Issue 24 · Jun 8–14, 2026
Every candidate
All 3,588 papers were scored from their abstracts by gpt-5.6-luna; 1,337 were not skipped. Shown here: the top 150 of those, in score order. Picks are marked.
-
Models Take Notes at Prefill: KV Cache Can Be Editable and Composable
The paper argues that a transformer’s KV cache stores not just token representations, but downstream conclusions already conditioned on earlier fields. It shows that these cached “notes” can be edited after an input correction and repositioned or spliced into new contexts, achieving near-full-recompute behavior with much lower latency; an online vLLM test reports large time-to-first-token reductions.
This proposes a potentially important new view and capability for KV caches—editable, composable memoized computation—with unusually large reported latency gains and validation across models, modalities, and serving conditions.
-
MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling
MaxProof trains one model to generate, verify, critique, repair, and rank mathematical proofs, then uses population-based test-time search and tournament selection to choose a final proof. The reported system scores 35/42 on IMO 2025 and 36/42 on USAMO 2026, above the stated human gold-medal threshold on both competitions.
The claimed competition-level proof performance and unified generative-verifier test-time search would be a major result, but the abstract provides too little detail about proof validity, baselines, contamination controls, and evaluation methodology for a strong verdict.
-
SimWeaver: Zero-Shot RGB Sim-to-Real for Deformable Manipulation
SimWeaver trains RGB vision-language-action policies entirely in simulation and transfers them to real-world deformable-manipulation tasks without real-data fine-tuning or per-task calibration. Using 200 simulated demonstrations per task, it reports over 80% success per task and 91% average success across five tasks, with especially strong robustness to visual shifts in silk grasping and much lower trajectory-generation cost than real-data collection.
Zero-shot sim-to-real for visually complex deformable objects is a difficult and relatively unresolved problem, and the reported cross-task real-world success, distribution-shift robustness, and large data-collection cost reduction would be a substantial advance if independently validated.
-
InterleaveThinker: Reinforcing Agentic Interleaved Generation
InterleaveThinker adds a planner-and-critic multi-agent loop around existing image generators to produce text-image sequences, using supervised tuning and step-wise reinforcement learning to correct generation instructions. The authors report broad improvements across image generators, including results comparable to Nano Banana and GPT-5 on interleaved-generation benchmarks, plus gains on multi-step visual reasoning tasks.
The potentially important idea is turning otherwise single-image generators into interleaved generators through trajectory-level planning and criticism, but the abstract gives no quantitative details or independent evidence for the strong benchmark comparisons.
-
Latent Spatial Memory for Video World Models
The paper proposes storing a video world model’s persistent 3D scene memory directly in diffusion latent space rather than repeatedly rendering RGB images and re-encoding them. Its Mirage system lifts latent tokens into 3D using depth and generates novel views through latent-space warping, reportedly achieving up to 10.57× faster generation and 55× lower memory than explicit 3D-memory baselines while remaining competitive on world-model benchmarks.
This is a genuinely interesting systems and representation shift with unusually large claimed efficiency gains, but the abstract provides limited evidence beyond a few benchmarks and does not establish how broadly the approach generalizes.
-
FlashMemory-DeepSeek-V4: Lightning Index Ultra-Long Context via Lookahead Sparse Attention
The paper introduces Lookahead Sparse Attention, which uses a separately trained dual-encoder indexer to predict which historical KV-cache chunks will be needed and keep only those in GPU memory. On reported long-context benchmarks, the system retains roughly 13.5% of the KV cache while maintaining or slightly improving accuracy; at 1M tokens it claims 2.8× throughput and 2.7× concurrency gains. The notable design choice is training the indexer independently of the large backbone, avoiding the need to load the full model during indexer training.
The combination of learned query-critical KV retrieval, substantial 1M-context memory reduction, and backbone-free training is potentially important, but the abstract does not establish how robust the retrieval is across workloads or how the comparisons and serving gains were measured.
