Issue 22 · May 25–31, 2026
Every candidate
All 4,449 papers were scored from their abstracts by gpt-5.6-luna; 1,699 were not skipped. Shown here: the top 150 of those, in score order. Picks are marked.
-
Extreme dynamic symmetry enables omnidirectional and multifunctional robots
The paper introduces “dynamic symmetry,” designing a robot so its actuators can produce nearly uniform center-of-mass accelerations in every direction. Simulations across more than 1,000 morphologies and a physical 20-leg spherical robot suggest that this property improves tracking, robustness, energy efficiency, self-stabilization, and tolerance to actuator failures; the robot also demonstrates orientation-invariant movement and interaction while moving through difficult terrain. The main novelty is treating isotropic attainable dynamics—not just geometric symmetry—as a primary robotics design objective.
This presents a genuinely new and broadly applicable robot-design principle, backed by large-scale morphology studies and a physical system demonstrating capabilities that directly follow from near-extreme dynamic isotropy.
-
Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet
The authors train sparse autoencoders on the residual stream of Claude 3 Sonnet, extracting up to 34 million features that often correspond to recognizable concepts, including entities, sarcasm, coding errors, deception, and sycophancy. These features generalize across languages and even to images, and manipulating them can causally steer model behavior, though the feature set is incomplete and its faithfulness is not yet rigorously established.
This substantially scales mechanistic interpretability from small toy models to a production-scale multimodal-capable LLM and provides causal evidence that extracted semantic features influence behavior, despite important validation gaps.
-
Ultrasensitive voltage imaging reveals distinct electrical microdomains in neurons
The authors introduce ASAP7y, a genetically encoded voltage indicator designed for sensitive, subthreshold measurements with two-photon microscopy. Using it in Drosophila neurons, they map millisecond-scale voltage propagation along neurites, combine these measurements with EM-based reconstructions of 717 cell types, and find that morphology produces highly diverse electrical transformations and locally distinct feature-selective subcellular domains.
This combines a potentially enabling measurement advance with broad cell-type-scale analysis and evidence that single neurons perform parallel, spatially localized computations rather than acting as uniform integrators.
-
Crazyflow: An Accurate, GPU-Accelerated, Differentiable Drone Simulator in JAX
Crazyflow is an open-source JAX-based drone simulator that combines high physical fidelity, differentiability, GPU acceleration, and large-scale swarm simulation. It reports more than 10× higher single-drone speed than prior simulators, supports thousands of 4,000-drone swarms, and enables real-world demonstrations including sub-centimeter tracking and a recovery policy trained in 0.38 seconds while a drone was airborne. The main novelty is using this simulation speed for online, in-execution learning rather than only offline train-then-deploy workflows.
The combination of differentiable, accurate simulation at extreme parallel scale with a credible in-flight-from-scratch recovery-learning demonstration is an unusually large and assumption-challenging step for drone learning systems.
-
Why Far Looks Up: Probing Spatial Representation in Vision-Language Models
The paper analyzes whether vision-language models represent 3D spatial relations or exploit visual shortcuts, using minimal contrastive pairs and a new synthetic benchmark. Across model families, it finds that models entangle vertical image position with distance—a perspective bias that becomes stronger with scaling despite improved benchmark scores—and that better-separated spatial representations predict robustness.
The finding that scaling can improve spatial benchmark accuracy while worsening an intrinsic vertical-distance shortcut, together with a representation-level diagnostic and controlled benchmark, is substantially more informative than another benchmark or architecture tweak, though the abstract provides no quantitative scope for the claimed generality.
-
Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments
Qwen-VLA extends a vision-language model with a diffusion-transformer action decoder and trains it jointly on manipulation, navigation, human demonstrations, simulation, and trajectory data. It uses textual embodiment descriptions to condition one model across robot platforms and reports strong results across simulated and real-world manipulation, navigation, and out-of-distribution settings, including some zero-shot dynamic manipulation.
The broad unification of tasks, embodiments, and action formats is potentially important, but the abstract does not establish how much of the gains come from the model design versus scale and data, nor provide enough baseline or ablation detail for a strong recommendation.
-
SCOPE: Self-Play via Co-Evolving Policies for Open-Ended Tasks
SCOPE trains language models on open-ended tasks without curated prompts or an external judge by co-evolving a Challenger that creates document-grounded tasks and a Solver that answers them. A frozen initial model generates rubrics and evaluates responses; across three 7–8B models, the method reports gains of up to 10.4 points on eight open-ended benchmarks and up to 13.8 points on seven held-out short-form QA benchmarks, while matching or beating training on about 9,000 curated prompts.
The co-evolution of task difficulty with self-generated rubrics is a substantive approach to removing both curated data and external judges, and the reported cross-model and held-out-task gains are promising, but the abstract does not establish whether the results generalize beyond the tested benchmarks or how robust the self-judging process is.
-
ScientistOne: Towards Human-Level Autonomous Research via Chain-of-Evidence
The paper proposes Chain-of-Evidence, a system for making autonomous research agents link each claim to verifiable sources, code, and experimental results, and introduces an audit covering citations, scores, specification compliance, and method–code consistency. In an evaluation of 75 papers across five systems and five research tasks, existing agents showed substantial integrity failures, while ScientistOne reported no fabricated references, perfect score verification, and stronger method–code alignment; it also claims competitive results on several additional tasks.
The evidence-centered framing and measured finding that apparently strong autonomous research outputs routinely contain serious reproducibility failures are genuinely noteworthy, but the broad capability claims and system comparisons need closer scrutiny beyond the abstract.
-
MobileMoE: Scaling On-Device Mixture of Experts
MobileMoE develops sparse mixture-of-experts language models with 0.3–0.9B active parameters for smartphone deployment, optimizing expert structure under mobile memory and compute limits. The models reportedly match or beat dense on-device LLMs using 2–4× fewer FLOPs, and smartphone profiling shows 1.8–3.8× faster prefill and 2.2–3.4× faster decoding at comparable INT4 memory. The main new result is that moderately sparse, fine-grained MoE designs can be advantageous even below one billion active parameters, rather than only at very large model scales.
This is a credible and potentially important demonstration that MoE sparsity improves real smartphone inference at small model scales, with substantial reported speedups, though the abstract does not establish how broadly the scaling law or gains generalize.
-
AI Research Agents Narrow Scientific Exploration
The authors generate 219,655 research ideas using five AI-agent frameworks and five language models, then compare them with human-authored and later follow-on research. They find that current agents cluster near the starting literature, align less with future human work, and occupy lower-impact parts of the historical research space, suggesting they mostly elaborate locally rather than expand scientific exploration.
This is a large-scale empirical challenge to the assumption that research agents naturally broaden the search for ideas, though the abstract does not establish how robust the comparisons or impact measures are.
-
VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion
VideoMLA applies Multi-Head Latent Attention to causal video diffusion, replacing per-head KV caches with a shared low-rank content latent and a shared positional key. It cuts cached KV memory by 92.7% while retaining short-horizon quality and achieving the best reported long-horizon VBench score among the tested methods, with a 1.23x throughput gain on one B200. The notable analysis is that pretrained video attention is not actually low-rank; the learned MLA bottleneck itself sets the usable rank and can preserve quality despite poor spectral reconstruction.
This is a meaningful memory-efficiency direction for minute-scale video generation, and the finding that MLA works despite non-low-rank pretrained attention is genuinely non-obvious, but the abstract reports limited hardware and benchmark evidence and only a modest throughput improvement.
-
Colored Noise Diffusion Sampling
The paper proposes a training-free diffusion sampler that injects frequency-dependent, timestep-dependent colored noise instead of uniform white noise. It uses the observation that diffusion models typically establish low-frequency structure before high-frequency detail, and reports sizeable FID improvements across SiT, JiT, and FLUX without retraining, including 8.26→6.27 on ImageNet-256 for SiT-XL/2.
The frequency-aware inference-only sampler and the reported multi-architecture FID gains are genuinely interesting, but the abstract does not establish robustness across seeds, compute/sampling budgets, perceptual quality, or whether the gains hold beyond FID.
-
AutoScientists: Self-Organizing Agent Teams for Long-Running Scientific Experimentation
AutoScientists uses decentralized AI-agent teams that form around hypotheses, critique experiments before spending compute, and share both successful and failed directions during long-running research. Under matched budgets, it reports gains across biomedical ML, language-model training optimization, and protein fitness prediction, including 7 accepted GPT-training improvements versus 0 for a single-agent baseline and a 6.5% average Spearman improvement across 217 ProteinGym assays.
The combination of decentralized team formation, persistent experimental memory, and broad cross-domain gains is substantially more interesting than a routine agent scaffold, but the abstract does not establish how much of the improvement comes from orchestration versus extra effective search or domain-specific implementation details.
-
Everything at Every Scale: Scale-Invariant Diffusion with Continuous Super-Resolution
The paper introduces SKILD, a diffusion model that treats image generation and super-resolution as the same process of reversing information loss across spatial scales. Scale is built directly into the forward and reverse dynamics, so one unconditional model can generate images or perform continuous 2×–8× super-resolution simply by changing the starting timestep, without task-specific conditioning or retraining. It reports strong CIFAR-10 generation results, competitive ImageNet super-resolution, and recovery of higher-order correlations in critical Ising systems.
The unified, scale-coordinate formulation and claimed ability to handle continuous super-resolution from a single unconditional model are genuinely interesting, but the abstract provides limited detail and evidence short of establishing a major breakthrough.
-
Trust Region Q Adjoint Matching
The paper introduces TRQAM, an off-policy fine-tuning method for pretrained flow policies that explicitly constrains divergence from the original policy using a trust-region parameter and a closed-form path-space KL relationship. On 50 OGBench tasks, it reports a 68% overall offline-RL success rate versus 46% for the strongest baseline, and also improves offline-to-online performance.
The exact, adaptively controlled path-space trust region addresses a real stability problem in critic-guided flow-policy improvement, and the broad benchmark shows a substantial reported gain, though the result still appears to be a specialized algorithmic advance rather than a clearly field-changing direction.
-
Representation Forcing for Bottleneck-Free Unified Multimodal Models
The paper proposes Representation Forcing, where a unified multimodal model predicts intermediate visual representations before generating pixels, using those representations to guide pixel diffusion. This removes the need for a separately pretrained VAE and reportedly preserves image-generation quality while improving image understanding, but the abstract gives no quantitative details about the scale of the gains.
The native, intermediate-representation pathway is a potentially important direction for end-to-end unified multimodal models, but the abstract does not provide enough results to establish that it matches VAE-based systems broadly or represents a major capability jump.
-
Draft-OPD: On-Policy Distillation for Speculative Draft Models
Draft-OPD trains speculative-decoding draft models on states generated by their own proposals, while using target-model assistance to keep rollouts stable. It replays positions where proposals were accepted or rejected so the drafter learns from the errors that actually limit speculation; experiments report over 5× lossless acceleration for reasoning models, with 23% and 13% gains over EAGLE-3 and DFlash.
The on-policy training formulation targets a real limitation of existing draft-model SFT and reports substantial inference-speed gains, but the abstract lacks enough detail about absolute latency, model scales, and breadth of validation for a strong verdict.
-
When Does LeJEPA Learn a World Model?
This paper analyzes when LeJEPA’s alignment-plus-Gaussian-regularization objective can recover the true latent variables of a world from nonlinear observations. It proves linear identifiability for worlds with stationary additive-noise dynamics, shows that Gaussian latent distributions are uniquely compatible with this guarantee, and connects the resulting representation to optimal latent-space planning. Experiments, including pixel-based robotic control and latent dimensions up to 1024, support the theoretical results.
The Gaussian-uniqueness theorem and explicit link from a representation-learning objective to identifiable, planning-suitable world models are genuinely notable, though the practical significance depends on how restrictive the transition and observation assumptions are.
