about robustness, baselines, and whether the result generalizes beyond the highlighted model.
-
Learning Gaussian Graphical Models from a Glauber Trajectory Without Mixing
The paper learns the conditional-independence graph of a sparse Gaussian model from one temporally correlated Glauber-dynamics trajectory, rather than from independent samples. Its polynomial-time method normalizes conditional variances, tests edges using short update windows, and uses robust aggregation to handle dependence; notably, the required trajectory length is independent of the chain’s mixing time.
Avoiding any dependence on mixing time while recovering graph structure from a single dependent trajectory addresses a real theoretical gap, but the abstract gives no concrete sample-complexity or empirical scale to establish how substantial the advance is.
-
Decodable Is Not Grounded: A Vision-Ablation Arbiter for VLM Spatial Reasoning
The paper argues that successfully probing or steering a VLM’s hidden representation does not show that the model is using the image. Using a blank-image causal control across 14 VLMs, it finds three cases: genuinely image-grounded reasoning, image-independent priors, and representations whose decoded answer has the wrong sign when deployed; horizontal, vertical, and depth reasoning reportedly fall into these categories. It also studies whether simple rotations or trained low-rank edits can correct the deployment mismatch.
The decodable-versus-deployed distinction and the surprisingly inverted spatial representations are non-obvious and supported across many models, but the work is primarily a diagnostic methodology rather than a demonstrated capability or efficiency breakthrough.
-
Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers
The paper uses a recurrent-depth Transformer to perform several latent reasoning steps in parallel, while training each latent position against a corresponding gold chain-of-thought token. Its LOTUS model reportedly matches explicit-CoT reasoning at 3B parameters, reduces thought-phase latency by 2.5–6.9×, and produces latent states that can be decoded into gold or alternative valid reasoning steps.
This is a genuinely relevant combination of looped computation and supervised latent reasoning, with a potentially important efficiency gain, but the abstract does not provide enough benchmark detail or breadth to justify a strong verdict.
-
Reinforcement Learning-Based Control for an Inline Skating Humanoid Robot
The authors train a reinforcement-learning controller for a humanoid robot fitted with passive inline skates, without demonstrations or motion priors. The learned policy develops edge-driven skating strategies, transfers zero-shot to a physical Booster T1, and enables turning, balance recovery, and perturbation rejection; simulations report up to a 50% lower cost of transport than walking.
This is a genuinely unusual real-robot locomotion capability with an emergent control strategy and claimed efficiency gain, but the abstract lacks enough quantitative hardware results and comparisons to justify a strong recommendation.
-
Introspective Coupling: Self-Explanation Training Tracks Behavioral Change Despite Fixed Supervision
The authors train language models to explain which input features affected their behavior, using counterfactual explanations generated from earlier checkpoints or other models. Surprisingly, the explanations often become more faithful to the model’s current behavior than to the fixed behavior represented in the training data, and they continue to track behavior changes caused by other post-training objectives. The effect appears across sycophancy and refusal tasks and persists with noisy labels.
The potentially important finding is that fixed counterfactual explanation data can induce explanations that update alongside the model’s changing behavior, suggesting a scalable route to introspection, but the abstract gives no quantitative results or detail on scope and limitations.
-
Prototype Language Models
PRISM replaces the usual dense prediction pathway with a sparse, non-negative mixture of learned prototypes tied to neighborhoods of training examples. Across 130M–1.6B-parameter models, it reportedly retains near-baseline accuracy while making attribution much faster, and enables prototype-level corrections or behavior removal without finetuning.
The combination of prototype-based language modeling, scalable training-data attribution, and editable behavior is a genuinely interesting direction, but the abstract lacks enough methodological and comparative detail to justify a stronger verdict.
-
AnchorSplat: Fast and Structure Consistent Detail Synthesis for Gaussian Splatting
AnchorSplat refines existing 3D Gaussian Splatting assets directly in 3D, without the original multi-view images or costly image-space optimization. Its point-anchor mechanism aims to preserve geometric consistency, while a single-pass Gaussian multiplication replaces iterative densification; the authors report much higher throughput and zero-shot transfer to generated assets and real scans.
The source-free, single-pass 3D-native refinement setup and claimed 100,000× speedup are potentially important, but the abstract provides no quantitative quality results and the benchmark is introduced by the paper itself.
-
RoboWorld: Fast and Reliable Neural Simulators for Generalist Robot Policy Evaluation
RoboWorld uses a fast video world model to simulate robot-policy rollouts and a vision-language model to score task progress. Its Step Forcing training scheme mixes anchored context with one-step self-generated context to make long-horizon predictions more reliable; the resulting scores correlate strongly with real-world evaluations (Pearson r=0.989, Spearman ρ=0.970).
The combination of self-consistent long-horizon video simulation and progress-aware policy scoring could make large-scale robot evaluation substantially cheaper, but the abstract gives too little detail about task breadth, baselines, and how robust the striking correlations are.
-
GSRQ: Gain-Shape Residual Quantization for Sub-1-bit KV Cache
The paper proposes Gain-Shape K-means, which separates vector magnitude from direction during codebook learning to reduce centroid shrinkage in residual quantization. Its GSRQ KV-cache method reportedly improves LLaMA-3-8B LongBench accuracy at 1 bit from 11.34 to 33.54, suggesting a substantial gain for long-context inference under severe memory constraints.
The gain-shape codebook-learning idea and the large reported 1-bit improvement are notable, but the abstract provides evidence from only one model and benchmark family, so the result needs verification before warranting a strong recommendation.
-
ROSA: A Robotics Foundation Model Serving System for Robot Factories
ROSA treats robot-factory inference as a shared distributed-serving problem rather than giving each robot its own GPU. It pools server GPUs across robots, supports multi-model pipelines and task-specific service requirements, and schedules requests for overall factory productivity; experiments on real robots and synthetic workloads report up to 12.06× higher productivity than dedicated serving.