-
The Signs Were Always There: Training-Free Concept Detection and Steering in Raw Transformer Dimensions
This paper argues that raw transformer hidden-state dimensions already provide a useful, training-free basis for detecting and steering concepts, without sparse dictionaries or learned probes. Across language, vision, and audio models, sign patterns reportedly detect hundreds of concepts with AUC up to 0.99, preserve much of language-model accuracy when magnitudes are discarded, and enable closed-loop steering in four language models. It also claims that concept “readers” and “writers” use different dimensions, suggesting a new way to study transformer representations.
The claim that standard-basis signs are broadly causal and nearly as informative as learned representation rotations would overturn common assumptions in mechanistic interpretability, but the unusually sweeping results and incomplete steering release warrant verification before treating it as a major finding.
-
Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models
Embodied-R1.5 is an 8B-parameter model combining embodied reasoning, planning, grounding, correction, and pointing, trained on more than 15B tokens with automated data pipelines and balanced RL. Its Planner-Grounder-Corrector loop is intended to support long-horizon execution and self-correction; the authors report leading results across 16 of 24 embodied-VLM benchmarks, strong VLA fine-tuning results, and zero-shot real-robot demonstrations.
The unified closed-loop embodied model and broad real-robot evaluation could be important, but the abstract provides no quantitative margins or details sufficient to distinguish a genuinely new capability from a large, well-engineered system and optimistic SOTA claims.
-
MiniMax Sparse Attention
The paper introduces a block-sparse attention mechanism for million-token contexts. An index branch selects different key-value blocks for each GQA group, while an optimized GPU kernel computes exact attention only over those blocks; in a 109B multimodal model, it reportedly matches GQA quality while reducing attention compute 28.4x and achieving 14.2x prefill and 7.6x decoding speedups at 1M context on H800 GPUs.
The combination of group-specific block selection, model–kernel co-design, and reported million-token speedups is unusually consequential, but the abstract does not provide enough task-level quality results or comparisons to judge whether the claimed efficiency translates broadly beyond the reported setup.
-
ABot-Earth 0.5: Generative 3D Earth Model
ABot-Earth 0.5 generates large 3D environments directly in a 3D Gaussian Splatting representation, using only georeferenced satellite imagery as conditioning. It claims generation at under 10 minutes per square kilometer, with hierarchical level of detail for web-based interactive viewing and potential UAV-navigation simulation. The notable idea is combining satellite-conditioned generative reconstruction with scalable, directly generated 3DGS scenes rather than reconstructing each environment through conventional pipelines.
The direct generative 3DGS formulation and claimed large-area, low-cost reconstruction could be important, but the abstract provides no quantitative fidelity, coverage, or sim-to-real/navigation results to substantiate the broader claims.
-
Dynamic trajectory cues drive sequenced integration in approach detectors
The study shows that changing luminance alone can make both humans and fruit flies perceive an object as approaching or retreating, even without the usual cue of image expansion. In flies, the authors identify neurons that respond to both luminance-based and expansion-based approach, and find that the cues are integrated synergistically only when luminance changes come first—the temporal order expected from real approaching objects.
It links a previously unrecognized cross-species approach percept to identified neural circuitry and reveals a biologically grounded, sequence-sensitive mechanism for integrating correlated motion cues.
-
Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling
Bebop studies why multi-token prediction (MTP) speculative decoding becomes less effective during LLM reinforcement-learning training. It links lower acceptance rates to increased entropy fluctuations, uses probabilistic rejection sampling, and introduces an end-to-end total-variation loss targeted directly at acceptance; the authors report up to 95% acceptance, 25% extra inference throughput, and 1.8× end-to-end acceleration across reasoning, coding, and agentic RL workloads. They also find that training the MTP component before RL can maintain speedups without expensive online updates.
The entropy-based analysis and loss designed around rejection-sampling acceptance are more than routine speculative-decoding tweaks, with potentially important RL-training speedups, but the abstract provides mainly best-case gains and insufficient detail to establish broad or robust superiority.