-
Optimal Gap-Dependent Regret for Private Stochastic Decision-Theoretic Online Learning
This paper resolves an open problem on the optimal gap-dependent regret of stochastic full-information online learning under pure event-level differential privacy. It gives a horizon-free block-based algorithm using exponentially growing blocks, random prefixes, and the exponential mechanism, achieving regret O(log K/Δ_min + log K/ε), matching the known lower bound up to constants for every horizon.
It provides a concrete algorithm and proof that close the optimal-rate gap for a COLT open problem, with a privacy-dependent regret term matching the lower bound rather than merely improving a benchmark.
-
Large-scale volumetric two-photon calcium imaging enables cellular-resolution mesoscopic mapping of marmoset cortical areas
The authors combine Light Beads Microscopy with improved calcium-indicator targeting to image up to 96,000 neurons simultaneously across a cellular-resolution, 4 × 4 × 0.3 mm volume in marmoset cortex at multi-hertz rates. In auditory cortex they observe fine-grained frequency tuning within broad tonotopic maps; in visual area MT, direction preferences form continuous, columnar, pinwheel-like organization, while noise correlations persist across distances larger than several columns.
This is a substantial technology and scale advance for primate systems neuroscience, paired with nontrivial evidence that functional interactions and tuning organization span multiple cortical columns, though the abstract does not establish a major revision of cortical theory.
-
VisualThink-VLA: Visual Intermediate Reasoning for Effective and Low-Latency Vision-Language-Action Policies
The paper replaces textual chain-of-thought in vision-language-action policies with compact visual-evidence tokens, trained through selective routing and a 754.7k-example supervision resource. It reports similar or better task success across several benchmarks and real-robot tests while reducing BridgeData V2 reasoning latency from 8.377 seconds to 0.367 seconds, a 22.8× reduction.
The visual intermediate-reasoning interface addresses a real obstacle for closed-loop robot control and the reported latency reduction is substantial, but the abstract lacks enough detailed success-rate and ablation results to establish a field-changing advance.
-
BenchEvolver: Frontier Task Synthesis via Solution-Centric Evolution
BenchEvolver makes harder coding problems by evolving executable reference solutions, then generating matching problem statements and tests, rather than asking a model to invent tasks from scratch. On LiveCodeBench and SciCode it produces valid, substantially more difficult tasks; reinforcement learning on these evolved tasks improved held-out coding performance for gpt-oss-20b by 8.7 and 8.3 Pass@1 points over seed-only training.
The solution-centric, executable-semantics approach is a meaningful alternative to human-authored or free-form generated benchmarks, with credible difficulty and training gains, but the evidence is concentrated in coding tasks and a relatively small curated benchmark.
-
One Click per Cell Type Suffices: Training-free Group Interaction for Cell Instance Segmentation
The paper proposes Group Prompting, where one click for each cell type is expanded into prompts that segment all cells of that type, reducing interaction from one prompt per instance to one per type. Its training-free Chain-of-Prompts method uses SAM features to find likely same-type cells and chooses spatially distant prompts; across 11 benchmarks, it reportedly retains over 90–95% of per-instance prompting performance across unseen cell types and imaging modalities.
The O(N)-to-O(T) interaction reduction and claimed zero-shot cross-modality generalization are genuinely interesting, but the abstract provides too few quantitative details to establish how broadly the strong performance claims hold.
-
Gemini Embedding 2: A Native Multimodal Embedding Model from Gemini
Gemini Embedding 2 maps text, images, audio, and video—including interleaved combinations—into one shared embedding space. It uses large-scale contrastive, multitask, multistage training and reports strong retrieval results across multimodal, multilingual, code, and specialized-domain tasks, sometimes exceeding modality-specific embedding models.
A unified embedding model spanning four modalities and arbitrary interleaved inputs could be broadly useful for search and RAG, but the abstract provides limited comparative and experimental detail to establish a field-changing advance.
-
FRAPPE: Full Input, Residual Output Autoencoding with Projection Pursuit Encoder
FRAPPE is a learned image-compression architecture whose encoder uses the full input to predict successive residuals, with independent projection-based latent channels ordered by importance. This gives variable bitrate without sequential encoding, and the authors report that at roughly 0.1 bpp it beats AVIF perceptually while encoding 47× faster, enabling claimed CPU-only 1080p/30fps operation.
The combination of naturally ordered latent channels and a parallel, CPU-friendly encoder addresses a real weakness of learned codecs, but the abstract provides limited comparative evidence beyond one compression regime and AVIF comparison.
-
Recursive Flow Matching
The paper introduces Recursive Flow Matching, which trains generative dynamics models to remain self-consistent across different numerical discretization scales. It reports accurate one- to four-step forecasting for scientific systems, claiming up to 20× speedups over diffusion-based emulators and more than 15% lower MSE than standard flow matching.
The combination of recursive scale-consistency and near-solver-quality few-step generation could be important for real-time physics emulation, but the abstract lacks benchmark details and enough evidence to justify a strong recommendation.
-
Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players
This paper builds an interactive video world model for multiple independently controlled agents. It combines permutation-symmetric simplex-based agent encodings, hub-mediated sparse attention to reduce cross-agent computation, and causal diffusion distillation with KV caching for 24-FPS rollout; it reportedly generalizes from two to four players without retraining while preserving action responsiveness and consistency.
The combination of order-independent multi-agent identity, linearized cross-agent attention, and real-time causal generation is a substantive direction for interactive world models, but the abstract gives no quantitative results or evidence beyond a limited two-to-four-player setting.
-
Intrinsic space-time couplings governing multi-scale cortical dynamics
The authors introduce conformal immersion microscopy, which they say can record activity across the entire dorsal cortex at millisecond timing and roughly 100 µm spatial resolution, including single-trial activity above 100 Hz. Using these recordings and multiscale analysis, they propose that cortex is organized into frequency-dependent dynamical elements with characteristic coherence lengths, propagation speeds, and boundaries that remain stable across activity regimes and many perturbations, while some boundaries can be selectively overridden.
The combination of cortex-wide, fast imaging and a proposed scale-invariant dynamical organization is unusually ambitious and potentially important, but the abstract gives few quantitative results and the extraordinary measurement and robustness claims need careful validation.
-
Chenonceau: mesoscopic MRI deep phenotyping of a post-mortem human brain at 7 and 11.7 Tesla
The authors scanned one post-mortem human brain at 7 and 11.7 T for 8,000 hours, combining anatomical, diffusion, and quantitative MRI at 100–200 μm isotropic resolution. They release a whole-brain, multimodal mesoscopic atlas that includes cortical lamination, myelo- and cyto-architectural information, and estimates of intracortical connectivity. The main contribution is an unusually comprehensive bridge between in-vivo MRI and microscopic post-mortem datasets, rather than a new modeling method.
This is an unusually ambitious and potentially foundational open dataset, with whole-brain mesoscopic multimodal coverage and intracortical connectivity, but it is based on a single specimen and the abstract provides limited evidence about validation or scientific discoveries enabled by it.
-
Your Agents Are Aging Too: Agent Lifespan Engineering for Deployed Systems
This paper proposes AgingBench, a longitudinal benchmark for testing how persistent agents degrade as their memories, retrieved context, facts, and maintenance state evolve over many sessions. Across 7 scenarios and 14 models, it reports that aging is mechanism-specific: behavioral performance can look stable while factual precision declines, and similar errors may require different repairs depending on whether writing, retrieval, or utilization failed.
The lifespan-engineering framing and finding that clean behavioral evaluations can hide memory-related degradation are genuinely useful, but this is primarily a benchmark and diagnostic framework with limited abstract-level evidence of broad practical impact.
-
GE-Sim 2.0: A Roadmap Towards Comprehensive Closed-loop Video World Simulators for Robotic Manipulation
GE-Sim 2.0 is an action-conditioned video simulator for robot manipulation, trained on large-scale real robot data. It adds latent-state decoding for downstream policies, an automated rollout evaluator that supplies rewards, and faster long-horizon simulation; the authors report better world-model leaderboard performance and some transfer gains when policies are trained with its rollouts.
The combination of action-faithful video simulation, automatic success judging, and policy training could make closed-loop world models practically useful, but the abstract gives few quantitative details about real-world gains, task breadth, or whether simulator-trained policies generalize beyond the reported evaluations.
-
LLaVA-OneVision-2: Towards Next-Generation Perceptual Intelligence
LLaVA-OneVision-2 proposes a video tokenization scheme that uses compressed-video bit costs and motion residuals to adaptively allocate tokens to salient temporal and spatial content, alongside a shared 3D positional encoding. The authors report large gains in fine-grained temporal grounding and improvements across video, spatial, and tracking benchmarks, with especially dramatic results on their newly introduced JumpScore benchmark.
The codec-aware adaptive tokenization is a genuinely interesting direction for efficient long-video understanding, but the strongest result is on a new author-designed benchmark and the abstract provides limited evidence that the gains generalize beyond it.
-
MobileGym: A Verifiable and Highly Parallel Simulation Platform for Mobile GUI Agent Research
MobileGym is a browser-based, controllable simulator for mobile GUI agents, with deterministic JSON-state judging, dense rewards, and hundreds of parallel instances for online RL. It includes 416 parameterized tasks across 28 apps; in a GRPO case study, training in simulation improved test performance by 12.8 percentage points, with 95.1% of that gain retained on a 59-task real-device subset. The main contribution is making verifiable, scalable RL training for everyday mobile apps practical without reproducing proprietary backends.
The combination of structured deterministic judging, scalable parallel rollouts, and a reasonably convincing sim-to-real result addresses a real bottleneck in mobile-agent research, though it is primarily an infrastructure and benchmark contribution rather than a fundamentally new agent method.
-
Self-Improving Language Models with Bidirectional Evolutionary Search
The paper proposes Bidirectional Evolutionary Search, which combines forward search over candidate reasoning trajectories with backward decomposition into checkable subgoals. Evolution operators recombine partial trajectories to explore solutions outside the model’s usual sampling distribution, while subgoal feedback provides denser guidance; experiments report gains on post-training and three problem-solving benchmarks, though the abstract gives few quantitative details.
The combination of evolutionary trajectory search and backward, verifiable goal decomposition is a meaningful alternative to standard best-of-N and tree search, but the abstract does not establish how large or robust the gains are.
-
Why Larger Models Learn More: Effects of Capacity, Interference, and Rare-Task Retention
The paper argues that larger models learn rare and complex tasks not merely because they can represent them, but because they suffer less interference from frequent, easy tasks. Synthetic experiments and OLMo pretraining from 4M to 4B parameters suggest that small models devote limited neuron resources to common tasks, whose ongoing updates overwrite slowly learned features for rare tasks; larger models can isolate these features more effectively.
It offers a concrete, experimentally supported explanation for a central scaling phenomenon—reduced gradient interference and better retention of rare-task features—though the evidence is mainly synthetic and uses a controlled set of novel tasks rather than broad real-world training.
-
DynaFLIP: Rethinking Robotics Perception via Tri-Modal-Dynamics Guided Representation
DynaFLIP pretrains an image-only robot perception encoder using aligned image, language, and 3D-flow triplets from human and robot videos. It uses a simplex-volume objective plus cosine and contrastive losses to make representations capture both scene content and motion, and reports improved downstream manipulation policies, including up to 22.5% gains in out-of-distribution settings across simulation and real-world experiments.
The upstreaming of action-relevant dynamics into a reusable visual backbone via tri-modal geometric supervision is a substantive idea with promising OOD manipulation results, but the abstract gives too little detail about baselines, task breadth, and absolute performance for a strong verdict.
-
PhoneWorld: Scaling Phone-Use Agent Environments
PhoneWorld is a pipeline for turning real mobile GUI trajectories into controllable mock Android apps, executable tasks, automatic verifiers, and training data. Its current version covers 34 apps across 16 domains, and adding its supervision improves several phone-agent benchmarks substantially, while broader app coverage appears more valuable than simply adding data within existing apps.
The potentially reusable, scalable recipe for generating phone-use environments—and the result that app diversity matters more than extra data at fixed budget—makes this worth a look, but the evidence is still from a work-in-progress system and likely depends on synthetic mock-app fidelity.
