Issue 38 · Sep 14–20, 2026
Every candidate
All 3,476 papers were scored from their abstracts by gpt-5.6-luna; 1,422 were not skipped. Shown here: the top 150 of those, in score order. Picks are marked.
-
A Horizon-Independent Regret Bound for Optimistic Hedge in General-Sum Games
The paper proves that standard Optimistic Hedge with a constant learning rate can achieve constant individual regret in self-play general-sum games, under expected loss-vector feedback, rather than the previously known logarithmic bound. Consequently, time-averaged play approaches a coarse correlated equilibrium at rate O(1/T); the proof uses an analytic recurrence and a nonconstructive Noetherianity argument, leaving the dependence on the number of players and actions implicit.
If the stated assumptions are sufficiently general, this removes the horizon dependence for the canonical Optimistic Hedge algorithm and overturns the expectation that modified regularization or higher-order prediction is needed for constant regret, although the nonconstructive proof and implicit parameter dependence limit practical interpretation.
-
The Exponential Price of Determinism in Nonsmooth Nonconvex Optimization
This paper proves that deterministic first-order methods for finding Goldstein stationary points of nonsmooth, nonconvex Lipschitz functions require oracle complexity exponential in the dimension, roughly $(1/\varepsilon)^{\Omega(d)}$. This closes an open gap between prior linear deterministic lower bounds and dimension-free randomized algorithms, and extends the separation to related stationarity, descent-direction, and smoothing tasks.
It resolves an explicit open problem with a striking exponential separation between deterministic and randomized optimization, potentially changing how algorithm designers view randomization in nonsmooth nonconvex problems.
-
An explicit solution of the five-expert prediction PDE and the exact optimality set of COMB
The paper gives an explicit solution to the stationary prediction-with-expert-advice PDE for five experts, extending the known four-expert solution across three regions. It proves that the direction (1,0,1,0,0) is optimal throughout the ordered sector, while the commonly conjectured COMB strategy (1,0,1,0,1) is optimal only on the lower-dimensional set where x₁=x₂ and x₃=x₄, thereby disproving the COMB optimality conjecture. The key inequalities are verified with computer-assisted exact rational certificates.
This is an exact, assumption-breaking result in online prediction theory that resolves and disproves a longstanding optimality conjecture, with unusually strong computer-assisted verification.
-
Radio-Frequency Convolutional Neural Networks
The paper uses the frequency mixer already present in wireless radios to perform convolution directly in the analog frequency domain, mapping channels and weights onto radio-frequency tones. Experiments show inference with CNNs up to 26.4 million parameters and nine layers across wireless-signal classification, image classification, and controllable image generation, with near-full-precision results and reported energy as low as 0.72 fJ per MAC. The unusual idea is to turn communication hardware into a shared neural-network accelerator rather than adding a separate digital compute chip.
Repurposing ubiquitous RF mixers for multi-channel CNN computation is a genuinely novel hardware direction, and the reported scale, task diversity, and very low energy cost make it more than a minor analog-accelerator variant.
-
Efficient Robust Learning at the Information-Theoretic Limit
The paper gives a polynomial-time algorithm for robustly learning Boolean concept classes that achieves the information-theoretically optimal error of η+ε, assuming access to an empirical-risk-minimization oracle. It also removes the oracle requirement for classes with suitable sandwiching polynomials under hypercontractive distributions, yielding the first efficient algorithm achieving η+ε error for Gaussian halfspaces. The main technical idea is to combine robust learning with several no-regret learning procedures.
It resolves a stated open computational problem and reaches the optimal η+ε error—rather than the usual 2η+ε barrier—with a concrete efficient result for Gaussian halfspaces.
-
An Eternal Irradiance Camera
The authors build an omnidirectional camera that measures low-frequency illumination rather than detailed images. Using the observation that spherical irradiance is well approximated by the first seven spherical-harmonic degrees, their 49-detector FluxCam reconstructs ambient illumination and operates wirelessly at 30 Hz using only harvested light, with applications including rotation sensing, workspace monitoring, relighting, and solar-panel orientation.
This is a genuinely different camera design that trades visual detail for extreme energy efficiency and inherent privacy, with a physical prototype demonstrating battery-free wireless operation rather than only a theoretical proposal.
-
Optimal High-Order Methods for Solving Monotone Variational Inequalities
The paper proposes a second-order algorithm for smooth monotone variational inequalities that reportedly achieves an O(T^-2.5) convergence rate, matching the known lower bound. It also gives a unified pth-order method with rate O(T^{-(3p-1)/2}) for every order p, claiming optimal complexity across the hierarchy.
If the stated bounds and assumptions hold, this closes a major complexity gap for monotone variational inequalities and establishes optimal rates for all derivative orders, rather than offering a minor algorithmic refinement.
-
RecreationWorld: Scalable and Verifiable Environments for Hybrid Computer-Use Agents
RecreationWorld trains and evaluates hybrid computer-use agents that alternate between GUI exploration, coding, execution, and visual verification. It uses running applications as behavioral oracles across Ubuntu, macOS, Windows, Android, and the Web, and reports that training on recreation trajectories transfers to other coding and computer-use tasks. Its 250-task benchmark shows a large gap between reproducing static UI structure and reproducing interactions or computed behavior: GPT-6 Astra scores 58.1% overall but passes every programmatic test on only 2.8% of tasks.
The cross-platform recreation setup and execution-grounded verification are a meaningful direction for hybrid computer-use agents, and the strong failure on behavioral fidelity is a useful non-obvious finding, but much of the contribution is a benchmark and training infrastructure rather than a demonstrated capability breakthrough.
-
CodeMidas: Scaling Agentic Coding RL Environments from Code Itself
CodeMidas uses agents to turn existing source code directly into executable reinforcement-learning tasks, without relying on issues, commits, or other development artifacts. It generates behavioral specifications, tests, and validation rollouts, producing 5,545 tasks across 3,185 codebases, and training on them improves several coding-agent benchmarks, including gains of 17% on ProgramBench and 11.7% on DeepSWE.
The source-code-only approach to scaling verified coding RL environments is a meaningful and broadly useful direction, with sizable cross-benchmark gains, but the abstract does not establish whether the generated tasks are genuinely novel or how much of the improvement comes from data scale and task overlap.
-
SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness
SoL-Pi uses automated, recursive search over coding-agent harness designs, selecting improvements to action execution, context compaction, observation processing, and delegated reading. On 51 EdgeBench tasks, it reportedly matches the performance of Pi across GPT-5.6 Sol and Opus 5 while reducing recorded token traffic by 44.7–49.0% and API cost by roughly one third.
The combination of auto-discovered harness improvements with near-constant coding-agent performance and roughly 45–50% lower token usage is practically important, but the evidence is limited to one 51-task evaluation and the abstract does not establish broad generalization.
-
AthenaZero: A low-inertia, bimanual robot for dynamic manipulation
AthenaZero is a bimanual robot built with quasi-direct-drive actuators and remotely located transmissions to reduce endpoint inertia while retaining torque and control authority. Its effective endpoint mass is reported to be roughly human-like and about ten times lower than conventional manipulators, enabling throwing above 30 m/s and catching and batting above 14 m/s, including robot–robot and human–robot interactions.
The combination of unusually low reflected inertia, high dynamic capability, and demonstrated high-speed physical interaction represents a meaningful advance beyond standard robot manipulation demos.
-
Observing and evading quantum back-action on a kilogram-scale oscillator
The authors use LIGO’s 40-kg test masses to create a narrow-band optomechanical mode whose motion is dominated by radiation-pressure back-action. By engineering correlations in the quantum state entering the interferometer’s dark port, they cancel part of the measurement imprecision and back-action, suppressing the observed motion near resonance by about 47%.
This is a credible demonstration of quantum back-action evasion and reservoir engineering at kilogram scale, using a gravitational-wave detector rather than a small laboratory oscillator, with a substantial measured suppression of quantum noise.
-
DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression
DeepSeek-V4.1-Flash is a 552B-parameter multimodal MoE model designed for million-token, input-heavy agent workloads. Its main contribution is aggressive KV-cache reduction: cross-layer reuse, FP4 caching, and a deployment strategy reportedly reduce always-resident KV storage to 890 bytes per token and persistent storage to about one-eighth of the prior model, while using fewer active parameters during prefill. The abstract claims substantially better performance than DeepSeek-V4-Flash, but gives no benchmark numbers.
The combination of cross-layer KV reuse, FP4 caching, and bounded replay could materially change the economics of million-token agent serving, but the abstract lacks quantitative quality, throughput, and ablation results needed to establish how much of the claimed advance is real.
-
The Neural Forcing for Three-Dimensional Incompressible Navier-Stokes finite time blowup
The paper proposes using neural networks and differentiable Navier–Stokes simulations or PPO to discover external force trajectories that drive 3D incompressible flow toward vorticity growth. It then claims a mathematical argument transferring these numerical candidates to continuum dynamics and proving finite-time loss of smoothness with positive probability under a neural output distribution. The result concerns forced, rather than the standard unforced, Navier–Stokes problem.
A valid continuum-level finite-time blowup result would be extraordinary, but the abstract gives no concrete conditions, forcing regularity, numerical validation, or proof details, and the probabilistic closure claim makes the headline result difficult to assess.
-
VC-Attention: Value Smoothing and Softmax Casting for Low-bit Attention
VC-Attention is a training-free low-bit attention implementation for long-sequence video diffusion Transformers. It reduces value-quantization error by clustering value tokens, quantizing mean-subtracted residuals, and reconstructing the mean from the softmax row sum; it also replaces the usual high-precision exponential with a fused log-domain probability casting operation. The authors report 1.46–1.59× attention speedups on datacenter GPUs, 2.3–3.6× on workstation GPUs, and measurable end-to-end video-generation speedups while improving fidelity over low-bit baselines.
The combination of value-token reordering with mean-residual quantization and fused softmax casting addresses two real deployment bottlenecks and reports substantial cross-GPU speedups, but it remains a specialized, hardware-dependent attention optimization rather than a clearly new general model capability.
-
Omni-Streaming Thinking
The paper addresses premature commitment in streaming audio-visual models: a model may infer something from early video and keep treating it as true even when later audio contradicts it. OST represents evidence, forecasts, and tentative claims separately, verifies claims over specified future intervals, and weakens dependent reasoning states when contradictions appear. With lightweight adaptation of a frozen Qwen3-Omni model, it reports over 10% relative average gains across five benchmarks and a large improvement on a controlled audio-visual contradiction test.
The explicit pending-claim and modality-specific verification mechanism is a genuinely relevant approach to a central streaming multimodal failure mode, with promising but still abstract-level evidence and unclear robustness beyond the reported benchmarks.
-
Grounded Action Model: 3D Grounding as a Foundation for Robotics
The paper builds robot action models around explicit 3D, metric grounding: language, points, or boxes are converted into an object-centric representation that is combined with robot state to predict action chunks. Compared with existing policies, it reports better robustness to relocated targets, scene randomization, and visual shifts, including real-robot tests, and can act as a low-level controller for high-level planners on long-horizon tasks.
The explicit use of 3D object grounding as the central interface for a robot foundation model, together with perturbation and real-robot evidence, is a meaningful direction, but the reported gains are not yet large or broad enough to make this a must-read.
-
Taming Long-form Text-to-Speech
The paper introduces Localized Attention-Constrained Inference (LACI), an inference-only mechanism that detects likely TTS failures, rolls back to their onset, and regenerates locally with temporary constraints. On long prompts, it reduces worst-of-10 WER for Qwen3-TTS-0.6B from 35.2% to 3.4%, and substantially improves voice-cloning consistency on 120-second references; it also proposes a sliding-window similarity metric that reveals failures hidden by global speaker-similarity scores.
This is a potentially broadly useful inference-time solution to a major long-form TTS reliability failure, with unusually large reported gains and a diagnostic metric, although the evidence is limited to the abstract and worst-of-N evaluations.
-
Verifiable Social Reasoning for LLM Assistants
The paper introduces Fuse, a multi-agent simulation in which an assistant must infer a hidden social motive from a user’s subjective account, while the simulator provides ground-truth motives by construction. A human study with 24,000 annotations validates the simulations, and tests across 12 LLMs show that user framing can systematically bias models, models may need more details than humans, and longer conversations do not reliably improve reasoning. The framework and 21,000-example dataset are released.
The user-mediated setup offers a relatively new and verifiable way to study social reasoning, with substantial human validation and several non-obvious findings, though it remains primarily an evaluation framework rather than a demonstrated capability advance.