The factory-level scheduling objective and shared GPU-pool architecture are a meaningful reframing with potentially large deployment benefits, but the abstract gives only a maximum improvement figure and insufficient detail about workloads, baselines, and scaling to justify a strong verdict.
-
QuasiMoTTo: Quasi-Monte Carlo Test-Time Scaling
QuasiMoTTo replaces independent parallel language-model samples with correlated samples generated using quasi-Monte Carlo sequences. Each sample still has the correct marginal model distribution, but the batch covers possible outputs more efficiently, achieving similar pass@k results with 25–47% fewer samples across four reasoning benchmarks and similar GRPO performance with 50% fewer training steps. The paper also introduces an unbiased bootstrap estimator for pass@k under correlated sampling.
This is a genuinely interesting inference-time scaling idea with potentially meaningful compute savings, but the abstract lacks enough detail on model scales, tasks, and robustness to justify a strong verdict.
-
Can Language Models Actually Retrieve In-Context? Drowning in Documents at Million Token Scale
This paper studies whether a language model can retrieve from a million-token corpus directly, rather than using a separate vector index. It introduces BlockSearch, a 0.6B model with length-generalizing and sparse-attention changes, and argues that retrieval fails at long contexts because irrelevant documents dilute attention; its fixes match dense retrieval on MS MARCO and NQ and reportedly outperform it by 3× on LIMIT.
The combination of million-token in-context retrieval, an interpretable attention-dilution failure mode, and strong gains on a task with a different similarity notion is a genuinely non-obvious direction, but the abstract provides too few quantitative and scalability details for a strong recommendation.
-
SCAPE: Accurate and Efficient LLM Training with Extreme Sparse Communication
SCAPE is a distributed optimizer that sparsifies communication using Adam's first-moment statistics rather than raw gradients. It also aligns mask generation with optimizer sharding, overlaps mask synchronization with computation, and reconstructs second-moment updates from the same sparse buffer. On 32 GH200 GPUs, the authors report stable quality at 90–99% sparsity, up to 43.3% lower end-to-end training time for a 500M-parameter Llama model, and up to 3.26× faster steps for a 1.8B model.
Extreme communication sparsity with near-unchanged training quality and substantial measured speedups is genuinely useful, but the evidence is limited to relatively small models and one hardware/system configuration, so the generality of the gains is not yet established.
-
Towards Load-Aware Prefill Deflection for Disaggregated LLM Serving
This paper proposes letting underutilized decode GPUs temporarily perform chunked prefill work when prefill GPUs are overloaded, while respecting the decode workload’s token-latency SLO. On a 2-prefill/2-decode A100 setup using DeepSeek-V2-Lite and production-style traces, the scheduler reportedly cuts P95 time-to-first-token by up to 81% and improves SLO attainment by up to 79%, largely by avoiding prefill queues and KV-cache transfers.
The asymmetric use of decode capacity to deflect prefill is a meaningful systems idea with large reported tail-latency gains, but the evidence is limited to a small cluster configuration and trace-based evaluation, so it falls short of strong.
-
SUNTA: Hierarchical Video Prediction with Surprise-based Chunking
SUNTA chooses temporal chunk boundaries in hierarchical video-prediction models from prediction surprise rather than fixed lengths or similarity. It uses decoupled training to avoid hierarchical collapse and an internal-consistency signal to estimate surprise during imagined, open-loop rollouts; on 2D and 3D environments, it reportedly preserves prediction quality for 250 timesteps while baselines degrade within about 10.
The prediction-error-driven temporal abstraction and its rollout-time surprise mechanism are a non-obvious approach with a potentially large long-horizon benefit, but the abstract gives few quantitative details and the dramatic result is limited to unspecified 2D/3D environments.
-
An Optimisation Framework for the Well-Conditioned Training of Physics-Informed Neural Networks
The paper introduces DSGNAR, a sketched Gauss–Newton optimizer for PINNs that adaptively controls regularization and step size to address their ill-conditioned loss landscapes. On a broad set of PDEs, it reports dramatically lower errors—up to five to eight orders of magnitude better than prior results—and much faster convergence, including near-roundoff accuracy on Burgers’ equation in single-digit seconds.
The reported accuracy and speed gains would be highly consequential for PINNs, but the abstract provides little detail about computational cost, baselines, problem setups, or whether the comparisons are genuinely like-for-like, so the extraordinary claims warrant verification before elevating this to strong.
-
One More Time: Revisiting Neural Quantum States from a Reinforcement Learning Perspective
The paper reframes variational energy minimization for autoregressive neural quantum states as an advantage-policy-gradient problem and proposes Proximal Wavefunction Optimization (PWO), a trust-region method that clips amplitude probability ratios and phase updates without matrix inversion. On Ising and frustrated J1-J2 systems, it reportedly improves stability and wall-clock convergence over Adam, minSR, and SPRING, and is demonstrated on a 1.5B-parameter RWKV-7 wavefunction model.
The RL/trust-region formulation and unusually large-scale NQS demonstration are genuinely notable, but the abstract gives no quantitative gains or details sufficient to establish that the method is broadly superior rather than carefully tuned.
-
DecompRL: Solving Harder Problems by Learning Modular Code Generation
DecompRL trains language models to split programming problems into independently solvable modules, generate multiple implementations for each module, and recombine them into many full programs. The resulting combinatorial search can produce up to k^n candidates while shifting much of the cost from GPU generation to cheaper CPU evaluation; on LiveCodeBench and CodeContests, the authors report roughly 50x lower GPU token cost and better performance than standard and diversity-focused RL at high test-time budgets.