-
Linearizing Vision Transformer with Test-Time Training
The paper converts a pretrained Softmax-attention vision transformer into a linear-complexity Test-Time Training (TTT) architecture, rather than training a linear-attention model from scratch. Key instance normalization and a locality module help preserve the original model’s representations; applied to Stable Diffusion 3.5, the converted model reportedly matches a fine-tuned Softmax model after just one hour of tuning and speeds inference by 1.32× at 1K and 1.47× at 2K resolution.
The pretrained-model conversion strategy and structural alignment between Softmax attention and TTT are genuinely interesting, but the reported gains are moderate and the abstract gives limited evidence beyond one diffusion model and a vague quality comparison.
-
When and How Long? The Readout-Mediator Angle in Temporal Reasoning
The paper shows that a representation can make calendar-date information linearly decodable while being irrelevant to the model’s actual computation. In language models from 1.5B to 9B parameters, a causal subspace identified with DAS was nearly orthogonal to the probe direction; circuit analysis suggests attention and MLP components use a separate pathway to transform absolute dates into durations. The authors frame this mismatch as “readout-mediator orthogonality” and report similar preliminary findings for spatial displacement and symbolic arithmetic.
This is a non-obvious, causally supported challenge to the assumption that successful linear probes reveal the representations a model actually uses, with replication across model scales and families, but the demonstrated tasks and broader generality remain limited.
-
Citation Grounding Measures the Oracle: Graph Coverage Determines Reported LLM Hallucination Rates in Law
The paper shows that citation-grounding scores for legal LLM outputs are driven mainly by how complete the reference citation graph is, not by the model’s actual citation quality. On 400 answers from four commercial models, a sparse graph suggested 15–21% hallucinated citations, while a denser snapshot scored the same answers at 98.9–99.9%; simulations and an external registry indicate that the sparse graph mostly produced false positives and that neither graph reliably separates the models.
This is a non-obvious, empirically supported warning that a seemingly objective LLM evaluation metric can measure oracle coverage rather than model quality, though the evidence is limited to one legal citation setting and a modest response sample.
-
Where to Look: Can Foundation Models Reach a Target Viewpoint Through Active Exploration?
The paper introduces Target Viewpoint Reproduction, where an agent must actively move and rotate through a simulated indoor environment until its camera matches a target image. Existing models perform poorly (7.8–12.0% success), while trajectory-based visual-action fine-tuning raises a 9B model to about 51%; the analysis shows that using visual history and translating the body are major difficulties.
This is a useful, concrete test of active 3D spatial reasoning with a notable improvement and an informative failure analysis, though it remains a simulation benchmark and the core training recipe is largely standard post-training.
-
LiveBrowseComp: Are Search Agents Searching, or Just Verifying What They Already Know?
The paper tests whether search agents discover evidence or mainly use web search to confirm information already stored in the model. On BrowseComp, agents answer up to 44.5% of questions without tools, often formulate queries from internal hypotheses, and degrade when supporting evidence is removed; the authors introduce LiveBrowseComp, a 335-question benchmark built around recently published, non-salient facts, where closed-book accuracy falls below 2% and model rankings change substantially.
The intrinsic-knowledge-dependence diagnosis and temporally grounded benchmark expose a non-obvious weakness in current search-agent evaluations, though the work is primarily an evaluation and benchmark paper rather than a new agent capability or method.
-
From Pixels to Words -- Towards Native One-Vision Models at Scale
The paper presents NEO-ov, a vision-language model trained end-to-end without a separate image encoder, adapter, or post-hoc fusion module. It extends this “native” design beyond single images to multi-image and video understanding, claiming competitive performance with modular VLMs and stronger fine-grained perception, alongside architectural analyses and training recipes.
A genuinely notable architectural direction—scaling native pixel-to-word modeling to video and multi-image inputs—but the abstract gives no model scale, benchmark numbers, or comparisons strong enough to establish a major advance.
-
VLM3: Vision Language Models Are Native 3D Learners
The paper argues that standard vision-language models can learn a broad set of 3D tasks without specialized 3D architectures, losses, or heavy augmentation. VLM3 instead uses focal-length normalization, text-based pixel references, and suitable data scaling, and reportedly improves depth estimation while reaching expert-model performance on correspondence, camera pose, and object-level 3D understanding.
The potentially important result is that simple prompting and data-design choices may replace much of the specialized machinery used for 3D vision, but the abstract gives too few quantitative and experimental details to support a stronger verdict.
-
Channel-wise Vector Quantization
The paper replaces spatial patch tokens with channel-wise discrete tokens, interpreting the channels as progressively finer levels of visual detail. It then generates images by predicting these channels sequentially, from global structure to fine attributes, and reports full codebook utilization plus competitive text-to-image scores. The main novelty is changing the autoregressive axis from spatial patches to feature channels.
The channel-wise representation and next-channel generation order are a genuinely different design direction, but the abstract gives limited comparative evidence that it produces a major capability or efficiency gain.
-
Task-Focused Memorization for Multimodal Agents
The paper treats long-term multimodal memory selection as a learned policy rather than a fixed compression or retrieval rule. TaskMem first learns to preserve faithful information, then adapts a lightweight policy during deployment using task-derived rewards; on three streaming video benchmarks, it improves memory-only VQA accuracy by 5.3–7.0% over the base system.
The task-conditioned, post-deployment learning formulation for deciding what an agent should remember is a meaningful idea, and the gains span three benchmarks, but the abstract does not establish how broadly it transfers or how strong the baselines and online-learning costs are.
-
Do Language Models Need Sleep? Offline Recurrence for Improved Online Inference
The paper gives language models a periodic “sleep” phase: instead of keeping an ever-growing KV cache, they repeatedly process accumulated context offline and consolidate it into fast weights in SSM blocks. On synthetic long-horizon tasks and a math-reasoning task, longer sleep improves performance—especially on problems requiring deeper reasoning—while keeping wake-time inference latency unchanged.
The sleep/consolidation framing is a genuinely interesting approach to trading offline computation for persistent memory and deeper reasoning, but the abstract provides no quantitative comparisons or evidence of scale beyond controlled tasks and one math benchmark.
-
Masking Stale Observations Helps Search Agents -- Until It Doesn't: A Regime Map and Its Mechanism
The paper studies masking older retrieved observations during long-horizon search-agent trajectories, testing multiple model sizes and retrievers on offline and live-web benchmarks. It finds that masking helps most for strong retrieval paired with mid-capacity models, but provides little benefit with weak retrieval and can hurt saturated models; the proposed explanation is a tradeoff between freeing context for more tool calls and deleting evidence the model still needs.
The regime-dependent inverted-U pattern and interaction between retriever quality and model filtering ability are useful, non-obvious findings, though the intervention itself is simple and the abstract provides no quantitative effect sizes suggesting a major advance.
-
Learn from Weaknesses: Automated Domain Specialization for Small Computer-Use Agents
LearnWeak specializes small computer-use agents by first identifying where a student agent fails, then generating targeted tasks and supervision for those weaknesses using a stronger reference agent. It also separates planning errors from execution errors during training. On OSWorld, it reports sizable average gains of 11.6 and 11.1 percentage points over two 7–8B baselines across eight domains, outperforming generic autonomous trajectory-generation approaches.
The student-aware, error-targeted data generation and training approach is a meaningful step beyond generic synthetic specialization, and the reported double-digit gains are notable, but the evidence is limited to one benchmark and the abstract does not establish broader generality.
-
OCC-RAG: Optimal Cognitive Core for Faithful Question Answering
The paper introduces compact language models specialized for retrieval-grounded, multi-hop question answering rather than broad memorized knowledge. It trains 0.6B and 1.7B models on over three million synthetic examples emphasizing context faithfulness, citation through quoted evidence, and calibrated abstention; the authors report matching or exceeding models 2–6 times larger on several reasoning, faithfulness, and refusal benchmarks.
The combination of small task-specialized models, synthetic multi-hop training, grounded reasoning traces, and abstention is useful and reportedly competitive with much larger models, but the abstract gives no numerical results or details sufficient to judge whether this is a substantial advance over existing RAG distillation and specialized-SLM work.
-
CUA-Gym: Scaling Verifiable Training Environments and Tasks for Computer-Use Agents
CUA-Gym presents a pipeline for generating computer-use RL tasks together with executable web environments and deterministic reward functions. It produces 32,112 verified training examples across 110 environments; models trained on them reach 62.1% and 72.6% on OSWorld-Verified and also improve on WebArena, suggesting some transfer beyond the synthesized environments.
The coordinated generation and verification of tasks, environments, and rewards addresses a real bottleneck for computer-use agents, and the reported scale and cross-benchmark gains are substantial, though the abstract does not establish how much comes from the pipeline versus straightforward data and training scaling.
-
RayDer: Scalable Self-Supervised Novel View Synthesis from Real-World Video
RayDer is a single feed-forward transformer that jointly estimates cameras, reconstructs scenes, and renders novel views, trained self-supervised on unconstrained real-world video. It treats changing content as a nuisance state rather than attempting dynamic-scene reconstruction, and reports smooth scaling with data and compute plus strong zero-shot performance on multiple benchmarks.
The unified design and use of dynamic video purely as supervision could make scalable self-supervised static NVS substantially more practical, but the abstract gives no quantitative results or details sufficient to establish a major breakthrough.
-
Learning to Search and Searching to Learn for Generalization in Planning
The paper trains a relational GNN heuristic through weighted A* search, using the resulting search experience as Q-learning data. The learned heuristic can then act as a standalone policy, and the authors report striking zero-shot scaling—for example, training on Blocksworld problems with fewer than 30 blocks reportedly transfers to 488-block instances without search—along with results across Sokoban, PushWorld, The Witness, and IPC benchmarks.
The search-to-learning loop and especially the claimed 30-to-488-block transfer are genuinely interesting and assumption-challenging, but the abstract gives too few quantitative comparisons and robustness details to justify a strong recommendation.
-
True Self-Avoiding Walk for Accelerating Markov-Chain Monte Carlo Integration
The paper uses a true self-avoiding, history-dependent walk that penalizes states and transitions that have already been visited too often, while preserving a target Markov chain’s stationary distribution. For finite state spaces, it claims the state and edge counts deviate from their stationary expectations by only O(√log t) almost surely, yielding an empirical integral error of O(√log t/t), rather than the usual O(1/√t) fluctuation scale.
The claimed near-deterministic balancing of MCMC occupation counts is a genuinely surprising improvement over ordinary sampling noise, but the abstract provides only asymptotic finite-state guarantees and no evidence yet about constants, practical speedups, or scalability.
-
The Variance Brain Foundation Models Forgot: Third-Order Statistics Predict Cognition Where Billion-Parameter Models Fail
The paper argues that fMRI foundation models learn dominant second-order signal statistics while discarding higher-order structure—especially third-order co-skewness—that is more predictive of cognition. A simple linear projection designed to preserve co-skewness, followed by functional-connectivity estimation, reportedly beats three pretrained brain foundation models across datasets and parcellations, and targeted fine-tuning closes the gap without changing the architecture.
The potentially important result is that a specific failure mode of large fMRI Transformers—destroying cognition-relevant third-order statistics—could explain why scaling hurts, but the abstract provides no quantitative effect sizes or details sufficient to justify a stronger verdict.
-
Parallax: Parameterized Local Linear Attention for Language Modeling
Parallax replaces the usual local-constant approximation in attention with a parameterized local-linear estimator, using a learned projector instead of an expensive numerical solver. The authors report faster or comparable decoding to FlashAttention and consistent perplexity gains when pretraining 0.6B- and 1.7B-parameter models, including under parameter- and compute-matched comparisons; they also find that the gains depend strongly on the Muon optimizer.
The combination of a genuinely different attention estimator, hardware-aware implementation, and apparent architecture–optimizer interaction is noteworthy, but the evidence is limited to relatively small model scales and the abstract gives no quantitative size for the claimed gains.
