Issue 39 · Sep 21–27, 2026
Every candidate
All 4,417 papers were scored from their abstracts by gpt-5.6-luna; 2,143 were not skipped. Shown here: the top 150 of those, in score order. Picks are marked.
-
Post-Training Leaves Behavioral Shadows on Unrelated Decisions
The paper shows that a model can acquire capabilities from post-training through task-unrelated supervision: a student is trained only to predict one of two ordinary words selected from prompts where a post-trained teacher’s preference differs subtly from its shared ancestor. With 5,664 such prompt-word pairs and no coding examples, logits, or teacher parameters, Qwen2.5-1.5B improves by 5.34 percentage points on HumanEval+, with similar transfer reported for science, commonsense, and reading-comprehension tasks.
The striking result is that a single task-irrelevant teacher word can transmit measurable capabilities from post-training, suggesting that model updates leave a composable behavioral signature and challenging the assumption that capability transfer requires task-relevant data or richer teacher supervision.
-
Recursive self-improvement of AI research agents
AIDE^2 repeatedly rewrites and evaluates its own AI research-agent code, retaining changes that improve performance on hidden research tasks. Over an 8-day autonomous run it found seven successive improvements, including a new search strategy and mechanisms for compressing and managing context; the resulting agents transferred to four held-out tasks, including out-of-distribution weather forecasting, and matched or exceeded a strong human-engineered agent. Reward hacking also decreased without being directly optimized, though the abstract gives limited detail on experimental scale and controls.
This presents a genuinely important recursive self-improvement loop with multiple successive self-modifications and apparent cross-domain transfer, rather than a one-off agent tweak, although the claims still need scrutiny of evaluation design and reproducibility.
-
Minimax-Optimality of Posterior Sampling for Reinforcement Learning
This paper proves that unmodified posterior sampling for reinforcement learning is minimax-optimal in Bayesian regret, even when the prior over rewards and transitions is arbitrary and correlated. For finite-horizon tabular MDPs it obtains the standard \(\widetilde{O}(\sqrt{SAH^3K})\) rate, and extends the result to linear-mixture MDPs with rate \(\widetilde{O}(d\sqrt{H^3K})\). The key technical device is a shared empirical transition reference plus a Bellman variance argument that controls the mismatch between sampled-model and true continuation values without an extra state-space factor.
It reportedly resolves a long-standing question by showing that vanilla PSRL—not a modified algorithm and without structural prior assumptions—achieves minimax regret in both tabular and linear-mixture settings.
-
Disaggregated Quantization: Specializing LLM Prefill and Decode
The paper treats LLM prefill and autoregressive decode as separate inference problems, using different weight formats, activation precision, and storage strategies for each phase. Separate compute-native prefiller weights improve prompt-processing speed and substantially recover accuracy at extremely low decode bitwidths, including 32.5–35.3 point gains on 1-bit Qwen3.8-27B; streaming the prefiller from SSD yields a reported 1.78× time-to-first-token improvement at 8K context.
Phase-specific quantization and checkpoint disaggregation is a genuinely useful systems/training direction, supported by unusually large low-bit accuracy gains, latency results, and evaluations across model sizes and serving frameworks.
-
Adversarially Robust PAC Learning with Optimal VC Rates
This paper proves that adversarial robustness need not increase the distribution-free sample complexity of PAC learning: for any perturbation map, a VC class can be learned with the same optimal rates as ordinary PAC learning, in both realizable and agnostic settings. It also gives an improved first-order agnostic bound and introduces a simple “binomial-bagging” proof and algorithmic principle, improving exponentially over prior bounds.
The result overturns the expectation that arbitrary adversarial perturbations necessarily impose a statistical price, while matching classical lower bounds and substantially improving prior theory.
-
Prediction with Expert Advice: Anytime Regret with Many Experts Matches the Fixed-Time Constant
This paper studies online prediction when the time horizon is unknown and the regret guarantee must hold at every time, rather than only at a preselected final time. It gives an anytime algorithm whose regret is asymptotically the optimal fixed-horizon constant, approaching \(\sqrt{t\ln n/2}\) up to a vanishing \(O(\sqrt{\ln\ln n/\ln n})\) factor, rather than the previously known bound with a \(\sqrt{2}\) loss.
It resolves a longstanding open question by showing that requiring a simultaneous anytime guarantee does not fundamentally worsen the optimal asymptotic regret constant.
-
A New Gap Sequence for Shellsort: RL-Driven Algorithm Discovery Beyond $N^{4/3}$
The paper uses reinforcement learning and exact execution feedback to search directly over valid Shellsort gap generators, rather than fitting a hand-designed family. It reports a sparse sequence that performs best among seven classical baselines on 25 large sorting tasks and proves an asymptotic upper bound with exponent about 1.0243, matching a recent lower bound up to polylogarithmic factors—well below the long-standing $N^{4/3}$ bound for comparable sparse sequences.
If the proof and experiments hold up, this combines a genuinely unusual use of self-supervised RL for algorithm discovery with a major theoretical improvement on a decades-old Shellsort problem, though the abstract provides limited detail for independently assessing the delicate asymptotic claim.
-
Direct voltage imaging of peripheral axons in vivo reveals ongoing nociceptor firing in healthy mice
The authors use genetically encoded voltage indicators and targeted widefield imaging to record membrane-voltage dynamics directly from corneal nociceptor terminals and parent axons in living mice. They find that many terminals show spontaneous, sodium-channel-dependent action-potential-like firing even without stimulation; this activity is synchronized across branches, originates distal to the trigeminal ganglion, and depends on NaV1.8 and HCN/Ih conductances.
Direct voltage imaging of intact peripheral axons is a major technical advance, and the observation of coordinated ongoing nociceptor firing in healthy tissue—localized to peripheral terminals rather than the soma—is both unexpected and biologically consequential.
-
Boolean threshold functions, neuron capacity, and memory retrieval
This paper gives an extremely precise asymptotic count of Boolean threshold functions, showing that a single threshold neuron stores about n² − log₂(n!) bits, with an error far smaller than in previous results. It also proves that up to n−1 random memories, almost all collections have no unintended Boolean ground states, establishing a sharp memory-retrieval threshold and resolving several conjectures and open problems.
This is a substantial theoretical advance on neural capacity and associative-memory retrieval, with near-exact asymptotics and multiple conjecture/open-problem resolutions rather than an incremental model improvement.
-
Ceiling of a Task: When Can a Transformer Succeed Without Its Chain of Thought?
This paper defines a task’s “ceiling” as the best accuracy achievable by a transformer using only shallow computation, and calls tasks whose ceiling is below perfect accuracy inherently serial. It proves that, for such tasks, chain-of-thought content is computationally necessary rather than decorative: removing or scrambling the chain reduces performance to the shallow-computation ceiling, while the final answer can be produced from the completed chain in one pass. Experiments on finite-group problems, MATH-500, and AIME support the theory, including large accuracy drops after chain removal or token shuffling.
It offers a potentially general theoretical explanation for when chain-of-thought is genuinely required, backed by impossibility results and experiments spanning synthetic tasks and open reasoning models.
-
Double Descent and Malign Overfitting in Diffusion Models
The paper argues that diffusion models do not exhibit the usual benign overfitting seen in many overparameterized regressors. Using a random-features theory and U-Net experiments on CelebA, it shows that test loss begins worsening around model size p~n, even though the formal interpolation peak is delayed to p~nm when each example uses m noise realizations. The mechanism is growing bias toward the empirical score, which memorizes the training set; ridge regularization or early stopping restores the benefits of large models.
This gives a non-obvious, theoretically grounded explanation for why overparameterization can be harmful in diffusion models and identifies a scaling law and practical regularization implication that challenge the standard benign-overfitting picture.
-
Breaking Pore Symmetry to Prolong DNA Residence in Solid-State Nanopores
The study shows that tilting a solid-state nanopore can dramatically slow the passage of 500-base-pair DNA without chemical surface modifications or active trapping. In 13 nm SiNx pores, tilted membranes produced DNA events averaging 11.1 seconds at 600 mV—over four orders of magnitude longer than typical reported dwell times—and simulations attribute this to field redistribution and increased off-axis wall contact.
A simple geometric change allegedly produces an enormous increase in nanopore observation time, with experimental controls and simulations supporting a non-obvious transport mechanism.
-
GAE: Learning a Geometry-Native Latent Space for 3D-Consistent World Generation
The paper builds a compact latent representation by reparameterizing features from a geometry foundation model, so that the same latent can decode appearance, depth, camera poses, and 3D point maps. Using this latent in an otherwise standard conditional flow generator improves video quality and 3D consistency, reducing FVD by 12.7–23.1% and roughly halving camera-trajectory error on two datasets.
The paper presents a meaningful representation-level intervention—making geometry native to the generative latent rather than adding geometry as an auxiliary output—with controlled comparisons and sizable gains, though the evidence is limited to the abstract’s two benchmarks and lacks detail on generality.
-
Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents
The paper replaces fixed, write-time summaries of agent experience with read-time curation: it keeps raw trajectories and synthesizes a compact memory tailored to the current task. On ALFWorld, WebShop, and τ²-bench, this approach improves success rates over the strongest write-time memory baselines by 16.2, 16.3, and 3.9 percentage points; even an untrained curator performs competitively, suggesting that query-conditioned memory construction is itself important.
The reversal of the usual write-time memory design is a meaningful and somewhat surprising idea, supported by sizable gains across three agent benchmarks, though the abstract does not establish robustness beyond these evaluations or explain the implementation and cost tradeoffs.
-
JEV-as-a-Judge: Accept When Confident, Escalate When Unsure
The paper proposes using a decision-only language model judge that outputs label probabilities, accepting high-confidence decisions and sending uncertain cases to a slower reasoning judge. On text-comparable verdicts, it reportedly reaches within three points of GPT-6 at 0.36% of the cost and 0.15-second median latency; the confidence-based cascade is claimed to slightly exceed GPT-6 accuracy at 41% of the cost, though it struggles on math, code, logic, adversarial style, and reference-free prose.
The confidence-routed cheap-judge cascade offers a potentially important cost and latency reduction, but the abstract gives limited detail about JEV, evaluation breadth, and the credibility of comparisons against the unspecified GPT-6, so it merits inspection rather than being a clear must-read.
-
Harness-Zero: Harness Distillation via Agent-as-Harness
Harness-Zero distills the benefits of a specialized agent harness into a model’s weights. An auxiliary agent uses the optimized harness to correct the student’s actions in the target harness’s action space, and the resulting trajectories are used for fine-tuning; across knowledge work, tool use, and science tasks, success rises from 23.3% to 44.3% even after removing the specialized harness, with 82.3% recovery of tested harness-induced behaviors.
The idea of transferring harness behavior into model weights, together with the reported near-doubling of task success and performance exceeding deployment with the harness attached, is genuinely notable, but the abstract provides too little detail about tasks, baselines, and evaluation scale for a stronger verdict.
-
WorldCrafter: Consistent Video World Model with Implicit 3D-aware Memory
WorldCrafter adds a camera-queryable, implicitly 3D-aware memory to a video world model, allowing past observations to be compressed into view-specific tokens before generation. The authors report better long-horizon scene consistency and camera control during minute-scale interactive exploration from a single image or text prompt, without explicit depth correspondences.
The view-conditioned implicit memory is a meaningful approach to a central limitation of video world models, but the abstract gives no quantitative results or details about the evaluation breadth, so the claimed gains are not yet credible enough for a strong recommendation.
-
The Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks
This paper introduces “taste” as an agent’s ability to choose well between alternative directions during a long-horizon task, before seeing which direction succeeds. Taste-Bench automatically extracts such decision forks from parallel or internally branching agent trajectories, and finds that even the best evaluated model answers only 59.7% correctly; later-arriving evidence is especially difficult, and more reasoning budget does not help. Distilling outcome-informed judgments into a student improves decisions on unseen tasks and raises performance on held-out SWE-bench Pro tasks.
The fork-based, outcome-grounded measurement of intermediate agent judgment is a non-obvious framing with a potentially useful training signal, and the finding that extra reasoning does not solve delayed-evidence decisions is notable, but the abstract lacks quantitative detail on generalization and end-to-end gains.
-
A Spectral Theory of Grokking: Weight Decay induces Feature Learning
The paper develops a spectral dynamical theory of grokking in which weight decay leaves residual errors after memorization; those residuals drive the neural tangent kernel toward task-relevant feature directions, producing delayed generalization. Its reduced model predicts that grokking time scales inversely with the product of learning rate and weight decay and identifies a critical decay regime beyond which generalization fails. Experiments on modular addition across large learning-rate/decay grids support the predicted phase structure in homogeneous MLPs and qualitatively in a Transformer.
It offers a mechanistic, quantitatively testable explanation for grokking that links post-memorization feature learning to phase transitions and scaling laws, with unusually broad parameter-sweep evidence and partial transfer beyond the exactly analyzed architecture.