-
Attention Amnesia in Hybrid LLMs: When CoT Fine-Tuning Breaks Long-Range Recall, and How to Fix It
The paper reports that chain-of-thought fine-tuning can severely damage long-context retrieval in hybrid linear-attention LLMs, apparently by changing the query/key projections that support long-range routing. It proposes restoring those projections from the pre-fine-tuning checkpoint, optionally with a Procrustes adjustment; this reportedly recovers much of the lost 256K-context retrieval without retraining while retaining reasoning ability.
The striking cross-architecture failure mode and zero-training QK restoration are worth checking, but the evidence is currently limited to abstract-level claims and synthetic needle-retrieval results on a relatively specialized class of models.
-
Pythagoras-Prover: Advancing Efficient Formal Proving via Augmented Lean Formalisation
Pythagoras-Prover develops compute-efficient Lean theorem provers, including 4B/32B autoregressive models and an experimental diffusion-based prover. It combines curriculum training, filtering of long proof traces, and Augmented Lean Formalisation (ALF), which generates mutated formal problems to expand limited verified data. The 4B model reportedly reaches 86.1% pass@32 on MiniF2F-Test—above a 671B-parameter prior model—and the 32B model reaches 93.0%, though the evidence is concentrated on formal-math benchmarks.
The striking parameter-efficiency result and ALF data-augmentation approach merit attention, but the claims rely mainly on MiniF2F/PutnamBench results and the abstract gives limited evidence that the gains generalize beyond these benchmarks.
-
Retrieve, Don't Retrain: Extending Vision Language Action Models to New Tasks at Test Time
The paper trains a vision-language-action policy once using demonstrations from a target robot and a cheaper source embodiment, then adds new tasks at deployment by indexing source-side demonstrations for retrieval rather than fine-tuning. Retrieved trajectories provide task-progress and motion priors, while a video-generation-based policy helps maintain visual consistency; experiments report gains on PushT, RoboTwin 2.0, and a real robot, including unseen tasks and goal angles.
Replacing per-task robot adaptation with retrieval over cross-embodiment demonstrations is a meaningful deployment idea, but the abstract gives no quantitative scale or robustness evidence sufficient for a strong recommendation.
-
SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning
SpatialClaw gives a vision-language model a stateful Python environment containing spatial perception and geometry tools. Instead of making one complete tool call or analysis plan, the model executes code cells step by step, reusing prior results and adapting to intermediate text and visual outputs. Across 20 static and dynamic 3D/4D benchmarks, it reports 59.9% average accuracy—11.2 points above a recent spatial agent—with gains across six VLM backbones without task-specific adaptation.
The stepwise, stateful code interface is a meaningful and plausibly general design for open-ended spatial reasoning, and the broad evaluation with a large reported gain makes it worth a closer look, though the abstract does not establish a major capability breakthrough beyond improved agent orchestration.
-
Lip Forcing: Few-Step Autoregressive Diffusion for Real-time Lip Synchronization
Lip Forcing distills a large bidirectional video-diffusion lip-sync model into causal autoregressive students that generate each video chunk in two denoising steps. It uses an analysis of classifier-free-guidance trajectories to select a synchronization-focused training window, schedule, and reward, achieving 31 FPS with a 1.3B model and reporting up to 39.8× speedup over the 14B teacher while retaining comparable reference fidelity.
The combination of trajectory-based distillation and causal few-step video diffusion appears technically distinctive and delivers a meaningful real-time speed breakthrough, but the abstract does not provide enough quantitative quality or robustness results to justify a strong verdict.
-
i1: A Simple and Fully Open Recipe for Strong Text-to-Image Models
The authors run 300+ controlled experiments on text-to-image diffusion training and inference, then use the findings to train a fully open 3B-parameter model from public data. Their model is competitive with leading systems on five benchmarks and improves the previous best fully open model by 29.5 percentage points on average, with weights, code, and data-processing pipelines released.