The modular recombination strategy is a genuinely nonstandard way to overcome near-zero-probability solutions and could materially improve test-time scaling, but the abstract gives limited quantitative detail and evidence from only two coding benchmarks.
-
Benign Overfitting Does Not Occur in Diffusion Models
This paper argues that diffusion models do not share the benign-overfitting and double-descent behavior known from some regression settings. Its theory says that, unless the sample size is exponential in the data dimension, fitting the training data damages population loss because score matching lacks the covariance–target alignment that can make regression overfitting benign; time smoothness and early stopping instead act as implicit regularizers, with supporting high-dimensional image experiments.
The claim that diffusion-model generalization is fundamentally unlike regression overparameterization is potentially important and counterintuitive, but the abstract indicates that the strongest results rely on simplified assumptions and provides limited detail about the breadth of the empirical validation.
-
DREAMSTEER: Latent World Models Can Steer VLA Policies During Deployment Without Any Finetuning
DREAMSTEER steers a pretrained vision-language-action policy at deployment time by sampling action chunks and motion primitives, predicting their outcomes with a latent world model, and selecting them with a language-conditioned value model. On four real-world manipulation benchmarks with unseen objects, it raises success from 23.75% to 66.25% and instruction-following accuracy from 38.75% to 56.25%, without finetuning or changing policy parameters.
The combination of latent imagined rollouts and language-conditioned value steering yields a large reported deployment-time robustness gain without target-environment demonstrations, though the abstract lacks comparisons, scale, and ablations needed to establish a broadly important advance.
-
Variable Bit-width Quantization: Learning Per-Group Precision for "Bigger-but-Smaller" Language Models
The paper learns a separate bit-width for each group of 64 weights, rather than assigning one precision globally or per layer. It finds highly uneven allocations—many groups use 1 bit—and reports that this can preserve the quality of much larger models at substantially lower storage, with increasing decode-speed gains at larger scales. It also argues that later layers partially repair quantization errors introduced by earlier layers.
Learned fine-grained precision allocation plus the reported bigger-but-smaller scaling and speed results are genuinely interesting, but the evidence is mainly abstract-level claims on limited training/evaluation settings and the method’s scalability beyond 1.5B parameters remains unresolved.
-
Spectral Rewiring for Exploration, Purification, and Model Merging
The paper proposes Subspace-Aligned Rewiring (SAR), which decomposes post-training parameter updates and keeps the components aligned with the base model’s spectral structure while removing other directions. The authors claim that very small “reasoning cores” (down to about 0.58% of parameters) retain over 99% of performance, improve test-time exploration in math and coding, reduce interference in mixed-domain training, and improve model merging across experts.
The spectral filtering view of reasoning updates and its claimed benefits for exploration and model merging are potentially substantial, but the abstract provides limited quantitative detail and relies partly on an in-house model, so the broad claims need verification.
-
GALOSH: Blind, Training-Free Denoising of Raw Bayer and sRGB Images by Parallel-Friendly Local Shrinkage
GALOSH is a training-free, blind image denoiser that replaces BM3D-style content search with fixed, local operations: noise estimation, a generalized Anscombe transform, Walsh–Hadamard shrinkage, and chrominance regression. On four real-noise benchmarks it reportedly outperforms blind BM3D/NLM variants, approaches trained models on raw images, and runs 7–650× faster than the compared deep-learning systems, including practical CPU and INT16 implementations. Its main novelty is combining competitive denoising with a regular, search-free computation graph suitable for fixed-latency or streaming hardware.
The search-free, multi-domain denoising design and claimed CPU/fixed-point practicality are genuinely interesting, but the abstract lacks detailed quantitative comparisons and the advance appears focused on a specialized classical vision problem rather than a broadly transformative method.
-
WSA$_1$: a 3D-Centric World-Spatial-Action Model for Generalizable Robot Control
WSA₁ is a robot foundation model that explicitly represents 3D scene state, predicts how that state will change, and jointly models those changes with robot actions. Using 6,000 hours of demonstrations, including only 1,000 hours from real robots, it reports 93% success on RoboTwin2.0 and an average 20% improvement over prior models on unspecified real-world control tasks.
The explicit coupling of 3D world-state transitions and actions, together with strong claimed generalization from limited real-robot data, is potentially important, but the abstract lacks enough detail about tasks, baselines, and ablations to justify a stronger verdict.
-
Learning Biophysical Models of Large-Scale Multineuronal Data to Enable Precise Neurostimulation
The authors infer cell- and circuit-level Hodgkin–Huxley parameters from a few minutes of extracellular recordings on a 512-electrode macaque retina array, using differentiable simulation and simulation-based inference. The fitted models predicted responses to previously unseen multi-electrode stimulation patterns with 90.6% accuracy, potentially replacing hours of empirical stimulus testing. The notable contribution is extracting biophysical models suitable for stimulation prediction from scalable extracellular data rather than invasive intracellular recordings.
Inferring useful multi-compartment biophysical models from extracellular population recordings is a genuinely interesting capability with implications for neurostimulation, but the evidence is currently limited to an isolated macaque-retina preparation and an underspecified accuracy metric.
-
Critical period plasticity enables credit assignment
This study proposes that the same instructive signals that later train a circuit first help build its task-relevant connectivity during a two-day developmental critical period. In zebrafish olivocerebellar circuits, olivary inputs shaped maturation of cerebellar projections to coactivated targets; theory and simulations suggest this preconfigured architecture enables more effective credit assignment later, and that reduced plasticity after development can improve robustness.
The paper offers a genuinely interesting developmental solution to credit assignment, with biological observations plus modeling, but the evidence is confined to one circuit and the abstract gives limited quantitative validation.