-
Bend the Clock: Predicting Ahead to Beat Latency in Event-Based Object Detection
This paper argues that event-based detectors should be evaluated against the scene state when their predictions become available, rather than when the input was observed. ChronoFuse uses causal cross-time feature fusion to predict that future state without seeing future observations, adding only 0.84 ms latency; it recovers most latency-related accuracy loss and improves sAP from 2.25 to 20.95 on an extreme-motion flying dataset.
The availability-time evaluation framing and causal prediction approach address a practically important failure mode in high-speed robotics, with unusually large gains under extreme motion and results across multiple event-camera datasets.
-
Where Hallucinations Live: A Cross-Architecture Circuit in VQ-Tokenized Vision-Language Models
The paper uses activation patching across 25 VQ-tokenized vision-language models to identify an early-layer attention-routing circuit associated with object hallucinations. Replacing CLIP features with VQ tokens installs the circuit, while an otherwise matched MLP control does not; ablating the circuit reduces open-ended hallucinations by 31% on CHAIR, unlike tuned VCD or DoLA decoding. The main new claim is that this failure is tied to VQ-based architecture and pretraining rather than generic calibration.
The cross-model causal analysis and controlled architectural swap offer a specific, potentially general explanation for VQ-VLM hallucinations, with an intervention that appears to work where standard decoding fixes do not.
-
AD-WM: Action-Discriminative World Models for Counterfactual Model Predictive Control
AD-WM trains latent world models to preserve differences between the outcomes of alternative actions, rather than only minimizing factual next-state prediction error. It adds inverse-dynamics and action-recovery objectives during training, substantially improving hard-start simulated control (3.7% to 52.0% success) and zero-shot Franka pick-and-place (42.2% to 71.1%) without changing the MPC procedure at test time.
The paper identifies a consequential mismatch between standard world-model training and counterfactual action selection, proposes a targeted remedy, and reports unusually large gains across simulation plus a real-robot transfer result.
-
A Long-Legged, Direct-Drive Monopedal Robot Achieves Exceptional Jump Height
The authors build a 281 g, single-legged robot that jumps 7.6 m high with a 190 ms stance phase, using a direct-drive motor rather than springs or other elastic energy storage. Its fabric-wrap transmission provides variable mechanical advantage to keep the motor near peak power while supporting the long leg, and onboard propellers provide ground and in-flight attitude control; outdoor tests include vertical, tilted, and stabilized jumps.
This combines an unusually large demonstrated capability jump with a genuinely distinctive spring-free transmission and credible outdoor validation, making it more than an incremental legged-robot design.
-
Mandela-Bench: Multimodal Models Remember Canonical Images Instead of Seeing Them
Mandela-Bench tests whether multimodal models verify what they actually see in familiar historical images, rather than relying on memorized canonical versions. Across 36 models and 1,507 factual edits, models often identify removed public figures from memory and accept contradictory images as authentic; supplying facts helps only when the recognizable composition is removed, suggesting recognition of the canonical image can override visual verification.
This is a well-controlled, broad evaluation exposing a non-obvious failure mode—memory-driven recognition overriding observation—that matters for multimodal reliability and is more than a standard benchmark result.
-
The Decomposition Tax: LLM Pipelines Lose Up to 40 Accuracy Points at Their Own Interfaces
The paper measures how much accuracy is lost when an LLM pipeline passes an intermediate representation from one stage to the next without showing the original problem again. Across 21 open-weight models and two math benchmarks, these interfaces can cause very large drops—up to 40.5 percentage points—and the loss depends strongly on prompt wording and whether relationships between quantities are preserved. Re-showing the original problem to the stage after the lossy interface, and explicitly asking quantity-listing stages to retain relationships, often repairs the damage; a preregistered rule for predicting where the loss occurs did not hold.
This reports a large, counterintuitive failure mode in modular LLM reasoning, with broad paired-model experiments and concrete interventions rather than just a benchmark improvement.
-
Oracle-Efficient Online Classification with Stochastic Inputs and Adversarial Outputs
The paper studies online binary classification where contexts are stochastic but the adversary can choose losses adaptively. It shows that a simple Gaussian-perturbed Follow-the-Perturbed-Leader method achieves near-optimal regret using only one optimization-oracle call per round, without enumerating the experts; for infinite classes, the bound depends on VC dimension. This resolves a prior open problem by showing that the computational cost of hybrid classification can match ordinary statistical learning.
It gives a simple oracle-efficient algorithm with optimal regret and resolves an explicit open problem, establishing a non-obvious computational equivalence between online hybrid classification and statistical learning.
-
YuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality
YuE2 uses one autoregressive–non-autoregressive mixture-of-transformers to plan a song as a readable symbolic score, convert that plan to semantic music tokens, and render a complete audio recording. The authors report better expert preferences and benchmark scores than public baselines, near-parity with Suno v5 under best-of-8 selection, plus score editing and zero-shot cover generation; auxiliary models provide symbolic and semantic supervision from unaligned recordings.
The unified symbolic-planning-to-audio pipeline and demonstrated editability are genuinely interesting, but the strongest comparisons rely on self-reported listening studies, candidate selection, and a technical report, so the claims need closer validation before warranting a strong verdict.
-
Coding Agents for Generalized Task and Motion Planning Problems
This paper tests whether coding agents can build generalized task-and-motion planners by interacting with a simulator, writing programs, and refining them across test cases. Across 28 simulated environments and 98,000 evaluation episodes, the agents achieve 56–95% mean success versus 47% for available hand-engineered planners, while using about an order of magnitude less computation per instance as object counts increase. The notable result is that general-purpose coding agents can discover useful, reusable TAMP strategies with little TAMP-specific engineering.
The broad evaluation and strong scaling result make automated program synthesis for generalized TAMP a genuinely interesting direction, but this is primarily evidence that existing coding agents are effective baselines rather than a new planning method or demonstrated real-world robotic capability.
-
VideoGen-Agent: Reinforcing Video Generation Agents
VideoGen-Agent trains a multimodal agent to improve video generation by deciding when and how to use augmentation, generation, and verification tools through multi-turn interactions. On a 600-prompt benchmark spanning identity, physical consistency, procedural knowledge, composition, and temporal structure, it raises performance from 56.5 to 75.6; replacing its generation tools with stronger ones further raises results to 86.1 without retraining the agent.
The combination of reinforcement-trained tool use with video generation, including transfer to improved downstream generators, is a meaningful direction, but the evidence relies heavily on a new 600-prompt benchmark and abstract-level claims rather than broad independent validation.
-
Verifiable Hidden Dynamics Play: Generating Agentic RL Environments from Solved Mechanisms
VHD-Play generates stateful, tool-based RL environments by first sampling and solving an underlying mathematical mechanism, then rendering that mechanism into an interactive environment with an inherited executable evaluator. The authors report training gains from 0.204 to 0.815 on a five-family diagnostic, transfer to eight unseen mechanism families and several external agent benchmarks, and near-perfect performance on their e-commerce evaluation. A notable analysis suggests that the main learning challenge is interacting with hidden evolving state, rather than solving the underlying written problem.
The solved-mechanism-to-verifiable-environment pipeline and the finding that stateful hidden dynamics drive much of the difficulty are genuinely interesting, but the unusually large gains come primarily from abstract-level claims in a technical report and need careful scrutiny of benchmark contamination, task diversity, and external evaluation.
-
RRSI: Regularized Recursive Self-Improvement of Agent Harnesses
RRSI automatically evolves an LLM agent’s prompts, control flow, tools, memory, and context handling, while regularizing the search to avoid benchmark-specific overfitting. Its proposal and selection mechanisms encourage reusable edits, prune ineffective or costly changes, and reduce the number of policy tokens by 30%; across eight benchmarks, it reports gains of up to 4.7 points on held-out out-of-distribution tasks and up to 14.1 points on the evolution benchmark.
The notable contribution is treating agent-harness evolution as recursive self-improvement with explicit anti-overfitting and cost regularization, supported by multi-benchmark in- and out-of-distribution results, though the gains and generality depend on details of the evolution setup and evaluation.
-
A discrete Connexin26+ neural crest lineage emerges in mid-life and mediates enhanced central brainstem responses to elevated CO2 levels for deep breathing.
The study identifies a small population of about 20 Connexin26-positive neural-crest-derived cells that appears in the ventral brainstem during middle age and disappears in old age. Genetically removing Connexin26 from these cells reduces local ATP release and weakens the increase in tidal volume during elevated CO2 by about 40%, linking an age-dependent cell population to breathing control. The paper also proposes a speculative evolutionary connection to ancestral CO2 tolerance and eusocial care.
The reported age-specific emergence and disappearance of a tiny neural-crest lineage with a large effect on CO2-driven breathing is highly unusual and mechanistically interesting, but the abstract provides limited validation beyond genetic ablation and includes substantial speculative evolutionary claims.
-
A prefrontal cortex-hypothalamus circuit for heart rate control in non-human primates
In monkeys, the authors identify a ventrolateral prefrontal area (a12) that causally controls heart rate through a specific projection to the lateral hypothalamus. Electrical, chemogenetic, and anatomical experiments show bidirectional and partly state-dependent control of cardiac activity, while ultrasound stimulation of virally sensitized a12 neurons produces related effects and also changes risky decision-making. The work links cognitive circuitry to autonomic control and demonstrates a primate sonogenetic manipulation of that circuit.
The combination of a newly localized prefrontal–hypothalamic heart-control pathway, causal cognitive/autonomic effects in primates, and a targeted ultrasound-based perturbation is unusually compelling and could open a new direction for noninvasive circuit control.
-
Neurodegeneration-inducing macromolecules exit the brain via nanovascular conduits formed by reticular fibroblasts
The authors report previously undescribed ~2-micron conduits made by reticular fibroblasts that form a mesh around cerebral vessels and extend through the meninges and CSF space. In mice and human tissue, fluorescent tau and amyloid-beta appear to leave the brain through these conduits, with arteriolar and venular routes contributing similarly—challenging the idea that glymphatic flow is the dominant clearance mechanism. They also propose that impaired transfer through the network could contribute to cerebral amyloid angiopathy and neurodegeneration.
This could substantially revise the anatomy and mechanism of brain waste clearance, but the evidence is from a preprint abstract with limited quantitative and causal validation, so the striking interpretation warrants scrutiny rather than a strong recommendation.
-
Rethinking Automated Voice Similarity by Shifting from EER to Embedding Geometry
The paper tests whether speaker-verification accuracy is a good proxy for human judgments of voice similarity. It finds that training objective matters more than EER: margin-based classification models align poorly with human perception, while prototype-based metric learning aligns better, apparently because human voice perception is effectively low-dimensional. A dimensionality bottleneck increases the reported human-alignment correlation from 0.08 to 0.74, and effective embedding dimensionality correlates strongly with perceptual alignment.
The claimed decoupling of verification accuracy from perceptual voice similarity, together with a geometric explanation and large bottleneck effect, is genuinely non-obvious, but the abstract does not establish how broad or robust the human evaluation and model comparisons are.
-
RoboFoundry: System-as-Policy Evolution for Self-Learning Embodied Agents
RoboFoundry treats an embodied robot’s supporting software—context management, persistent memory, skills, recovery, and action interfaces—as a single system that can evolve from execution traces. It validates task-specific changes, promotes recurring improvements, and uses a shared semantic interface to transfer capabilities across robots; the abstract reports large gains on several benchmarks and some zero-shot, online real-robot transfer.
The system-as-policy framing and persistent, validated self-improvement across memory and skills are potentially substantial, but the unusually large benchmark claims are difficult to assess from the abstract and need scrutiny of baselines, ablations, failure cases, and real-world scale.
-
Training Object Permanence in World Models
The paper introduces WROP, a procedurally generated training corpus and evaluation exam built around object permanence, solidity, and related cognitive tasks for video world models. Training a 16B-parameter continuation model on 1.5M samples yields the top score among continuation models and third overall in a blind pairwise Elo comparison of 14 video models. The main novelty is treating specific physical-cognition priors as explicit, scalable training targets rather than relying on them to emerge from generic video data.
It proposes a potentially useful new direction for training physical reasoning in video models, but the evidence is largely on a bespoke synthetic benchmark and does not yet establish transfer to real-world physical understanding or broadly improved world modeling.
-
Anchored Extra-Proximal Methods: Optimal Higher-Order Methods for Monotone Inclusion Problems
The paper introduces Anchored Extra-Proximal methods for composite monotone inclusion problems, extending extragradient-style optimization to second and higher orders through Taylor approximations in an implicit update. It proves a \widetilde{O}(\varepsilon^{-2/(3p-1)}) oracle complexity for order-p methods and matches it with a general deterministic lower bound, establishing optimal dependence on accuracy up to logarithmic factors.
The matching lower bound and optimal higher-order complexity appear to establish a genuinely new frontier for monotone inclusion optimization, though the result is specialized theoretical optimization rather than a demonstrated practical ML capability.