-
KaiNinja: Extending Native 3D Generators to the Part Level
KaiNinja extends a native single-image 3D generator to produce separate, editable object parts rather than one fused mesh. It introduces a dual-volume version of the O-Voxel representation, which can model the shared interfaces between touching parts, and reports better whole-object fidelity plus improved part-level F-scores without a separate segmentation stage.
The dual-volume representation appears to address a real structural limitation in native 3D generation and enables a useful new capability, but the abstract provides limited detail about evaluation breadth and baseline strength, so it falls short of a strong verdict.
-
The Other Half of the Memory Wall: Serving 35B MoEs from SSD with Trained Routing Prediction
Edge0 runs a 35B mixture-of-experts model on a 24GB consumer machine by storing expert weights on SSD and predicting the next layer’s routing decisions one token ahead, allowing expert reads to overlap with computation. The predicted experts are used directly rather than as a speculative cache, and a recovery LoRA compensates for quality loss from int4 quantization and replacing the original router; the authors report 20 tokens/s with only 3GiB of active memory and quality within a few points of the fp16 model across five benchmarks.
The combination of learned one-step-ahead routing with SSD streaming could make large MoEs practical on consumer hardware, but the abstract gives limited detail about hardware, workloads, latency variability, and benchmark results, so the unusually strong systems claims need verification.
-
Zing-0.5: Toward Playable Worlds with Real-Time Joint Action and Text Control
Zing-0.5 is a 5B autoregressive video/world model that lets users steer an ongoing generated environment with both keyboard input and online text commands. It uses event-level distillation and streaming generation to reach 832×480 at 24 FPS, with a reported cost of about $0.009 per stream-minute, and demonstrates changing an event without restarting navigation. Evaluation on 158 WBench Navigation cases reports scores of 81.0 overall and 88.5 for consistency, though the abstract gives limited detail about comparisons and breadth of testing.
The combination of real-time playable video generation, persistent joint keyboard/text control, and unusually low serving cost is a meaningful direction, but the evidence is mostly a single benchmark and demonstration rather than a clearly established capability jump.
-
Video DeltaNet: A Video-Native Hybrid Attention for Livestream Video Generation
Video DeltaNet replaces much of the expensive dense attention in a video diffusion model with a hybrid of local Softmax attention and bidirectional linear memory. Its frame-level memory update and staged teacher-alignment training let a MiniMax H3 variant generate 14.3-second, 768p videos in 6.70 seconds on eight B200 GPUs, with the authors reporting a 14.5× speedup over a 50-step dense-attention baseline.
The video-native linear-memory design and reported serving speed are potentially important, but the abstract gives no video-quality or quality-at-equal-compute results, and the headline speedup also changes the denoising step count.
-
Long-horizon autoformalization of a core theorem underlying MIP* = RE
The authors introduce FormalFlow, a multi-agent workflow for long-horizon Lean formalization that uses shared blueprints, nested planning, proof generation, and iterative review to limit statement drift and composition failures. They use it to produce a machine-checked Lean 4 proof of the quantum soundness of the classical low individual-degree test, a core component of the MIP* = RE proof, generating 126,367 lines of code in 63 days and correcting side conditions while retaining the published error bound under revised assumptions.
This is a substantial demonstration that coordinated proving agents can formalize a technically deep, long-horizon theorem rather than merely solve isolated benchmark lemmas, with a large verified artifact and a concrete workflow for scaling mathematical formalization.
-
The First-Order Oracle Complexity of Lipschitz Convex Optimization in Nondual Settings
This paper resolves an open question about whether a favorable mismatch between feasible-set and objective geometry can improve nonsmooth convex optimization rates. In particular, it obtains a near-/T rate for Euclidean-Lipschitz convex optimization over the -ball, instead of the usual 1/sqrt(T) rate, using a new online-learning game tied to sequential fat-shattering dimension; it also matches known lower bounds up to logarithmic factors and derives related geometric results.
The near-1/T rate in a nonsmooth setting and the resolution of a COLT open problem are genuinely non-obvious, but the result is primarily foundational theory with limited immediate impact on practical ML.
-
Modality-Autoregressive World-Action Models
ModAR is a world-action model that sequentially predicts several future modalities—such as point tracks, DINO features, and depth—before predicting robot actions, rather than relying mainly on future RGB frames. In experiments, these modalities improve performance, while future RGB adds no consistent benefit; the model slightly beats a video-pretrained baseline (75% vs. 72% success) using about 20× fewer training FLOPs, and performs well on three real-world bimanual tasks.
The combination of modality-autoregressive prediction and the finding that geometric, semantic, and motion representations can matter more than generated RGB is a substantive, somewhat surprising direction, though the evidence is limited to the reported tasks and modestly separated headline results.
-
Region-Level Policy Optimization for Fine-grained MLLM Perception
The paper separates fine-grained visual perception into coarse region localization and high-resolution recognition. It trains a lightweight proposal network with region-level reinforcement learning, using a frozen MLLM to score whether proposed regions contain evidence needed for the answer, then applies sparse high-resolution encoding to those regions. Across six benchmarks and four MLLM backbones, it reportedly achieves better accuracy at every visual-token budget and matches or exceeds the largest-budget model with about four times fewer tokens.
The answer-conditioned region-selection objective and reported roughly 4× reduction in visual tokens are genuinely useful, but the abstract gives no numerical results or details sufficient to establish whether the gains are robust or substantially beyond existing adaptive-resolution methods.
-
The Router Within: Eliciting Native Skill Routing from a Frozen LLM
Gavel uses a frozen LLM’s internal activations to choose among external skills, avoiding putting all skill descriptions in the context or relying on a separate retriever. Two learned linear projections produce a fast initial ranking, after which the model’s own likelihood and yes/no judgments refine the shortlist; it transfers zero-shot across several benchmarks and reportedly beats much larger retrieval-based systems by up to 13.4 points on written tasks and 21.9 points when skills are needed mid-rollout.
The potentially important idea is that a frozen model’s native representations can route a large skill library without skill text or a substantial external router, but the evidence is mainly benchmark claims and includes a new simulated benchmark, so it merits inspection rather than being an automatic must-read.
-
BVB: Benchmarking Agentic Video Understanding via Programmatic Reconstruction in Blender
BVB evaluates video-understanding agents by requiring them to reconstruct source videos as animated Blender scenes through code, rather than answer questions alone. Across 51 configurations, the strongest system achieved high perceptual similarity but preserved only 53.7% of the source video’s verified spatiotemporal facts; extra reasoning improved visual resemblance without resolving this semantic gap. Human preferences correlated strongly with the proposed perceptual metric.
The programmatic-reconstruction framing is a genuinely nonstandard way to test video understanding, and the finding that agents can make plausible-looking videos while missing many factual details is informative, though this is primarily a benchmark proposal rather than a capability breakthrough.
-
Dream-RSI: Recursive Self-Improvement through Evolving Worlds
Dream-RSI treats the history of an AI coding agent’s discoveries as a replayable simulator of the search process. It uses this simulator to test and improve exploration policies with cheap off-policy feedback, then deploys the updated policy online in an iterative loop; experiments in algorithm design, mathematical optimization, and GPU kernel engineering reportedly show similar or better discoveries at lower cost.
The replay-based approach to recursively improving exploration is a meaningful idea, but the abstract gives no quantitative results, baselines, or evidence that the simulator remains useful beyond the settings tested.
-
Occupancy Network-Guided Autonomous Robotic Partial Nephrectomy
The authors built a vision-guided robotic system for partial nephrectomy that reconstructs the kidney, tumor, and surgical margin from a single-view partial point cloud, then updates its plan as tissue is manipulated and cut. In eight patient-derived hydrogel phantom procedures, it autonomously made 77 electrosurgical cuts, achieved negative margins in every case, and had a mean absolute margin error of 1.61 ± 0.48 mm; the main novelty is closed-loop tumor resection rather than surface-only robotic manipulation.
This is a substantial new robotic capability—autonomous adaptive tumor removal with margin control—but the evidence is limited to eight hydrogel phantom trials, so it is not yet strong evidence of clinical surgical autonomy.
-
TAPe+ML: A Compact Structured Representation for Multi-Task Computer Vision
The paper presents TAPe+ML v3, a hand-structured visual representation that encodes contours, background, object locations, and prototypes before modular recognition, rather than relying primarily on pixel-processing networks. The authors report fewer than 100,000 parameters while achieving 65.3 COCO box mAP50-95, 58.4 mask mAP50-95, and competitive classification results, plus tests on video and distribution shift.
The claimed combination of sub-100k parameters with strong COCO detection and segmentation results would be a major result if independently verified, but the abstract gives too little detail about training, compute, comparisons, and evaluation to justify a strong recommendation.
-
OPEN-1B: A Fully Auditable Training Run
The paper presents a training procedure intended to make a distributed language-model run exactly auditable across different hardware. It fixes the ordering of GPU reductions, data-parallel batches, and collective communication so that independent auditors can replay individual training steps bit-for-bit; the authors release a 1B-parameter model, all checkpoints and data, code, and verification tools.
Bitwise, heterogeneous-hardware verification of arbitrary steps in a large distributed training run is a genuinely unusual transparency direction, but the abstract gives little quantitative evidence about practicality, overhead, auditor coverage, or whether the released system works at meaningful scale.
-
Single-electron detection in a high-purity Ge device
The paper presents a cryogenic high-purity germanium detector that measures the phonon energy from charge drift with an integrated NbSi transition-edge sensor. It reports sensitivity to a single electron–hole pair, potentially reaching eV-scale thresholds while rejecting low-energy backgrounds better than existing cryogenic detectors, enabling sub-100 eV recoil discrimination for dark-matter and neutrino experiments.
Single-charge sensitivity in a large-volume, high-purity Ge detector is a genuinely notable instrumentation advance, but the abstract gives limited quantitative validation of background rejection, scalability, and recoil discrimination.
-
What Does Privileged Information Add to On-Policy Self-Distillation?
This paper separates the benefit of on-policy self-distillation from the benefit of giving the teacher a worked solution or answer. Across 5,319 matched math problems and two small language models, reference-free distillation explains much of the improvement, while complete solution traces provide only modest extra gains and can hurt when the student uses long thinking rollouts. The results suggest that distillation mainly transfers capabilities already shared between direct-response and reasoning modes, rather than simply transferring more solution information.
The matched intervention cleanly challenges the assumption that privileged solutions are the main source of OPSD gains, but the evidence is limited to small models and mathematical reasoning, with only modest measured reference benefits.
-
Rethinking Critic Learning in PPO: Understanding and Mitigating Value Flattening
The paper identifies a failure mode in PPO critics called “value flattening”: critic estimates remain too similar across intermediate states even when Monte Carlo returns vary substantially. It attributes this to variance effects in the value loss and redundant updates from correlated states, then proposes applying value supervision only to a few well-separated states per response; experiments on Qwen3-Base suggest that supervising just three states can improve policy learning across several model sizes and evaluations.
The diagnosis of value flattening and the simple sparse-critic-training intervention are potentially useful for PPO-based LLM training, but the abstract provides limited quantitative evidence and the method may be an incremental loss-application strategy rather than a major new direction.
-
Confidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents
The paper proposes XConf, which estimates an LM’s confidence from its record of similar past attempts, including stated confidence, outcomes, and post-hoc lessons. On nine benchmarks and four models, it reportedly matches or beats ten-sample self-consistency for confidence discrimination while using one generation, and improves selective prediction on agent tasks when low-confidence cases are abstained from.
The experience-based confidence estimator is a meaningful alternative to inference-only calibration and appears substantially cheaper, but the abstract does not establish how much depends on retrieval quality, accumulated data, or comparisons with other outcome-based calibration methods.
-
JEPA-Anything: Learning Predictive Models across Different Worlds
JEPA-Anything proposes orthogonal predictive factorization: it splits latent targets into complementary factors, predicts them through separate pathways, and recombines them in a shared JEPA-style model. Across seven domains, it reports consistent gains on matched dynamics tasks, lower long-horizon molecular forecasting error, and a 34.8% reduction in intervention-prediction error on Interventional Pong; it also presents biological and physical case studies using the learned factors.
The cross-domain factorized world-modeling principle is a genuinely interesting direction with unusually broad experiments, but the abstract does not establish that the gains transfer beyond selected benchmarks or that the biological discoveries are caused by the method rather than exploratory case studies.
-
Robot World Models Are Not Invariant to How the Actions Are Written
The paper shows that a robot world model trained with absolute joint targets can behave very differently when given the same commands expressed as state-relative deltas, even though the two encodings are almost perfectly reconstructible from the joint state. Across three datasets and two robot morphologies, retrieval and goal-conditioned action selection degrade sharply under this re-parameterization, with latent predictions sometimes becoming nearly orthogonal. It proposes testing action-encoding invariance and using objective-level averaging plus a disagreement penalty, which substantially improves worst-case agreement in the reported experiments.