-
SANA-Streaming: Real-time Streaming Video Editing with Hybrid Diffusion Transformer
SANA-Streaming presents a video-to-video editing system designed for interactive use, combining a hybrid diffusion-transformer architecture, cycle-reverse training regularization, and Blackwell-specific kernel fusion and quantization. It reports 1280×704 editing at 24 end-to-end FPS on a single RTX 5090, although the abstract gives few details about visual-quality gains or comparisons with prior systems.
Real-time high-resolution streaming editing on a consumer GPU is a meaningful capability advance, but the individual techniques appear incremental and the abstract lacks quantitative baselines and broader evidence to justify a strong verdict.
-
How can embedding models bind concepts?
The paper analyzes why CLIP-like vision-language embeddings often recognize the right concepts but fail to bind them to the correct objects, such as matching each color with its shape. It finds that scene embeddings are additive over object representations, while CLIP uses a high-complexity binding function; in controlled transformers, broader training coverage leads to low-complexity, multiplicative binding that generalizes to unseen combinations.
It offers a potentially useful mechanistic explanation for compositional failures in embedding models and links successful binding to multiplicative interactions and data coverage, but the evidence is mainly from controlled models and the abstract gives no scale or quantitative results.
-
dMoE: dLLMs with Learnable Block Experts
dMoE adapts mixture-of-experts routing for diffusion language models, where many tokens are decoded together and token-wise routing activates too many experts. It aggregates token expert preferences within each decoding block, reducing uniquely activated experts from 69.5 to 14.6 while retaining 99.11% of baseline performance, cutting memory use by roughly 77–80% and improving end-to-end latency by 1.14–1.66×. The main contribution is a block-level routing strategy designed around the inference pattern of diffusion LLMs.
The block-level MoE formulation addresses a specific and important efficiency mismatch in diffusion LLM inference, with large reported memory and routing reductions, but the abstract provides limited detail on evaluation breadth and the work is explicitly preliminary.
-
Native Audio-Visual Alignment for Generation
NAVA generates audio and video jointly by first aligning the modalities in a dedicated interaction space, then combining them for shared denoising under textual or other context. Its additional timbre-in-context mechanism links reference speaker characteristics to speech segments, and experiments report improved synchronization, video quality, and timbre control with a 6.3B-parameter model.
The align-then-fuse design is a meaningful architectural direction for synchronized audio-video generation, but the abstract gives no quantitative results or evidence that the gains substantially exceed existing systems.
-
Unified Neural Scaling Laws
The paper proposes a single functional form for predicting how neural-network performance changes when model size, data, optimization, inference steps, compute, and hyperparameters vary together. It reports more accurate extrapolations than prior scaling-law forms across vision, language, math, and reinforcement-learning tasks, but the abstract gives no quantitative results or details about the form.
Jointly modeling and extrapolating many scaling dimensions across substantially different tasks could be broadly useful, but the abstract lacks the numerical evidence needed to judge whether the claimed improvement is genuinely large or robust.
-
YoCausal: How Far is Video Generation from World Model? A Causality Perspective
YoCausal evaluates whether video diffusion models understand causal structure or merely learn temporal regularities. It uses temporally reversed real-world videos as counterfactuals and introduces two metrics that separate arrow-of-time detection from causal reasoning; tests on 13 models find that strong temporal-direction sensitivity does not translate into human-level causal cognition.
The real-video counterfactual protocol and separation of temporal bias from causality address an important evaluation gap, with the finding that arrow-of-time perception is insufficiently causal being meaningfully non-obvious, though this is primarily a benchmark/evaluation paper and the abstract gives limited quantitative evidence.
-
How and What to Imagine? Visual Thinking in Unified Multimodal Models for Cross-View Spatial Reasoning
The paper trains unified multimodal models to generate and use intermediate images for cross-view spatial reasoning. Its View Dropout method hides portions of an input view from the answer pathway while leaving them available to the intermediate image pathway, encouraging the model to rely on visual thought; among several renderings, panoramic images generalize best from synthetic training scenes to five real-world benchmarks.
The targeted intervention for making image-based reasoning causally useful, plus the finding that panoramic visual thoughts are more learnable and informative than other renderings, is a meaningful idea, though the abstract gives no quantitative gains or evidence of broad impact.
-
EarlyTom: Early Token Compression Completes Fast Video Understanding
EarlyTom compresses video tokens inside the vision encoder rather than only after encoding, reducing computation before the language model stage. It is training-free and reports up to 2.65× lower time-to-first-token and 61% fewer FLOPs on LLaVA-OneVision-7B on an A100, with accuracy close to the full-token baseline. It also uses a decoupled spatial token-selection method.
The paper targets an important but often overlooked deployment bottleneck—vision-encoder cost—and reports a substantial latency improvement, though the abstract provides evidence only for one model and hardware setup and the core idea is an extension of token compression methods.
-
Closed-Form Node Classification with Exact Graph Unlearning
The paper replaces gradient-trained GNNs with a homophily-routed family of closed-form graph predictors: propagation plus Ridge regression for assortative graphs, and a layer-wise closed-form refinement model for heterophilous graphs. Across 14 benchmarks, these methods reportedly match or exceed vanilla shallow GNNs while supporting exact retraining-equivalent updates after labels, features, edges, nodes, or subgraphs change; localized updates are reported to be 21–45× faster than full re-solving and about 10^6× faster than gradient retraining.
The combination of competitive closed-form graph learning with formal exact graph unlearning and large update-speed gains is a genuinely interesting direction, but the headline results are mainly against vanilla GNN baselines rather than the strongest modern or deeply tuned methods.
-
How Should LLMs Consume High-Quality Data? Optimal Data Scheduling via Quality-Aware Functional Scaling Laws
The paper derives optimal schedules for data quality and batch size in a feature-space regression model, identifying different strategies for noise-limited and signal-limited training. It turns this into Drop-Stable-Rampup: reduce the batch size when data quality changes, maintain it low to accumulate signal, then increase it late to reduce terminal noise; experiments on 600M and 15B models report sizable gains over standard decay schedules, especially on math and code tasks.
The combination of a quality-aware scaling-law analysis and a simple, potentially general training-schedule intervention with unusually large reported gains is worth checking, though the theory is limited to a regression model and the empirical claims need scrutiny.
-
Automated Benchmark Auditing for AI Agents and Large Language Models
The paper presents Auto Benchmark Audit, an agent-based system that checks LLM and agent benchmark tasks for ambiguities, environment dependencies, grading flaws, and incorrect answers. Auditing 168 benchmarks across nine domains, it reports critical problems in over 25% of tasks; removing them changes model rankings and raises measured performance on SWE-bench Verified and Terminal-Bench 2 by roughly 10%.
The notable contribution is evidence that benchmark defects are common enough to materially distort agent rankings, backed by a fairly broad audit and expert/independent validation, though the core auditing framework is still an agentic tool rather than a demonstrated leap in model capability.
-
Agile Online Model Selection: Resolving Adaptation Lag via Safeguarded Large Learning Rates
The paper proposes an online model-selection method that permits very large learning rates—up to order T—while using a post-hoc penalty to discard unstable updates. It claims to retain near-optimal worst-case regret while adapting to abrupt distribution shifts in a few rounds rather than hundreds, supported by experiments on synthetic data and 11 real-world datasets.
The combination of safeguarded, potentially linear-scale learning rates with a logarithmic cumulative penalty addresses a real robustness-versus-adaptation trade-off, but the abstract does not establish a field-changing capability or enough detail to warrant strong.
-
Cast a Wider Net: Coordinated Pass@K Policy Optimization for Code Reasoning
The paper replaces independent pass@4 sampling with coordinated exploration: a planner proposes four different high-level strategies, and a solver produces one code solution for each. Training rewards strategy sets only when at least one attempt passes the verifier; across three coding benchmarks, this improves pass@4 in six of nine model–benchmark settings, including a gain from 0.588 to 0.748 on Qwen3.5-9B LiveCodeBench-v6.
The explicit optimization of strategy diversity for verifier-based pass@K is a meaningful departure from independent sampling, supported by multi-benchmark results and one large reported gain, but the evidence is limited to K=4 code generation and does not yet establish broad generality.
-
Can Retrieval Heads See Images? Multimodal Retrieval Heads in Long-Context Vision-Language Models
The paper identifies a small subset of attention heads in long-context vision-language models that selectively retrieve relevant text and image evidence. Masking the top 5% of these heads causes very large drops on document and slide question answering, while using their attention scores directly improves visual document retrieval without additional training. The heads are partly shared across modalities but adapt to context length and the type of evidence being retrieved.
The combination of a modality-general retrieval-head analysis, causal ablations, and a useful zero-training retrieval application is notably more than an incremental attention study, but the work-in-progress status and reliance on a few reported benchmarks make it premature for a strong verdict.
-
Algorithmic Monocultures in Hiring
The paper studies 3 million applicants and 4 million applications screened by algorithms from the same hiring vendor. It finds substantial racial disparities and unusually homogeneous decisions: applicants applying to ten jobs were rejected from all of them more often than chance would predict, suggesting that using a common vendor reduces the value of applying broadly.
The large-scale evidence for correlated, vendor-wide hiring outcomes—and the implication that broad application may not overcome algorithmic rejection—is a non-obvious finding with meaningful real-world consequences, though it is primarily an empirical fairness study rather than a new AI method.
-
Hurwitz Quaternion Multiplicative Quantization for KV Cache Compression
The paper introduces a calibration-free KV-cache quantizer that represents groups of four values as quaternion directions and uses a structured 24-element Hurwitz rotation group to expand a small random codebook. Across five LLMs, it reports roughly fp16 perplexity at about 5 bits, and substantially better robustness than naive int4 on outlier-heavy models; it also claims competitive downstream accuracy with KIVI at fewer bits and up to 5.05× cache compression.
The combination of multiplicative quaternion codebooks, no calibration, and strong quality at roughly 4–5-bit KV caches is a genuinely interesting efficiency result, but the evidence is limited to the abstract’s reported model and task evaluations and needs closer inspection of comparisons, memory overhead, and implementation practicality.
-
Primitive Subspaces Mediate Few-Shot Transfer in VLAs
The paper trains vision-language-action policies on explicitly segmented motion primitives, then uses a few demonstrations to compose behavior for unseen tasks without updating the weights. Primitive-aware training reportedly reaches the performance of fine-tuned models with three demonstrations instead of ten, across two VLA architectures, multiple seeds, and two datasets; targeted hidden-state ablations suggest the primitive representations are causally involved. It also identifies a potentially important evaluation error in chunked-policy action-range tests that can greatly inflate failure rates.
The cross-architecture, cross-dataset 3× few-shot improvement and causal representation ablation are unusually substantive, but the result is still demonstrated on a limited set of benchmark manipulation tasks and lacks enough absolute performance and generalization detail for a strong verdict.
-
Repetition Mismatch: Why Data Mixture Experiments Don't Scale and How to Fix Them
The paper argues that small-scale data-mixture studies fail mainly because scarce high-quality data is repeated at a different rate than it will be at the target training budget. By matching repetition rates during subsampling, the authors recover the near-optimal two-source mixture using only 1/16 of the target tokens, whereas uncontrolled experiments require several horizons and much more compute; similar gains hold with three sources, though multiple experiments are needed.
This identifies a specific, actionable reason why mixture extrapolation fails and reports a potentially large reduction in tuning cost, but the evidence is limited to relatively small 757M-parameter experiments and a few synthetic mixture settings.
-
Topologically Consistent Multi-view 3D Head Reconstruction via Coarse-Guided Layered Surface Sampling
SHELLS reconstructs dense, semantically corresponding 3D head meshes from multi-view images using a coarse mesh to guide layered surface sampling, rather than independently sampling features for every vertex. On 18k-vertex meshes, it reportedly cuts inference memory by 88%, improves registration error by 21–29%, and runs 3.5× faster; it also transfers from synthetic-only training to real captures.
The resolution-decoupled layered sampling design and strong efficiency gains are meaningful, while synthetic-to-real generalization is notably non-obvious, but the contribution remains a specialized 3D reconstruction advance rather than a broad new direction.