The unusually broad controlled study plus a large improvement over prior fully open models makes this worth knowing about, but the core result is a carefully optimized diffusion recipe rather than a clearly new modeling paradigm, and the abstract gives limited detail on benchmark comparability.
-
Toward Generalist Autonomous Research via Hypothesis-Tree Refinement
Arbor is an autonomous-research framework that maintains a persistent tree of hypotheses, experiments, artifacts, and lessons, using a coordinator to assign work to isolated executor agents. On six optimization tasks spanning model training, harness engineering, and data synthesis, it reportedly beats Codex and Claude Code under matched budgets, with over 2.5× their average held-out improvement; it also reports 86.36% Any Medal on MLE-Bench Lite with GPT-5.5.
Persistent cross-experiment hypothesis management for long-horizon agentic research is a meaningful design direction, but the evidence is limited to six tasks and benchmark comparisons, so the broad generality and size of the claimed advantage need closer scrutiny.
-
End-to-End Context Compression at Scale
The paper develops Latent Context Language Models, encoder–decoder systems that compress long token sequences into latent embeddings before a decoder processes them. After large-scale architecture searches and continued pretraining—0.6B encoders and 4B decoders trained on over 350B tokens at multiple compression ratios—the authors report improved tradeoffs among task quality, compression speed, and memory, including adaptive retrieval of relevant context for long-horizon agents.
This is a potentially important alternative to KV-cache compression with unusually large-scale training and a claimed Pareto improvement, but the abstract gives no quantitative results or comparisons strong enough to justify a strong verdict.
-
VideoMDM: Towards 3D Human Motion Generation From 2D Supervision
VideoMDM trains a 3D human-motion diffusion model using only accurate 2D poses from monocular video, with a pretrained 2D-to-3D lifter supplying noisy depth and a reprojection loss providing supervision. The authors show that a depth-weighted 2D loss can match 3D supervision in expectation, and report an FID of 0.88 on HumanML3D versus 0.54 for fully 3D-supervised MDM, plus positive human evaluations on real-video datasets.
It offers a credible route to learning coherent 3D motion priors without expensive 3D labels, but the remaining performance gap, reliance on a pretrained 3D lifter, and limited abstract-level evidence make it promising rather than indispensable.
-
FastContext: Training Efficient Repository Explorer for Coding Agents
FastContext separates repository search from code generation by using a dedicated 4B–30B exploration model that performs parallel tool calls and returns concise file-and-line citations. Trained from reference trajectories with task-specific rewards, it reportedly improves Mini-SWE-Agent resolution by up to 5.5% while cutting token use by up to 60% across several SWE-bench-style evaluations.
The explicit separation of repository exploration from solving, combined with specialized smaller models and substantial reported context savings, is a meaningful agent-design idea, but the abstract does not establish how broadly the gains generalize or how much comes from the training setup.
-
Track2View: 4D-Consistent Camera-Controlled Video Generation via Paired 3D Point Tracks
Track2View uses sparse, paired 3D point tracks between source and target camera views to condition a video diffusion transformer. The explicit, temporally continuous correspondences improve preservation of scene dynamics and camera control across static and dynamic videos; on a 400-video benchmark, it reports 30–65% lower rotation error and 61–72% lower translation error than prior methods.
The explicit use of paired 3D point trajectories as a general camera-control interface is a meaningful alternative to pose embeddings and implicit correspondence, with reasonably broad benchmark evidence, but the abstract does not establish a major capability jump or generalization beyond the curated multi-camera setting.
-
JoyAI-VL-Interaction: Real-Time Vision-Language Interaction Intelligence
The paper presents an 8B vision-language model that continuously monitors video, decides whether to remain silent or respond, and can delegate difficult cases to a larger background model. It also releases a deployable streaming system with interchangeable speech, memory, UI, and agent components, claiming emergent abilities such as screen guidance and slide-based lecturing; human raters reportedly preferred it to two commercial video-call assistants across six scenarios.