-
Much stronger coarse-to-fine visual processing in primate superior colliculus than primary visual cortex neurons
Using matched visual stimuli and the same animals, the authors compared coarse-to-fine processing in primate superior colliculus (SC) and primary visual cortex (V1). They find that SC response latencies are strongly determined by image spatial frequency—coarse patterns arrive fastest—whereas V1 latencies are more closely tied to response sensitivity and broadband tuning. The result suggests that coarse-to-fine dynamics are especially prominent in the collicular circuits that guide gaze shifts, rather than simply being inherited from cortex.
The matched SC–V1 comparison challenges the expectation that the SC mainly inherits efficient visual processing from V1, but the abstract provides no quantitative effect sizes or broader behavioral validation, so it merits a look without being an obvious must-read.
-
Synaptic activity controls local exposure of an 'eat-me' signal via ANO3-ITPR1 signaling
The authors develop COMPASS-seq, a CRISPR-screening method for linking genes to phenotypes in specific subcellular compartments, and use it to study synaptic pruning. They identify ANO3 and ITPR1 as a postsynaptic calcium-signaling system that exposes phosphatidylserine locally, marking synapses for glial engulfment without requiring caspase activation; inhibiting ANO3 increases synapse numbers in vivo.
This proposes a non-apoptotic, activity-linked mechanism for synaptic eat-me signaling and introduces a potentially broadly useful compartment-resolved CRISPR screen, but the abstract provides limited detail on the breadth and strength of the in vivo evidence.
-
Dual thalamic drive defines the intra-amygdala wiring complexity
Using molecularly defined amygdala subdivisions, the authors mapped mouse connectivity and electrophysiology and found that information does not mainly flow through the traditional lateral-to-basolateral-to-central serial pathway. Instead, separate thalamic inputs appear to drive partly distinct routes, with the anterior basolateral nucleus contributing mainly to outputs outside the amygdala; the authors report a similar organization in humans.
It presents a potentially important revision of a canonical amygdala circuit model, including an apparent cross-species parallel, but the abstract provides limited quantitative detail and no direct evidence yet that the wiring changes behavioral or computational conclusions.
-
Progressive loss of independence in neuronal representations predicts cognitive decline
The study finds that, in a rhesus macaque model of Alzheimer’s-related disease, visual neurons in area V4 progressively lose the ability to represent different features independently. This “feature confusion” predicts cross-feature biases in visual choices, and a similar behavioral signature distinguishes people with mild cognitive impairment from age-matched controls. The main contribution is linking a specific population-coding failure to an observable cognitive symptom across species.
The proposed loss of independence in feature representations is a concrete, mechanistic account of cognitive confusion with cross-species behavioral evidence, but the abstract provides too few quantitative details to justify a strong recommendation.
-
dOPSD: On-Policy Self-Distillation for Diffusion Language Models
The paper proposes dOPSD, a self-distillation method for diffusion language models that uses later, more-complete denoising states from the model’s own trajectory as privileged supervision for earlier masked positions. This provides dense, on-policy training signals without reference answers or tractable sequence likelihoods, and reportedly improves math reasoning and out-of-domain code generation on Dream and LLaDA.
The trajectory-based source of privileged information is a genuinely relevant idea for post-training diffusion LMs, but the abstract gives no quantitative results or broad evaluation to establish how substantial the gains are.
-
Discrete Diffusion Language Models for Interactive Radiology Report Drafting
The paper fine-tunes a 26B mixture-of-experts discrete diffusion language model for medical visual question answering and compares it with a same-size autoregressive model using the same LoRA setup. It reports comparable or better judged performance and 3.5–4.4× faster decoding, while enabling bidirectional infilling: a radiologist can edit parts of a report and ask the model to complete the text between them. The main novelty is applying diffusion’s native any-order editing to interactive report drafting, although the reported evaluation is on VQA datasets rather than real drafting workflows.
The combination of competitive quality, substantially faster decoding, and native arbitrary-span infilling is a genuinely useful capability difference from autoregressive models, but the evidence is limited because the abstract does not report direct radiology-report drafting or clinician evaluations.
-
MANCE: Manifold Aware Concept Erasure
The paper proposes constraining concept-removal updates to the estimated manifold of natural representations, rather than modifying representations in arbitrary directions. Its MANCE variants are evaluated across 119 text and vision settings, with MANCE++ reportedly improving the tradeoff between removing the target concept and preserving unrelated information, especially for nonlinear erasure.
The manifold-constrained intervention is a plausible but meaningful new framing, supported by broad experiments, though the contribution remains a specialized improvement to concept-erasure methods rather than a clearly general breakthrough.
-
Are Performance-Optimization Benchmarks Reliably Measuring Coding Agents?
The authors replay 740 repository-level code-optimization tasks across four machine types and find substantial cross-machine validity failures, especially for SWE-Perf. They also show that leaderboard rankings change with the scoring rule, while public submissions already match or beat the reference patch on 85.3% of replay-valid GSO and SWE-fficiency tasks, suggesting strong saturation and limited discrimination in current benchmarks.
This is more than a routine benchmark critique: concrete cross-machine replays reveal fragile validity, ranking reversals, and severe task saturation that undermine common evidence of coding-agent progress, though it does not introduce a new agent capability or method.
-
Flex-Forcing: Towards a Unified Autoregressive and Bidirectional Video Diffusion Model
Flex-Forcing trains a video diffusion model to support multiple inference schedules rather than committing to either fully bidirectional diffusion or strictly causal autoregressive generation. It uses flexible temporal and denoising-step chunking, enabling global planning across chunks, autoregressive synthesis within chunks, and adaptation to different compute budgets. The claimed gains are better quality and long-video stability with faster inference, but the abstract gives no quantitative results.
The unified, non-strictly-causal scheduling framework is a meaningful idea for controllable video-generation efficiency, but the abstract lacks numbers or enough experimental detail to establish a major advance.