-
Consolidating Rewarded Perturbations for LLM Post-Training
This paper studies whether the gains from rewarded weight perturbations can be merged into one model rather than requiring an ensemble of perturbed models at inference time. Its CoRP method uses reward-weighted, compatibility-aware aggregation plus validation, without backpropagation; across five models and five tasks, it reports an average 8.1-point improvement while using one-tenth the perturbation budget and only one test-time forward pass.
The low-rank structure found consistently across 25 model-task pairs and the substantial efficiency/capability tradeoff make this more than a routine post-training tweak, but the abstract lacks enough baseline and task-specific detail to justify a strong verdict.
-
On the Relationship Between Activation Outliers and Feature Death in Sparse Autoencoders
The paper explains why sparse autoencoders sometimes learn many features that never activate: large mean-shifted activation dimensions bias feature pre-activations negative when their directions oppose the activation mean. A simple mean-centering step removes this effect, and the proposed outlier measure predicts feature death across 454 model-layer combinations spanning language, vision, protein, and genomic models.
This gives a concrete, cross-domain mechanism for a major SAE failure mode and a simple principled fix, but its impact is mainly on interpretability tooling rather than core model capability.
-
Generate in Reconstruction Space, Match in Semantic Space: Transport Geometry for One-Step Generation
The paper studies one-step image generation by matching generated and real images in the space of frozen self-supervised features, using Sinkhorn divergence as a tractable transport objective. On ImageNet, semantically structured features reportedly reduce FID by 39×, and the authors argue this works because removing nuisance details makes distribution matching statistically easier; they also show that optimizing in the evaluator’s Inception space can improve FID while harming actual sample quality and training stability.
The combination of transport-based one-step generation with a concrete explanation based on feature-space geometry is substantially more than a routine loss tweak, and the finding that Inception-space optimization can metric-hack FID is non-obvious, but the abstract does not provide enough absolute results or breadth to warrant strong.
-
Task Structure Reverses Layerwise State Encoding in Sequence Models
This paper shows that where sequence information is represented across layers depends strongly on the task, not just the architecture. Across Transformers, recurrent models, and Mamba variants, Parity and Dyck-k induce opposite layerwise patterns, while non-commutative S3 permutation composition behaves like Parity, suggesting that computational structure—prefix updates versus stack-like processing—matters more than commutativity. Causal ablations and activation patching also show that probe-readable state is not always the computation’s bottleneck, especially in pretrained models.
It offers a fairly broad, causally tested challenge to the idea that layerwise representations are primarily architectural, and separates readable state from computational necessity, though the claims remain confined to synthetic sequence tasks and mechanistic analyses.
-
ResearchMath-14K: Scaling Research-Level Mathematics via Agents
The authors compile 14,056 research-level mathematics problems and generate 220K model-produced reasoning traces, including attempts that are incomplete or wrong. They find that newer open models increasingly produce references—and fabricated references—and that filtering these traces before fine-tuning improves Qwen3 models by an average of 9.2 points across sizes. The main contribution is a resource and an empirical observation that imperfect open-problem attempts can still provide useful training signal.
The dataset itself is mainly a benchmark/resource contribution, but the quantified rise in fabricated references and evidence that filtered unsuccessful research attempts improve mathematical reasoning make it worth a closer look despite limited abstract-level validation.
-
RT-Lynx: Putting GEMM Sparsity in the Right Place for Diffusion Models
RT-Lynx argues that Diffusion Transformers are better suited to activation sparsity than weight sparsity: their activations are naturally sparse and tolerate N:M pruning with less quality loss. It combines activation sparsification, error compensation, and specialized CUDA kernels, reporting preserved generation quality and up to 1.55× speedup in linear layers across several diffusion models.
The shift from pruning DiT weights to exploiting activation sparsity is a meaningful systems and modeling insight, but the abstract lacks end-to-end latency, energy, or generation-throughput results needed to establish a major practical advance.
-
UniSteer: Text-Guided Flow Matching in Activation Space for Versatile LLM Steering
UniSteer trains a conditional flow-matching model to transform an LLM’s residual-stream activations according to natural-language descriptions, rather than learning one fixed steering vector or module per behavior. The same model is used for behavioral and truthfulness control, compositional constraints, fine-grained concept steering, and classification from activations, but the abstract gives no quantitative results or comparisons.
The universal text-conditioned activation transport formulation is a notable departure from fixed-vector steering, but the broad claims are supported only by unspecified experiments and lack the numbers needed for a stronger recommendation.
-
GrepSeek: Training Search Agents for Direct Corpus Interaction
GrepSeek trains a compact language-model search agent to interact directly with a text corpus through executable shell commands rather than relying solely on a conventional retriever. It combines answer-aware trajectory generation with GRPO refinement and introduces sharded parallel execution that reportedly speeds up shell-based search by up to 7.6× while preserving exact results; across seven QA benchmarks, it achieves the best aggregate F1 and Exact Match, though lexical search struggles with paraphrases.
Direct corpus interaction is a meaningful alternative search-agent design with a practical systems contribution, but the abstract gives no absolute benchmark numbers or evidence of a major capability gain beyond competitive QA performance.
-
LVSA: Training-Free Sparse Attention for Long Video Diffusion
LVSA is a training-free, model-agnostic sparse-attention pattern for extending video diffusion beyond its training horizon. It combines local windows with rotating global tokens and reports roughly 3x lower attention compute, longer generation on a single GPU, and improved resistance to frozen or repetitive outputs across Wan and HunyuanVideo models. The authors also introduce VQeval to detect looping failures that existing video metrics may reward.
The combination of training-free long-context sparse attention, meaningful multi-model speedups, and enabling otherwise-OOM long-video generation is practically important, though it appears primarily to be a strong systems/attention-pattern advance rather than a clearly new research direction.
-
Pantheon360: Taming Digital Twin Generation via 3D-Aware 360° Video Diffusion
Pantheon360 generates controllable 360° videos from sparse 360° observations for building digital twins. It reconstructs an explicit 3D cache from the input and uses it as a geometric scaffold, letting diffusion refine appearance while preserving consistency along user-defined camera paths.
The explicit 3D-cache-plus-360°-diffusion design addresses a real weakness of perspective video generation, but the abstract provides no quantitative results or comparisons to substantiate its broad superiority claims.
-
DecMem: Towards Minute-Long Consistent World Generation with Decoupled Memory
The paper proposes DecMem, a memory architecture for long-horizon video generation. It separates memory into a sparse global component for accessing distant history and an anchored local component for stable near-term extrapolation, claiming minute-long videos with improved consistency and efficiency.
Long-duration, spatially consistent video generation is an important capability and the decoupled global/local memory design is a meaningful architectural direction, but the abstract gives no quantitative results or details sufficient to establish how large the improvement is.
-
On-Policy Adversarial Flow Distillation for Autoregressive Video Generation
The paper proposes distilling a black-box video teacher into an autoregressive student using on-policy rollouts rather than teacher denoising traces or latent scores. A prompt-paired discriminator estimates whether clean teacher videos are better than student-generated videos, and that feedback is converted into dense flow-matching updates on the student’s noised states. Experiments on two causal video-generation families reportedly improve motion and physics while retaining general quality, but the abstract gives no quantitative gains.
The combination of on-policy adversarial feedback with forward-process flow supervision is a plausible and useful interface for heterogeneous or proprietary video teachers, but the lack of reported numbers makes the claimed impact difficult to assess.
-
Balancing Fidelity and Diversity in Diffusion Models via Symmetric Attention Decomposition: Hopfield Perspective
This paper interprets transformer attention through a Hopfield-network lens by splitting the pre-softmax QKᵀ matrix into symmetric and skew-symmetric components. It argues that the symmetric part shapes feature-retrieval stability while the skew-symmetric part induces circulation, and uses this perspective to relate attention dynamics to the fidelity–diversity trade-off in diffusion generation. The authors propose controlling that trade-off by modifying the circulation component, though the abstract gives few details about the size or breadth of the experimental gains.
The symmetric/skew-symmetric attention decomposition offers a potentially useful new mechanistic framing and control mechanism, but the abstract provides only correlations and no quantitative evidence that the proposed knob produces a substantial improvement.
-
Agent Explorative Policy Optimization for Multimodal Agentic Reasoning
The paper identifies a failure mode in reinforcement learning for multimodal tool-using agents: tool calls are infrequent, and tool-using attempts often fail together, providing little useful learning signal. AXPO keeps the reasoning prefix fixed and resamples the tool call and continuation for these failed attempts, with uncertainty-based selection; on nine benchmarks, it improves Qwen3-VL-Thinking over GRPO by about 1.8 percentage points and lets an 8B model exceed a 32B base model on Pass@4.
The targeted resampling strategy addresses a concrete and plausible tool-use RL pathology, but the reported gains are modest and the evidence is limited to one model family and benchmark suite.
-
A Universal Cliff and a Design Fingerprint: Cross-Section Defect Detection Under LLM Orchestration
The paper tests whether multi-agent orchestration preserves an LLM’s ability to detect contradictions between distant sections of a document. Across ten models, detection reportedly drops by at least two-thirds when work is partitioned, even though private worker records often contain enough information to reconstruct the defect; the authors also find that stronger alignment in one model family shifts behavior toward fewer misses but more false alarms. The result suggests that recomposing agent outputs can structurally hide cross-section defects, while integrated-report confidence provides little warning.
The claimed, cross-provider orchestration cliff is a non-obvious systems finding with practical importance, but the evidence is limited by the small model set, a specialized defect class, and unclear breadth of task and replication.
-
BigMac: Breaking the Pareto Frontier of Compute and Memory in Multimodal LLM Training
BigMac reorganizes multimodal LLM training by nesting encoder and generator computation inside the LLM pipeline while preserving dependency correctness. The authors claim this makes encoder/generator activation memory constant, retains the compute efficiency of an unlimited-memory setup, and yields 1.08–1.9× training speedups with stable memory as batch size grows across several workloads.
The dependency-safe nested pipeline could represent a genuinely useful systems-level way to improve both memory and throughput, but the abstract provides too few experimental details to establish that it truly breaks the compute–memory tradeoff across realistic baselines.
-
'Si'multaneous 'S'patial-'T'emporal Message Passing for Dynamic Graph Representation Learning
SiST-GNN combines a node’s current features with a recurrent representation of its past and performs spatial and temporal message passing in one augmented graph-convolution step, rather than applying the two operations sequentially. On 14 model–dataset combinations, the authors report very large link-prediction gains over prior methods and comparable performance to event-based continuous-time models on newly constructed node-classification tasks.
The joint message-passing formulation is a meaningful departure from the usual temporal-first/spatial-first decomposition, but the unusually large benchmark gains and discretised classification tasks need verification before warranting a strong recommendation.
-
Llamion Technical Report
Llamion converts Orion-14B into a Llama-compatible architecture using parameter mapping plus cross-architecture distillation, rather than retraining the model from scratch. The authors report recovering Orion’s benchmark behavior with only about 123M tokens on one A100, while preserving capabilities such as Python coding and 200K-token context handling that were absent from the transfer data; the released base model reportedly scores 66.87% on KoMMLU.
The potentially important result is that a capable model can be converted between substantially different normalization and architectural conventions with very little data or compute, but the abstract provides limited independent validation beyond a few benchmarks and strong self-reported claims.
-
Paris 2.0: A Decentralized Diffusion Model for Video Generation
Paris 2.0 trains a video diffusion model using decentralized computation rather than a single centralized GPU cluster. In low-resolution text-to-video experiments with matched total compute and data, it reports roughly halving FVD from 561.04 to 279.01 while also improving text-video similarity and aesthetic scores.
Decentralized pretraining of temporally coherent video is a genuinely interesting direction, and the reported twofold FVD improvement is notable, but the abstract gives too little information about scale, datasets, decentralization protocol, and reproducibility to justify a strong recommendation.