Proactive, internally gated real-time visual interaction is a relevant and potentially important direction, but the abstract provides few quantitative details and the broad capability and user-preference claims are not yet strongly substantiated.
-
TopoCap: Learning Topology-Agnostic Motion Priors for Monocular Video-to-Animation
TopoCap learns a shared motion representation across many different skeletal structures, then predicts that representation from monocular video and decodes it for a specified target rig. The authors claim zero-shot retargeting across unseen topologies—including bipeds, quadrupeds, hexapods, and non-animal objects—without test-time optimization, supported by a 5,000-topology, 2-million-frame dataset.
Topology-agnostic monocular motion capture would be a genuinely useful change in generality, but the abstract gives no quantitative results or detail sufficient to establish that the broad zero-shot claims work reliably.
-
The Program Is Still There: A Conservation Law for Program Discovery
The paper studies black-box program discovery, where a search algorithm can only observe a candidate program’s score. It claims a lower bound exponential in a problem’s “coupling width,” and presents a conservation law saying that structural information supplied to the search trades directly against the remaining search effort. A deterministic implementation reportedly recovers generators for 2,383 of 3,914 sequences, including 244 of 256 elementary cellular automata, and validates them by predicting held-out continuation data.
The claimed lower bound and information-versus-search tradeoff could be a genuinely useful framework for understanding program synthesis, but the abstract is unusually sweeping and does not provide enough detail to assess the theorem’s assumptions, the meaning of coupling width, or how broad the empirical comparison really is.
-
ANEForge: Python for direct computation on the Apple Neural Engine
ANEForge is a Python compiler/runtime that directly targets Apple’s Neural Engine rather than relying on CoreML, compiling lazy tensor graphs into single ANE programs. It supports quantized and sparse weights, resident recurrent state, fused attention, and unusually claims training primitives—forward, backward, and optimizer updates—on the ANE itself; demonstrations include ResNet-18, a sentence encoder, ViT, and a Stable Diffusion U-Net, with a 0.33 ms ResNet-18 forward pass.
Direct, general-purpose programming of the normally opaque ANE—and especially reported backward and optimizer execution—would be a genuinely useful new capability, but the abstract gives limited comparative evidence and does not establish real-world gains over GPU/CoreML baselines.
-
A Fully Endovascular Neural Interface
The authors developed a sub-1-mm³ neural implant that is delivered through a blood vessel, expands against the vessel wall, and receives power and telemetry wirelessly via ultrasound without needing a particular orientation. In rabbits, carotid-implanted devices stimulated autonomic circuits and modulated blood pressure, demonstrating a less invasive route to targeted neural stimulation.
A fully endovascular, wireless neural interface with a demonstrated physiological effect is a genuinely new and potentially important platform, but the evidence is still limited to a small-animal stimulation experiment and does not yet establish durable safety, recording capability, or broad applicability.
-
EurekAgent: Agent Environment Engineering is All You Need For Autonomous Scientific Discovery
EurekAgent argues that the main bottleneck in autonomous scientific discovery is increasingly the design of the agent’s environment rather than its workflow. It combines sandboxed permissions, Git/filesystem-based artifact sharing, budget-aware exploration, and lightweight human intervention, reporting new results across mathematics, kernel engineering, and machine learning—including a 26-circle-packing result found for under $11 in API costs.
The environment-engineering framing and unusually low-cost circle-packing result are genuinely worth checking, but the abstract gives too few comparative details to establish how broadly or substantially it advances autonomous discovery.
-
See What I See, Know What I Think: Dense Latent Communication Across Heterogeneous Agents
This paper studies communication between agents built from different language models by transferring KV caches instead of decoded text. It proposes a lightweight cache transformation trained first to reconstruct the sender’s internal state and then to support generation; across Qwen3 model pairs, it reportedly preserves both visual/contextual information and reasoning signals, including when the receiver has no input, while using roughly 2–3× less compute than text communication.
Cross-model latent alignment that works even for context-unaware receivers is a meaningful advance over typical homogeneous or steering-only cache transfer, but the abstract lacks quantitative details and independent evidence of broad generality.