-
What Does the Rank Buy? A Spectral and Distributional Analysis of Low-Rank Adaptation
This paper argues that LoRA rank does not necessarily control model capacity when each low-rank factor has its own norm budget. In the analyzed linear setting, all ranks r≥1 can achieve the same nuclear-norm-bounded class, so rank-independent complexity and distribution-shift bounds result; rank instead determines which spectral directions can be changed and the budget required to cancel pretrained components. A joint product-norm constraint restores rank-sensitive complexity only for well-spread feature distributions.
The paper offers a potentially important reinterpretation of what LoRA rank controls and derives sharp theoretical consequences, but the claims are limited to an idealized linear model and norm-budget assumptions, without evidence that they explain practical nonlinear LoRA behavior.
-
Memory as a cache: Exact context reuse and deletion by construction
SMem redesigns transformer memory so each fixed-size context block is encoded independently and later combined through cross-attention, rather than entangling every cached token with its full prefix. This makes blocks exactly reusable, order-independent with respect to edits, and deletable in time proportional to the removed block; experiments report much better extrapolation beyond the training context, lower cache costs, faster decoding, and large deletion speedups, with comparable perplexity at some model sizes.
Exact, path-independent cache composition and deletion is a genuinely new systems/architecture direction with credible multi-scale efficiency and extrapolation results, but the evidence is still largely from controlled benchmarks and the perplexity tradeoff and limited retrieval accuracy make it premature to call a field-changing result.
-
MassAlloc Attention: Let Attention Allocate Its Own Compute
The paper introduces MALA, a fused attention kernel that computes attention scores for all legal causal positions but adaptively skips low-contribution post-score work based on normalized attention mass. It reports near-FullAttn outputs, gradients, perplexity, and task performance, while achieving up to 2.2× forward, 3.0× backward, and 1.6× decoding speedups at long context, with similar results in models from 0.6B to 32B parameters.
This is a credible new direction for attention efficiency—distribution-adaptive allocation with a shared training/inference tolerance—and the reported matched-work, long-context, and multi-scale model results suggest more than a routine kernel optimization, though the claims still warrant checking against implementation details and broader workloads.
-
Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs
This paper studies whether a Transformer can process two linearly combined text streams as a superposition of the streams’ separate predictions. It reports that this approximate linearity is present in pretrained models, weakens with further pretraining, can be restored by lightweight fine-tuning, and can be exploited to decode two continuations from one forward pass.
The proposed superposition-linearity hypothesis and single-pass disentangled generation are genuinely interesting, but the abstract gives no quantitative results, model scale, or experimental detail to establish how robust or useful the effect is.
-
CARE: Experience-Guided Atomic Corrective Execution for Vision-Language-Action Policies
CARE improves recovery in vision-language-action robot policies by collecting real execution failures, modeling where failures occur within task stages, and synthesizing corrective demonstrations from those distributions. Its inference system uses 3D monitoring to trigger small atomic corrections or re-operations, reporting roughly 14.5-point gains in simulation and 15.9-point gains on real dual-arm tasks, along with a benchmark for intermediate failure recovery.
The combination of experience-driven failure modeling and atomic corrective execution addresses a major practical weakness of VLA policies, with substantial reported real-robot gains, but the abstract does not establish how broad or durable the improvement is relative to strong recovery baselines.
-
PACT: From Credit Assignment to Critic Alignment
The paper gives a formal characterization of token-level credit assignment using three conditions—completeness, prefix consistency, and neutrality—and uses it to relate OPD, RLOO, and GAE to policy-gradient credit. It then proposes PACT, which updates the actor before training an importance-corrected critic; on agentic math reasoning it reports sizable gains over PPO and GRPO, while improving SWE-bench Verified by 2–4 percentage points.
The combination of a purported uniqueness result for token-level credit with a policy-aligned critic and substantial reported gains is worth checking, but the abstract does not establish how broad or rigorously validated the results are.
-
CompoWorld: Compositional Environment Scaling for General Agents
CompoWorld builds scalable training environments by composing reusable, typed services into multi-service workflows, rather than generating tasks inside one environment. It creates 448 services and 10,130 tools, generates verified cross-service trajectories, and combines supervised fine-tuning with a reward focused on incomplete task criteria; the resulting Qwen3.6-35B-A3B improves its backbone by 9.17 points across eight benchmarks and reportedly beats several agent baselines on AutomationBench.
The compositional, cross-service environment-generation approach is a plausible new direction for training general tool-using agents, but the abstract gives limited detail about benchmark validity, baselines, and how much of the gain comes from synthetic data or training scale.
-
TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations
TrackEverything represents a long video as persistent 3D scene tracks rather than repeatedly tracking 2D points in short windows. It deduplicates observations of the same surface, focuses trajectory refinement on dynamic points, and uses sampled 3D feature correlations to reduce memory; the authors report tracking all visible points for over 1,000 frames within 40 GB of GPU memory and a large short-clip gain on TAPVid-3D.
The persistent, deduplicated 3D representation addresses a real long-horizon dense-tracking bottleneck and claims a substantial capability and memory improvement, but the abstract provides limited comparative and scaling detail for such a strong result.
-
CoWindow Attention: Full Causal Coverage Is a Collective Property
CoWA assigns different KV heads complementary long-range windows, while sharing local and prefix-sink context, so the head ensemble collectively covers the full causal history without every head duplicating it. In controlled experiments it nearly matches dense attention on recall, perplexity, and downstream long-context evaluations while reporting large operator-level gains at 128K context: 7.4–8.6× faster training operations, 3× faster decoding, and 7.6× lower decoding memory.
The collective-coverage framing is a substantive alternative to per-head full attention, and the reported scaling and long-context results plus large efficiency gains make it worth checking, though the evidence is still abstract-level and the broad downstream equivalence claims need verification.
-
Learning to Learn from Context: Synthetic Training from Perturbed Public Documents
The paper builds an automated training-data pipeline that lightly rewrites public documents, generates document-dependent questions and answers, and filters out examples that do not truly require the supplied context. Training on roughly 10,000 such examples improves a Qwen3.6-35B-A3B model on CL-bench from 13.7% to 24.6% after SFT and rubric-based RL, with transfer to long-context understanding, instruction following, and reasoning but little change in coding or factual knowledge. The main novelty is using perturbed public documents to reduce the benefit of memorization while synthesizing context-learning supervision without human annotation.
The sizeable context-learning gain from only 10k automatically generated examples and the explicit attempt to separate context use from memorization are worth examining, but the evidence is centered on one benchmark and the core synthetic-data strategy is a plausible extension of existing approaches rather than an obviously new paradigm.
-
Displacement Geometry Captures Platonic Shared Reality Across Models and Modalities
The paper argues that independently trained models do not necessarily place corresponding examples at the same absolute coordinates, but do share the directions between examples after a simple rotation/reflection alignment. Across 44 vision and language encoders, it reports a split between linearly aligned semantic structure and model-specific capability structure, with fine-tuning and relational distillation affecting these components in opposite ways. It also proposes “Shadow Casting,” which transfers a source model’s capabilities through a cached forward pass rather than fine-tuning, and reports a promising SHADOWCLIP proof of concept.
The displacement-based account of cross-model representation similarity and cache-based capability transfer are genuinely interesting directions, but the abstract gives too few quantitative details to establish that the broad geometric claims and practical transfer gains are robust enough for a strong recommendation.
-
Provably Complete Generalized Planning with LLMs
The paper uses an LLM to generate generalized planning programs in Lean, along with formal proofs that those programs solve every instance satisfying a supplied domain specification. A semantic-preserving PDDL-to-Lean translation lets Lean’s kernel check the proofs; the method succeeds on 12 of 13 benchmark domains, moving beyond prior work that only tested plans on sampled instances.
Automatically producing kernel-checked completeness proofs for LLM-generated generalized plans is a meaningful shift from empirical test coverage to formal guarantees, although the evidence is limited to 13 benchmark domains and depends on the supplied specifications.
-
Scaling Density Functional Theory with Gaussian Splatting
The paper introduces GS-DFT, which represents molecular orbitals as optimizable 3D Gaussian clouds rather than using a fixed atom-centered basis. The Gaussian positions, widths, and coefficients are optimized directly against the DFT energy, with new screened density-fitting and differentiable orthogonalization procedures. The authors report conventional large-basis accuracy with fewer parameters, improved treatment of stretched bonds and anions, and calculations on systems up to 2,742 atoms using one four-GPU node.
This is a genuinely unusual, potentially scalable reformulation of DFT with an impressive claimed system size, but the abstract provides limited quantitative comparisons and the computational robustness of the approach needs close inspection.
-
MemBodied: Recurrent Associative Memory for Vision-Language-Action Models
MemBodied adds a fixed-size episodic memory to vision-language-action robot policies, combining an interaction-based associative state with a compact representation of the initial scene. On five history-dependent manipulation tasks, it substantially outperforms stateless and recurrent-memory baselines, while using much less additional capacity; it also improves performance on the fully observable LIBERO-Long suite.
The fixed-size memory design targets a real limitation of VLA policies and reports large gains over recurrent and context-expansion alternatives, but the evidence is mainly benchmark-based and the abstract does not establish broad generality beyond the evaluated manipulation suites.
-
Fine Until Fine-Tuned: Repeated Solutions Make Reasoning Fragile
The paper finds that repeatedly training on the same worked reasoning solutions can make a model’s reasoning highly vulnerable to later fine-tuning, even when that later training is unrelated to reasoning. Models trained on repeated solutions initially match models trained on many unique solutions, but later instruction tuning reduces their math performance from about 95% to 86% or as low as 59%; using fresh solutions or modest replay prevents the degradation. The lost ability is largely recoverable with a few reasoning-training updates, suggesting suppression or interference rather than permanent forgetting.
This is a non-obvious and practically important claim about data repetition creating latent fragility during continued training, supported by comparisons across repetition schemes, recovery tests, and multiple models/tasks, though the abstract does not establish how broadly it generalizes.
-
Charge-screening dominates signal transduction in electrochemical nucleic acid sensors for aminoglycosides
This study argues that aminoglycoside electrochemical aptamer sensors do not mainly signal through the expected target-induced DNA shape change. Using electrochemical, optical, binding, structural, and fluorescence measurements, the authors find similar responses from unrelated DNA sequences and attribute the signal primarily to nonspecific electrostatic charge screening by the positively charged antibiotics.
It challenges the standard mechanistic explanation for a widely used sensor class and provides multimodal evidence for a substantially different, more general signaling mechanism, though it is outside the reader’s core areas and is a preprint without quantitative detail in the abstract.
-
Lifespan EEG Reference Charts Reveal Aperiodic Confounds in Beta-Band Biomarkers Across Neurological and Neuropsychiatric Disorders
The authors build lifespan EEG reference charts for beta power and frequency using data from 22,094 people aged 1–100, separating periodic beta activity from the aperiodic 1/f background. They report non-linear beta development and show that correcting for the aperiodic component reverses the apparent developmental trajectory of beta frequency; it also changes clinical interpretations, including revealing increased adjusted beta power in schizophrenia despite reduced absolute beta power.
The large normative dataset and striking demonstration that aperiodic activity can reverse developmental and disorder-related beta conclusions make this more than a routine EEG biomarker study, although the preprint abstract does not establish cohort comparability or robustness well enough for a strong verdict.
-
On the Diffusibility of High-Dimensional Latents
The paper studies diffusion or flow-matching models operating in high-dimensional latent spaces produced by reconstruction-capable visual encoders. It argues that reconstruction fine-tuning makes these representations effectively low-dimensional, so standard velocity prediction wastes capacity modeling noise directions off the signal manifold; switching to clean-data ($x_0$) prediction improves text-to-image generation across several encoders.
The geometric explanation for why velocity prediction becomes inefficient in high-dimensional, effectively low-dimensional latents is a non-obvious and potentially reusable insight, though the abstract provides limited detail on the size and breadth of the reported gains.
-
InternW0-$Δ$: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data
InternW0-Δ is a generalist robot manipulation model that jointly learns visual dynamics and actions from more than 20,000 hours of heterogeneous robot and human demonstration data. It combines video and action experts, frozen semantic guidance, distilled 4D geometric priors, and a training-only mechanism called Causal Imprint intended to expose future-relevant scene changes to the action policy without rolling out video at inference.
The unified world-action architecture, large open multimodal training corpus, and future-supervision mechanism could represent a meaningful direction for generalist manipulation, but the abstract gives no quantitative results or details sufficient to establish that the claimed cross-platform gains are substantial.
-
World Action Agent: Harnessing VLMs for Robot Manipulation via World Action Rehearsal
World Action Agent gives a vision-language model an interactive visual workspace for robot manipulation rather than asking it only to predict constraints or generate programs. It combines automatically selected contact views, editable action rehearsal, visual correction during execution, and reusable skills learned from demonstrations; on LIBERO-Pro it reports 75.6% average success, transfers the skills to robosuite without additional learning, and improves a smaller model’s out-of-domain success from 1.7% to 43.3% through harness-trace fine-tuning.
The unified action-rehearsal and in-view-correction harness, plus transfer of learned skills across simulators and supervision of a smaller VLM, is a substantive systems direction, but the evidence is currently limited to abstract-level claims, one main benchmark, and a work-in-progress report.