-
Quantized Keys Steal Attention: Bias Correction for KV-Cache Compression in Video Diffusion
The paper identifies a systematic softmax-attention bias caused by quantizing keys in the KV cache: quantization noise makes cached keys attract too much attention, reducing attention to the current video chunk. It derives an on-the-fly correction based on quantization step sizes and query norms, reporting that INT2 caches on three video diffusion models recover near-BF16 quality and can outperform INT4 while using half the memory.
The Jensen-bias diagnosis and nearly free correction appear genuinely useful for long-context video generation, but the abstract gives no quantitative results or details sufficient to establish how broadly the claimed INT2 gains hold.
-
Rethinking the Multilingual Reasoning Gap with Layer Swap
The authors train matched Qwen3-8B specialists to reason either in the input language or via English across six languages and several task types. Contrary to earlier reports, the native-language reasoning deficit is only 1.9–3.5% under comparable multilingual supervision; analysis finds shared middle-layer updates, and swapping the English specialist’s middle layers into native specialists closes most of the remaining gap while retaining native-language chain-of-thought.
The combination of a smaller-than-expected multilingual reasoning gap and a simple layer-swapping intervention suggesting language-agnostic reasoning is genuinely interesting, but the abstract lacks detailed per-task results and comparisons needed to establish how broadly or robustly the effect holds.
-
Once-For-All: A Train-Once and Select-Anytime Framework for Multimodal Instruction Tuning
The paper proposes training a lightweight data selector once, using clustering in frozen CLIP embedding space, and reusing it across different multimodal instruction datasets and VLM backbones. Selecting 15% of the examples reportedly preserves 98.3% of full-data performance across 10 benchmarks, and transfers to an unseen dataset where it exceeds full-data training by 10.6%.
Reusable, model- and dataset-independent data selection with substantial claimed compression is a meaningful direction, but the evidence is limited to two datasets and the abstract does not establish how it compares with strong selection baselines or explain the surprising transfer gain.
-
Forward Pass Domain Adaptation (Without Cross-Layer Backpropagation)
The paper proposes training selected transformer layers using a single output-level error signal, without backpropagating through the model or constructing an autograd graph. On three 7–8B model families, it reports 2.7–3.2× higher throughput and about 40% lower peak memory than standard fine-tuning, while preserving off-domain benchmark performance; a layerwise diagnostic predicts where this approximation works.
The combination of a no-backprop adaptation method, a simple viability diagnostic, and substantial reported efficiency gains is genuinely notable, but the evidence is limited to three models with few quantitative quality results and depends on an empirical gradient approximation that needs closer validation.
-
MAIGO: Mitigating Lost-in-Conversation with History-Cleaned On-Policy Self-Distillation
The paper argues that multi-turn performance drops partly because a model’s earlier assistant responses remain in context and propagate mistakes. MAIGO trains with cleaned histories and self-generated references, improving Qwen2.5-7B-Instruct SHARDED accuracy from 52.8% to 66.1% and substantially narrowing the gap to full-context performance, without inference-time scaffolding or external rewards.
The history-cleaning/self-distillation formulation and large reduction in the multi-turn gap are genuinely interesting, but the evidence is limited in the abstract to one model and a specialized paired-view protocol, so it does not yet merit a strong verdict.
-
The Future of Facts: Tracing the Factual Generation-Verification Gap
This paper studies how language models learn to generate factual answers versus judge whether answers are correct, across acquisition, continual learning, and factual updating. Across multiple model families and scales, verification appears before generation, survives continued training better, and can remain split between old and new answers after an update. The work’s main contribution is a cross-model account of these training dynamics, including the surprising “multi-verse” state after factual updates.
The recurring generation–verification asymmetry and conflicting-answer state could inform self-improvement and knowledge-editing methods, but the abstract provides limited quantitative evidence for how large or practically important the effects are.
-
Uni-LaViRA: Language-Vision-Robot Actions Translation for Unified Embodied Navigation
Uni-LaViRA treats embodied navigation as translating language into semantic directional commands and vision into pixel-level targets, using a pretrained multimodal model rather than training on robot trajectories. It adds checklist-style memory and rollback after failed actions, and reports zero-shot results across four navigation tasks and four robot types, including a UAV, with performance claimed to match some heavily trained navigation models.
The structurally unified, zero-shot formulation across substantially different navigation tasks and robot embodiments is notable, but the abstract provides limited baseline, experimental, and real-world deployment detail to support the broad performance claims.
-
S-Cheetah: A Novel Quadrupedal Robot with a 3-DOF Active Spine Learning Agile Locomotion
The paper introduces S-Cheetah, a quadruped with a serial 3-DOF active spine that can flex, bend laterally, and rotate axially, together with reinforcement learning rewards and curricula designed to make the robot use the spine. On hardware, it reports 6.9 m/s galloping, 7.2 rad/s in-place turning, and an emergent feline-like self-righting behavior that lands the robot on its feet from arbitrary orientations. The main novelty is demonstrating a fully spatial, actively controlled spine rather than a reduced-DOF design, though the abstract does not quantify improvements against strong baselines.
A real quadruped with a 3-DOF active spine and emergent aerial self-righting is a notable capability advance, but the strength of the claimed benefit is difficult to assess without baseline comparisons, hardware details, and ablations in the abstract.
-
Extracting Small Translation Specialists from LLMs by Aggressively Pruning Experts
The paper identifies and removes mixture-of-experts components that appear unnecessary for machine translation, producing much smaller translation-specialized models without changing the remaining weights. It reports that pruning 50% of experts causes negligible quality loss, 70% causes minor loss, and brief fine-tuning can recover baseline performance after pruning 75% or more, sometimes nearly 90%.
The combination of task-specific expert separability and extreme, training-free MoE compression is potentially important, but the abstract does not provide enough model, language, or benchmark detail to establish how broadly the result generalizes.
-
SPRINT: Efficient Spectral Priors for Humanoid Athletic Sprints
SPRINT uses frequency-domain priors learned from only five human locomotion sequences to generate feasible humanoid joint trajectories across speeds, including speeds beyond the reference data. A policy guided by these priors reportedly transfers directly from simulation to a Unitree G1 and reaches 6 m/s in real-world sprinting, with gait transitions and human-like motion.
The combination of highly data-efficient spectral motion priors and a reported 6 m/s zero-shot real-world humanoid sprint is genuinely notable, but the abstract lacks comparative results and details needed to establish how robust or reproducible the capability is.
-
Spectral Guidance for Flexible and Efficient Control of Diffusion Models
The paper learns a spectral basis of features that remain informative during diffusion denoising, then uses that basis to apply different guidance signals—labels, CLIP embeddings, or masks—without retraining the model or backpropagating through the denoiser at sampling time. On CIFAR-10, it reports a 37-point conditional-accuracy improvement over the strongest training-free baseline and 4× faster sampling, while identifying a time window when guidance is most effective.
The shared spectral representation for flexible, fast, model-independent guidance is a genuinely interesting direction, but the evidence is mainly a single CIFAR-10 result and the practical generality of the claimed gains is not yet established.
-
Thinking as Compression: Your Reasoning Model is Secretly a Context Compressor
The paper treats an LLM’s reasoning trace as a compressed representation of a long input, rather than using a separate compressor. A constrained, reward-trained version produces more compact and controllable traces, reportedly improving average F1 by 17.4% at 4× compression and 23.4% at 8× over the strongest baseline across four long-context QA benchmarks.
Using reasoning itself as an information bottleneck is a genuinely interesting reframing with large reported gains, but the abstract does not establish how broadly it works, how much extra inference or training it requires, or whether the results transfer beyond QA benchmarks.
-
OpenURMA: A Clean-Room Open Implementation of the Unified Bus Protocol
OpenURMA is an open, clean-room implementation of Huawei’s Unified Bus protocol, provided as FPGA RTL, a SystemC simulator, and a gem5 scaffold, with matched OpenRoCE baselines. On a 64-byte remote fetch, the authors report about 500 ns end-to-end latency—4.37× lower than RoCE—and 2.8× higher throughput while using roughly 14% of an Alveo U50’s LUTs.
The protocol abstraction and open implementation could materially affect datacenter and AI-system interconnect design, but the evidence is centered on one microbenchmark and an FPGA implementation rather than broad real-system validation.
-
When RL Suppresses Its Own Vocabulary: Recovering Reasoning Diversity in Puzzle-to-Math Transfer
The authors study why RL with verifiable rewards transfers from constraint-satisfaction puzzles to hard mathematics in a 7B language model. They find that puzzle training creates reusable reasoning primitives, while subsequent RL composes them into longer chains but suppresses exploratory behaviors such as hypothesizing and backtracking; rewarding diverse correct rollouts restores those behaviors and improves OlymMATH-Hard pass@32 by another 7 percentage points. The full puzzle-only training recipe raises performance from 16% to 36% without math problems in post-training.
The analysis of RL-induced loss of reasoning diversity and the corresponding diversity reward are genuinely interesting, with a substantial reported gain, but the evidence is limited to one 7B model and benchmark setting, so it does not yet warrant a strong verdict.
-
Structured interactions improve distributed coordination beyond model scaling in a real-world multi-robot system
This study compares improving each robot’s neural network with changing how robots communicate in a 10-robot transport-and-mapping system. A modular hierarchical communication topology improved normalized performance by 47 points, while doubling hidden size improved it by at most 9 points; replications in SMAC showed the same pattern, though the hardware study was small and task-specific.
The striking result is that restructuring multi-robot interactions vastly outperformed scaling individual models, but the limited physical experiments and narrow tested setting make this an intriguing result rather than a broadly established design principle.
-
Kronecker Embeddings: Byte-Level Structured Token Representations for Parameter-Efficient Language Models
The paper replaces the large learned token-embedding table with a deterministic byte-position encoding followed by a small learned projection, reducing input-side trainable parameters by 91–94%. In a 124M GPT-2 experiment, this representation reportedly achieves about 2.5% lower validation loss, reaches the baseline’s loss in 1.43× fewer steps, and improves robustness to typos, while a runtime implementation reconstructs embeddings from a few megabytes of storage. The main tradeoff is that byte similarity can group semantically unrelated words, leaving more disambiguation to later layers.
The byte-level factorization is a genuinely interesting alternative to standard embedding tables, with an unexpectedly positive small-scale language-model result and substantial parameter/storage savings, but the core training evidence is limited to one 124M model and does not yet establish frontier-scale generality.
-
VLA-Pro: Cross-Task Procedural Memory Transfer for Vision-Language-Action Models
VLA-Pro stores task-specific LoRA adapters as procedural memories during training, then retrieves and dynamically combines relevant memories from the current multimodal context to control a robot on new tasks. Across RoboTwin, RLBench, and real-world manipulation, it reports large cross-task generalization gains, including a real-world success increase from 5.8% to 65.0%.
The procedural-memory view of LoRA adapters and its large reported gains on unseen manipulation tasks are genuinely interesting, but the abstract does not establish how broadly the method works or how much the dramatic real-world improvement depends on a weak baseline.
-
Planning with the Views
The paper studies whether vision-language models can plan sequences of camera movements to reach a desired viewpoint in 3D scenes. It finds that models can understand individual view changes but struggle to compose them over long plans, and proposes distilling self-exploration trajectories into a view graph and supervised training tasks; this raises Qwen2.5-VL-7B’s task success from 2.5% to 47.8% on their benchmark.
The view-graph distillation approach and large reported improvement address a real compositional-planning failure in VLMs, but the evidence is concentrated on one point-cloud benchmark and may not establish broader 3D planning generality.
-
OccamToken: Efficient VLM Inference with Training-Free and Budget-Adaptive Token Pruning
OccamToken is a training-free VLM token-pruning method that uses register-token attention as a relative reference instead of ranking all visual tokens and keeping a fixed top-K set. It adapts pruning to both image redundancy and the user query, reportedly reducing LLaVA-NeXT’s 2,880 visual tokens to about 40 while retaining over 93% of baseline accuracy across several VLMs.
The register-anchored, budget-adaptive pruning idea and extreme claimed compression are potentially important, but the abstract gives insufficient detail about absolute accuracy, task breadth, and comparisons to strong pruning baselines for a stronger verdict.