-
Autonomous Scientific Discovery via Iterative Meta-Reflection
DiscoPER is an LLM-based system that autonomously generates and tests hypotheses by writing and running code over multimodal ecological data, without a predefined research question. Its distinctive mechanism is periodic “second-order” analysis of its prior discoveries to identify patterns, confounds, and gaps, and it recovers 8 of 9 literature-grounded patterns on a new benchmark with a 72.7% hypothesis-support rate.
The self-analysis loop for redirecting open-ended scientific search is a meaningful idea, but the evidence rests mainly on one new, narrow benchmark and the abstract does not establish reliable real-world discovery beyond recovery of known ecological patterns.
-
QVal: Cheaply Evaluating Dense Supervision Signals for Long-Horizon LLM Agents
QVal proposes a training-free way to compare dense reward or supervision signals for long-horizon LLM agents. It scores signals by how well they rank actions according to the estimated Q-values of a strong reference policy, avoiding expensive downstream training comparisons. Across 21 methods, four environments, and six model backbones, simple prompting baselines consistently outperform many recent dense-supervision methods, with results clustering by methodological family.
The reusable evaluation framework is useful, and the broad finding that simple prompting beats more elaborate dense-supervision methods is non-obvious, though this remains primarily a benchmark and its conclusions depend on the choice of reference policy and environments.
-
ShardNet: Training Neural Controllers with Hard, Non-Convex Constraints
ShardNet embeds a differentiable projection mechanism into neural controllers so they satisfy nonconvex unions of polyhedral safety constraints by construction, rather than optimizing safety as a soft objective. The authors also propose a way to train ReLU value functions for verification, reporting 100% safety on verified double-integrator sets, lower objective loss, and safe sets over three times larger than prior verification methods.
The combination of hard nonconvex constraint enforcement and verifiable ReLU value functions is genuinely interesting, but the evidence is limited to double-integrator benchmarks and the abstract provides few quantitative details beyond broad comparative claims.
-
Certified Speculative Execution for Untrusted AI Agents
The paper proposes Certificate-Gated Prefix Acceptance, which lets an untrusted learned policy or LLM draft multiple actions while a trusted verifier accepts only certified-feasible prefixes and hands the remainder to a solver. Across adversarial and six frozen LLM proposal sources, it reports zero constraint violations, much lower regret than unguarded drafting, and a 2.96× speedup on a deployment-scale unit-commitment problem while maintaining a fixed guarantee.
The combination of certified prefix verification, conformal regret gating, and speculative execution is a substantial and potentially reusable way to obtain LLM speed without trusting its actions, but the evidence is concentrated in constrained sequential decision problems and one main deployment setting rather than demonstrating broad generality.
-
ELASTIC: Efficiently Learning to Adaptively Scale Test-Time Compute for Generative Control Policies
ELASTIC learns a state-dependent schedule for robot policy inference, deciding when to spend compute on more denoising refinement versus sampling more parallel action candidates. Trained as a meta-policy around a frozen generative control policy, it outperforms fixed and single-axis scaling at matched compute and achieves best-of-10 success on a real robot while reducing latency by 34%.
The adaptive allocation of sequential refinement and parallel exploration is a meaningful test-time scaling idea with real-robot evidence, but the abstract does not establish broad generality or a sufficiently large capability jump for a strong verdict.
-
Think While You Map: Asynchronous Vision-Language Agents for Incremental 3D Scene Graphs
The paper builds an incremental 3D scene graph in parallel with slower vision-language reasoning, so a robot or agent can query the scene while it is still exploring. Its voxel-based mapper maintains object identities, while background VLM agents add attributes, spatial relations, and semantic loop closures; a frame scheduler reduces VLM calls. The system reportedly improves visual grounding by 15.3–18.8 A@0.25 across three benchmarks and matches or exceeds prior segmentation results.
The asynchronous coupling of online mapping and deferred semantic refinement, together with large reported grounding gains, is a meaningful systems direction, but the abstract does not establish a major real-world capability jump or broad deployment evidence.
-
GAIA: Geometry-Adaptive Operator Learning for Forward and Inverse Problems
GAIA is a geometry-conditioned neural operator that represents domain boundaries and interior fields as tokens, then uses them to adapt integral-transform kernels locally. It aims to handle forward, boundary-value, and inverse PDE problems across varying geometries in one model without retraining or iterative inversion; across seven 2D/3D benchmarks, it reports leading results on all inverse and BVP tasks, including 64% lower median error for airfoil reconstruction and 27% for electrical impedance tomography versus the next amortized method.
A unified, geometry-adaptive operator for forward and inverse problems without per-task retraining is a substantive direction, supported by broad experiments and sizable reported error reductions, but the abstract does not establish how much comes from the architecture versus benchmark or baseline choices.
-
FurnitureVLA: Learning Long-Horizon Bimanual Furniture Assembly with Vision-Language-Action Model
FurnitureVLA studies real-scale, bimanual furniture assembly with a vision-language-action model, using simulated expert data and VR teleoperation demonstrations on a Kinova robot. Its VLA predicts both actions and continuous task progress to switch among up to seven subtasks, raising simulated success from 48% to 80% across three furniture types; real-world performance falls by 16% on the hardest task.
The combination of long-horizon bimanual assembly, progress-based subtask control, and real-scale robot validation is a meaningful capability demonstration, though the abstract does not establish broad generalization or unusually strong real-world success rates.
-
SE(2) Navigation Mesh
The paper extends navigation meshes to encode whether a region is traversable for different robot headings, rather than assuming yaw-invariant traversability. It builds a layered SE(2) graph, introduces a hierarchical A*-based planner, and incrementally updates the representation from streaming point clouds. Simulations report over 50% more captured traversable area than standard NavMeshes, while real-robot experiments demonstrate online generation and navigation in several environments.