-
Program-Verified Self-Evolution for Vision-Language Models
VQS trains vision-language models from unlabeled images by having the model build structured records such as scene graphs or chart tables, then using fixed programs to generate questions and answers from those records. Rather than voting on full answers, a vision-language checker verifies short extracted claims; this raises human-rated answer correctness from 76% with majority voting to 94% and improves Qwen3-VL by up to 3.84 points over three self-training rounds across ten benchmarks.
The program-generated, claim-level verification scheme is a substantive alternative to noisy self-training labels and is supported by broad benchmark results, but the reported capability gains remain moderate rather than transformative.
-
Learning to Discover Interesting Mathematics
The paper proposes ranking automatically generated theorems by an intrinsic notion of interestingness: how difficult their proofs are relative to the length of their statements. A 27B model predicts proof difficulty, and optimizing this score reportedly produces more novel, less Mathlib-overlapping theorems while supporting iterative expansion of a formal library. The work also claims that this metric correlates with downstream usefulness, though the abstract gives limited detail about that evaluation.
The proof-difficulty-per-statement metric and its use for steering self-expanding formal mathematics are genuinely interesting, but the abstract provides limited evidence that the metric reliably captures human mathematical value or that the generated results are substantively useful.
-
RoboFollow: Unveiling the Instruction Following Mirage in Embodied Agents
RoboFollow argues that high success rates in embodied agents can mask poor instruction following when each visual scene has only one plausible task, making language unnecessary. It introduces a benchmark with multi-branch scenes and hierarchical tests that separate language intent understanding from motor execution, and reports that nine VLA/WAM policies—and several attempted mitigations—perform well on simple settings but fail as visual and linguistic ambiguity increase.
The low-scene-entropy diagnosis and controlled evaluation protocol expose a plausible, consequential weakness in current embodied-agent evaluations, but this is primarily a benchmark-based diagnosis with limited evidence from the abstract rather than a demonstrated capability breakthrough.
-
InternW0: A Foundational Physical World Model for Efficient Real-World Interactions
InternW0 is a multimodal robot world model that predicts visual futures while producing continuous actions at a faster rate, using cached context and observation-conditioned updates instead of recomputing predictions after every action change. It is trained on about 7,200 hours of robot and egocentric data and tested on long scientific workflows, including metal–organic framework synthesis and force-aware dexterous pipetting, but the abstract gives no quantitative results or comparisons.
The asynchronous video-to-action design, heterogeneous embodiment support, and demonstrations on extended contact-rich laboratory workflows are plausibly important, but the lack of numerical results and baselines makes the claimed advance difficult to assess.
-
Morphometric Imitation: From Morphology and Contact Aware Hand Retargeting to Sim-to-Real Visuomotor Policy
The paper proposes a three-stage pipeline for transferring human hand-object demonstrations to robot hands: morphology- and contact-aware kinematic retargeting, residual RL for dynamically feasible motions, and distillation into visuomotor policies. Across three robot hands and ten interactions, the retargeting improves contact matching and downstream success; on a Sharpa hand, the resulting policies reach 89.3% success in 300 real-world trials across 30 objects and 10 categories without object-specific deployment training.
The combination of contact-preserving morphometric retargeting with residual RL and zero-shot real-world visuomotor deployment addresses a central dexterous-manipulation bottleneck, with substantial multi-object hardware results, but the abstract does not establish a clearly paradigm-shifting capability or broad generalization beyond the tested setup.
-
Rate-distortion optimization for full-reference image quality metrics via stochastic Hessian estimates
The paper uses the Hessian of full-reference image-quality metrics such as MS-SSIM and LPIPS to construct an input-dependent quadratic distortion that can be optimized block-by-block during video encoding. Approximating the Hessian with either block-diagonal or diagonal terms enables codec-compatible rate-distortion optimization, yielding reported 14.2–36.7% BD-rate savings under the target metrics on Kodak and CLIC with VVC, at a 10–30% encoding-complexity cost.
The notable contribution is turning non-decomposable perceptual metrics into blockwise RDO objectives without decoder changes, with unusually large reported gains, though the evidence is limited to two image datasets and the abstract gives little detail on baselines or perceptual tradeoffs.
-
All-in-One Multilingual Scene Text Recognition with Script-aware Mixture-of-Experts
The paper builds a single multilingual scene-text recognizer covering 229 languages across 10 scripts. Its ScriptMoE model uses a shared visual encoder, script-routed decoder experts, and a shared expert; on the authors’ benchmarks it improves recognition accuracy by 1.31 points and raises PP-OCRv5 end-to-end F1 from 65.71% to 80.89%, slightly exceeding a much larger VLM.
The script-aware sparse decoder appears to deliver a substantial practical gain and a much cheaper alternative to VLM-based multilingual recognition, but the evidence relies heavily on synthetic data and an author-assembled benchmark, making the generality of the result uncertain.
-
Strategically Diverse Sampling for Self-Training
This paper argues that self-training benefits more from diverse problem-solving strategies than from repeatedly sampling similar correct answers. It introduces GROOT, which samples distinct paths through a hierarchy of approaches, and compares it with a simpler verbalized-sampling method; across programming and next-chapter prediction, strategically diverse data improves difficult-task performance and initialization for RL and test-time scaling. A particularly surprising claim is that diverse but incorrect traces from a 4B model outperform IID distillation from a 235B teacher.
The strategic-diversity framing and the reported success of incorrect small-model traces challenge standard assumptions about self-training, but the abstract gives limited quantitative and breadth-of-evidence detail, so this merits inspection rather than a strong recommendation.
-
Generalization Dynamics of LM Pre-training
This paper argues that language models do not smoothly progress from shallow pattern matching to robust generalization during pre-training. Using a toy evaluation suite, the authors find abrupt switches between relying on memorized or in-context patterns and using more general reasoning, and propose that competing circuits in a capacity-limited model and the recent data mix drive these switches. They use the phenomenon to select better intermediate checkpoints and pre-training data, though the abstract gives few quantitative details about the scale or robustness of the evidence.
The proposed mode-hopping view of pre-training generalization is genuinely non-obvious and potentially important, but the evidence is described mainly through a toy suite with no reported effect sizes or broad-model validation.
-
Hunger and sleep recruit distinct brain systems to form spatial memory
In food-deprived rats, spatial memories formed while hungry were later retained even without sleep, whereas feeding and staying awake prevented robust memory. This hunger-dependent memory required NPY signaling and the retrosplenial cortex, but not the hippocampus during encoding, consolidation, or retrieval, unlike sleep-dependent spatial memory.
The finding that hunger can support hippocampus-independent spatial memory through an NPY–retrosplenial pathway is genuinely surprising and potentially important, but the abstract provides limited detail and the evidence appears confined to pharmacological tests in rats.
-
Uranus: Building the Next-Generation Simulation Infrastructure for Embodied AI
Uranus is a learned robot simulator that conditions an autoregressive diffusion video model on incoming future joint-position trajectories and generates synchronized multi-view observations online. It supports open-ended rollouts rather than fixed-length episodes and reportedly reaches 24 FPS after optimization across multiple robot embodiments and camera setups; code, weights, and data are released.
The combination of streaming, trajectory-conditioned generation and real-time multi-embodiment simulation could be practically important, but the abstract provides no quantitative comparisons or evidence that it substantially outperforms existing learned simulators.
-
Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents
Jev-Mem separates fast, lightweight memory-management decisions from slower LLM reasoning. Its controller handles memory organization, retrieval routing, graph traversal, budget allocation, and stopping, while the LLM is used mainly for difficult reasoning and synthesis; on LoCoMo it reports an 11% relative quality gain, 6.6× faster construction, and 36.7% lower query latency than competing memory systems.
The explicit System-One control plane for adaptive memory operations is a meaningful systems direction with notable reported efficiency gains, but the evidence is limited to one benchmark and LLM-as-a-Judge results, so it is not strong enough to prioritize highly.
-
Agensh: Scaling Organizational Intelligence to 1,024 Agents
Agensh replaces a central multi-agent orchestrator with an asynchronous shared workspace, messaging layer, and shared context, allowing workers to claim tasks, exchange findings, verify work, and merge progress themselves. On five difficult ProgramBench tasks, increasing GPT-5.6-sol workers from 1 to 128 improved mean test-pass rate from 19.31% to 28.78%; on pandoc, 1,024 workers improved it from 33.89% to 55.06% and reached comparable performance sooner. The main new result is empirical scaling of a self-organized agent organization to 1,024 workers, though the evidence is limited to a small task set.
The decentralized coordination design and unusually large agent scale are worth noting, but the abstract provides limited evidence beyond a handful of coding tasks and does not establish that the gains come from a broadly useful organizational principle rather than more parallel sampling and compute.
-
EmbodiedSWE: Coding Agents for Long Horizon Dexterous Robotics
The paper uses coding agents to write and iteratively refine robot-control programs for difficult, long-horizon manipulation tasks, then turns verified solutions into many simulated demonstrations for training vision-language-action policies. The agents can solve contact-rich and deformable-object tasks and transfer solutions across embodiments, while diversification of generated trajectories improves held-out generalization; a policy trained only on these simulated demonstrations also completes one task on a real robot.
The combination of coding-agent task solving with scalable, diversified supervision for general robot policies is a meaningful direction, but the evidence is mainly simulation-based with only a single real-robot demonstration and no quantitative results in the abstract.
-
Tactile-JEPA: Topology-Aware Self-Supervised Representation Learning for Distributed Tactile Sensors
Tactile-JEPA pretrains encoders for distributed electronic-skin sensors by predicting masked sensor embeddings, using the sensors’ irregular connectivity graph and separate local/global masking scales. Across three datasets covering different sensor types, robot setups, and configurations, it improves force estimation by 6.3% and in-hand orientation prediction by 20.8%, with additional gains in downstream policy learning.
The topology-aware, dual-scale self-supervised treatment of sparse tactile arrays is a meaningful adaptation beyond standard visual SSL, and the gains across three diverse datasets make it worth a closer look, though it appears to be an incremental advance rather than a major capability jump.
-
ExplorationBench: Measuring AI Systems' Exploration in Verifiable Alien Worlds
ExplorationBench evaluates whether AI systems can discover and use rules that conflict with familiar knowledge. It places systems in executable “alien” sandboxes with flawed manuals, feedback, and tools, then tests them on held-out tasks; the authors report that systems can learn unfamiliar rules, but exploration quality varies and can even undo prior gains. The main contribution is a benchmark designed to separate genuine exploration from retrieval of pretrained knowledge.
The alien-world setup and held-out verification address a real weakness in current exploration evaluations, but this is still primarily a benchmark paper and the abstract gives too little quantitative evidence to justify a stronger recommendation.
-
AURA: Uncertainty-Routed Activation Editing for Acoustic Grounding in Speech Foundation Models
AURA freezes an encoder–decoder speech model and edits a sparse set of decoder cross-attention activations, routing those edits from uncertainty signals such as diffuse or abruptly shifting attention. Across four stress-test datasets, it reportedly cuts hallucinations on non-speech audio from 89.18% to 1.94% and reaches near-LoRA WER on imperfect speech labels with about 500 times fewer trainable parameters.
The combination of uncertainty-routed activation editing and a very large reduction in non-speech hallucinations is genuinely interesting, but the abstract gives limited detail about baselines, model scale, and robustness behind the unusually strong headline result.
-
Dexterous Robot Manipulation from Human Demonstrations via Contact-Anchored Retargeting and Residual Policy Learning
This paper converts human motion-capture demonstrations into dexterous robot manipulation policies by preserving the demonstrated finger–object contact structure rather than copying human joint motion. It uses simulated physics refinement to infer contacts and forces, contact-anchored retargeting across different hand morphologies, and a shared residual RL policy to repair kinematic and dynamic inconsistencies. The method reports large success-rate improvements across 25,454 single-hand and 25 dual-hand trajectories, transfers to four robot hands, and performs four bimanual tasks on hardware without real-robot training data.
The morphology-independent contact-anchored retargeting plus shared residual repair is a genuinely interesting approach, and the cross-hand transfer and zero-real-data hardware results are substantial, but the abstract does not establish how much comes from the method versus simulation assumptions or provide enough comparative detail for a strong verdict.
-
DexTacWAM: A Visuo-Tactile World-Action Model for Dexterous Manipulation
DexTacWAM extends vision-based world-action models with independently encoded fingertip tactile signals for bimanual dexterous manipulation. It predicts future visual and contact states jointly, substantially outperforming the strongest baseline across six contact-rich tasks (70.6 vs. 38.0), while using about 100 demonstrations per task for adapting a pretrained video model and preserving visual prediction quality. The ablations suggest that modeling the evolution of contact—not merely conditioning actions on tactile input—is the main source of the improvement.
The joint visuo-tactile world-modeling formulation and large reported manipulation gains are genuinely interesting, but the evidence is limited to one 22-DoF platform and six tasks, so it is not yet clearly a broadly validated breakthrough.