This identifies a non-obvious and potentially important failure mode in action-conditioned world models, backed by large cross-dataset degradation, but the abstract does not establish how broadly it affects modern architectures or real robot performance.
-
Discovering Physical Representation Languages
The paper proposes learning the representation system behind physical measurements before learning the physical laws themselves. From anonymous controlled experiments, it aims to recover structures such as measurement types, refinement relations, primal-dual correspondences, and gauge-like ambiguities, with identifiability results and experiments on Maxwell and RLC systems. The abstract claims polynomial-time recovery, minimax-optimal measurement bounds, scaling to tens of thousands of cells, and robustness to nonlinear, nonlocal, and hysteretic regimes, but gives few concrete quantitative details.
The representation-language-discovery framing is genuinely unusual and potentially important for scientific ML, but the very broad claims are supported only vaguely in the abstract, with no specific recovery rates, comparisons, or experimental results to establish that the framework works beyond curated physical systems.
-
Agora: Git as Shared Memory for Collective AutoResearch
Agora uses Git as an append-only shared research memory for multiple language-model agents, recording experiments and linking results in a searchable DAG. In a nearly 12-day run, 13 unsupervised agents made 1,703 contributions and improved initialization of a frozen attention–state-space model, with later diversity-oriented views prompting exploration of neglected methods. The system also produced many reproductions, but the paper does not yet provide a matched comparison showing that shared memory improves discovery or compute efficiency.
The Git-backed, multi-agent research workflow and observed shift toward diverse exploration are genuinely interesting, but the central benefit is not causally evaluated and the technical result comes from one specialized task.
-
Fathom: Per-Query Read Depth for Sparse Decoding over Offloaded KV Caches
Fathom accelerates attention retrieval from very large, host-resident KV caches by letting each query read a different number of bits from each key channel. Using bit-plane storage and a variance-weighted reverse-water-filling policy, it reports up to 1.67× faster GPU-time scans than several 136-bit baselines; at equal speed to a lower-read baseline, it reads 18% fewer bytes with lower attention error, while matching exact top-k decoding on RULER-style tasks and reaching comparable agreement on coding-agent traces at 92 bits.
Per-query, per-channel bit-depth allocation is a genuinely useful systems idea for the host-memory KV-cache bottleneck, with meaningful speed and accuracy results, but the evidence appears concentrated on one model and a specialized serving regime rather than demonstrating a broadly transformative advance.
-
When Agents Slow Down: Understanding LLM Agents' Test-Time Strategies via Elo-per-token Analysis
The paper introduces Elo-per-token, a way to compare progress during long-running agent sessions by rating the best solution found at each token budget across tasks with different score scales. Across four agents and four open-ended benchmarks, agents initially use tokens more effectively than independent sampling but eventually suffer diminishing returns; splitting compute across multiple sessions substantially outperforms one long session, while human contestants continue improving faster on a shared programming task.
The analysis offers a useful, fairly novel way to study agent test-time scaling and reports a non-obvious practical result—parallel restarts can beat prolonged deliberation—but the evidence is still limited to a small set of agents and benchmarks, making this more of a promising empirical framework than a field-changing result.
-
Fingers as Legs: Learning Self-Supported Locomotion and Manipulation with an Anthropomorphic Hand
The authors build an untethered anthropomorphic robotic hand that uses its fingers both to crawl and to manipulate objects. Reinforcement-learning policies trained in a simulator calibrated to hardware enable crawling, steering, fall recovery, keyboard operation without vision, and visually guided object pushing, while retaining the hand’s existing finger design and position controller. The main novelty is demonstrating a compact mobile manipulator with no separate legs or locomotion mechanism.
The hardware demonstrates a genuinely unusual form of locomotion-manipulation reuse and several real-world behaviors, but the abstract provides limited evidence about speed, robustness, generality, or comparison to alternative robot designs.
-
ModaLens: Measuring Image Sensitivity in Report-Conditioned Medical VLMs
ModaLens introduces a paired image-swap audit for testing whether report-conditioned medical VLMs actually use the image when answering radiology questions. On 3,199 paired MIMIC-CXR cases, MedGemma-27B changed its answers much less when a report was available (4.26% versus 20.94% without one), suggesting that reports can mask image sensitivity; the effect also appeared in two other model lineages. The authors provide prompts, code, and run-level records, while noting that report-derived labels cannot establish visual correctness.
The paired image-swap protocol exposes a non-obvious failure mode—report availability can make a medical VLM appear less image-dependent—with substantial, clustered evidence and replication, though it is primarily an audit metric rather than a demonstrated capability advance.
-
IntBMoE: Integrating Block-Level Conditioning into Expert Composition for Full-Participation Mixture-of-Experts
IntBMoE proposes composing all experts within a layer into a small set of block-specific experts, then sparsely routing each token to only a few blocks. This aims to provide full expert participation without dense inference cost or input-dependent parameter materialization; a dual-path multiplicative gating mechanism is also introduced. The method reportedly improves image classification, language modeling, and recommendation results, and yields a 2.4% relative online UVCTR gain in a large-scale deployment under a 60 ms latency budget.
The explicit decoupling of expert participation, execution, and materialization is a meaningful MoE design idea with real production evidence, but the abstract gives few quantitative comparisons and does not yet establish a major capability or efficiency jump.
-
GAVEL: Graph World Models for Verified and Efficient Long-Horizon LLM Task Planning
GAVEL augments an LLM robot planner with an explicit graph-based world model that tracks object relations, action preconditions/effects, and probabilistic beliefs about hidden object locations. It verifies and repairs executable plans before acting, using LLM replanning only when semantic reasoning is needed, and achieves large reported gains on BEHAVIOR-1K: single-task success rises from 41.2% to 91.8% and multi-task success from 19.9% to 92.6% with Qwen3-8B.
The combination of symbolic graph verification, local plan repair, and distributional search reasoning addresses a real weakness of long-horizon LLM planning and reports unusually large gains, but the evidence is limited to benchmark evaluation and the core ingredients are adjacent to established neuro-symbolic planning ideas.
-
Algorithmic Collusion and the Complexity of Information-Value-Free Equilibria
This paper studies information-value-free correlated equilibria in succinct games, where players cannot gain from following recommendations compared with committing to a fixed action. It proves that computing such equilibria is PPAD-complete even for constant-accuracy approximations in several compact game classes, with an exponential query lower bound, while the coarser version has an FPTAS at ordinary precision but remains tied to hard problems at exponentially high precision. These results show that the tractability of this equilibrium concept in explicitly represented games does not extend to realistic large games or imply efficient learning dynamics.
The combination of constant-approximation PPAD-hardness, an unconditional query lower bound, and a nuanced complexity gap for no-regret-compatible equilibria is a substantial theoretical result relevant to algorithmic learning and collusion, though it is primarily a specialized game-theory complexity paper.
-
iTBS induces a suppression-rebound pattern in human cortical excitability
The authors measured cortical responses during intracranial intermittent theta-burst stimulation (iTBS) in 20 neurosurgical patients. Rather than producing immediate excitation, iTBS caused about a minute of reduced reactivity in connected cortical regions, followed by a rebound increase in excitability and changes in excitation/inhibition balance; a similar pattern was observed in three patients receiving transcranial magnetic iTBS.
This directly challenges the common LTP-like interpretation of iTBS and uses unusually direct human cortical measurements, but the rebound findings are from a small, specialized patient sample and need broader replication.
-
ImpossibleRubrics: Stress-Testing Generated Rubrics as Reward Signals
The paper introduces ImpossibleRubrics, a benchmark for testing whether model-generated grading rubrics reward honest acknowledgments of impossible tasks or answers that confidently fabricate unsupported conclusions. Across 169 impossible tasks, 11 rubric generators were exploitable 8–26% of the time, and a generic rubric encouraging decisiveness was exploited 64% of the time—often worse than tailored rubrics. The central finding is that rubric specificity can expose the claim an attacker should fabricate, rather than making evaluation more reliable.
It reports a non-obvious failure mode for rubric-based reward and evaluation—tailoring and decisiveness can increase exploitability—but the evidence is confined to a 169-task benchmark and does not yet establish broad impact across real training or grading systems.
-
HuRo: Robotizing Human Videos for Scalable VLA Pretraining
HuRo converts diverse human manipulation videos into robot-aligned observations and retargeted action trajectories, including inferred intermediate signals, producing 630K episodes for VLA pretraining. On four real-world tasks, adding more robotized video data raises completion from 51.5% to 80.3% and out-of-distribution completion from 34.9% to 72.2%; visual robotization mainly improves robustness, while retargeted actions provide additional gains.
The combination of large-scale human-video robotization and action-level retargeting addresses a major data bottleneck, with substantial reported in-domain and OOD gains, but the evidence is limited to four tasks and the abstract does not establish how broadly the pipeline transfers.
-
PhysStream: Streaming Physics-Grounded Video Generation with Structured Scene Memory and Fine-Grained Motion Control
PhysStream is a causal, streaming image-to-video model for tabletop rigid-body scenes that allows users to inject sparse velocity changes while generation is underway. It maintains online positional and object-tracking maps as structured scene memory, and reports better physical consistency than baselines, including 33% lower motion-distribution distance and 12% lower trajectory error on synthetic tests, plus over 85% human preference in in-the-wild comparisons.
Mid-generation control using physically meaningful velocity increments and structured causal scene memory is a non-obvious capability, but the evidence is limited by the narrow tabletop setting, synthetic benchmarks, and few details about scale or generalization.
-
From Pretraining to Proficiency: Real-World Subtask RL for Long-Horizon Manipulation with Minimal Human Intervention
PARTS fine-tunes a pretrained robot policy by focusing reinforcement learning on the few subtasks where it fails, rather than collecting more demonstrations for the entire task. It uses learned selectors and success verifiers to trigger residual corrections and generate local rewards, and reports raising full-task success from 32% to 61% on bimanual YAM tasks and from 50% to 95% on Franka tasks with only tens of minutes of real-robot rollouts.
Targeted, locally rewarded residual RL with minimal human intervention addresses a real bottleneck in long-horizon manipulation and shows large gains under a fixed real-world rollout budget, though the evidence is limited to two robot settings and the abstract lacks enough detail to establish broad generality.
-
Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design
The paper lets a frozen language-model agent improve at professional graphic-design tasks by maintaining an external library of natural-language procedures. Procedures are added for recurring uncovered subtasks and revised when executions fail, with replay checks intended to prevent regressions; across five rounds of user-traffic data, the skill bank grew from 76 to 139 entries and reported execution success rose substantially without model retraining or human labels.
The combination of continual procedural-memory growth, failure-driven revision, and regression-gated replay is a meaningful adaptation strategy for long-horizon tool agents, but the evidence relies heavily on automatically graded design trajectories and a specialized setting, making the broad impact uncertain.
-
Register Tokens for Bounded-State Reasoning in Diffusion Language Models
This paper adds a small set of persistent continuous “register” tokens to diffusion language models. After each generation chunk, the text is discarded but the register hidden states are retained to carry reasoning progress into the next chunk; on LLaDA and Dream, this reportedly beats carrying the previous text, by up to 8.5 points on math and 19.5 on code, with additional gains from reinforcement learning.
The fixed-size latent state is a genuinely interesting alternative to growing textual context for iterative diffusion decoding, and the reported gains—especially on multi-chunk code—are non-obvious, though the abstract provides too little detail to establish robustness or generality.
-
LynnReal-Omni: Native multi-modal Video Generation for Agentic Visual Workflows
LynnReal-Omni is a large shared multimodal diffusion-transformer system that combines text, images, references, 3D renders, game states, structural controls, editing, restoration, and long-video generation in one video model. A 27B accelerated variant reportedly generates and decodes 22-frame 540p clips in 377 ms on one H100, but the abstract gives no comparative quality results or evidence that its unified design substantially improves fidelity or long-horizon coherence.
The broad native integration of agent-provided visual states with controllable video generation and the claimed near-real-time latency are worth noting, but the abstract provides mainly system scale and speed claims rather than convincing comparative evidence of a major capability advance.
-
ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement
ModularRSI evolves an agent’s execution harness rather than only its model or prompts. It compares successful and failed trajectories across 2,000 separate evolution tasks, identifies recurring problems, and updates five modules independently before integrating them; the authors report transfer to unseen tasks, other domains, and different foundation models on TB2.0 and SWE-Bench Verified.