Yaw-dependent polygonal navigation meshes with online updates address a real limitation for non-circular robots, and the combination of simulation and physical-robot results is credible, but the abstract does not establish a broad enough capability jump for a strong verdict.
-
Token Geometry
The paper argues that token embeddings and the language-model output head have unusually simple gradient geometry, and introduces Ember, an optimizer specialized for these matrices. Ember reportedly reduces optimizer memory from Adam’s O(2VD) to O(V+D), while working across supervised fine-tuning, reinforcement learning, and pretraining; the authors also claim token updates largely follow one-dimensional trajectories. The abstract gives no quantitative results or comparisons beyond these claims.
The combination of a potentially important optimizer-memory reduction and the claim that token optimization follows near-1D geometry is genuinely interesting, but the lack of numerical evidence and detail makes the broad scaling and training claims hard to trust without inspecting the paper.
-
PairCoder++: Pair Programming as a Universal Paradigm for Verified Code-Driven Multimodal and Structured-Artifact Generation
PairCoder uses two alternating agents: one writes code for a structured artifact, while the other inspects compiler diagnostics, execution results, and rendered outputs before revising it. Across 17 benchmarks and seven models, it substantially improves verifiable generation—for example, Blender scene executability rises from 0.20 to 0.78 and TikZ compilation improves by 10–30 points—but costs roughly 7× a single-model attempt. The gains depend on having a useful toolchain-based verification signal and largely disappear when that oracle is weak.
The broad, tool-grounded pair-programming recipe and large gains across multimodal structured-artifact tasks are more than an incremental prompting trick, though the core idea extends familiar iterative agent/tool feedback and comes with substantial cost.
-
Born Discrete, Made Smooth: Variational Formulation of Shallow Neural Networks
The paper replaces training shallow neural networks with a continuum variational problem over parameter densities, designed to be convex, regular, and solvable through a single linear system rather than iterative nonconvex optimization. It claims stability, nearly C3-regular optimal densities, a 1/α generalization bound, and an O(1/N) approximation rate for finite-width networks, while connecting kernel and feature-learning viewpoints.
The proposed convex/elliptic reformulation and direct linear-system solution could be a substantial theoretical alternative to mean-field and NTK analyses, but the abstract gives no concrete task, assumptions, comparisons, or empirical evidence to establish practical significance.
-
GRAFT: Grafted Reference Audio for Fine-grained Pronunciation in Zero-shot Text-to-Speech
GRAFT lets a zero-shot text-to-speech system use a short spoken example to control the pronunciation of a particular word, while generating the rest in the desired target voice. It reports 22–39% lower target-word phoneme error across five languages than the same text-only backbone, with better human preference and no apparent loss in speaker similarity or naturalness.
Per-word acoustic pronunciation control is a concrete and useful capability beyond ordinary text- or phoneme-conditioned TTS, and the multi-language listening and objective results are promising, but the abstract does not establish scale or robustness strongly enough for a strong verdict.
-
A Near-Linear-Time Solver for Graph $p$-Laplacian Semi-Supervised Learning via Continuation in $p$
This paper develops a damped chord-Newton continuation method for graph p-Laplacian semi-supervised learning, allowing each linearized solve to use a near-linear-time Laplacian solver rather than expensive sparse factorization. The implementation scales roughly linearly on graph families, is faster than a released solver, handles a 68-million-edge graph in minutes, and substantially improves MNIST accuracy over quadratic Laplacian SSL in the scarce-label setting.
The combination of continuation, damping, and near-linear Laplacian solvers addresses a real numerical bottleneck and has unusually strong scaling evidence, but the contribution is specialized to p-Laplacian SSL and the accuracy result is shown on only one dataset.
-
RADIO1D: Elastic Representations for Condensed Vision Modeling
RADIO1D studies how vision encoders change during vision-language-model training and finds that their features become more abstract and less spatially organized, with a few tokens carrying much of the image’s global meaning. It uses multi-teacher distillation and an autoencoder to turn images into adjustable-length 1D token sequences, aiming to reduce VLM compute while preserving or improving scene understanding and retrieval.
The potentially important insight is that VLM vision features may not need dense, spatially coherent patch tokens, but the abstract gives no quantitative results or scaling evidence to establish how large the efficiency and accuracy gains are.
-
Light Coils: MRI with Fully Optical Data and Power Transmission
The paper presents an MRI receive-coil architecture that replaces most electrical cabling with optical links: optical modulation carries each coil’s signal, power-over-fiber drives the front-end electronics, and optical control handles coil detuning. A single-channel coil produced in-vivo brain images at 3T with comparable SNR to a conventional coaxial setup, while a four-channel wavelength-multiplexed array demonstrated parallel imaging and low inter-channel coupling.
The fully optical combination of signal transmission, power delivery, and detuning is a genuinely interesting route to much denser MRI arrays, supported by human imaging and a four-channel demonstration, but the scale and practical advantage remain early-stage.
-
MechMath Agent Team: LLM Driven Agents for Mathematical Research
The paper presents MMAT, a multi-agent system for mathematical research that separates coordination, execution, and auxiliary capabilities, with agents for knowledge management and natural-language and formal proof generation. It reports solving 11 problems across several areas of mathematics during a two-month deployment, with formally certified proofs, but the abstract gives no details about the problems, proofs, comparison to prior systems, or how much human intervention was required.
The claimed end-to-end solution of 11 open mathematical problems would be highly notable, but the abstract provides too little evidence to distinguish genuine autonomous mathematical discovery from substantial human guidance or routine formalization.
-
Noise-Aware Synthesis of Quantum LDPC Encoder Circuits via Two-Sided Hamming Descent
The paper treats quantum LDPC encoder construction as a linear-reversible circuit resynthesis problem rather than relying on fixed algebraic procedures. Its two-sided Hamming descent and noise-aware selection pipeline reportedly cuts encoder gate counts by 53.8% on average, reduces routed two-qubit depth by up to 71%, and improves noisy state-preparation success across several CSS LDPC code families.