-
GINIO: A Geometric SO(3)-Equivariant Interface for Neural Inertial Odometry
GINIO is a geometric interface for learned inertial-odometry measurements that enforces the correct SO(3) transformation laws for both predicted motion and uncertainty under arbitrary IMU mounting rotations. Its Last-Frame Alignment preprocessing is claimed to be equivalent to world-frame training for equivariant models, while experiments across several odometry architectures report large gains, including a reduction in unseen remounting error on Fetch from 8.15 m to 0.50 m without retraining and substantially lower compute than a competing equivariant method.
The combination of tensorially correct uncertainty, a simple provably equivalent alignment procedure, and strong zero-retraining robustness to physical sensor remounting is notably more than an incremental odometry tweak, though the evidence is still limited to the abstract’s reported benchmarks.
-
From Decorative to Load-Bearing: Task Difficulty Shapes the Causal Role of Chain-of-Thought
The paper proposes continuation-based causal testing for measuring whether a model’s written chain-of-thought actually constrains its answer: corrupt or truncate a reasoning prefix, then force the model to continue. Across three language models and several reasoning benchmarks, easy tasks often show silent bypass of corrupted reasoning, while difficult tasks more often propagate errors from the corrupted step; task difficulty explains far more variation than the type of perturbation. Reasoning-focused RL reduces error propagation, but hidden-state probes and activation steering suggest that these behavioral modes are easier to detect than to control.
The causal continuation test and the strong task-difficulty gradient challenge the assumption that readable CoT is uniformly load-bearing, though the conclusions remain behavioral and are most directly relevant to CoT oversight rather than general reasoning capability.
-
The Limits of Simulated Societies: How Post-Training and Survey Fine-Tuning Erase Cross-Cultural Variance
This paper evaluates whether LLMs reproduce not just average survey answers but the diversity of opinions within different populations. Across World Values Survey data and several base, instruction-tuned, DPO, and GRPO models, it finds that post-training sharply compresses response variation, with especially severe loss for non-WEIRD populations such as Nigeria; higher sampling temperature and distribution-shaped rewards do not reliably fix the problem. The authors call this failure mode “consensus collapse” and propose measuring dispersion retention alongside point accuracy.
The combination of a concrete dispersion metric and evidence that alignment improves point accuracy while erasing culturally specific variance is a non-obvious and potentially important critique of LLM-based social simulation, though the conclusions are based on one survey source and a limited model set.
-
Statistical Gains from Looped Estimation under Parameter Budgets
The paper studies parameter sharing through looping: a single learned update is repeatedly applied, so model depth and computation can grow without increasing the number of distinct trainable parameters. It gives approximation and statistical-risk bounds for looped neural networks and Transformers, claiming that with sufficiently large fixed parameter budgets, increasing the number of iterations can drive worst-case risk toward zero while optimally sized untied models remain bounded away from zero; the results cover several likelihood and regression settings.
The theoretically broad claim that recurrent parameter sharing can yield vanishing worst-case risk at fixed parameter count is potentially important and challenges parameter-count comparisons, but its practical significance depends on computation, numerical precision, and the exact model-class assumptions behind the asymptotic results.
-
You Should Be Properly Scoring Your Odometry
The paper argues that odometry should be evaluated using proper scoring rules that combine trajectory error with the estimator’s reported covariance, rather than using position errors alone. Applying this framework to four LiDAR–inertial filters, the authors find severe and systematic overconfidence—including cases reporting centimeter-level uncertainty despite kilometer-scale errors—and trace it to filters overestimating the independent information provided by LiDAR measurements. They also introduce a one-sided pairwise test that can reveal overconfidence between estimators without ground truth, implemented in the open-source smfeval framework.
The combination of proper scoring for odometry, a ground-truth-free consistency test, and a striking diagnosis of widespread catastrophic overconfidence is more than an incremental metric change, though the underlying scoring-rule idea is established and the abstract gives limited detail on validation breadth.
-
Sample, Simulate, Select: Physics-in-the-Loop Text-to-Motion for Humanoids Without Training
The paper uses a frozen text-to-motion model to generate several candidate motions, retargets them to a Unitree G1, simulates each with the deployment tracking controller and full rigid-body dynamics, then selects the motion that executes best—without additional training. On 4,184 prompts, this raises upright execution from 80.5% to 89.5%; hardware-gate passes increase from 33 to 85 in a 200-prompt subset, and all 177 selected clips complete on the real robot with simulation-tracking correlation of 0.94. The results also show that simple fall prediction is insufficient for ranking candidates, while low-pelvis motions remain a systematic failure mode.
The notable contribution is a compelling training-free, physics-in-the-loop selection strategy with substantial simulated and real-robot gains, though it remains a constrained inference-time wrapper around existing models and controllers rather than a broadly demonstrated new capability.
-
Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents
The paper proposes Growing Harness, which starts with a controller-free agent scaffold and incrementally turns repeated task failures into reusable executable control code. Function-level traces identify where failures occur, an optimizer jointly repairs those regions, and a success-first gate prevents regressions. On BrowseComp-Plus and WebArena-Verified, it reportedly matches or nearly matches the best results while reducing LLM calls by 76–91.8% and inference cost by 74.4–98.6%; notably, WebArena performance stays around 45% across model sizes while a tool-calling baseline drops to 6.7% with a 4B model.
Learning a persistent executable harness from failures, rather than repeatedly reconstructing control in context, is a substantial and non-obvious agent-design direction with strong reported cost and small-model robustness results, though the evidence is limited to two web-agent benchmarks and needs scrutiny of task specialization and comparison fairness.
-
On Preference Coverage Collapse from Hindsight Relabeling in Multi-Objective Reinforcement Learning
The paper studies hindsight relabeling for preference-conditioned multi-objective reinforcement learning and finds that relabeling transitions with the preferences the agent actually achieved often causes severe performance degradation. Across 36 algorithm–environment settings, it identifies a new failure mode—Preference Coverage Collapse—in which the critic loses coverage of much of the preference space, and proposes `her_mix`, which largely restores coverage and performance with one fixed mixing parameter.
This is a non-obvious, broadly tested failure analysis that challenges a natural extension of HER and offers a simple cross-setting mitigation, though it is primarily an empirical MORL result rather than a major capability advance.
-
Change detection with conformal martingales: new optimal constructions, and suboptimality of existing methods
This paper gives a distribution-free theory for conformal martingale methods in sequential changepoint detection when both the pre- and post-change distributions are unknown. It proves that commonly used constructions can have delays as large as Ω(T) under false-alarm control and Ω(√ARL) under average-run-length control, then introduces conformal e-processes and e-detectors with minimax-optimal logarithmic delays, supported by simulations.
The paper identifies a substantial, theoretically demonstrated efficiency failure in existing conformal changepoint methods and replaces them with provably optimal constructions, though the advance is specialized to sequential statistical detection rather than a broad AI capability.
-
What Do Tabular Foundation Models Compute In Context? In-Situ Representation Refinement through Attention-Gated Updates
The paper argues that tabular in-context learning works partly by iteratively refining the internal representations of support examples, rather than merely reading them with attention. It turns this idea into RefineICL, an attention-gated, FFN-free model with low-rank feature interactions and typed memory; the model reports higher AMLB29 and TabArena performance than TabPFN-3, while using less inference memory, and an intervention across 72 episodes supports the role of intermediate support-state updates.
The combination of a mechanistic account of tabular ICL, a direct intervention, and a competitive FFN-free architecture is notably more substantive than a routine tabular-model tweak, though the evidence is still limited to reported benchmarks and may reflect benchmark-informed continuation.
-
Exact Bayes Regret and Asymptotic Optimality in High-Dimensional Gaussian Bandits
This paper derives exact proportional-dimension Bayes-regret curves for Gaussian linear bandits when the horizon and dimension grow together. It proves that posterior-mean greedy selection is asymptotically Bayes-optimal, while Thompson sampling has strictly higher leading regret—approaching roughly twice greedy’s instantaneous regret at long horizons—and characterizes the limiting decision-loss distribution.
The uniform policy-level asymptotics and result that greedy beats Thompson sampling in this high-dimensional regime are a substantive, counterintuitive theoretical contribution, though its assumptions and primarily asymptotic scope limit practical impact.
-
Where Does Exactly-Once Live? Model, Harness, and Tool-Contract Effects on Duplicate Side Effects in LLM Agents
The paper studies duplicate side effects when LLM agents retry tool calls whose outcomes are uncertain. In a deterministic sandbox covering six services, multiple models, harnesses, and fault modes, it finds that models can handle lost acknowledgements when a read-back is possible, but duplicate actions are common when writes may still be in flight or may have been redelivered; explicit idempotency keys are much more reliable than model or harness behavior. It also proves that verification alone cannot guarantee exactly-once execution under unbounded late commits.
The model-versus-harness-versus-tool-contract decomposition, combined with broad fault-injection results and a limitation theorem for verification-only policies, gives a useful non-obvious design rule: exactly-once behavior generally belongs in tool contracts, especially through idempotency keys, rather than being left to the agent.
-
Policy as Code: A Coroutine-Bridge Harness for Fast-Reasoning Reliability on CAR-bench
The paper describes an agent harness where the model emits a Python coroutine that pauses for tool results and resumes, reducing dependent tool-use tasks from one model call per exchange to a median of two calls. On CAR-bench it achieved 60.0% Pass^3, 4.5× the organizer baseline, with low latency and cost, and reproduced the score on GPT-5.5; benchmark-specific policies are implemented directly in the tool layer rather than via prompts.
The coroutine-based separation of model calls from tool round-trips and executable policy layer are genuinely interesting systems ideas with strong reported benchmark results, but the evidence is concentrated on one challenge and may depend heavily on benchmark-specific policy engineering.
-
Reasoning Instructions Can Break Answer Decoding in Vision--Language Models
The paper shows that a common way of evaluating vision-language models with chain-of-thought prompts—reading answer-label probabilities immediately after adding a reasoning cue—can severely underestimate accuracy. On ScienceQA, Qwen2.5-VL-7B falls from 80.76% to 45.48%, while label permutations, probes, and free generation indicate that the answer remains encoded in the hidden state but is not yet expressed as an answer token.
This identifies a concrete and potentially widespread evaluation-interface failure, supported by large accuracy changes and diagnostics, though it is mainly a methodological correction rather than a new model capability.
-
The Alignment Illusion in Multimodal Large Language Models
The paper shows that common layer-wise visual–text similarity metrics can remain high even after the visual input to a multimodal language model is replaced with noise and task performance collapses. Across 13 models, it attributes this “alignment illusion” to shared language-model projections that create largely one-dimensional, weight-induced similarity, and proposes a principal-angle gap that better tracks visual corruption and task accuracy.
It challenges a widely used interpretation of multimodal representation-alignment analyses with broad cross-model interventions and a plausible diagnostic replacement, but the abstract provides no detailed quantitative results establishing how much the new metric improves understanding or prediction.
-
Audit Before You Commit: Locating Belief Failures in Active Identification for One-Shot Manipulation
This paper proposes auditing two separate prerequisites before a robot commits to an irreversible action: whether its belief still covers the true action-relevant state, and whether its predicted failure model matches realized failures. In insertion experiments across seven task families, simulations, and a physical robot, the audit exposed a small tap-boundary model error whose correction reduced failure from 0.354 to 0.112 and transferred to another engine, while identifying a different execution-model problem in a third.
The useful new contribution is a concrete failure-localization framework showing that sharper beliefs and conformal calibration can still produce confidently wrong commitments, supported by cross-task, cross-engine, and physical-robot evidence.
-
Adapting Personalized Speech Enhancement for Low-Latency Audio-Visual Target-Speaker Extraction
The paper adapts a personalized speech-enhancement model for online audio-visual target-speaker extraction by adding mouth-region features to speaker conditioning and jointly fine-tuning the visual and reconstruction modules. This reduces target-confusion errors from 46% to 1.6% on two-speaker mixtures, operates with no lookahead and 20 ms delay, and improves listening quality over an online autoregressive extractor on recorded meetings while preserving the target when alone.
The combination of personalized enhancement with visual mouth conditioning produces a large, practically relevant reduction in target confusion and is validated with low-latency tests and meeting-corpus listening studies, though it remains a focused system advance rather than an obviously field-defining new direction.
-
Causal Retention in Interactive Agents: Interface Factorization and Selective Adaptation
This paper introduces “causal retention,” asking whether an agent’s frozen internal state preserves the mechanism needed to answer a fixed intervention probe after the environment changes, rather than merely retaining task performance or source-domain information. It develops conditions for when retention is possible and proposes Causal Core, which selectively writes and reads mechanism-relevant information; experiments report large gains on changed delays and actuator effect-sign recovery, including 0.948 accuracy from only five target states per actuator.
The distinction between task sufficiency, source-domain decodability, and retention of causal mechanisms is a genuinely useful framing, supported by theory and cross-domain experiments, but the practical generality and independence of the proposed mechanism state remain uncertain.