The benchmark-disjoint, contrastive, modular approach targets a real weakness of recursive harness improvement and claims cross-task and cross-model transfer, but the abstract gives no quantitative gains or evidence that the modularization materially outperforms simpler harness optimization.
-
G-ray: Ray-Level Relative Geometric Position Encoding in Multi-View Vision Transformers under Camera Heterogeneity
G-ray replaces image-plane rotary positional coordinates with camera-local ray angles, aiming to make relative geometric cues consistent across cameras with different fields of view or projection models. Across several 3D reconstruction and novel-view synthesis setups, it improves multiple host encodings and reportedly reduces pointmap relative error by 45.8% versus MapAnything; notably, a model trained only on pinhole images can process mixed pinhole and non-pinhole inputs without retraining.
The projection-invariant ray-level positional encoding and zero-shot handling of heterogeneous cameras are genuinely useful, but the contribution remains a focused encoding change whose broad impact is not yet established by the abstract alone.
-
Hub-Spectral Activation of Latent Multimodal Knowledge
The paper introduces Hub-Spectral Activation, a closed-form procedure that extracts cross-modal structure from two separately trained hub connections in frozen multimodal representations. Without target-pair supervision, optimization, or backbone updates, it uses covariance spectra and reliability-weighted matching to improve retrieval and prototype classification, reporting large average gains across 19 retrieval and 11 classification relations on ImageBind and LanguageBind.
The combination of zero-shot hub-based knowledge recovery, a closed-form spectral method, and large gains across many relations is genuinely notable, but the abstract does not establish how broadly the approach generalizes beyond these specific hub models and benchmarks.
-
Large Language Models Develop Belief State Geometry In-Context
The authors test whether LLM activations represent the Bayesian belief state needed to predict sequences from hidden Markov models. Across six open-source models and 40 HMMs, the latent-state posterior is highly linearly decodable from residual activations, and patching or steering the corresponding subspace affects predictions in ways suggesting it is functionally used. The result connects in-context learning in production LLMs to approximate Bayesian inference, though the evidence is limited to controlled HMM settings.
This offers unusually direct representational and causal evidence that LLMs form usable Bayesian belief states during in-context learning, but the controlled HMM setup and lack of broader task validation make it a promising mechanistic result rather than a must-read breakthrough.
-
Not All Layers Need Tuning: Diagnosing and Directing Adaptation in Vision-Language-Action Models
The paper studies where different kinds of deployment shifts require adaptation in vision-language-action models, finding that appearance and novel-object shifts mainly affect vision features and action heads, while instruction shifts affect the language backbone. It uses ten unlabeled target observations to estimate per-region adaptation needs, then assigns variable-rank LoRA adapters and freezes regions predicted to be stable. Across five VLAs, the method reportedly ranks adaptation needs well and matches or beats uniform LoRA; on an xArm-7 it matches full fine-tuning with only 0.04% as many trainable parameters.
The potentially important contribution is diagnosing which VLA components need adaptation before training and using that diagnosis for highly sparse, shift-specific adaptation, supported by results across five architectures and both simulation and a physical robot, though the claims need scrutiny around evaluation breadth and robustness.
-
Reinforcement Learning for Real-Time Vision-Language-Action Policies
The paper adds a lightweight, fast action-editing policy around a slower pretrained vision-language-action model. The large model proposes action chunks, while the editor uses the latest observation to make reactive corrections, enabling reinforcement-learning fine-tuning despite inference delay. It reports improvement from 42% to 97% average performance across four dynamic real-world manipulation tasks using only 10 minutes of online robot data, and the best result in all 10 Kinetix environments.
The separation of expressive but slow VLA planning from fast RL-trained action correction directly addresses a major deployment bottleneck, and the large gains on dynamic real-robot tasks are notable, though the abstract does not establish how broadly the method transfers or how strong the comparisons are.
-
When a High Score Is an Illusion: Certifying Genuine versus Repackaged Forecasting Skill
The paper shows that forecast-ranking scores can look strong simply because forecasts and outcomes reuse the same reference observations, creating artificial association even when the forecasts themselves are unchanged. It derives conditions for eliminating this effect, proposes unbiased three-trajectory and complete U-statistic estimators, and reports that the artifact explains 91.4–94.5% of shared scores in a Beijing air-quality study; using independent references also greatly reduces false rejections in a null experiment.
This identifies a potentially substantial and underappreciated flaw in comparative forecast evaluation and gives formal correction methods, but the evidence is concentrated in one application and the broader impact depends on how widely these rank-based evaluation setups are used.
-
Ask the Tool, Don't Guess: Agent Tool Calls Hold Their Progress, and the Serving System Should Read It
This paper argues that LLM serving systems should not predict how long an agent's tool call will take; instead, the tool should report its progress while running. Across four agent corpora, the authors find that tools often expose useful progress signals, and integrating these signals into a production serving engine reduces p90 time to first token after tool completion by about 21% versus LRU cache management, nearly matching an oracle.
The key idea—having tools explicitly communicate progress so KV-cache eviction and restoration decisions can be made from current state rather than unreliable duration predictions—is a practical and non-obvious serving-system change backed by broad measurements and a sizable latency improvement, though its impact depends on tool ecosystems adopting the interface.
-
Taming the Agentic RAN: Stability-Guaranteed Arbitration of Autonomous AI Agents in O-RAN
The paper studies instability caused when independently designed AI agents control shared resources in an O-RAN system. It introduces AURA, an arbitration layer using feasibility constraints, minimum dwell times, and deadbands, and proves convergence; on an OpenAirInterface testbed it reduces resource oscillations from 8.4 to 0.4 PRBs and nearly eliminates cross-slice starvation without harming the protected latency SLA.
The combination of runtime arbitration, stability guarantees, and live O-RAN evidence is a meaningful treatment of multi-agent control interactions, though the contribution is specialized and builds on familiar control mechanisms such as deadbands and dwell times.
-
Why Pretraining Fails to Share Cross-Lingual Knowledge
The authors pretrain 360M- and 7B-parameter language models and investigate why knowledge learned in one language transfers poorly to another. In a controlled experiment, two identical copies of a language with disjoint token vocabularies still fail to share knowledge, suggesting that separate token spaces themselves create compartmentalization; mapping languages into a shared word-level token space improves transfer, recovering up to 12.6% of native-language learning efficiency, or 14× the baseline.
The matched-language control isolates token-space separation as a concrete cause of poor cross-lingual transfer, and the reported improvement is substantial, though the proposed shared-space mapping and evidence may not yet establish broad practical generality.
-
Form Over Content In Gradient-Based Data Attribution Methods
The paper tests what gradient-similarity data attribution methods actually capture by independently changing benchmark task and answer format. Across model checkpoints, scales, and families, gradient alignment tracks shared answer format rather than shared task semantics: examples with the same format align substantially (cosine near 0.4), while the same task in different formats shows essentially no alignment. The authors also find that LESS, a gradient-based data-selection method, over-selects examples matching the target’s answer format.
This is a broadly tested, non-obvious result that challenges the semantic interpretation of gradient attribution and has direct implications for data selection, though it is primarily a diagnostic finding rather than a demonstrated capability or major method advance.
-
The Output-Space Hypothesis: Enumerative Equivalence Checking for Tensor Programs
The paper proposes checking tensor-program correctness one output element at a time, symbolically proving that the element matches the reference for all possible inputs, rather than relying on random-input differential testing. Its system, Dirigo, found bugs in 600 of 6,988 CUDA kernels previously judged correct, detecting 97.3% of those bugs within two minutes.
The output-space quantifier reversal is a genuinely interesting verification strategy with a striking bug-discovery rate, but the evidence is currently limited to a public corpus of AI-generated CUDA kernels and the abstract does not establish broader scalability or comparison with competing verifiers.
-
Beyond the Stability--Plasticity Frontier in Streaming Target Speaker Extraction
The paper studies how streaming target-speaker extraction should update its internal representation when the target is silent, masked, or acoustically mismatched with enrollment. It shows that common update heuristics face a stability–plasticity tradeoff, then introduces a meta-trained 41k-parameter anchored fast-weights memory that improves severe-mismatch extraction by 3 dB while retaining near-static-enrollment performance after 30 seconds of target absence, with less than 5% runtime overhead.
The combination of a measured failure frontier and a compact, closed-loop-trained memory that reportedly beats it is a substantive advance for robust streaming speaker extraction, but the abstract does not establish broad real-world validation beyond the reported configuration study.
-
Noctif3R: Feed-Forward Monocular Real-Time SLAM for Photon-Limited Scenes on Embedded Hardware
Noctif3R studies monocular SLAM when images are near the sensor-noise floor and proposes SYS, a feed-forward low-light pointmap pipeline with an explicit gate that refuses to output meaningless poses. On a calibrated noise ladder and dark robot videos, most real-time baselines produce apparently complete but uninformative trajectories, whereas SYS tracks only in usable segments and achieves lower error when it does track. Its embedded implementation also reports higher throughput, lower memory use, and lower energy per pose on a Jetson AGX Orin.
The combination of photon-limited evaluation, explicit failure refusal, and an embedded real-time implementation is genuinely useful and exposes a serious failure mode in existing SLAM systems, but the method appears to be a specialized pipeline with limited coverage rather than a broadly demonstrated breakthrough.
-
SynthDemo-RL: Breaking the Zero-Reward Barrier in VLA Adaptation with LLM-Guided Synthetic Demonstrations
SynthDemo-RL uses simulator-privileged state to generate successful manipulation trajectories, distills them into a vision-language-action policy, and then applies PPO with sparse binary rewards. On perturbed LIBERO-PRO tasks where direct PPO leaves 17 of 27 initially failed tasks at zero success, the proposed pipeline reportedly rescues all 27 and achieves about 97% average success without new human demonstrations; it also reports results on RoboTwin and a physical-robot transfer test.
The compelling result is eliminating zero-reward failure cases through privileged-state synthetic demonstrations rather than additional teleoperation, but the evidence is still largely benchmark-specific and the core teacher-distillation idea may depend heavily on simulator access and task design.
-
Near-Optimal Acceleration for Smooth $\ell_p$ / $\ell_q$ Nondual Convex First-Order Oracle Optimization
This paper develops a first-order optimization method for Hölder-smooth convex functions over mismatched _p/_q geometries. It combines movement bounds from convex-set chasing with Hölder descent to obtain near-optimal convergence rates in high dimensions, including a rate of roughly T^{-(2κ-1)} for (_p,_q)=(1,2), or cubic decay T^{-3} in the standard smooth case. The authors claim this resolves a COLT 2015 open problem up to logarithmic factors.
The combination of convex-set-chasing geometry with acceleration and the claimed resolution of a long-standing open problem are substantial, but the result is primarily a specialized theoretical advance rather than a demonstrated practical optimization breakthrough.
-
LLaDA-PRM: A Bidirectional Step-Level Reasoning Evaluator
This paper studies process-reward models that judge individual reasoning steps. It finds that allowing a bidirectional evaluator to inspect later steps substantially improves step-level correctness judgments, with an 8B model outperforming a 34B causal baseline by 10–11 F1 points on two benchmarks and remaining useful for incomplete traces and training-data selection.
The controlled causal-versus-bidirectional comparison and large gains from using downstream context suggest a meaningful new direction for process supervision, but the evidence is still confined to benchmark evaluation and a relatively specialized evaluator setting.
-
ZIL: Zero-shot Image-to-LiDAR Registration
ZIL is a single vision-and-point-cloud foundation model for estimating camera pose from an image and a LiDAR scan, even when they come from different times or domains. It combines pose regression with 3D-coordinate prediction and finds that cross-dataset zero-shot performance requires normalizing camera intrinsics and the LiDAR vertical-axis origin; trained on seven datasets and 1.4M frames, it reports large gains over prior methods across in-domain and zero-shot tests.
Cross-domain, non-synchronized image–LiDAR registration with one model and reportedly very large zero-shot gains is a meaningful advance, but the abstract provides no detailed benchmark numbers or real-world deployment evidence to justify a strong verdict.
-
Are Coreset Selection Methods Worth Their Cost?
This paper evaluates coreset selection using an end-to-end wall-clock budget that includes both subset-selection time and downstream training, rather than comparing accuracy at a fixed subset size. Across datasets and selectors, class-balanced or repeated random sampling—and often simply training on all data for fewer epochs—beats sophisticated selection methods under equal budgets; selection is dominated by a full-dataset scan and only becomes worthwhile when subsets are reused. The authors also identify nine implementation issues in a common codebase, including one that changes Herding results by nearly six accuracy points.