-
One Token per Multimodal Evidence: Latent Memory for Resource-Constrained QA
The paper compresses each text or image evidence item into a single latent token, retrieves those latent tokens, and feeds them directly to a pretrained LLM or VLM instead of passing raw passages or images. A small compressor is trained jointly for reconstruction, retrieval, and generation; across text and multimodal QA benchmarks, the method reportedly matches strong RAG baselines while using 3–10× fewer generator tokens, with the best image-grounded results on WebQA.
The unified latent-memory interface and substantial claimed token reduction are worth examining, but the abstract reports mostly competitive rather than clearly superior QA quality and leaves key details about training cost, latent compatibility, and robustness unspecified.
-
APPO: Agentic Procedural Policy Optimization
APPO proposes branching agentic RL rollouts at fine-grained, influential token-level decision points rather than only at tool calls or fixed workflow boundaries. Its branching score combines token uncertainty with the likelihood gains of later continuations, and a procedure-level advantage distributes credit across resulting rollouts; experiments on 13 benchmarks reportedly improve strong baselines by nearly 4 points without increasing tool calls.
The token-level branching and credit-assignment framing is a meaningful departure from standard tool-call-level agentic RL, but the abstract gives limited detail and only an aggregate improvement, making this worth noting rather than a must-read.
-
MoVerse: Real-Time Video World Modeling with Panoramic Gaussian Scaffold
MoVerse turns a single narrow-view image into an interactively navigable 3D scene. It first generates a gravity-aligned 360° panorama, lifts it into a persistent Gaussian representation, and uses a distilled causal video renderer to produce user-controlled camera views at 8 FPS on one RTX 4090. The notable design is the separation of world construction from fast observation rendering, combining explicit 3D memory with generative video synthesis.
The single-image-to-real-time, freely navigable world-modeling pipeline is a substantial and timely direction, but the abstract gives no quality, generalization, or comparison results beyond the 8 FPS claim, so it does not yet merit strong priority.
-
AHA-WAM:Asynchronous Horizon-Adaptive World-Action Modeling with Observation-Guided Context Routing
AHA-WAM separates slow visual world modeling from fast action generation instead of running both at the same temporal rate. Its video model maintains long-horizon scene context, while an action model repeatedly queries that context for responsive control; on RoboTwin it reports 92.8% success and on four real tasks 78.3%, with 24.17 Hz control and a 4.59× speedup over Fast-WAM without robot-data pretraining.
The asynchronous world-planner/action-executor design and reported real-time, real-robot results are meaningfully interesting, but the abstract provides limited baseline and task detail, so the claimed gains are not yet enough for a strong recommendation.
-
DeNovoSWE: Scaling Long-Horizon Environments for Generating Entire Repositories from Scratch
DeNovoSWE is a 4,818-example dataset for training agents to generate complete software repositories from documentation rather than modify isolated files. It uses an automated sandboxed workflow with decomposition, critique/repair, and difficulty-based filtering; fine-tuning Qwen3-30B-A3B reportedly raises performance on BeyondSWE-Doc2Repo from 5.8% to 47.2%.
The nearly eightfold benchmark improvement suggests whole-repository generation may benefit dramatically from targeted long-horizon data, but this is primarily a synthetic dataset paper and the abstract gives too little detail to establish how general or contamination-free the result is.
-
Rethinking the Role of Efficient Attention in Hybrid Architectures
This paper studies hybrid language models that mix full attention with sliding-window or recurrent efficient-attention layers. It finds that full attention performs most long-range retrieval, while efficient attention mainly changes how quickly retrieval abilities develop; unexpectedly, larger sliding windows can delay the emergence of retrieval heads. Based on this analysis, applying NoPE only to full-attention layers improves long-context performance in a small-window hybrid without much short-context degradation.
The mechanistic claim that efficient attention influences optimization speed rather than eventual long-context capability, plus the counterintuitive large-window laziness effect and targeted NoPE intervention, is more than a routine architecture tweak, though the abstract provides limited quantitative evidence about scale and generality.