-
MARS Policy: Multimodality Only When It Matters
MARS adaptively turns stochastic, multimodal action generation on only during task phases where multiple behaviors are useful, using deterministic prediction elsewhere. Across 8 simulated and 4 real-world manipulation tasks, it reports a 16.67% success-rate improvement and 83.20% lower inference latency than its comparison policies, while also training more efficiently on nearly deterministic tasks.
The phase-adaptive use of stochasticity is a plausible and relatively new way to retain multimodal behavior without paying generative-policy costs, and the reported real-world latency reduction is substantial, though the abstract does not specify baselines or experimental breadth enough for a strong verdict.
-
Déjà View: Looping Transformers for Multi-View 3D Reconstruction
DéjàView replaces a deep, independently parameterized reconstruction transformer with one transformer block reused for multiple refinement steps. The looped model can trade inference compute for quality and reportedly matches or beats much larger baselines across five varied 3D reconstruction benchmarks, while using far fewer parameters; weight sharing also outperforms an equivalent untied iterative model.
The explicit finding that recurrent weight sharing is not just cheaper but a better inductive bias for multi-view 3D reconstruction is genuinely interesting, though the abstract lacks quantitative gains and details needed to justify a strong recommendation.
-
Beyond 3D VQAs: Injecting 3D Spatial Priors into Vision-Language Models for Enhanced Geometric Reasoning
GASP trains a VLM to learn geometric structure directly, rather than relying on 3D visual-question-answering data. It adds layer-wise correspondence supervision and depth-consistency training from video-scene geometry, reportedly raising internal correspondence accuracy from below 5% to over 70% and improving spatial-reasoning benchmarks by 18.2% and 29.0%.
This is a meaningful and somewhat surprising alternative to 3D-VQA fine-tuning, with large reported gains and an interpretable geometric mechanism, but the abstract alone does not establish how broad or robust the results are.
-
Self-Trained Verification for Training- and Test-Time Self-Improvement
The paper trains a verifier to identify errors in a model’s own solutions by giving it access to the reference solution during supervision, then uses that verifier for both test-time refinement and further generator training. It reports roughly doubled accuracy on hard math, a rise from 1.5% to 21% on scientific reasoning, and additional gains in standalone pass@1 after verifier-guided RL training.
The reference-conditioned self-training idea and its use to improve both verification loops and the generator are plausibly important, but the abstract lacks enough detail about models, benchmarks, baselines, and scaling to justify a strong recommendation despite unusually large reported gains.
-
MonoPhysics: Estimating Geometry, Appearance, and Physical Parameters from Monocular Videos
MonoPhysics estimates a deformable object’s 3D geometry, appearance, and physical properties from a single monocular video by combining differentiable MPM simulation with 3D Gaussian Splatting. It introduces scale alignment, physics-guided geometry refinement, and a differentiable position representation, and reports performance comparable to multi-view methods on existing and newly collected elastic/plastic-object data.
Monocular recovery of both object geometry and material physics is substantially harder than standard multi-view inverse simulation, and matching multi-view performance would be important, but the abstract gives no quantitative results or details about scene and motion limitations.
-
ElasticMem: Latent Memory as a Learnable Resource for LLM Agents
ElasticMem treats an agent’s long-term memory as an adaptive latent resource rather than a fixed number of retrieved text snippets or latent tokens. It retrieves memories from the reasoner’s hidden state, learns how much latent capacity to allocate to each one, and trains this memory-use policy using task rewards; the abstract reports large relative gains on memory QA and ALFWorld while reducing token cost.
The combination of query-dependent latent memory budgets and reward-trained memory use is a meaningful departure from fixed retrieval interfaces, but the evidence is limited by benchmark-only results and ambiguous relative-improvement numbers without absolute scores or broader comparisons.
-
Feat2Go: Visual Feature-Grounded Value Estimation for Embodied Reinforcement Learning
Feat2Go uses a pretrained visual world model to estimate continuous task progress from patch-level similarity to subgoals, then trains a value model to reshape sparse terminal rewards for VLA reinforcement learning. It reports very large gains on manipulation benchmarks—for example, raising OpenVLAOFT’s out-of-distribution success on ManiSkill3 from 17.5% to 82.9%—and works with PPO and GRPO without manually designed rewards.
The combination of visual-world-model progress estimation and learned value-based reward shaping is a meaningful idea, and the reported gains are unusually large, but the abstract does not establish how broadly the method transfers or how much depends on benchmark-specific assumptions.
-
Wall-OSS-0.5 Technical Report
Wall-OSS-0.5 is a 4B vision-language-action model pretrained on over 20 robot embodiments, using a combination of discrete action prediction, multimodal learning, and continuous flow matching. Unlike most VLA reports, it evaluates the pretrained checkpoint directly on physical robots and reports useful zero-shot behavior, including progress on a held-out deformable-manipulation task; after fine-tuning, it outperforms π_0.5 on 15 tasks. The main claim is that large-scale VLA pretraining can produce executable robot capability rather than merely a better initialization.
The direct zero-shot real-robot evaluation across many embodiments addresses an important neglected question and could change how VLA pretraining is judged, but the abstract gives limited detail on task protocols, baselines, and the strength of the reported zero-shot results.
-
Few-Shot Resampling for Scalable Statistically-Sound Data Mining
The paper introduces FewRS, a resampling method for testing the statistical significance of data-mining discoveries while analyzing far fewer resampled datasets. It derives a bound on the maximum deviation of quality statistics and reports up to 100× lower runtime than existing resampling approaches in pattern mining and network-analysis experiments, while retaining statistical power.
The combination of rigorous false-discovery guarantees with reductions of up to two orders of magnitude in resampling cost is a substantial and broadly applicable efficiency result, though the abstract lacks enough experimental detail to justify a stronger verdict.
-
Fixed-Point Masked Generative Modeling
The paper replaces part of masked generative models’ repeated denoising computation with a shared-layer fixed-point solver, using cross-step consistency training and a three-state warm start across refinement steps. Its CoFRe system claims substantially lower parameter, training, and memory costs, while improving low-budget generation quality on text and images; it can also convert pretrained models with short fine-tuning rather than retraining from scratch.
Adaptive fixed-point computation with warm-started state reuse is a genuinely interesting direction for reducing masked-generation cost, and the reported gains are unusually large, but the abstract does not provide enough baseline, scale, or ablation detail to justify a strong verdict.
-
mRNAutilus: Multi-Objective-Guided Discrete Generation of mRNA with Optimized Therapeutic Properties
mRNAutilus uses a masked discrete-diffusion model together with Monte Carlo tree search to generate complete mRNA transcripts, jointly optimizing coding sequences and UTRs for stability, translation, and expression. The authors report very large expression gains for luciferase, strong results on Spike mRNAs, and improved expression in prime-editing and targeted protein-degradation applications, including designs made without target-specific experimental optimization.
Joint de novo generation of coding regions and UTRs under multiple objectives is a substantive direction, and the reported hundreds-fold expression gains are striking, but the abstract gives too little detail about experimental breadth, validation, and baseline comparability to justify a strong verdict.
-
Graph Neural Networks Are Not Continuous Across Graph Resolutions
This paper argues that standard graph neural networks can produce very different representations for graphs that describe the same object at different resolutions, because common message-passing schemes lack continuity across graph scales. It identifies a structural cause, proposes an architectural modification intended to enforce cross-resolution consistency, and reports theoretical and numerical validation across multiple experiments.
The potentially important contribution is a theoretical challenge to the assumption that GNN representations vary smoothly across graph resolutions, together with a principled fix, but the abstract gives no quantitative results and the overlap note warrants checking what is genuinely new.
-
Is Zero-Shot Super-Resolution Possible in Operator Learning?
This paper studies whether neural operators trained on coarse grids can reliably predict on finer grids without retraining. It shows that zero-shot super-resolution is information-theoretically impossible even for a simple rank-one linear operator, while output Hölder smoothness is sufficient for it and yields generalization bounds; experiments support the theoretical failure modes.
The impossibility result challenges broad claims about zero-shot super-resolution and gives a concrete condition under which it can work, but the abstract provides limited detail about the scope and practical impact of the bounds and experiments.
-
HiGS: A Hierarchical Rendering Architecture for Real-Time 3D Gaussian Splatting
HiGS decouples the spatial partitioning scale from the rasterization scale in 3D Gaussian Splatting: it bins Gaussians using coarse macro-tiles, then renders fine tiles within them. This reduces serialization in dense regions and reportedly achieves up to 15.8× the rendering speed of original 3DGS while preserving exact front-to-back compositing.
The separation of partitioning and rasterization scales is a substantive systems idea with a potentially large speedup, but the abstract gives only an upper-bound result and limited evidence across scenes, so it falls short of a strong recommendation.
-
Physical Object Understanding with a Physically Controllable World Model
The paper trains an autoregressive probabilistic world model that can estimate different visual and dynamical variables conditioned on one another. Its generated futures reveal motion correlations that can be used to discover objects and articulated parts, manipulate them in 3D, and reason about physical relationships in tasks such as Visual Jenga.
The combination of arbitrary conditional inference, emergent object discovery from sampled futures, and physical interaction reasoning is a genuinely interesting direction, but the abstract gives no quantitative comparisons or scale details to establish a strong result.
-
Harmonic: Hierarchical State Space Models for Efficient Long-Context Language Modeling
Harmonic is a hierarchical recurrent state-space language model with three timescales, where slower levels process prediction errors from faster levels rather than their hidden states. On enwiki8 and WikiText-103, it reportedly improves with context length relative to similarly sized Transformers and Mamba, trains at 64K tokens within 80GB, and at 1B parameters avoids the severe long-context degradation seen in a TinyLlama variant with RoPE.
The multiscale error-feedback design and reported long-context gains are genuinely interesting, but the evidence is concentrated on small language-modeling benchmarks and makes potentially misleading claims about RoPE limits, so the headline results need verification.
-
SKIP: Sparse Keyframe Interpolation Paradigm for Efficient Embodied World Models
SKIP generates only robot-relevant keyframes in a world-model rollout, then predicts gaps and interpolates the missing frames conditioned on actions. On LIBERO, it is 4.16× faster than dense video generation, reduces FVD by 89%, and produces training data that nearly preserves policy performance, including on a real robot; dense frame-by-frame generation largely fails when used in the same way.
The event-preserving sparse-to-dense design and large reported efficiency and policy-training gains are genuinely interesting, but the evidence is concentrated on LIBERO and one real-robot setting, making this promising rather than a clear field-changing result.
-
Certificates without Electrons? Theory and Evidence on Impacts from AI-Driven Power Demand
The paper combines a game-theoretic model with a difference-in-differences analysis to study how AI data-center demand affects local electricity systems. It argues that matching annual consumption with RECs is insufficient because renewable generation may not coincide with demand: AI growth is associated with more fossil generation, higher wholesale prices, and more outages, while on-site generation and storage reduce or reverse some effects. The proposed remedies are spatially distributed inference, relocating workloads, and colocated generation/storage rather than REC-only procurement.
The timing-wedge framing and claimed empirical link between LLM releases, data-center growth, and grid reliability are unusually relevant, though the abstract does not provide enough methodological detail to assess whether the causal evidence is strong.
-
Ryze: Evidence-Enriched Data Synthesis from Biomedical Papers
Ryze automatically turns biomedical papers into multimodal QA training data that preserves links among figures, tables, captions, extracted structure, and explanatory text. Its post-training pipeline produces an 8B biomedical VLM for under $200, reportedly improving Qwen3-VL-8B by 12.6 percentage points on LAB-Bench and exceeding GPT-5.2 by 3.8 points.
The evidence-preserving synthesis pipeline and unusually low-cost domain specialization are genuinely interesting, but the abstract gives results on only one benchmark and provides too little detail to establish that the gains generalize or that the comparison with GPT-5.2 is robust.