The broad, auditable cost accounting produces a non-obvious negative result that challenges standard coreset evaluation, but it remains a benchmark and the conclusions are limited to the tested datasets, selectors, and training setups.
-
Algebraic Consistency Alone Does Not Certify Temporal Structure in Latent Action Models
This paper argues that additive composition and antisymmetry errors do not demonstrate that a latent action code captures temporal action structure. Reconstruction can make the code resemble a generic difference between frame features, satisfying the algebra even when temporal pairings are destroyed. Across five domains, unconstrained and pairing-destroyed models achieve similarly low errors, with no consistent downstream or action-decoding benefit from real temporal pairing; a proposed contrastive repair provides only marginal separation.
It presents a concrete, cross-domain falsification of a commonly used representation-quality certificate and proposes stronger validation tests, though the result mainly diagnoses a niche evaluation practice rather than demonstrating a new capability.
-
Stressor-Selective Sympathetic Preganglionic Modules for Organ-Biased Control
Using transcriptomics, activity mapping, tracing, and perturbation in mice, the authors identify spinal sympathetic preganglionic neuron subtypes that are recruited in stressor-specific combinations. A cold-responsive Nts+ population projects through specific sympathetic pathways, preferentially mobilizes fat in females, and is required for cold tolerance during fasting, suggesting that the spinal cord organizes organ-selective autonomic responses rather than relaying a generic stress command.
This is a substantive cell-type and circuit-level account of stressor- and organ-selective autonomic control with causal functional evidence, though the abstract does not establish how general the modules are beyond the mouse examples tested.
-
A single cortical circuit implements fast and slow visual search
Using intracranial EEG from 19 patients performing easy color and difficult orientation visual-search tasks, the authors find that both tasks engage the same cortical regions, frequencies, and processing hierarchy. Neural activity in the slower task appears to be a time-stretched version of the faster task, with alpha oscillations coordinating network timing and local high-frequency activity. The result argues that at least these two search modes may differ mainly in processing speed rather than relying on separate dedicated circuits.
This is a potentially important challenge to the longstanding idea of distinct fast- and slow-search circuits, supported by intracranial recordings and a specific temporal-scaling mechanism, but the evidence is limited to two tasks in 19 patients and the broader claim remains untested.
-
NaV1.7-dependent peripheral sensitization drives chronic pain in Parkinson's disease
Using a brain-restricted Parkinson’s disease model, the authors show that nigrostriatal degeneration causes lasting hyperexcitability in peripheral sensory neurons through dysregulated NaV1.7 sodium channels. Safinamide and interventions targeting NaV1.7 or its CRMP2 regulatory mechanism reversed or prevented pain-like behaviors, while human Parkinson’s DRG transcriptomes showed selective changes in excitability-related genes.
The paper presents a non-obvious brain-to-peripheral mechanism for Parkinsonian pain, supported by pharmacology, genetic prevention over 30 weeks, and human DRG data, but the main causal evidence remains preclinical and based on pain-like behaviors.
-
The cerebellum exploits a thalamic carrier frequency to refine motor behaviors
In behaving mice, the authors identify a 1–5 Hz delta rhythm in thalamic and cerebellothalamocortical activity during nonmoving periods. They show that CaV3.1-dependent resonance in thalamic neurons promotes transmission of this cerebellar rhythm to cortex, and that selectively disrupting CaV3.1 impairs motor learning more than motor execution. The work proposes that low-frequency thalamic oscillations act as a carrier for cerebellar signals during offline motor learning.
The combination of circuit recordings, closed-loop stimulation, dynamic clamp, and cell-type-specific knockout supports a non-obvious mechanistic link between a thalamic resonance frequency and cerebellum-dependent motor learning, though the result remains confined to a mouse circuit and a specific behavioral context.
-
FLAT: Resampling Image and Text into 1D Flexible-Length Aligned Transmodal Tokens for Retrieval and Generation
FLAT jointly trains a shared image-text encoder with text-to-image and image-to-text decoders, rather than freezing a representation model before generation. It represents both modalities as aligned variable-length 1D token sequences, using nested dropout so shorter or longer prefixes can serve as representations; the authors report competitive retrieval, captioning, and generation results, plus interpolation, latent arithmetic, and composed retrieval.
The unified, flexible-length representation that is directly usable for both retrieval and generation is a meaningful design direction, but the abstract gives limited comparative detail and the reported gains do not yet establish a major capability or efficiency breakthrough.
-
An Empirical Study of Harness Design for Coding Agents
The paper systematically varies planning, action-space, and context-management components in a fixed coding-agent harness, evaluating 176 configurations across four models on SWE-Bench Verified and Terminal-Bench 2.1. It finds that context management mainly prevents context overflow under tight budgets, rule-based elision followed by summarization is the most efficient strategy, planning helps weaker models’ accuracy but mainly saves cost for stronger ones, and bash-only interfaces can be cheaper for models with strong bash skills.
This is a relatively rigorous component-level study with useful model- and budget-dependent findings, though it remains an empirical harness ablation paper rather than a fundamentally new agent design or capability advance.
-
How Lossless Is Lossless Speculative Decoding? The Role of Numerical Precision in Orthrus
The paper reproduces Orthrus, a speculative-decoding system claimed to generate exactly the same token trajectory as its autoregressive backbone, and tests that claim under BF16 and FP32. With BF16, exact trajectory matching occurs on only 43–45% of 1,190 prompts, while FP32 matches every tested prompt; nevertheless, standard language-model benchmarks show no consistent degradation from the divergent trajectories.
This is a useful and somewhat surprising replication finding: numerical precision can invalidate an ostensibly lossless decoding guarantee even when downstream benchmark performance appears unchanged.
-
VABench: Measuring Embodied Spatial Intelligence through Visual Demonstrations, Active Perception, and Metric Control
VA-Bench evaluates whether vision-language models can complete an observe–reason–act–revise loop in physical manipulation, using RGB demonstrations, active camera selection, metric Cartesian commands, and execution feedback without privileged state or learned action heads. Across 14 task families, models localize targets well but achieve only 53.9% average task success; active viewpoint selection nearly doubles success in one comparison, while held-out geometry and long-horizon composition expose major failures.
This is primarily a benchmark, but it provides useful non-obvious evidence that active perception materially improves embodied MLLM control while localization ability does not translate into reliable execution, with substantial failures in geometric transfer and long-horizon tasks.
-
When EOS Tokens Disagree: Understanding Length Inflation in On-Policy Distillation
This paper analyzes why on-policy distillation can make student responses unnecessarily long or hit the generation limit. It identifies a mismatch between the EOS tokens preferred by the base student and post-trained teacher: merely sharing the same stopping-token set is insufficient, while treating equivalent EOS tokens as one semantic stopping action reduces the problem across Qwen3, Llama, and Gemma. The authors also find a separate late-training source of length inflation that termination handling does not explain.
It offers a concrete, cross-model diagnosis and mitigation for a poorly understood OPD failure mode, though the contribution is relatively narrow and the abstract provides no quantitative scale of improvement.
-
Steady-State Convergence of Stochastic Approximation
This paper develops a theory for the steady-state behavior of constant-stepsize stochastic approximation when the noise is Markovian and multiplicative, including cases where the mean update is locally nondifferentiable. It proves an optimal-rate Gaussian approximation in smooth settings and shows that nondifferentiability can increase the leading bias from order α to order √α; applications include Markovian linear stochastic approximation and asynchronous Q-learning, with a regime-agnostic bias-reduction method.
The unified reduction framework and the √α bias in locally nondifferentiable stochastic approximation address important gaps in existing constant-stepsize theory, but the abstract provides limited detail about the breadth and empirical impact of the results.
-
A Language-Guided Multimodal Foundation Model for Zero-Shot and Multi-Task Brain Signal Analysis
METIS is a multimodal brain-signal foundation model trained to align neural recordings with language, using more than 70,000 hours from 11,000 subjects and 20 datasets. The authors report zero-shot performance across 12 datasets that exceeds a generalist baseline by 20.9% on average, along with gains in few-shot learning and cross-dataset transfer, sometimes matching supervised task-specific models without fine-tuning.
The scale, language-guided alignment, and claimed zero-shot cross-task generalization are potentially important, but the abstract gives few details about signal modalities, baselines, leakage controls, and evaluation design, so the unusually large gains need verification before warranting a strong recommendation.
-
Quenched Ensemble Sampling
The paper introduces Quenched Ensemble Sampling, a nested-sampling-style particle method that replaces the hard energy constraint with a repulsive boundary potential, allowing gradient-based sampling while retaining a monotonic path through energy levels. Experiments claim it can cross first-order phase transitions where tempering methods stall, estimate marginal likelihoods for Bayesian neural networks, and traverse a high-dimensional lattice field theory while estimating its partition function.
The combination of quenched nested sampling with scalable gradient-based constrained sampling is a substantive and potentially broadly useful direction, but the abstract gives no quantitative comparisons or evidence that the claimed advantages persist beyond selected synthetic and physics examples.
-
Touch2Trace: Tactile-Driven Imitation Learning for Dexterous Cable Tracing
Touch2Trace trains a tactile-only imitation-learning policy for repeatedly pinching and feeding a cable through a dexterous hand. Using self-supervised pretraining for a 32×32 tactile sensor and a lightweight transformer policy, it improves mean tracing distance from 0.2 cm to 20.1 cm and success from 0% to 93% versus proprioception alone, including transfer to unseen cables and routing conditions.
The combination of tactile-only control, large real-world gains, and zero-shot transfer is notable, but the contribution is centered on a narrow manipulation task and an empirical recipe rather than a clearly general new method.
-
Constant Swap Regret in General-Sum Games via Two-Scale Higher-Order Optimism
This paper proposes uncoupled, deterministic learning dynamics for finite general-sum games that achieve swap regret bounded by a constant over time when all players use the method, rather than regret growing with the horizon. It combines the Blum–Mansour framework with optimism, a specialized normalization map, and a two-scale higher-order predictor; a switching wrapper also gives adversarial robustness with the usual square-root regret rate.
Horizon-independent swap regret in self-play for general-sum games is a genuinely non-obvious theoretical result, but the abstract provides no comparisons, lower-bound context, or detailed evidence to establish how broadly important the construction is.
-
Symbolic Separation: Grounding Deep Agents in Knowledge Graphs for Trustworthy Operational Data Analytics
The paper gives a tool-using analytical agent access to operational data only through an ontology-constrained virtual knowledge graph, with deterministic validation before queries execute. On 49.9 TB of supercomputer telemetry, this reportedly raises end-to-end task success from 43% to 86%, avoids silent integrity errors, and reduces token use by 2.4×, allowing a smaller local model to beat a larger non-symbolic baseline.
The large gain from enforcing semantic graph-level separation is notable and potentially broadly useful for reliable agentic analytics, but the abstract gives limited detail on task breadth, ontology construction, and comparison rigor, and the core idea is related to existing semantic-layer and neurosymbolic text-to-query systems.
-
Escape-Aware Control Barrier Functions for Quadrotor Safety under Body-Rate Limits
The paper argues that standard state-only control barrier functions can falsely certify quadrotor states as safe because they ignore the thrust direction and body-rate limits needed to reorient before braking. It derives the size of this certification gap, introduces an escape barrier using the current state and previously applied input, and reports perfect completion in 550 closed-loop test episodes, outperforming stopping-distance and backup-CBF baselines in directional margin and clearance.
The combination of a closed-form characterization of a previously overlooked safety gap and an analytically tractable escape-aware barrier is a substantive control contribution, but the evidence appears limited to quadrotor experiments whose real-world validation and generality are unclear.
-
METALICA: METAdynamics and repLICA exchange for enhanced diffusion sampling
METALICA combines metadynamics with replica exchange across the noise levels of a pretrained diffusion model to sample rare conformational states. Unlike batch-based sequential control, it uses long-running chains whose exploration improves with runtime rather than particle-memory budget; on a toy system and a protein-unfolding task, it finds an unfolded basin and a second free-energy minimum that a comparable sequential-control run misses.
The level-parallel replica-exchange formulation is a non-obvious way to turn diffusion sampling into long-chain rare-event exploration, but the evidence is limited to a toy target and one protein example, with no quantitative details in the abstract about breadth or accuracy.
-
Missing Bridges: Composition-Aware Active Imitation Learning
This paper proposes AALT, an active imitation-learning method that chooses expert demonstrations for their ability to connect many reusable behaviors, rather than only for local policy information gain. It builds a latent topology of hub states and bridge transitions, then plans through that topology with a diffusion policy; in a simulated 72-task UR5e retrieval domain, it reportedly reaches all tasks with only 3 additional demonstrations and 5 transitions, while a strong baseline reaches 88.6% after 20 demonstrations and 98 transitions.