-
Causeway: Restoring Task Accessibility for Instruction Switching in VLA Policies
Causeway addresses a failure mode in vision-language-action policies: a new instruction can become impossible after a previous task leaves the robot in an unusual physical state. At inference time, it back-propagates through the frozen action decoder to modify an internal action-stream representation toward a specified re-entry pose, letting the VLA generate its own recovery motion without retraining or a separate controller. Across multiple models, switch timings, task pairs, simulation settings, and a real xArm, it raises switching success from 3–26% to 47–65%.
This is a nontrivial, training-free way to restore task accessibility during instruction switching, with large gains across architectures and some real-robot validation, but the dependence on a supplied re-entry pose and the limited benchmark scope keep it below strong.
-
Co-design of trajectory and morphology for a vertical jump-climbing robot
The authors jointly optimize a robot’s body morphology and motion trajectory to enable dynamic vertical climbing rather than conservative contact-based gaiting. Their 290 g robot launches upward, reorients during an aerial phase using its rear leg as an inertial tail, and continuously climbs at 0.375 m/s while using spined grippers for adhesion; it can also move on the ground and transition onto vertical surfaces.
This demonstrates a genuinely unusual climbing capability—a bounding gait with an aerial phase—and combines morphology and trajectory co-design, but the evidence is from a specialized small prototype and does not yet establish broad robustness or generality.
-
Stale-Document Poisoning: When Outdated Retrieval Overrides Correct Model Answers
This paper studies cases where retrieved documents are outdated and cause a model to change a correct answer into an incorrect one. On 317 verified knowledge reversals across medicine, law, software, and platform policy, stale evidence flipped 30–37% of answers in two model families without explicit trust instructions, and 66–75% with such instructions; up-to-date evidence was followed in 97–100% of trials. The authors show that telling models when old evidence stopped applying largely fixes the problem, while a date-aware reranker gives smaller gains.
It identifies and measures a consequential RAG failure mode that challenges the assumption that external retrieval reliably corrects outdated model knowledge, with cross-domain evidence and a plausible temporal-applicability intervention, but the proposed mitigation remains modest and the result is not yet a demonstrated broad solution.
-
Recipe-Matching, Not Equivalence
The paper argues that MathNet-Retrieve rewards reproducing the benchmark’s data-generation recipe rather than genuinely finding equivalent math problems. In controlled experiments, training on LLM-generated examples and benchmark-specific near-miss negatives produces large gains on the benchmark, but much smaller or negative gains on generator-free duplicates and cross-language tests; similar effects appear, more weakly, on other retrieval benchmarks. The authors release generator-free evaluations, near-miss tests, and trained models to measure this artifact.
This is a well-controlled and non-obvious demonstration that a retrieval benchmark can measure recipe matching and even trade off against real generalization, though the evidence is concentrated in a few benchmarks and the practical scope is not yet clear.
-
Reasoning Concentrates Errors, and Self-Consistency Never Notices
This paper studies whether sampling multiple reasoning traces makes self-consistency more reliable. Across 74,944 samples, it reports that reasoning makes wrong answers more concentrated and less diverse, so independently sampled errors agree more often; across 280 combinations of models, datasets, and confidence-weighting methods, none improves on ordinary majority voting after correction. The paper argues that confidence signals should be judged by whether they improve decisions, not merely by whether they discriminate correct from incorrect answers.
The large-scale finding that reasoning can systematically concentrate correlated errors and undermine self-consistency challenges a common assumption, but the abstract omits the specific models, tasks, and absolute performance effects needed to judge how general and consequential it is.
-
Depth Laws for the Precision Floor of Trained Neural Networks: Amplification, Residual Scaling, and a Quantization-Aware Training Paradox
This paper develops a theory for when numerical noise or quantization causes a trained network’s accuracy to collapse, linking the precision threshold to a predictive-amplification quantity that grows with depth. Across MLPs, CNNs, ViTs, and pretrained language models, it reports that residual scaling and pre-normalization can remove the usual depth penalty, while quantization-aware training helps shallow models more than deep ones and can actually steepen the depth dependence.
The proposed depth law and the counterintuitive finding that QAT’s relative benefit worsens with depth are potentially useful and broadly tested, but the abstract does not establish whether these laws generalize beyond the reported architectures and quantization setups.
-
Toward On-Chip Training of Spiking Neural Networks for Dense Event-Based Vision
The paper introduces DELL, a local-learning method for training fully spiking networks on dense event-camera tasks without propagating gradients through the entire network and time sequence. On DSEC optical flow, it cuts peak training memory by 39.6% and improves endpoint error from 1.941 to 1.670 pixels versus end-to-end BPTT, while also producing competitive semantic segmentation results; learnable local heads are key to making local supervision work for dense prediction.
This is a substantive extension of local SNN learning from classification to dense optical flow and segmentation, with a meaningful memory reduction and an unexpected accuracy improvement over BPTT, but the evidence is confined to a narrow event-vision benchmark and segmentation still trails end-to-end training.
-
How to Reduce Whisper Hallucination
This paper fine-tunes Whisper using hallucinated phrases in two contrasting forms: as text to suppress when there is no speech, and as genuine speech to transcribe when those phrases are actually spoken. Across 33 matched fine-tuning comparisons and multilingual tests, this reduces emissions on silence while improving recovery of real spoken phrases; the best model lowers hallucinations on room tone from 61.9% to 2.4% and raises phrase recovery from 69.8% to 82.7%.
The paper reports a simple but non-obvious way to resolve the suppression–recall tradeoff, with unusually broad matched-pair evidence and large practical gains, though the method remains a targeted fine-tuning strategy rather than a clearly new model direction.
-
Retrieved but Not Delivered: Multimodal Memory Delivery for Long-Term Agents
This paper separates multimodal-agent memory into retrieval and a previously under-measured “delivery” stage: deciding which retrieved content reaches the model and how it is represented. On MemLens, restoring original visual content improves accuracy by 13.87 points with a fixed evidence set, while perfecting retrieval adds only 2.31 points; the proposed DeliverMem also improves results across two benchmarks and uses far less input without training or changing stored memories.
The controlled finding that post-retrieval delivery matters substantially more than retrieval itself is a non-obvious and potentially useful reframing, but the evidence is limited to two benchmarks and the abstract does not establish broad generality beyond this evaluation setup.
-
GAUGE: Group-Wise View-Inconsistency Rectification for Feed-Forward 4D Tracking
GAUGE is a training-free post-processing method for monocular feed-forward 4D point trackers. It identifies motion groups and corrects their per-frame radial scale and translation using a small number of metric anchors; across eight trackers, it reports 15.1–62.6% lower endpoint error, substantially outperforming using the same anchors for fine-tuning.
The paper offers a potentially general geometric explanation for residual tracking error and a model-agnostic correction with large reported gains, but the method depends on external metric anchors and the abstract does not establish robustness across anchor placement, motion complexity, or evaluation settings.
-
Does CoT-Pass@k Really Check the CoT? A Multilingual Mathematical Audit
The paper audits CoT-Pass@k, a metric intended to distinguish correct answers reached through sound reasoning from lucky guesses. Across English, Turkish, and Portuguese math tasks, it corrupts reasoning chains and answers separately and finds that LLM judges usually accept flawed chains, with verdicts driven mainly by whether the chain agrees with the final answer. As a result, the metric’s gap from ordinary Pass@k can be small, unstable, and highly dependent on sampling and token budgets.
This is a useful and somewhat assumption-challenging audit showing that a prominent reasoning metric may not actually verify reasoning, though it is still an evaluation study rather than a new capability or method.
-
Staying on the Attractor: Supervising Neural Surrogates of 3D Turbulence Where They Leave It
The paper trains neural surrogates for 3D turbulence not only on normal simulation states, but also on states where the surrogate itself is likely to fail. It selects such states through rollout failures, targeted surrogate attacks, and dynamical perturbations, then uses offline DNS to provide corrective targets. On a 128^3 turbulence task, this increases median time to failure from 21 to 721 steps and improves long-horizon statistics over the tested baselines.
The off-attractor supervision framing is a substantive and broadly relevant approach to distribution shift in learned dynamical systems, with a large reported stability gain, but the evidence is currently limited to one turbulence setting and abstract-level comparisons.
-
Constraints Are Graphs, Not Chains: Exact Decoding for Diffusion Language Models
The paper introduces FactorDLM, a training-free constrained decoder for diffusion language models that represents constraints as factor graphs rather than sequential automata. It uses variable elimination to enforce relational constraints alongside syntax, with complexity governed by graph induced width; across nine benchmarks and three backbones, it reports 100% constraint satisfaction with 0.4–6.9% projection overhead and faster repeated-query solving than CP-SAT.
The factor-graph formulation is a genuinely nonstandard alternative to automaton-based decoding and appears to enable exact nonlocal constraints with practical overhead, but the advance is specialized to constrained diffusion decoding and the abstract does not establish broad impact beyond the reported benchmarks.
-
Train Together or Merge Later? Unifying VLA Experts via a Shared Action Interface
PolicyWeave trains separate vision-language-action experts while forcing them to keep a shared action interface, then merges only the task-relevant LoRA updates selected from the visual-language context. On GR00T N1.5 across 18 RoboCasa365 tasks, preserving the interface improves average success for standard merging from 17.0% to 52.8%, while PolicyWeave reaches 64.7% with limited SFT and 74.1% after task-specific RL, with supporting results on LIBERO-10 and a real Piper arm.
The combination of a shared action interface and context-dependent sparse expert merging addresses a concrete failure mode in multi-task VLA policies and shows a large empirical gain, though the evidence is still centered on a few manipulation benchmarks and one robot platform.
-
The Statistical Benefits of Multiple Responses for Learning from Demonstrations
This paper studies learning from demonstrations when each prompt produces multiple candidate responses but the demonstrator may be imperfect. In a finite reward-class setting, it proves that using any k≥2 candidates changes the accuracy dependence from 1/ε² to 1/ε, and under standard evaluation also improves the dependence on a reward class of size N from log N to log N/log k. Matching bounds and a greedy multiplicative-weights algorithm show these gains do not require assumptions about demonstrator quality, while robust evaluation retains the accuracy gain but may lose the log-k benefit.
The claimed change from quadratic to linear accuracy dependence from merely allowing multiple responses is a substantial and somewhat surprising theoretical result, but its impact is currently limited by the abstract’s finite reward-class model and lack of demonstrated practical relevance.
-
Let CSP Be Your ANCHOR: Adaptive Crystal Search over Frozen Structure Priors
ANCHOR separates crystal discovery into two stages: a composition-search policy optimized with multi-objective rewards, and a frozen crystal-structure-prediction model that generates and evaluates structures. Its adaptive novelty reward and GRPO policy substantially improve unique stable-material discovery on the reported evaluator, raising MSUN from 11.4% to 47.6% and transferring across multiple CSP backbones, while direct fine-tuning of joint generation models performs poorly. The gains remain dependent on the underlying stability potential, so the key contribution is the search/prior decomposition rather than an evaluator-independent breakthrough in crystal prediction.
The separation of composition exploration from a frozen physical structure prior, supported by large cross-backbone gains and an analysis of why direct reward fine-tuning is limited, is a genuinely non-obvious and potentially reusable approach, though the results are tied to the evaluation potential and need independent validation.
-
FoldAttention: Declared-Reference Softmax for Fast Decode and Deterministic Backward
FoldAttention replaces online softmax’s dynamically changing normalization reference with a fixed, predeclared reference, making attention contributions final as they are computed. This enables Hopper GPU kernels to skip sufficiently small key/value reads, combine split-KV partials without rescaling, and perform deterministic gradient reductions; reported H100 speedups reach 1.36–2.30× on real-model decode, with up to 1.46× end-to-end Qwen3-8B decode improvement and comparable likelihood/accuracy.
The fixed-reference additive formulation is a non-obvious systems/algorithmic reframing with substantial reported decode and deterministic-backward gains, but the evidence is hardware-specific and the claims need closer scrutiny around numerical accuracy, cutoff behavior, and broader adoption.
-
EfficientAgent: What Makes KV Cache Offloading Work for Concurrent Agents?
This paper argues that KV-cache offloading for concurrent agents only helps when host memory can retain the reusable context of the whole agent pool, which it calls the reuse working set. It introduces a stack-distance model to size that tier and a policy that avoids writing large cache refills when the tier is too small; on SWE-bench Verified, sizing to the predicted working set reduced recomputed prompt tokens by 93% and end-to-end time by 39%, with results across three GPUs and two models.
The reuse-working-set framing and predictive capacity-sizing result offer a non-obvious systems insight for agent serving, supported by substantial cross-hardware and cross-model experiments, though the advance is specialized to inference infrastructure rather than a broad capability jump.
-
Annealed Sinkhorn with Momentum: Certified Unregularized Optimal Transport in Linear Memory
The paper reformulates unregularized discrete optimal transport using a Bregman Douglas–Rachford/Sinkhorn-style iteration that implicitly anneals the regularization temperature, adds momentum, and avoids explicitly storing the quadratic-size transport plan. It also provides a linear-memory primal–dual stopping certificate. On image color-transfer tasks, the method reaches million-pixel problem sizes with reported relative duality gaps around 1.6–2.8%, and is substantially faster than a comparison method on a 1024×1024 example.