The large, cross-family circuit reductions make this more than a routine compiler tweak, but it remains a specialized quantum-circuit optimization result whose generality and baseline strength need closer inspection.
-
Retinal resuscitation in post-mortem eyes
The authors built a perfusion system that cannulates the ophthalmic artery and maintains excised pig and human eyes after death. In eyes perfused within 30 minutes, retinal structure and cell viability were preserved for up to 24 hours, while light responses reportedly persisted for up to 10 hours post-mortem. The main advance is an ex vivo platform for studying and potentially preserving whole eyes, not restored vision in an organism.
The persistence or recovery of retinal light responses hours after death is highly surprising and potentially important for retinal biology and transplantation, but the abstract gives limited functional detail and does not demonstrate useful vision or clinical transplantation.
-
A CSF Proteomic Clock Reveals Opposing Brain-Aging Programs and Predicts Neurological Disease Progression
The authors build a 249-protein cerebrospinal-fluid aging clock from more than 10,000 samples and validate it across cohorts and proteomic platforms. The clock is elevated in neurological disease and predicts cognitive decline, imaging progression, and dementia conversion; its proteins separate into inflammatory/vascular pro-aging programs and neuronal/synaptic maintenance programs, with a smaller 30-protein panel retaining similar performance.
The unusually large, cross-platform CSF dataset and the decomposition into opposing aging programs make this more than a routine biomarker, but the abstract supports association and prediction rather than a causal explanation or demonstrated clinical utility.
-
Curiosity shapes brain-like architectures and functions
The authors measured what infants choose to explore, then used similar curiosity-driven sampling to train recurrent networks under biological constraints. They report that this combination led to brain-like synaptic development and adult-scale network structure, while also producing compositional generalization across 20 tasks. The central claim is that curiosity may help cause cognitive and architectural complexity rather than merely emerge from it.
The curiosity-as-a driver-of-brain-like organization framing is genuinely interesting and potentially important, but the abstract gives too few quantitative comparisons or controls to support its broad emergence claims.
-
ELDR: Expert-Locality-Aware Decode Routing for PD-Disaggregated MoE Serving
ELDR routes requests in prefill-decode-disaggregated MoE serving based not just on current worker load, but on which experts each request is predicted to use. It derives expert signatures from prefill activations, groups similar signatures offline, and routes each request to a lightly loaded worker likely to already have the needed experts; a cache preserves these signatures under prefix caching. In vLLM experiments on up to 40 GPUs, this reduced median time per output token by 5.9–13.9% across three MoE models and two workloads without changing outputs.
This is a concrete and reasonably general systems insight—expert locality is an omitted factor in MoE load balancing—with multi-model, multi-workload evidence, though the gains are moderate and the core routing approach is an incremental serving optimization rather than a major new direction.
-
BrainJanus: A Unified Model for Understanding and Generation across Brain, Vision, and Language
BrainJanus proposes a single autoregressive model that converts neural activity, images, and text into shared discrete tokens, supporting both brain encoding (image/text to brain) and decoding (brain to image/text). The authors report strong benchmark results, zero-shot generalization, and preservation of brain topography, but the abstract gives no quantitative results or detail on datasets and experimental scale.
The unified any-to-any treatment of brain signals, vision, and language is a notably broader direction than separate encoding/decoding models, but the abstract provides insufficient quantitative evidence to justify a stronger recommendation.
-
GUICrafter: Weakly-Supervised GUI Agent Leveraging Massive Unannotated Screenshots
GUICrafter trains a GUI agent in two stages: it first learns visual grounding from large collections of unlabeled screenshots and webpages, then uses a small amount of labeled interaction data with reinforcement learning for calibration. The authors report performance comparable to or better than UI-TARS while using 0.1% as much data, and better results than prior methods when the annotation budget is matched.
The potentially important result is extreme reduction in annotated GUI-agent data through weak supervision, but the abstract gives no quantitative results, task coverage, or details sufficient to judge whether the gain is robust.
-
Walking in the Implicit: Interactive World Exploration via Neural Scene Representation
The paper replaces frame-by-frame latent video rollout with a fixed-size, renderable neural scene state. Its NeuWorld system learns this state from sparse posed views, evolves it based on future camera paths and retrieved history, and renders observations deterministically; the authors report better long-horizon consistency and efficiency without pretrained video models or separate 3D reconstruction modules.
The scene-state factorization is a meaningful direction for interactive world models, but the abstract gives no quantitative comparisons or evidence that the claimed consistency and efficiency gains are substantial.
-
AVTok: 1D Unified Tokenization for Holistic Audio-Video Generation
AVTok proposes a single 1D latent representation and shared codebook for paired audio-video, using a dual-stream transformer with modality-specific queries and staged training to handle the imbalance between modalities. The tokenizer reportedly supports reconstruction and downstream audio-to-video, video-to-audio, and joint generation, but the abstract gives no quantitative results or comparisons.
The unified audio-video tokenizer is a meaningful architectural direction for multimodal generation, but the abstract provides insufficient quantitative evidence to establish a substantial advantage over dual-branch approaches.
-
Nemotron-Labs-Diffusion-Image: Advancing Masked Discrete Diffusion for High-Resolution Image Synthesis
This paper presents a masked discrete diffusion text-to-image model that can revise tokens after they have been unmasked, addressing the usual one-way commitment of masked generation. It also introduces a grouped cross-entropy loss that gives partial learning signal to tokens near the target in embedding space, plus a fused implementation to reduce memory use; the abstract reports strong GenEval, DPG, and HPSv3 scores, but does not provide comparative numbers or ablations.