-
Hybrid Verified Decoding: Learning to Allocate Verification in Speculative Decoding
The paper predicts how many tokens a cached draft will be accepted for before verification, then chooses between cache-based drafting and a learned model-based drafter. Across three language models and sixteen datasets, this hybrid strategy reportedly beats EAGLE3 in all tested agentic-workflow settings, with an average 2.73× speedup. It also analyzes why prompt structure creates useful cache matches and finds that high-value drafts are concentrated in a small portion of the draft space.
The payoff-guided runtime choice between fundamentally different speculative-drafting sources is a meaningful efficiency direction with strong headline results, but the abstract lacks latency details and independent evidence that the gains generalize beyond the emphasized agentic workloads.
-
Revise, Don't Freeze: Sampler-Matched Training for Self-Correcting Masked Diffusion Language Models
The paper introduces D3IM, a parameter-free sampler for masked diffusion language models that can revise already visible tokens instead of freezing or remasking them. It also proposes SCOPE, a lightweight post-training method to counteract the model’s tendency to preserve its own mistakes; on LLaDA-8B, the combination produces sizable gains on GSM8K, MATH-500, HumanEval, and MBPP, especially with more denoising steps.
Direct visible-token revision is a meaningful and relatively principled use of diffusion models’ iterative prediction ability, with large reported math and coding gains, but the abstract provides results for only one model and a limited set of benchmarks, so it falls short of a strong recommendation.
-
STARFISH: faST Accuracy Recovery in pruned networks From Internal State Healing
STARFISH restores accuracy after weight pruning by training the pruned model to match the dense model’s internal representations, using only a tiny unlabeled calibration set. On reported ViT/ImageNet experiments, it substantially outperforms existing recovery methods, including recovering 82% of dense-model accuracy after removing 75% of DeiT-B’s weights with just 0.4% of the training images.
The combination of internal-state healing and extremely small unlabeled calibration data appears genuinely useful, especially under aggressive pruning, but the abstract gives limited evidence beyond selected ViT/ImageNet results and an ‘up to’ comparison.
-
FlexLink: Decoupling Control and Data Beams for Next-Generation Wideband Networks
FlexLink proposes a delay-phased-array architecture that distributes signal energy across both frequency and spatial directions, allowing separate control and data beams from the same array. A custom 4–7 GHz prototype reportedly supports multiple narrow beams without the usual gain loss or extra power, nearly doubling spectral efficiency relative to conventional phased arrays.
The hardware-prototyped decoupling of control and data beams could address a fundamental phased-array limitation, but the abstract provides limited detail on the experimental scope and comparison behind the nearly-twofold gain.
-
LEGS: Fine-Tuning Teleop-Free VLAs for Humanoid Loco-manipulation in an Embodied Gaussian Splatting World
LEGS trains humanoid vision-language-action policies without human teleoperation by combining a mesh-based robot and objects with a photorealistic 3D Gaussian Splatting scene, then generating demonstrations from procedural motion primitives. On a Unitree G1, three VLA backbones matched or outperformed teleoperation-trained policies on three loco-manipulation tasks, while re-rendering the same motions in new scenes improved robustness to appearance shifts at over 15× lower scene-coverage cost.
The notable result is credible-looking teleop-free transfer for humanoid loco-manipulation, with photorealistic scene capture and cheap re-rendered augmentation apparently outperforming teleoperation under scene and object shifts, but the evidence is limited to one robot, three tasks, and headline claims without detailed success rates in the abstract.
-
An Enigma of Artificial Reason: Investigating the Production-Evaluation Gap in Large Reasoning Models
This paper tests whether large reasoning models can detect flawed reasoning when the final answer is correct, using a dataset designed to separate reasoning evaluation from solving. Frontier models reportedly perform near-perfectly on producing solutions but score as low as 48% at identifying trivial flaws, and analyses suggest they confirm the answer and rationalize the steps rather than verify them. Probing and causal patching further implicate the final-answer representation in this bias.
The production-versus-evaluation gap and causal evidence for answer-confirmation bias are a genuinely non-obvious finding about reasoning models, but the abstract provides limited detail on dataset scale, model breadth, and robustness, so it merits a closer look rather than a strong recommendation.
-
Emergent Transfer of a Physics Foundation Model from Simulation to Laboratory Turbulence
The authors fine-tune a continuum-dynamics foundation model using at most three Rayleigh–Taylor instability simulations, then apply it without experimental training data to laboratory measurements. The model reproduces long-term simulated dynamics, shifts toward the experimentally observed mixing-growth regime, and generalizes to stable stratification, suggesting that initial conditions may explain part of the simulation–experiment discrepancy.
The zero-shot transfer from DNS to noisy laboratory turbulence and to an unseen physical regime is a genuinely interesting result, but the abstract gives too little quantitative detail about accuracy, baselines, and experimental breadth to justify a stronger verdict.
-
Noninvasive Diagnostic Ultrasound-Guided Focused Ultrasound Enables Selective, Reversible Inhibition of Peripheral Nociceptive Fibers and Prevents Acute Pain
The study uses diagnostic-ultrasound guidance to focus ultrasound noninvasively on the rat sciatic nerve before incision. In rats, this reduced thermal pain hypersensitivity for up to 16 weeks and mechanical hypersensitivity for up to four weeks, while apparently preserving motor and non-pain sensory reflexes. The notable result is selective, reversible nociceptive blockade without an injected anesthetic, though the evidence is still limited to an animal model.
Selective long-lasting analgesia from transcutaneous focused ultrasound would be a meaningful alternative to conventional nerve blocks, but the abstract provides only preclinical behavioral evidence and limited detail on durability, safety, mechanism, and replication.
-
The slow-acting G-protein Gz defines the duration of circadian rest time
This study identifies the slow-acting G-protein Gz as a regulator of how long mice remain in the nightly rest phase. Removing Gz shortened rest from about 10 to 7.5 hours, while genetic rescue restored it; the effect tracked changes in SCN cAMP-PKA activity and occurred without disrupting the 24-hour rhythm itself.
It provides a relatively specific mechanism for allocating circadian rest duration—distinct from setting the clock period—but the evidence is currently limited to mouse genetics and the abstract does not establish broader relevance.
-
Visual speech supports phonetic attunement in the absence of early auditory input
The study compares neural speech encoding in hearing children and cochlear-implant users who did or did not receive auditory input during the first year of life. Auditory phonetic-feature encoding depended on early auditory exposure, while visually discriminable phonetic features were encoded even without early hearing, suggesting that early experience shapes phonetic specialization through a modality-flexible but experience-dependent process.
The cross-modal dissociation offers a potentially important refinement of sensitive-period theory, but the abstract provides no sample sizes, effect sizes, or methodological detail needed to judge how robust the conclusion is.
-
Causal and directional elements of global brain dynamics
The authors identify conserved, directional spatiotemporal patterns in spontaneous cortex-wide activity across cell types and frequency ranges, including fast dynamics measured with genetically encoded voltage sensors. Optogenetic experiments suggest these patterns have a causal source-to-sink hierarchy: stimulating source regions recruits the full pattern, whereas stimulating sink regions does not; manipulating and targeting these patterns also affected visual detection behavior.
This presents a potentially important causal framework for large-scale brain dynamics, supported by broad recording, perturbation, pharmacological, and behavioral experiments, but the abstract gives few quantitative details and the claims await validation beyond the reported preparation.
-
Local control of dopamine release in nucleus accumbens gates opioid withdrawal aversion
The study identifies a previously uncharacterized population of MOR- and Tshz1-expressing neurons in the nucleus accumbens that controls the affective, but not physical, symptoms of opioid withdrawal. In mice, manipulating these neurons altered withdrawal-related dopamine suppression and aversion learning, while mGluR8 activation reduced aversion, suggesting a potentially druggable circuit target.
The causal separation of affective and physical withdrawal through a specific accumbal cell population is a substantive mechanistic result, but the abstract provides no quantitative detail and the therapeutic claim remains preclinical.
-
Directed cortical connectivity inferred from neural energy metabolism
The paper proposes estimating the direction of communication between cortical regions from asymmetric metabolic costs, especially the greater energy use of postsynaptic neurons. Applying this to PET-derived metabolism and an average metabolic template for fMRI, the authors find a reproducible hierarchy in which sensory and attention systems drive higher-order networks, while default-mode and control systems mainly receive input; mitochondrial density and cortical lamination provide independent cellular correlates.
The metabolism-based route to whole-brain directed connectivity is a genuinely interesting idea with biologically grounded validation, but the abstract gives no quantitative comparisons or details establishing that the inferred direction is more than a plausible proxy for known cortical hierarchy.
-
A biophysical corticostriatal model predicts learning-dependent 7T fMRI dynamics and individual reward bias in humans
The authors link a biophysical corticostriatal model to human fMRI by converting simulated neural activity through a hemodynamic model and predicting behavior. The model makes a counterintuitive prediction: learning can increase prefrontal–striatal coherence at the LFP level while decreasing their BOLD correlation, and this pattern is observed in human data; it also fits individual differences in positive versus negative reward bias.
The cross-scale, mechanistic prediction that neural coupling and BOLD coupling can change in opposite directions is genuinely interesting and reportedly validated in humans, but the abstract lacks sample sizes and quantitative evidence needed for a stronger verdict.
-
Electrophysiological lag threads reveal a temporal hierarchy of the human cortex
The authors introduce lag-specific orthogonalization to estimate non-zero-lag relationships between cortical activity while reducing EEG/MEG volume-conduction artifacts. In resting-state MEG, they find reproducible alpha- and beta-band temporal ordering from occipital and sensorimotor areas toward frontal and temporal association cortex, with spatial similarities to much slower fMRI lag patterns; these patterns also vary with age, handedness, and task demands.
This combines a potentially useful methodological advance with an unexpected cross-modal result linking fast electrophysiological lag structure to slow fMRI propagation, but the abstract provides limited quantitative validation and the causal interpretation remains uncertain.
-
High sensitivity imaging of intracellular Cl- reveals the spatiotemporal evolution of GABAergic chloride fluxes in vitro and in vivo
The authors introduce iClima, a ratiometric genetically encoded chloride sensor designed to combine high affinity, pH robustness, and photostability. It detects GABA-related chloride transients, dendritic chloride hotspots, and sensory-evoked changes in vivo, and can be used alongside GCaMP6f to compare inhibition and excitation in the same neurons. In that experiment, somatic inhibitory drive was not orientation-tuned, whereas excitatory calcium responses were strongly tuned.
This is a potentially enabling imaging tool with in vivo dual measurement of inhibition and excitation, plus a non-obvious result about their differing orientation selectivity, but the abstract lacks quantitative validation and broader demonstrations needed for a strong recommendation.
-
Choice-driven remapping of action- and stimulus-anchored value in human single neurons
The authors recorded human single-neuron activity across vmPFC, ACC, preSMA, and amygdala while participants made action-based and stimulus-based reward choices. Neurons initially represented value across both domains, but after commitment their coding diverged: ACC preferentially represented action value, vmPFC stimulus value, and preSMA tracked chosen action identity and value regardless of domain. The main contribution is a proposed transition from shared valuation to domain-specific choice execution across prefrontal regions.
Human single-neuron evidence for a domain-general-to-segregated transition, including an ACC–vmPFC double dissociation, is a genuinely interesting result, but the abstract gives no sample sizes, effect magnitudes, or robustness analyses to justify a stronger recommendation.
-
A Toxic Tau-PFKFB3 Circuit Reduces F2,6BP Levels and Drives Neurodegeneration
This study proposes that Tau aggregates trap the metabolic enzyme PFKFB3, lowering neuronal F2,6BP and creating a feedback loop that worsens DNA damage, Tau phosphorylation, and aggregation. Across cultured neurons, hippocampal slices, and a Drosophila model, adding F2,6BP reportedly rescues several disease-related phenotypes, suggesting that this metabolite may coordinate metabolism, genome repair, and Tau proteostasis.
The potentially important and unexpected finding is that F2,6BP acts directly on both DNA-repair and Tau-proteostasis pathways, but the abstract provides no quantitative effect sizes or mammalian in vivo evidence, so the broad therapeutic claims need closer scrutiny.