The explicit focus on compositional value and bridge demonstrations is a meaningful departure from standard active imitation learning, with a striking demonstration-efficiency result, but the evidence is limited to one simulated task domain and the baseline and initial-data setup are not detailed.
-
Reaching Every Position Without Searching: Rotating Sparse Wiring on the Hypercube as a Substitute for Attention
The paper replaces dense attention with a fixed sparse communication pattern: sequence positions are mapped to hypercube vertices, and each layer connects along a different hypercube dimension. This guarantees that information from every position can reach every other in logarithmically many layers; on character-level language modeling, the sparse model reportedly matches or slightly beats dense attention while using far fewer links, parameters, and runtime. The authors also identify failed variants and argue that rotating through all dimensions, rather than the exact order, is the key factor.
The rotating hypercube wiring is a clean and potentially broadly useful alternative to attention with substantial claimed efficiency gains, but the evidence is limited to small character-level models and synthetic coverage tasks, so its importance for modern large-scale language modeling remains uncertain.
-
UMI-Bridge: Action-Anchored Latent Alignment across Human and Robot Manipulation Data
UMI-Bridge uses handheld human demonstrations as an action-supervised bridge between egocentric human video and robot manipulation data. Its latent model aligns examples by end-effector motion and gripper behavior rather than appearance, improving mean success from 73.3% to 91.7% in three real-robot tasks; it also reaches 85–90% on tasks learned from UMI data without task-specific robot demonstrations and can outperform a full-data robot-only baseline with only 25% as many robot demos.
The action-anchored intermediate-domain alignment and zero-task-specific-robot-demo transfer are genuinely interesting, but the evidence is limited to a few real-robot tasks and the abstract does not establish broad generality.
-
ForceDelta-VLA: Distilling Force-Conditioned ActionCorrections for Contact-Rich Manipulation
ForceDelta-VLA separates contact-dependent force corrections from a VLA’s task-level reference actions. It distills correction targets from a frozen force-conditioned teacher, then trains a lightweight policy to update actions from recent force history and robot state without regenerating full action chunks. On nine single-arm and bimanual contact-rich tasks, it reports 82.2% mean success versus 54.4% for ForceVLA, while reducing peak contact force by about 26% on successful trials.
The explicit, asynchronously applied correction policy is a meaningful approach to making force-aware VLA control more responsive and efficient, with a large reported gain, but the evidence is limited to the abstract and apparently a single experimental setup, so it falls short of a strong verdict.
-
DetAug: Obstacle-Blind Trajectory Augmentation for Zero-shot Obstacle Avoidance
DetAug augments free-space robot manipulation trajectories with obstacle-avoiding transit motions without using obstacle geometry, while labeling the augmentation parameters. At test time it samples these labels and executes the candidate with the lowest collision cost; on SafeLIBERO it reportedly improves collision-free success by over 20 percentage points over the next best method, and real-robot tests show it handles large detours better than inference-time steering.
The combination of obstacle-blind training augmentation and inference-time selection appears to provide a genuinely useful route to zero-shot avoidance, with unusually large benchmark gains and real-hardware evidence, but the abstract lacks absolute results and experimental detail needed for a strong verdict.
-
GroundingVLN: Reasoning and Acting with Grounding for Vision-Language Navigation
GroundingVLN connects language reasoning to navigation by grounding intermediate thoughts in image regions and producing a pixel-level spatial goal for a geometric planner. It introduces a dataset of temporally aligned grounded reasoning traces and an execution-aware reinforcement-learning method; the reported agent reaches 69.9% SR on R2R-CE and 75.1% on RxR-CE while using only 0.9% of the strongest baseline’s training data, with notably better cross-dataset transfer.
The shared grounding interface and especially the claimed dramatic data-efficiency and cross-dataset gains are genuinely interesting, but the abstract does not establish whether these results survive strong ablations and fair baseline comparisons.
-
WaveTLM: Reliable Time-Series Language Modeling through Task Compilation
WaveTLM treats time-series language modeling as a compilation problem rather than asking a language model to directly emit arbitrary answers. A task compiler converts the request and time-series evidence into typed task states, while task-specific executors produce valid tensors, labels, or structured records; on the proposed benchmark, this raises contract-valid outputs from 37.83% to 99.40% while reportedly preserving predictive performance across forecasting, imputation, classification, anomaly detection, and waveform analysis.
The typed compiler–executor framing and large reliability improvement are genuinely interesting, but the evidence is primarily an author-designed benchmark and the abstract does not establish how much of the gain transfers to realistic open-ended time-series use.
-
Using OCR Heads to Verbalize Image Semantics
The authors identify attention heads in four vision-language models that are causally important for OCR, then show these heads also verbalize semantic content from ordinary image regions. They derive a “verbalization lens” that exposes interpretable language-aligned concepts from image representations as early as layer 0, and use its inverse to causally edit visual concepts such as changing a tractor into a revolver.
This offers a notably concrete mechanistic account of how VLM image tokens become language-like, including causal head identification and concept editing, but the abstract provides few quantitative details about breadth, fidelity, or robustness.
-
How Model Growth, Recursion, and Boundary Operators Influence Scaling Exponents
The paper studies whether architecture, not just parameter count and data, can change pretraining scaling exponents. It finds that model-growth schemes based on looping—especially with weight sharing or expansion—can improve compute scaling, with a 7.4B growth model matching a 13B GPT-3 baseline on CORE using about 20× less compute; a simpler boundary operator gives a smaller benefit. The authors interpret this as making computational depth more usable, and report that increasing loop count with scale is compute-optimal in data-constrained multi-epoch training.
The reported change in scaling exponents and large compute-efficiency claim would be highly consequential if robust, but the abstract gives limited evidence about scale, baselines, and breadth beyond the cited benchmark, so it merits inspection rather than a strong recommendation.
-
Self-excited actuation enables adaptive and resilient flapping-wing flight
The authors build a flight-capable flapping-wing robot whose wings use insect-like asynchronous actuation: mechanical stretch activation sustains wingbeats without direct timing commands. In flight tests, the wings automatically adapt to changes in body resonance and react to collisions without sensing or control input, improving stability and performance in clutter compared with synchronous actuation.
This is a genuinely new embodied-control strategy for flying robots, with plausible real-world benefits from instant, sensorless reactions, but the abstract provides no quantitative results or detail on how large the gains are.
-
MiX: Micro-Inverted-Scaling for End-to-End Low-Bit Vision-Language Model Acceleration
The paper introduces MiX, a low-bit numerical format that reverses conventional microscaling: each value gets its own exponent while a group shares a mantissa. An adaptive MiX-MX scheme and custom accelerator reportedly match or outperform NVFP4 accuracy at 4.5 bits, while improving area efficiency by 25% and achieving 2.3–4.5× speedups with lower energy than a prior accelerator across several vision-language models.
The inverted scaling format is a genuinely distinctive hardware-aware quantization idea with potentially meaningful end-to-end gains, but the abstract provides limited detail about model coverage, accuracy margins, and baseline comparability, so the strong hardware claims warrant verification.
-
RotateIt! Fast and Reliable Single-Arm Garment Unfolding via Online-Adaptive Dynamic Rotation
RotateIt! uses dynamic axial rotation with a single robot arm to unfold garments, rather than repeated quasi-static moves or two-arm flinging. A grasp-ranking module chooses an effective anchor, while an online residual policy adjusts rotation speed and extent to time the release; simulation-trained policies transfer zero-shot to eight real garments, reaching 75.6% success and enabling subsequent autonomous folding.
The use of axial rotation as a primary single-arm garment-unfolding primitive is a distinctive capability with credible real-robot transfer, though the evidence is limited to a narrow manipulation task and abstract-level results.
-
Astronex-World 1.0: Real-Time Interactive World Model Foundation
Astronex-World is an open video world model that predicts controllable future frames from text or observations while incorporating camera motion, continuous actions, embodiment identity, and inserted text events. It provides both full-context and persistent causal generation, with KV caching and distillation enabling 832×480 at 24 fps in real time on a single L20 GPU; its reported benchmark scores are competitive with substantially larger models. The main contribution is a practical training and serving recipe for low-cost, interactive, action-conditioned video generation, though the abstract does not demonstrate downstream robot control or autonomous driving.
Real-time, action- and camera-controllable generation from a 5B model on one commodity GPU is a meaningful capability and efficiency result, but the evidence is mainly self-reported benchmark comparisons and lacks convincing real-world embodiment results.
-
Accelerating Sharded Data Parallelism at Scale with Federated Learning
The paper proposes splitting large sharded-data-parallel training jobs into loosely coupled groups, using FedAvg-style model aggregation between groups to reduce expensive cross-network communication. On a 512-A100 Llama 3.1 8B pretraining run, the authors report up to 8.04× higher data-processing throughput and 4.48 lower evaluation perplexity than conventional sharded-parallel baselines under the same hyperparameters.
The combination of federated aggregation with sharded parallelism could be a meaningful approach to scaling training across heterogeneous interconnects, but the abstract gives limited detail about baselines, convergence costs, and how broadly the large reported gains generalize.
-
Catch Me If You Can: Real-Time Feedback Denoising for Responsive VLAs
The paper adds high-frequency visual feedback to a diffusion-based VLA by keeping the final denoising step as a lightweight action-correction interface, instead of executing an entire action chunk open-loop. It reports similar performance to GR00T on static LIBERO tasks, but much higher success on dynamic simulation tasks (27.5% to 85.0%) and real-robot tasks (51% to 73%).
The two-timescale feedback design addresses a real limitation of open-loop diffusion policies and shows large dynamic-task gains, but the abstract provides limited detail about task breadth, baselines, and experimental controls.
-
Scaling Discovery through Test-Time Communication
The paper studies whether AI agents improve more by communicating during inference than by making independent parallel attempts. Using shared-directory communication without fixed roles, teams reportedly solve ARC-AGI-3 tasks at the rate of four times as many independent agents, with gains increasing at larger team sizes; similar benefits appear in polyomino packing and classifier compression, where four agents produced a 1,957-byte MNIST classifier with 99.4% accuracy. The advantage depends on having enough compute and clear progress feedback, so communication is not universally better.
The scaling result—that communication can make a small team outperform roughly four times as many independent attempts, including solving tasks no individual agent can solve—is a potentially important test-time inference finding, but the abstract does not establish how broadly it generalizes or rule out gains from additional effective search and coordination compute.
-
SWE-Proof: Can Language Models Resolve Real-World Issues with Machine-Checked Proofs?
The paper introduces Benchproofer, a pipeline that converts real SWE-bench issues into formally specified and machine-checked tasks, producing a 500-issue SWE-Proof benchmark. On two frontier models, formal verification finds counterexamples in roughly 25–50% of patches that pass tests, and a correct formal specification raises Opus 4.8 resolution from 85% to 95%; however, models generating their own specifications gain little because they often omit required behavior.
It offers a genuinely useful shift from incomplete test-based evaluation toward verified real-repository repairs and exposes specification faithfulness as the bottleneck, but the evidence is limited to two models and a newly constructed benchmark, so it is not yet a strong result.
-
Skel-WAM: A Hand-Skeleton-Conditioned World Action Model for Human-to-Robot Manipulation Transfer
Skel-WAM uses human hand-skeleton motion as a shared interface between human videos and robot demonstrations. Its visual and skeletal dynamics model learns from both sources, while a separate robot-trained action module converts the predictions into robot controls; across real and simulated bimanual tasks it reports substantial gains, including increasing success on unseen task variations from 38.89% to 86.11% with human-robot co-training.
The shared skeletal dynamics/action decomposition is a plausible new route for exploiting human video, and the large real-world gains are notable, but the abstract lacks enough detail about task difficulty, baselines, and ablations to warrant a strong verdict.
-
RAYA: Learning Where and When to Intervene for Robot Recovery
RAYA integrates a learned finite-horizon recoverability estimate directly into a constrained optimal controller, rather than reacting to predicted failures after the fact. It also learns to relax or reprioritize task objectives as recovery options diminish; the approach transfers across simulated disturbances and plant changes, and on a 35 g Crazyflie it completed 10/10 six-cycle missions under wind and severe model or motor-command failures, while three baselines failed every trial.
The controller-internal treatment of recoverability plus learned task reprioritization is a promising and fairly novel recovery direction with compelling small-scale hardware results, but the evidence is limited to 40 flights and a narrow set of scenarios, making a strong verdict premature.