The combination of an exact-OT interpretation, momentum-based annealing, and certified linear-memory computation enables unusually large transport problems, though the evidence is mainly on color-transfer benchmarks and does not yet establish broad superiority or practical exactness.
-
Lost with a Map: Conversational State and Behavioral Reliability in Language Models
The paper probes how instruction-tuned language models represent and use conversational state in task-oriented dialogue. Across eight models, it finds that structural information—active domains, slots, and requests—is most readable near the decision point, while exact slot values remain strongest near where they were mentioned; failures also arise when the model has the right constraint but does not apply it. A lightweight controller that reads structural signals and selectively edits the model’s action substantially improves MultiWOZ exact-query accuracy from 0.318 to 0.621, though task success rises more modestly from 0.272 to 0.371.
The separation of state structure, value resolution, and action deployment—validated with causal interventions and used to improve closed-loop behavior—is a meaningful departure from treating belief-state extraction as the sole route to reliable dialogue, but the evidence is still limited to standard benchmarks and modest end-task gains.
-
Quantization Error Is Spectrally Flat: A Single Random Probe Is a Calibrated, Data-Free Sensitivity Estimator, with Application to Budget-Targeted Mixed-Precision Quantization
The paper proposes using a single Gaussian random probe to estimate each tensor’s quantization error without calibration data, arguing that round-to-nearest error is close to spectrally flat. It propagates probes through the network to score candidate bit-widths and uses a knapsack solver for exact-budget mixed-precision allocation. Across models from 8B to 122B parameters, the method reportedly improves WikiText-2 perplexity over uniform 4-bit quantization by 3.5–13.6% on tested MoE models, while requiring only minutes of probe computation.
The combination of randomized, data-free propagated sensitivity estimation and budget-targeted quantization is a meaningful alternative to activation-calibration methods, with unusually broad scaling and nontrivial reported gains, but the abstract alone does not establish whether the spectral-flatness claim and comparisons will hold beyond the tested models and objectives.
-
Diffusion MRI Tractography Predicts Electrophysiological Connectivity and Explains Spectral Signatures of Evoked Potentials in the Human Brain
The study combines diffusion MRI with repeated electrical stimulation and SEEG recordings from 40 neurosurgical patients to relate structural pathways to causal electrophysiological communication across the human brain. Across roughly 150,000 electrode pairs, structural connections strongly predicted direct electrophysiological connectivity, while distinct time-frequency signatures separated direct from indirect propagation; thalamic involvement was associated with especially delayed and oscillatory responses.
The large-scale multimodal comparison and proposed spectral markers of direct versus polysynaptic signaling are genuinely interesting, but the abstract does not establish how well these findings generalize beyond implanted clinical cohorts or how robustly the connectivity predictions were validated.
-
Localized SARS-CoV-2 Infection Triggers a Tissue-Wide Antiviral Response and Functional Impairment of Olfactory Sensory Neurons
In a mouse model, the authors compared SARS-CoV-2 variants and found that Alpha and Beta infected olfactory epithelial support cells much more readily than Omicron, despite similar lung viral loads. Infection of these cells produced an interferon response across the olfactory epithelium, followed by persistent suppression of sensory-neuron odor-signaling and mitochondrial programs; odor stimulation also failed to induce normal neuronal activity markers, without widespread neuron death.
The combination of variant comparison, spatial and single-cell profiling, and functional odor assays supports a non-obvious mechanism in which localized support-cell infection causes prolonged, tissue-wide olfactory-neuron dysfunction without neuronal loss, though the evidence is limited to a mouse model and the abstract gives few quantitative results.
-
Striatal interneuron microcircuits gate reinforcement to stabilize adaptive choice
The study identifies a striatal inhibitory microcircuit that converts reward and punishment signals into longer-lasting changes in action policy. In mice performing probabilistic reversal learning, tyrosine hydroxylase interneurons disinhibited projection neurons and supported reinforcement, while somatostatin interneuron inhibition caused persistent, maladaptive choices accompanied by strengthened excitatory synapses onto projection neurons.
Causal evidence links outcome-responsive interneuron types to persistent adaptive or maladaptive choice policies, offering a non-obvious circuit-level mechanism for how single-trial reinforcement becomes longer-term behavioral change, though the abstract does not establish broad generality beyond this task and circuit.
-
The nanoscopic submyelin space mediates efficient K+ uptake and shapes nodal action potentials
The authors build a nanoscale electrodiffusion model of a myelinated node of Ranvier, including the narrow space beneath the myelin and Kir4.1 potassium channels. The model predicts that the action potential reverses polarity at the inner myelin membrane, causing Kir4.1 channels to rapidly buffer periaxonal K+ and thereby support high-frequency conduction. This proposes an electrical and ionic feedback role for juxtaparanodal channels beyond their conventional view as largely latent conductance.
The proposed K+ prebuffering mechanism and predicted action-potential inversion are highly non-obvious and potentially important for understanding myelinated axon computation, but the evidence is currently based on a model without apparent experimental validation or quantitative robustness results.
-
Apparent food selectivity reflects multiple non-food image properties
The authors tested whether an apparent food-selective region in the ventral visual pathway truly responds specifically to food, rather than to correlated image properties such as color, scene context, or material appearance. Using preregistered fMRI experiments with carefully matched food and non-food stimuli, they found that the response was not specifically driven by food when images were visually matched, and that gooey non-food objects could evoke similarly strong responses. The region is therefore better described as food-biased and partly sensitive to contextual or material features, challenging conclusions from prior natural-image studies.
This is a well-controlled, potentially assumption-changing re-evaluation of a claimed category-selective brain response, with evidence that natural-image confounds can substantially alter conclusions about neural selectivity.
-
An accumbal cholinergic ISR-HCN2 axis sustains maladaptive cue motivation
This study identifies nucleus accumbens cholinergic interneurons as a causal driver of persistent cue-induced seeking after cocaine and high-fat-reward experience. The authors link this behavior to an integrated stress response that increases HCN2 expression and pacemaker excitability in these neurons; suppressing the pathway, knocking down HCN2, or blocking HCN channels with ivabradine reduced maladaptive seeking while sparing sucrose seeking.
The causal ISR–HCN2 mechanism and pharmacological rescue across drug and non-drug reward paradigms are a meaningful, somewhat unexpected advance, but the evidence is still preclinical and the abstract does not establish whether this mechanism generalizes beyond the tested models.
-
Breaking the link between demyelination and axon loss: SARM1 inhibition as a neuroprotective strategy for multiple sclerosis
This study tests whether blocking SARM1, a molecular executor of axon degeneration, can protect neurons in mouse models of optic nerve injury and multiple sclerosis-like disease. Genetic or drug-based SARM1 inhibition preserved retinal ganglion cells, axons, visual electrophysiology, and reduced serum neurofilament, even though demyelination and broad CNS inflammation continued. The key finding is that axonal self-destruction can be uncoupled from demyelination in EAE, suggesting SARM1 inhibition as a neuroprotective add-on to immune and remyelinating treatments.
The mechanistic separation of axon loss from demyelination and neuroinflammation, supported by genetic, cell-restricted, pharmacological, structural, and functional experiments, is a meaningful preclinical result, though it remains limited to mouse models and does not yet establish therapeutic relevance in human MS.
-
Lifespan EEG reference charts disentangle alpha-band alterations in neurological and neuropsychiatric disorders
The authors build age- and lobe-specific EEG reference charts for alpha peak frequency and power using 22,094 people aged 1–100, separating oscillatory activity from the aperiodic 1/f background. In inpatient groups with neurological and psychiatric disorders, alpha power was reduced, but alpha peak frequency appeared slowed in raw spectra and accelerated after 1/f adjustment—suggesting that some reported alpha slowing may actually reflect changes in the aperiodic component.
The large lifespan normative dataset and reversal of the apparent alpha-frequency effect after 1/f correction are potentially important, but the abstract gives limited quantitative or independent validation evidence and the work is currently a preprint.
-
Complex KDA: Understanding and Enhancing the Expressivity of Kimi Delta Attention
The paper extends Kimi Delta Attention by allowing its channel gates to range from -1 to 1 and its delta coefficient to range from 0 to 2. This lets a single diagonal-plus-rank-one recurrent transition represent 2D rotations and, more generally, every orthogonal transition of that form, matching some state-tracking capabilities that otherwise require composing multiple delta-rule updates. Theoretical results and experiments show better length extrapolation on permutation-group tracking and periodic audio, while language-modeling performance is roughly comparable to KDA rather than clearly superior.
The gate-as-reflection construction provides a nontrivial expressivity improvement for efficient linear RNNs, supported by formal characterization and extrapolation experiments, though the practical language-modeling gains over KDA appear limited.
-
StudentBench: AI and human tutoring yield equivalent GRE learning gains
StudentBench evaluates AI tutoring against expert human tutoring and no tutoring on GRE Quantitative and Verbal questions, using 2,383 participants and over 175,000 student–AI messages. The authors report statistically equivalent learning gains between the best AI and human tutors, with AI outperforming human tutoring in five of seven domains and costing far less; they also analyze lesson plans, generated problems, pedagogy, engagement, and response speed.
The large controlled comparison and reported human-level learning gains at dramatically lower cost are notable, but the evidence is limited to short-term GRE learning and the abstract does not establish that the result generalizes to broader education or durable knowledge.
-
PUBG Ally: A Conversational Embodied Agent as an AI Teammate
PUBG Ally is a voice-enabled AI teammate that combines a language-model agent for perception, dialogue, memory, and high-level decisions with a faster control layer for gameplay. The team trained and refined it using nearly 39,000 real-player sessions, then deployed it in live service; surveyed players in 141 countries gave recommendations 25.1 percentage points more positive than negative. The notable contribution is an end-to-end, low-latency conversational agent operating with real players in a continuously changing game rather than a static benchmark.
The large-scale live deployment and integration of speech, tool use, memory, and real-time action are worth examining, but the abstract provides mostly engineering and subjective-preference evidence rather than a clear capability breakthrough or rigorous comparison.
-
D-JEPA: A Decision-Aligned Latent World Model
D-JEPA modifies latent world models so their geometry reflects which candidate action is more likely to succeed, not just which predicted future is closest in latent space. It learns ordinal, decision-relevant relations among competing futures from executed outcomes and reports gains across simulated control, manipulation, pretrained action models, physical robots, and autonomous driving, including 15.04 percentage points on RoboTwin and 17 points on physical-robot tasks.
The decision-local framing and ordinal alignment of pretrained predictive representations are a substantive control idea with broad reported evaluations, but the abstract does not establish how much the gains depend on task-specific supervision or how strong the comparisons are.
-
QwenGyre: An Elastic Reinforcement Learning Framework for Training xLong-Horizon Agents
QwenGyre is an online RL training system for agents whose rollouts last hundreds of interactions and hundreds of thousands of tokens. It dynamically shifts GPUs between rollout and optimization, and compresses branching trajectories by reconstructing histories, estimating partial progress, and removing redundant paths. On a reported 2.4T-parameter Qwen model with roughly 700K-token rollouts, it improves NL2RepoBench from 52.5% to 58.5% and reports up to 1.85× training speedups over baseline scheduling approaches.
The combination of elastic resource scheduling and trajectory deduplication targets a real bottleneck for long-horizon agent RL and shows a substantial reported gain, but the abstract provides limited evidence beyond one main benchmark and system-level speed comparisons.
-
TimeEvo: Failure-Driven Self-Evolution of a Time Series Agent
TimeEvo lets a time-series analysis agent improve its tool library from its own failures. It groups failures into capability gaps, creates evidence-based tools, and uses paired tests to admit only tools that help rather than silently degrading answers. Starting with no tools, it reportedly improves accuracy across ten tasks and three model backbones, and tools learned with a cheaper model transfer to stronger models.
The failure-driven, question-level tool admission process is a useful alternative to fixed tool libraries and generic self-revision, but the abstract gives no improvement magnitudes or details sufficient to establish a major advance.
-
Diffusion Reward Models
The paper models a response’s reward not as one scalar or a fixed-form distribution, but as a conditional distribution generated by a diffusion transformer. Its samples can represent multimodal judgments and support uncertainty-aware aggregation; experiments on five benchmarks and RLHF report competitive reward-modeling performance and improved downstream policies.
The diffusion-based conditional density formulation is a substantive departure from standard reward heads and could make uncertainty and disagreement useful in RLHF, but the abstract gives no quantitative gains or details establishing a large practical advantage.
-
FuseReg: Regularizing Layer Fusion Mitigates the Reconstruction-Generation Gap in Representation Autoencoders
FuseReg trains an RAE decoder on randomly sampled subsets of layers from a pretrained visual encoder, rather than committing to one fixed layer fusion. The resulting decoder can reconstruct from full, sparse, or single-layer representations without retraining, while also improving ImageNet-256 generation quality: the authors report 27–29% reductions in unguided gFID and gains across multiple encoder families.