The editable-token mechanism and vocabulary-aware training objective are meaningful ideas for improving discrete diffusion, but the abstract lacks the baseline comparisons and ablations needed to establish a major advance.
-
From SRA to Self-Flow: Data Augmentation or Self-Supervision?
This paper reexamines why Self-Flow improves diffusion-transformer training over SRA, testing whether cross-noise-level token interactions are actually necessary. An attention-separation ablation blocks those interactions while retaining dual-noise inputs, yet preserves or improves performance, suggesting the main benefit comes from noise-dimension data augmentation; the authors use this finding to design a combined alignment-and-augmentation method evaluated on ImageNet.
The non-obvious result is that Self-Flow’s proposed cross-noise-level information exchange may be unnecessary, with the gain instead explained by an augmentation effect, though the abstract gives no quantitative results or evidence beyond ImageNet experiments.
-
Seeing Is Not Sharing: Some Vision-Language Models Overestimate Common Ground in Asymmetric Dialogue
This paper tests whether vision-language models distinguish information that is merely available to both dialogue participants from information they have actually established as shared through interaction. Across 13,077 MapTask references and several models, task-relevant map content—whether shown as images or described in text—makes models over-predict that speakers are aligned, suggesting they use static visual or referential cues instead of tracking grounding through dialogue history.
The controlled finding that models confuse shared perceptual access with established common ground is a non-obvious and useful diagnosis for multimodal dialogue, though the evidence is based on a relatively narrow MapTask setting rather than broad real-world interaction.
-
SpheRoPE: Zero-Shot Optimization-Free 360 Panorama Generation with Spherical RoPE
SpheRoPE modifies the positional embeddings of pretrained diffusion transformers so they represent the spherical geometry and periodic boundaries of 360° panoramas, without fine-tuning or iterative optimization. Combined with a geometry-focused guidance method, it reportedly produces panoramas and panoramic video across Flux and LTX-Video backbones, with competitive results while retaining the base models’ generative flexibility.
The zero-shot replacement of standard RoPE with spherical, exactly periodic coordinates is a genuinely interesting way to address panorama seams without retraining, but the abstract gives no quantitative results or evidence that the modification preserves general generation quality.
-
Flash-BoN: Instant Drafts for Inference-Time Scaling in Diffusion Models
The paper argues that inference-time scaling for text-to-image models should spend more compute on generating diverse candidates rather than repeatedly verifying intermediate denoising states. Flash-BoN produces many cheap draft images using timestep truncation, layer skipping, and activation proxies, then verifies and fully refines the best candidates; it reports consistent wall-clock gains across three benchmarks and model scales, including +8% AUC at larger scales and improved compatibility with prompt optimization.
The wall-clock finding that simple Best-of-N can beat more elaborate guided search, plus the explicit inclusion of verifier overhead and multi-knob cheap drafting, is a non-obvious and potentially useful reframing, though the method largely combines known acceleration techniques and the abstract lacks detailed benchmark results.
-
Emergence of a Shared Canonical Object Frame from In-the-Wild Videos
The paper learns a shared canonical coordinate frame for object categories from in-the-wild videos, without manually labeled canonical poses. It uses noisy Structure-from-Motion camera geometry and a coarse category-agnostic mesh as a geometric bottleneck, then learns pixel-to-mesh correspondences across 160,000 videos. The resulting frame reaches competitive category-level pose estimation accuracy against methods trained with canonical-pose supervision.
The removal of canonical-pose annotation through a shared geometric bottleneck is a meaningful and potentially scalable idea, but the abstract gives no quantitative gains and only claims competitive rather than clearly superior results.
-
SA-VLA: State-aware tokenizer for improving Vision-Language-Action Models' performance
SA-VLA makes discrete action tokens depend on the robot’s current proprioceptive state, rather than mapping each token to one fixed continuous action. Its state adapter or cross-attention decoder substantially improves manipulation success on 12 RoboTwin tasks (0.29 to 0.56) and on three zero-shot sim-to-real tasks (0.15 to 0.33), while retaining the usual autoregressive or parallel token interface.
The state-conditioned codebook is a meaningful fix to a real limitation of discrete VLA action representations, and the reported gains are large, but the evidence is limited to a small task suite and the abstract does not establish robustness across robots, datasets, or stronger policy baselines.
-
Does Verbose Chain-of-Thought Really Help? In-Distribution Evidence that Content, Not Length, Matters
The paper tests whether longer chain-of-thought improves reasoning simply by giving the model more tokens to compute, or because those tokens contain useful reasoning and checking information. Across 25 models and controlled experiments on eight benchmarks, matched semantic traces show only modest gains from verbosity, while gains are larger when extra text contains useful intermediate values or validation; filler text provides no benefit. The evidence favors a content-based explanation, but suggests that neither pure token-count computation nor pure semantic content fully explains CoT effects.
The in-distribution pairing and graph-equivalent trace intervention are unusually careful ways to separate CoT content from length, with a somewhat surprising result that verbosity alone contributes little, though the gains are modest and the abstract does not establish a major capability or efficiency advance.
-
ReactiveBFM: Reactive Closed-Loop Motion Planning Towards Universal Humanoid Whole-Body Control
ReactiveBFM combines a generative motion planner with a behavior foundation model in a closed loop for humanoid control. It trains the planner on imperfect, physically encountered states to learn recovery behaviors, then uses asynchronous replanning and trajectory chunking to handle the latency mismatch between planning and high-frequency control. On a Unitree G1 it reports zero-shot reaching of moving targets and a 93.1% success rate under severe simulated perturbations, 28.6 points above open-loop cascades.
The combination of exposure-bias training, asynchronous replanning, and whole-body reactive control addresses a real limitation of humanoid behavior models, with physical deployment and a substantial perturbation result, but the abstract provides limited detail on task breadth and real-world quantitative performance.