-
NeuRIO: A Streaming Neural Estimator for Zero-Shot Sim-to-Real Multi-Robot Relative Inertial Odometry
NeuRIO is a streaming graph-based neural estimator for relative 6-DoF motion among multiple robots using inter-robot bearing/range measurements and IMUs. It is trained entirely in simulation with varied, persistent sensor corruptions, then transfers to real data without adaptation, achieving 14.1 cm position RMSE and 3.9° rotation RMSE across 24 sequences. Its shared graph operators support changing team sizes and observation graphs, with simulated updates remaining under 20 ms for teams of up to 400 robots.
The combination of zero-shot sim-to-real relative odometry, changing multi-robot graph sizes, and large-team computational scaling is genuinely interesting, but the abstract gives limited comparative and real-world evidence for judging how broadly meaningful the claimed gains are.
-
ZeroTouch: Tactile-Supervised Visual Contact Estimation for Contact-Rich Manipulation
ZeroTouch uses tactile measurements only during training to teach a wrist-camera model to infer dense contact deformation, six-axis forces/torques, and an appropriate grasp compression target. It substantially lowers normal-force prediction error and reports 80–95% success across several physical distribution shifts without tactile hardware at deployment, though the real-world evaluation is small.
The potentially important idea is transferring tactile supervision into vision so contact-rich manipulation can retain force-aware behavior without deployment-time tactile sensors, but the evidence is limited to 20 trials per condition and relatively narrow comparisons.
-
PSR: Predictive Sensorimotor Representation Learning for Contact-Rich Manipulation
The paper trains a multimodal Transformer to predict future contact dynamics from force, configuration, and interaction-history signals, then feeds hierarchical predictive representations into a visuomotor policy. Its PSR-VLA system reaches 91.7% success across six real-world contact-rich manipulation tasks, outperforming reported ForceVLA and π0.5 baselines by 19–30 percentage points.
The large real-world gains and active prediction of future contact dynamics make this more than a routine force-conditioning method, but evidence is limited to six tasks and the abstract does not establish how broadly the approach generalizes beyond the tested setup.
-
VLPSA: Vision-Language-Poisson-Safe Actions for Full-Body Safety of Learned Policies
VLPSA adds an online safety filter to vision-language-action robot policies without retraining them. It builds Poisson Safety Functions from perception and enforces full-body collision constraints—including grasped objects—through a control-barrier-function quadratic program, using multiple spatial resolutions for real-time operation. On SafeLIBERO it raises collision avoidance from 23.1% to 91.2% while retaining or improving task success, and it is demonstrated on a Franka robot around clutter, moving obstacles, and human interference.
The combination of online Poisson-derived barrier functions, full-body/object safety, and no VLA retraining addresses a real deployment bottleneck with a large reported collision-avoidance improvement, but the abstract provides limited detail on robustness, computational cost, and the breadth of physical evaluation.
-
Becoming a Fruit Ninja: Real-Time Probabilistic Kinodynamic Planning for Manipulator Projectile Interception
FRUITNINJA is an anytime GPU-based sampling planner for robot-arm interception of thrown objects, where the robot must satisfy both timing and blade pose/velocity constraints. It uses dynamically feasible cubic trajectories, parallel travel-time searches, and risk-aware ranking under uncertainty; in a calibrated Franka simulator it achieved 96.7% success for unobstructed tosses and 68.3% with five obstacles, substantially outperforming six baselines.
The combination of GPU-batched kinodynamic planning and uncertainty-aware interception targets a genuinely difficult real-time manipulation capability, but the evidence is limited to simulation rather than physical robot trials.
-
PixelART: Image-to-Layer Decomposition without Latents or Text-to-Image Pretraining
PixelART models image-to-layer decomposition directly in pixel space with a rectified-flow Transformer, rather than adapting a large text-to-image model or using an RGBA latent autoencoder. It adds timestep sampling that emphasizes the high-noise phase where layer assignment is decided, and reports comparable or better decomposition and reconstruction with over 80% fewer parameters, 98% lower latency, and 85% lower memory than Qwen-Image-Layered. The results also suggest that T2I pretraining contributes little to this specialized task.
The combination of a from-scratch pixel-space model, task-specific timestep analysis, and very large claimed efficiency gains is genuinely interesting, but the abstract provides limited absolute results and the advance is demonstrated on a relatively specialized image-editing task.
-
SmoLSTM: A Compact Vision-Language-Action Model with Recurrent Memory that Persists
SmoLSTM adds persistent recurrent memory to a compact vision-language-action policy, using a frozen 256M-parameter vision-language model and a small trainable LSTM control layer instead of expanding the observation history. Trained on 7,461 demonstrations spanning 140 tasks, it reaches 77.5% full-task success on a memory-focused LIBERO benchmark and 79.6% on standard LIBERO; resetting its memory drops success to 7.0%.
Persistent O(1)-storage episode memory in a compact VLA, together with the dramatic ablation drop and competitive multi-task results, is a genuinely interesting direction, though the evidence is confined to benchmark experiments and the underlying recurrent mechanism is not entirely new.
-
LLMs Anchor on Chief Complaint and Fail to Integrate Evidence in Sequential Clinical Triage
The study evaluates six LLMs on emergency-department triage as they receive a nurse–patient conversation one turn at a time, rather than only seeing the completed record. Across simulated and physician-authored conversations, model agreement with ESI labels falls sharply and remains poor: models anchor on the chief complaint, fail to integrate later evidence, and cannot be rescued by prompting or ensembling, while clinicians perform much better.
The sequential evaluation exposes a clinically important failure mode that retrospective benchmarks hide, with controlled perturbation results and a large clinician–model gap, but the evidence is limited by the small clinician-authored set and simulated data.
-
SparkDiffusion: Mitigating the High-Sparsity Trap --- A Unified Framework for up to $265\times$ Single-GPU Acceleration of Visual Generation
The paper argues that extremely sparse attention in video diffusion hits a “high-sparsity trap”: optimizing per-denoising-step losses can improve while final video quality stops improving or worsens. It proposes staged training—sparse architectural warm-up followed by terminal-aligned trajectory distillation—combined with FP8 and fused kernels, reporting up to 97% attention sparsity and claimed 265× end-to-end speedup for 720P generation on one RTX 5090.
The terminal-aligned explanation and combination of extreme sparsity with few-step distillation could be important, but the abstract provides no quantitative quality comparisons or details sufficient to validate the unusually large speedup claims.
-
Machine-Interpretable Information: Compiling Documents into Searchable and Readable Protocol States
The paper proposes compiling documents into a fixed 56-token neural state that supports retrieval, reasoning, and later reconstruction, rather than repeatedly feeding raw document text to a language model. A Writer produces the state and a Translator maps it into the embedding space of different frozen readers; on HotpotQA, a residual version reportedly reaches higher Exact Match than full-context processing at about 7% of the attention FLOPs, with some evidence of transferable entity representations.
The transferable, fixed-bandwidth document-state protocol is a genuinely nonstandard direction with potentially large efficiency benefits, but the evidence described is mainly one benchmark and abstract-level claims, so it is not yet strong enough to prioritize over better-validated work.
-
TaskAnchor: Grounding Task State in Reactive VLAs for Long-Horizon Manipulation
TaskAnchor augments pretrained reactive vision-language-action models with execution history to distinguish visually similar states that require different actions at different points in a task. It uses history-conditioned visual features and a scalar milestone-based task-stage signal, reportedly improving success by about 4.9–5.5× over π0.5 and X-VLA on RMBench, with gains on another benchmark and real robots at only 2.08 ms added latency per action chunk.
The task-state-aliasing framing and lightweight history grounding address a real weakness of reactive VLAs, and the large reported gains plus low latency are notable, but the abstract gives no absolute success rates, task breadth, or details establishing that the comparisons are robust.
-
Smoothed Analysis of Inconsistent A*
This paper studies why A* with inconsistent heuristics often works well despite worst-case exponential node re-expansions. By adding small random perturbations to edge weights, it proves a polynomial expected bound on the total number of iterations, and extends the analysis to Dijkstra’s algorithm on graphs with negative edge weights.
The first smoothed-analysis framework for inconsistent A* could explain a longstanding theory–practice gap, but the abstract gives no details about the perturbation model, assumptions, or how broadly the polynomial bound applies.
-
ARPP21 defines a TDP-43-independent aggregation pathway in amyotrophic lateral sclerosis
The study identifies ARPP21, an RNA-binding protein genetically linked to ALS, as forming fibrillar aggregates that are spatially and mechanistically separate from the better-known TDP-43 pathology. ARPP21 variants promoted phase separation, impaired motor-neuron viability, and induced tunnelling nanotubes that could support aggregate spread, suggesting a previously unrecognized ALS aggregation pathway.
The proposed TDP-43-independent pathology is a genuinely interesting and potentially important reframing of ALS heterogeneity, but the abstract gives limited information about human sample sizes and independent validation, so the evidence is not yet strong enough for a strong verdict.
-
An electronically steerable epidural ultrasound interface for deep brain neuromodulation in freely moving rats
The authors developed a small implantable epidural ultrasound array integrated with a custom ASIC that can electronically steer focused ultrasound to different deep-brain locations. In awake, freely moving rats, it stimulated the ventral tegmental area and increased dopamine release in the nucleus accumbens, with no inflammatory markers reported but some mild tissue damage from implantation.
This is a genuinely interesting hardware direction for chronic, steerable deep-brain neuromodulation in freely moving animals, but the evidence is still a proof of concept based on one target and limited physiological validation, with implantation-related tissue damage.
-
Divergence of Melanopsin's Intrinsic Properties and Downstream Actions
This study shows that melanopsin-driven cellular and behavioral responses do not always track the wavelengths that melanopsin absorbs most strongly. It argues that persistent melanopsin activity causes adaptation: strongly absorbed wavelengths can desensitize downstream responses, while weakly absorbed wavelengths may produce stronger net effects, and a quantitative model captures this behavior.
The proposed adaptation mechanism could revise how melanopsin stimulation is translated into physiological and behavioral effects, but the abstract provides no quantitative result details or scope sufficient for a stronger recommendation.
-
Brain organoid computing for robotic decision-making
The paper uses a cultured brain organoid as a reservoir-computing controller in a closed-loop humanoid robot, mapping sensory inputs to motor actions. The prototype reportedly performs real-world object grasping and laser-chasing tasks and shows cross-task adaptation, but the abstract gives no quantitative comparison, scale, or details about how much of the decision-making is performed by the organoid.
Using a brain organoid as a high-level controller for an embodied robot is an unusually novel direction, but the proof-of-concept claims are too weakly quantified to justify a strong recommendation.
-
Inter-regional interactions uncouple within-region inhibition stabilization and paradoxical responses
The paper analyzes multi-region excitatory–inhibitory network models to ask whether paradoxical responses reliably indicate local inhibition-stabilized dynamics. It finds that long-range coupling can create paradoxical responses in regions that are not locally inhibition-stabilized, or suppress them where local stabilization is present, so local perturbation responses do not uniquely reveal local circuit properties.
This is a theoretically important challenge to a common interpretation of paradoxical responses, but the evidence is limited to analytically tractable models and simulations rather than biological data or experiments.
-
High-Frequency Focused Ultrasound targeting the midbrain in the Guinea Pig Induces Activation and Plasticity of the Efferent System, and Protection Against Noise-Induced Hearing Loss
The study uses 20 MHz focused ultrasound aimed at the inferior colliculus in anesthetized guinea pigs and finds that its effects may be mediated largely through indirect cochlear and auditory-pathway activation rather than direct midbrain stimulation. Stimulation engaged medial olivocochlear efferent circuits, produced effects lasting at least two hours, and nearly prevented noise-induced hearing loss when delivered before acoustic trauma; gentamicin blocked these effects.
The combination of noninvasive ultrasound activation of auditory efferent circuitry and dramatic protection from acoustic injury is genuinely notable, but the evidence is limited to anesthetized guinea pigs and the proposed mechanism appears indirect rather than direct midbrain neuromodulation.
-
NeuroGate: waveform translation between brain
NeuroGate predicts the neural waveform in one brain region from recordings in another, rather than only estimating connectivity or shared latent variables. It is evaluated across 56 pathways and several signal types, and a toxin-based perturbation in the piriform–olfactory bulb circuit selectively disrupts predictions in the anatomically expected direction. The framework is also used to estimate recording redundancy and propose a perturbation-based test for directional translation.
The combination of cross-region waveform reconstruction with causal perturbation validation is a non-obvious and potentially useful direction, but the abstract gives no quantitative accuracy results or details sufficient to establish how much it improves over existing predictive models.