The random-subset regularization addresses a real reconstruction-versus-generation trade-off and reports substantial relative gFID improvements, but it remains a targeted training modification and the abstract gives no absolute metrics or broader evidence of general impact.
-
The Undetected Damage of Quantization on Retrieval and How to Fix It
The paper shows that quantization can preserve classification accuracy while changing 14–46% of top-1 retrieval results, a problem that aggregate ranking metrics may hide. It links these failures to the score margin between the top two candidates, derives a sufficient condition for preserving the top result, and uses the margin to guide selective extra precision: layer-wise precision in retrieval and full-precision routing for uncertain classification inputs.
The focus on quantization-induced retrieval instability and the margin-based, label-free precision allocation are a meaningful and plausibly broadly useful insight, but the abstract provides limited detail about datasets, baselines, and the breadth of validation.
-
Zeva-Ego: Egocentric Mid-Training with In-Context Causal Learning for Robot Manipulation
Zeva-Ego uses large-scale egocentric human video to mid-train a vision-language-action robot policy, with an action-centric encoder translating human visual transitions into robot-relevant supervision. It also uses in-context causal learning to adapt at deployment from action-effect feedback without updating model parameters. On RoboTwin, 10K hours of egocentric video raises success from 63.8% to 75.3%, comparable to 2K hours of robot demonstrations, while interaction-based adaptation improves success from 58% to 89% over four attempts.
The combination of egocentric-video mid-training with parameter-free, feedback-driven adaptation is a meaningful direction, and the reported human-video-to-robot data ratio and rapid improvement are non-obvious, though the abstract gives limited evidence about task breadth, baselines, and generalization.
-
InsertAnything: Generalizable Contact-Rich Precision Insertion from Simulation to Reality
The paper trains a reinforcement-learning insertion policy entirely in simulation using target poses and compact fingertip force feedback, then deploys it on real robots without demonstrations or real-world fine-tuning. It reports 95% success across eight unseen insertion tasks, including parts with 0.02 mm nominal clearance, and a 20/20 result on a peg-in-hole benchmark; the main novelty is broad sim-to-real and cross-geometry reuse of a single contact-rich policy.
Directly deploying one simulation-trained policy across tight-clearance, unseen real insertion tasks is a meaningful robotics capability, but the abstract provides limited detail about hardware, task diversity, and whether the benchmark results generalize beyond the reported setup.
-
Qwen-Audio-3.1-Realtime: Towards Reliable Agentic Voice Interaction
Qwen-Audio-3.1-Realtime combines audio-capable reasoning, tool use, policy optimization, and speech/action coordination in a real-time voice assistant. Relative to its 3.0 predecessor, it reports task success improving from 78.4% to 82.0%, while unwanted responses to background speech on a full-duplex benchmark fall from 73.0% to 13.0%. A separate prototype adds persistent tasks, memory, and foreground/background coordination.
The integrated training and coordination approach, especially the large reported reduction in reacting to background speech, is potentially important for reliable full-duplex voice agents, but the evidence is mainly an internal comparison and self-defined benchmarks rather than broad independent validation.
-
AC-DC: Adaptive Communication for Scalable Dynamic Average Consensus in Multi-Robot Ergodic Search
The paper introduces AC-DC, a peer-to-peer dynamic average-consensus method that adaptively chooses communication partners, timing, and which consensus-state coordinates to transmit under range, bandwidth, and interference constraints. In simulated multi-robot ergodic-search experiments with up to 80 robots, it reduces covariance-trace error AUC by 27.5% versus decentralized baselines while using 8.7x less modeled communication, and its payload nearly matches an ideal centralized system at 120 robots. The main novelty is coordinating selective, asynchronous communication and state transmission rather than requiring full consensus rounds.
The communication-efficiency gains and scaling result are unusually large and relevant to multi-robot coordination, but the evidence is limited to simulated trials and the abstract does not establish robustness across real networks or physical robots.
-
Silk-templated Nanostrips as Superprotonic Fibre Sensors
The paper uses the ordered nanofibrils of natural silk as a template for growing continuous metal-organic-framework nanostrips on fibres. The resulting proton-conducting fibres reportedly reach conductivity up to 40 S/cm, respond within milliseconds, remain mechanically robust, and support respiratory monitoring and distributed humidity mapping in textiles, including high-flow oxygen devices.
The combination of bio-templated MOF nanostrips, unusually fast proton transport, and textile-scale spatial humidity sensing is a meaningful materials and sensing advance, but the abstract alone does not establish how broadly the large conductivity claim generalizes or whether the demonstrations outperform competing wearable sensors beyond selected tests.
-
From Instrument-Mounted Demonstrations to In-Vivo Execution: Learning Bimanual Laparoscopic Appendectomy Without Robot-Collected Demonstrations
The paper mounts inertial, time-of-flight, and Hall sensors on ordinary laparoscopic instruments to record surgeons’ bimanual motions, then uses those demonstrations to train a diffusion-policy surgical robot without robot-collected training data. After simulator-based design selection and retraining on 849 in-vivo demonstrations, the robot completed appendectomy in 3 of 4 live rabbits with electrosurgery enabled, under surgeon-selected phase control. The main new result is showing a practical path from instrument-mounted demonstrations to autonomous robotic execution in vivo.
It demonstrates a genuinely important real-world capability—training a surgical robot from non-robotic surgeon demonstrations and executing on live animals—but the evidence is still limited to one procedure and four deployment animals, with substantial surgeon supervision.
-
PhyVisGen: Physically and Visually High-Fidelity Robotic Manipulation Data Generation
PhyVisGen is a simulation pipeline for generating robotic manipulation demonstrations that aims to reduce both physical and visual sim-to-real gaps. It uses Incremental Potential Contact for coupled arm–soft-gripper simulation and combines reconstructed real scenes with real-time path tracing; policies trained only on the resulting synthetic data reportedly reach 65–95% success on five real-robot tasks without real demonstrations or fine-tuning.
The combination of high-fidelity soft-contact simulation and appearance-preserving rendering, together with fairly strong synthetic-only real-robot results, is notable, but the abstract lacks task details, baselines, and evidence that the approach generalizes beyond a small evaluation.
-
Real-Time Atomic-Resolution Electron Phase Imaging without Probe Calibration via Ptychography-Supervised Learning
The paper trains a compact model on physics-based ptychographic reconstructions to predict phase-image patches directly from 4D-STEM diffraction data, avoiding explicit probe calibration and iterative reconstruction during acquisition. It reports roughly 0.27 ms per probe position (about 20,000 positions/s), with atomic-scale contrast, and claims transfer across materials, defocus conditions, and microscopes without fine-tuning.
Real-time, calibration-free atomic phase imaging would be a meaningful capability and the reported cross-material and cross-instrument transfer is notable, but the abstract provides limited quantitative validation beyond one teacher dataset and headline speed claims.
-
The Cartesian Hand: In-Hand Manipulation with All-Linear Fingers
The paper introduces a 7-DoF robotic hand built entirely from linear actuators: two independently controlled grippers hold separate parts of an object while translating fingertips create relative motion between them. It demonstrates tasks including opening threaded containers, pipetting, pumping, screwdriving, trigger use, and in-grasp reorientation on 35 objects, with procedures transferred from a fixed-base arm to a humanoid using two such hands. The main new idea is that substantial task-oriented in-hand manipulation can be obtained from simple, configuration-independent linear motions rather than a highly articulated human-like hand.
The mechanically simple hand and broad demonstrations suggest a genuinely different route to dexterous manipulation, but the abstract does not establish quantitative advantages, robustness, or generality beyond objects whose mechanisms match the hand’s structured linear primitives.
-
Skytopia: Monocular Drone Navigation with Action-Conditioned Latent World Models
Skytopia trains a monocular-drone navigation policy using an action-conditioned latent world model built around 3D Gaussian Splatting. Instead of running the learned predictor during control, it uses the representation learned for prediction and inverse motion recovery, reducing inference cost by 59.4%; one policy handles point-goal, image-goal, and goal-free navigation. It reports higher simulated success than all baselines and transfers without fine-tuning to indoor, outdoor, and woodland flights on a physical drone.
The idea of training with a predictive world model but discarding its predictor at deployment is a non-obvious and potentially useful reframing, supported by simulation and real-drone transfer, but the abstract gives limited baseline and real-world quantitative evidence for a stronger verdict.
-
Beyond Repeated Sampling: Learning Search Policies for LLM Reasoning
The paper replaces naive repeated sampling for LLM reasoning with semantic exploration: a model generates diverse concepts, hints, or strategies that condition a larger answer model. It also trains a small concept generator with reinforcement learning to maximize downstream solution success, reporting better math pass@k at the same answer-generation budget and transfer to unseen answer models, including a different model family.
The potentially reusable, model-agnostic search policy is a meaningful alternative to repeated sampling, but the abstract gives no quantitative scale, task breadth, or robustness details to justify a strong verdict.
-
CliffCompaction: Cost-Efficient Compaction for Long-Horizon Coding Agents
CliffCompaction is a context-management method for long-running coding agents that reduces context by truncating or dropping original content, rather than rewriting it. It avoids recursively compacting summaries, reporting up to 50% lower cost while maintaining or improving results on Terminal-Bench and KernelBench; with repeated sessions it claims substantial CUDA-kernel improvements and better test-time scaling efficiency.
The non-rewriting, non-recursive compaction design and reported cost–performance gains are potentially important for long-horizon agents, but the abstract provides limited experimental detail and makes unusually broad model-comparison claims that need verification.
-
DreamStream: Towards Policy-Oriented Generative Simulation for End-to-End Driving
DreamStream is a closed-loop generative driving simulator that uses a simulator-grounded video model to vary appearance while preserving scene layout and motion features important to driving policies. It introduces FDπ, a policy-feature-based sim-to-real metric, and uses the simulator to create Navhard-CL, an interactive and adversarial version of NAVSIM that reveals failures such as scorer bias and poor recovery behavior.
The policy-oriented notion of simulation fidelity, coupled with a generative closed-loop simulator and an interactive adversarial benchmark, is a meaningful direction for evaluating end-to-end driving, but the abstract provides limited evidence beyond metric comparisons and failure-case claims.
-
Tight Regret Bound for Online Inverse Linear Optimization via Multiscale Matrix Weights
This paper studies online inverse linear optimization, where a learner observes the action chosen by an unknown fixed linear utility and must recommend actions in changing feasible sets. It gives a horizon-free randomized algorithm using multiscale matrix multiplicative weights with expected cumulative utility regret O(\sqrt{d}), matching the known dimensional lower bound up to constants, though polynomial-time implementation remains open.
The horizon-independent, dimension-optimal regret bound is a substantial theoretical result, but the method is specialized and its unresolved computational efficiency limits its practical significance.
-
WTF?! Simulation-Free Reinforcement Learning with Wasserstein-Tilted Flow Maps
The paper proposes Wasserstein-Tilted Flow Maps (WTF), a way to fine-tune pretrained flow-based generative models by deterministically transporting individual samples toward higher reward, rather than reweighting the model distribution with a KL objective. It derives an optimal-control formulation that enables simulation-free fine-tuning and reports higher reward with comparable or better diversity on ImageNet and text-to-image tasks, using up to 280× less training compute and no post-hoc distillation.
The combination of a Wasserstein/optimal-control objective with simulation-free flow-map fine-tuning is a genuinely interesting alternative to standard KL reward tilting, and the claimed compute reduction is substantial, but the abstract lacks enough quantitative and comparative detail to establish that the large efficiency claim is broadly credible.
-
Realize What Matters: Principled Context Representation for Large-Scale Reasoning
This paper proposes principles for constructing useful representations of information that is too large and heterogeneous to fit in a model’s context, based on the idea of “relevance realization.” Its R3Con system reportedly beats nine strong baselines by 20 and 8.4 percentage points on two large-corpus reasoning benchmarks, and lets 4B/9B models outperform tested 35B models while reducing cost relative to Claude Sonnet 5.
The combination of a principled representation-design framework with large reported gains and much better small-model efficiency is notably interesting, but the abstract does not explain the principles, benchmarks, or evaluation details enough to establish that this is a broadly new advance rather than a strong task-specific orchestration system.
-
A Quasi-Direct-Drive Underactuated Asymmetric Hand for Dexterous and Efficient Grasping and Manipulation
The Berkeley QUAD Hand is an 11-DoF, 8-actuator anthropomorphic hand that combines individually driven quasi-direct-drive fingers with an asymmetric, compliant transmission for the ring finger. Under light loads the coupled fingers move together for dexterity, while under heavy loads the mechanism reallocates motion toward stronger grasp closure. Hardware tests achieved 29/33 Feix grasp types, very low backdrive force, and reportedly up to 96× lower sustained-load heat generation than a more uniformly actuated design.
The asymmetric load-dependent actuation strategy is a substantive hardware idea with promising dexterity, force-transparency, and thermal-efficiency results, but the abstract does not provide enough comparative detail to establish a major advance.