-
No benefit of subsampling in ensemble coding
The paper revisits whether visual ensemble perception is based on integrating all items or sampling only a subset. It argues that subsampling appears effective only when each item retains constant encoding precision as set size grows; under a fixed total resource budget, resource-rational full integration matches or exceeds subsampling with less total information, and fits several existing and new datasets equally well.
This directly challenges a prominent explanation of ensemble coding with a resource-constrained analysis and supporting behavioral-model comparisons, but the abstract provides limited detail on the breadth and strength of the empirical evidence.
-
SERINC1-dependent lipid preservation gates neocortical synaptogenesis and remote memory
The study identifies SERINC1 as an activity-induced lipid regulator in the anterior cingulate cortex during systems memory consolidation. In mice, disrupting Serinc1 reduced ACC synapse formation and selectively impaired remote memory, while restoring it rescued both; lipidomics suggests protection of DHA-containing phosphatidylserine and AA-containing phosphatidylinositol is involved.
It proposes a specific, localized neuro-lipidomic mechanism linking cortical synaptogenesis to remote memory and includes a targeted rescue, but the abstract provides limited quantitative and methodological detail and the result is still a bioRxiv preprint.
-
Countering Neural Activity Drift: Sustained Long-term Seizure Prediction Using an Evolutionary Machine-Learning Framework on Continuous Intracranial EEG
The study evaluates seizure prediction on nearly 665 hours of continuous intracranial EEG from 16 patients, rather than on artificially balanced offline segments. Offline AUROC around 90% fell to 57.9% in pseudo-prospective testing because of neural activity drift, but an evolutionary online model search predicted every seizure in 75% of patients and all but one in 87.5%; selecting channels from estimated epileptogenic networks also reduced contacts by about 79%.
The continuous online evaluation exposes a major offline-to-real-world performance gap and combines adaptation with substantial channel reduction, but the retrospective pseudo-prospective design and limited reporting of false-alarm and clinical utility metrics make the strong deployment claims premature.
-
Odour cueing during sleep augments trauma-focused psychotherapy
This study paired trauma-focused psychotherapy with repeated presentation of a therapy-associated odour during sleep. In healthy participants, cueing increased slow-wave activity and slow-oscillation effects, while in a randomized PTSD trial it was associated with faster and greater symptom improvement, especially for re-experiencing and negative mood/cognition.
The combination of sleep-dependent odour cueing with psychotherapy and convergent mechanistic and clinical findings is genuinely interesting, but the abstract gives no sample sizes or effect sizes, so the strength and reproducibility of the clinical benefit are unclear.
-
Mind2Dialogue: Training Human-Aware Language Models by Simulating User Mental States
Mind2Dialogue simulates users’ evolving beliefs, goals, and personal traits, then uses an oracle assistant with access to those hidden states to generate training supervision. Models distilled from this supervision reportedly improve personalization and theory-of-mind evaluations, including 26.6–40.9 percentage-point gains in preference-following over several instruction-tuned baselines, with some transfer to belief and action reasoning.
The shared latent mental-state simulator and privileged distillation setup are a meaningful approach to the supervision gap, but the evidence appears to rely mainly on internally generated conversations and synthetic evaluation, so it is not yet clear that the gains reflect real users or robust human-aware collaboration.
-
RetireOPD: Self-Retiring On-Policy Distillation for Agentic Reinforcement Learning
RetireOPD trains a privileged, skill-conditioned teacher with environment rewards, then uses token-level on-policy distillation to help a skill-free student before adaptively retiring the teacher when further supervision stops helping. On ALFWorld and WebShop, across Qwen2.5 models from 1.5B to 7B, it reports 11.8–19.0% gains over RL baselines and says the student consistently surpasses its teacher.
The adaptive, performance-based retirement of a privileged teacher is a plausible useful refinement, and the reported gains span two agentic tasks and several model sizes, but it remains an incremental training-recipe contribution without evidence yet of broader generality or a fundamentally new capability.
-
ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments
ScienceIDE converts scientific software repositories into executable training and evaluation environments, using expert-defined tasks and correctness criteria to generate verified interaction trajectories. The authors train several PhAI-IDE models on these environments and report improvements on held-out scientific-code repair plus selected code, reasoning, and knowledge benchmarks, suggesting some transfer beyond the scientific domain.
The repository-to-verified-environment pipeline could be a useful direction for training scientific coding agents, but the abstract gives no quantitative results, scale, or comparison details to establish that the claimed transfer is substantial.
-
AgentKV: Phase-Aware KV Eviction for Agentic LLMs
AgentKV is a KV-cache eviction method for agentic LLM serving that keeps separate query representatives for different generation phases—such as thinking, tool use, and action—instead of relying only on recent tokens. Across two models, six task domains, and three cache budgets, it reportedly improves task scores by about 5 points over prior eviction methods and raises output-token throughput by up to 1.8× versus full-KV SGLang, while supporting compressed KV state across turns.
The phase-specific attention analysis and persistent multi-turn cache design address a real weakness of recency-based eviction, with broad-enough reported evaluations and substantial throughput gains, though the core technique remains a targeted extension of existing KV-cache compression methods.
-
Nearly Minimax-Optimal Regret for Linear Contextual Bandits with Arbitrary Adaptive Action Sets
This paper derives nearly matching minimax regret bounds for stochastic linear contextual bandits when the available action menu can adapt to the interaction history and even depend on the unknown parameter. It improves the usual dependence on menu size from roughly $\sqrt{K}$ to $K^{1/4}$ when $K\le d$, and for large menus improves the standard $d\sqrt{T}$ scaling to $d^{3/4}\sqrt{T}$ over a broad regime, with matching lower bounds up to logarithmic factors.
The paper gives a genuinely sharper, nearly minimax characterization for a difficult adaptive-action-set model, including lower bounds that establish the improved dimension and menu-size dependence is intrinsic, though the result is primarily theoretical and specialized.
-
Strong and Compact Policies for Submodular Markov Decision Processes via LP-Based Submodular Orienteering
The paper develops an LP-based approximation algorithm for Submodular Orienteering using Sherali–Adams relaxations and Round-or-Cut, then extends the approach to Submodular Markov Decision Processes with stochastic transitions. It improves the polynomial-time approximation from linear in the horizon to roughly $H^{\varepsilon}$ for any fixed $\varepsilon>0$, while also identifying a trade-off between approximation quality and how much history policies use.
The substantial improvement from a horizon-linear approximation to subpolynomial dependence, together with a new LP-based framework that handles stochastic submodular MDPs, makes this worth a closer look, though it is primarily a theoretical approximation result rather than a demonstrated capability jump.
-
Near-Optimal Nonconvex Matrix Completion
This paper gives global recovery guarantees for nonconvex Riemannian gradient and Gauss–Newton methods for low-rank matrix completion. With a multiscale initialization and joint control of spectral error and incoherence, the methods recover an incoherent rank-r matrix using nearly optimal O(μnr) samples up to logarithmic factors, while RGN achieves eventual quadratic convergence. The main advance is removing the extra polynomial dependence on rank present in prior global analyses of common nonconvex methods.
Closing a long-standing sample-complexity gap for globally guaranteed nonconvex matrix completion is a substantive theoretical advance, but its practical impact and comparison with the strongest existing guarantees are not established in the abstract.
-
TIO-Former: Ultra-Lightweight 6-Directional ToF-Inertial Odometry for Nano-UAVs via a Streaming Causal Transformer
The paper presents a camera- and LiDAR-free 6-DoF odometry system for nano-UAVs using six orthogonal 8×8 ToF arrays and an IMU. Its causal transformer uses bounded local and compressed temporal memory, and real-flight tests report substantially lower position error than optical-flow and inertial baselines while running at 10.5 ms P95 latency and 6.3 MiB of memory on an edge RISC-V computer.
The combination of sparse six-direction ToF sensing, inertial feature routing, and constant-memory streaming inference addresses a real nano-UAV constraint with credible deployment results, but it is a specialized odometry system rather than a clearly field-changing advance.
-
DRT&R: Direct Radar Teach & Repeat
The paper presents a radar-only teach-and-repeat navigation system that keeps dense radar scans in local maps, preserving viewpoint-dependent artifacts that global maps tend to smooth away. Across 344 km of on-road and 20 km of off-road data, it reports localization within 4 cm in most conditions and 12 cm in sparse or geometrically degenerate areas, plus more than 10 km of real-time closed-loop operation on an off-road vehicle.
The combination of direct radar processing, local viewpoint-specific mapping, and demonstrated closed-loop off-road operation is a meaningful advance over standard radar localization, but it appears to be a strong systems integration result rather than an obviously new research direction.
-
REACT: A Fully Spiking State-Space Model for Real-Time Event-Driven Temporal Perception
REACT is a fully spiking state-space model that processes event-camera measurements individually, using the actual inter-event timing to update a complex-valued neuron state rather than grouping events into frames or time bins. It achieves nearly the best reported time-to-collision accuracy on EvTTC with 4.6 ms latency and no target-location input, while also supporting anytime prediction, sequence transfer, and INT8 quantization that substantially lowers estimated energy use.
The combination of event-by-event continuous-time spiking dynamics, competitive TTC accuracy without a target prior, and very low latency is a meaningful advance for reactive robotics, but the evidence is limited mainly to a small set of event-vision tasks and one benchmark.
-
ForwardDLO: Model-Based Bimanual Shape Matching of Unconstrained Deformable Linear Objects
ForwardDLO is a recurrent latent dynamics model for predicting how an unconstrained rope deforms when manipulated by two robot arms at arbitrary grasp points. Its per-segment predictions avoid expensive explicit segment-to-segment computation, allowing 8–22 times more candidate actions to be evaluated within the same planning budget; it achieves 13% lower open-loop prediction error than the strongest learned baseline and raises simulated routing success to 98%, while matching baselines on real-world shape matching.
The combination of accurate rope dynamics with much cheaper batched action evaluation is a meaningful planning advance, but the strongest 98% result is simulated and the demonstrated real-world capability remains relatively limited.
-
VLA-ULAP: Interleaving Cloud VLA Calls with Ultra-Lightweight Local Action Prediction at the Edge
VLA-ULAP uses a small local action predictor to fill in behavior between occasional calls to a large remote vision-language-action model. Its roughly 7.4M-parameter predictor runs in about 20 ms on a Jetson Orin Nano, cuts VLA calls by 49–77%, and maintains 95–98% of baseline success in simulation; physical SO-101 tests report similar success retention with roughly 48–62% lower latency or energy. The approach also improves some dynamic-task results by enabling faster local responses.
The cloud-VLA/local-policy interleaving design addresses a real deployment bottleneck and is supported by both simulation and hardware results, but it is an efficiency and systems advance rather than a clearly new capability or assumption-breaking result.
-
Open ultrasound foundation model for robust segmentation and clinical measurement across heterogeneous settings
The authors combine nearly 457,000 ultrasound images and 1.6 million expert masks from 53 datasets into SonoCorpus, then pretrain SonoBase for interactive segmentation across 24 clinical applications. It reportedly generalizes across organs, devices, operators, and countries, outperforming general and medical segmentation baselines, while producing clinically usable ejection-fraction and fetal measurements; five examples can adapt it to new settings. The checkpoints, training artifacts, splits, and code are released openly.
The unusually broad ultrasound-specific pretraining and demonstrated cross-device, cross-country robustness make this more than a routine medical segmentation model, but the abstract does not provide enough comparative detail to establish a field-changing advance.
-
Higher-order pruning of experts in mixture-of-experts language models
The paper introduces HOPE, a second-order method for pruning experts in mixture-of-experts language models that accounts for interactions between experts rather than treating their contributions as independent. Across three models up to 122B parameters and several task types, including agentic coding, it reportedly preserves quality better than existing pruning methods, especially at 50% pruning, with gains of up to 6.1% on agentic coding.
Accounting for cooperative expert interactions is a meaningful extension with potentially important memory savings, and the evaluation is broad, but it remains a specialized compression improvement rather than a clearly field-changing result.
-
Block Parallelism For Efficient Distributed Long-Context Diffusion Language Model Training
The paper introduces block parallelism for training block diffusion language models: ranks are assigned to target blocks, and context-sharded block parallelism additionally shards the shared clean prefix to reduce communication and memory. On long-context training, it reports 1.18–1.45× speedups at 256K tokens, up to 1.61× at 512K, and much larger gains for speculative-decoder training at 512K–1M contexts while maintaining or improving downstream pass rates.
This is a concrete new distributed-training dimension with substantial long-context speedups, though its importance is currently specialized to block diffusion models rather than broadly applicable to language-model training.