-
SearchSwarm: Towards Delegation Intelligence in Agentic LLMs for Long-Horizon Deep Research
The paper trains an LLM to make delegation decisions for long-horizon web research: decomposing tasks, deciding what to hand to subagents, and incorporating their results. A guided multi-agent harness generates demonstrations of these behaviors, which are then used for supervised fine-tuning; the resulting 30B-parameter MoE model reports 68.1 on BrowseComp and 73.3 on BrowseComp-ZH, allegedly best among similarly sized models.
Internalizing delegation behavior from harness-generated trajectories is a substantive and timely approach to context-limited agents, but the abstract gives little detail about baselines, ablations, or whether gains come from delegation itself rather than the broader training setup.
-
Lineage tracing and live-cell imaging reveal that NeuroD1 does not reprogram microglia into neurons
This study used virus-free lineage tracing, longitudinal two-photon imaging, and single-cell RNA sequencing to test whether NeuroD1 converts microglia into neurons in vivo. Across normal and injury conditions, NeuroD1-expressing cells remained microglia rather than becoming neurons, while sustained NeuroD1 expression promoted microglial apoptosis; the abstract’s final sentence appears to contain a contradiction or typo.
It directly challenges a prominent and controversial claim about in situ glia-to-neuron reprogramming using several complementary cell-fate methods, potentially changing how prior conversion results are interpreted.
-
HYDRA-X: Native Unified Multimodal Models with Holistic Visual Tokenizers
HYDRA-X uses one ViT-based visual tokenizer for both images and videos, combining hierarchical temporal compression with image/video teacher supervision. Its ablations suggest that frame-level causal temporal attention is better for reconstruction than full spatiotemporal attention, and it moves image/video editing interactions into the tokenizer’s latent space rather than the language model. The 7B system reportedly performs well on both understanding and generation, though the abstract gives no quantitative results.
The unified image-video tokenizer and latent-level editing formulation are plausibly meaningful directions for omni-modal models, but the abstract provides insufficient quantitative evidence to establish a major advance.
-
HarnessX: A Composable, Adaptive, and Evolvable Agent Harness Foundry
HarnessX treats an agent’s prompts, tools, memory, and control flow as a modular “harness” that can be automatically composed and evolved from execution traces. Its AEGIS system uses multi-agent search and reinforcement-learning-like feedback to update harnesses and training signals, reporting average gains of 14.5% across five agent benchmarks, with a maximum gain of 44%.
The potentially important idea is systematic, trace-driven evolution of the agent runtime rather than relying only on larger models, but the abstract does not establish how much of the gains come from new search, extra test-time compute, or benchmark-specific engineering.
-
World Tracing: Generative Pixel-Aligned Geometry Beyond the Visible
World Tracing represents a scene as an ordered stack of 3D points for every image pixel: the first point matches the visible surface, while later points model occluded geometry behind it. Its diffusion-transformer implementation jointly reconstructs observed surfaces and generates hidden structure, retaining pixel-to-3D correspondence for editing, novel-view video, and mesh-generation pipelines. The abstract reports gains over depth predictors and image-to-3D systems across object, scene, and dynamic benchmarks, but gives no quantitative results.
The per-pixel, multi-layer representation is a genuinely interesting way to combine geometric faithfulness with completion, but the abstract provides insufficient quantitative evidence to justify a strong recommendation.
-
Claw-SWE-Bench: A Benchmark for Evaluating OpenClaw-style Agent Harnesses on Coding Tasks
The paper introduces a multilingual SWE-bench-style protocol for comparing autonomous coding-agent harnesses under fixed prompts, budgets, workspace rules, patch extraction, and evaluation. In its experiments, adapter design raises OpenClaw’s Pass@1 from 19.1% to 73.4% with the same underlying model, while harness choice has nearly as much effect as model choice and systems with similar accuracy vary substantially in API cost. The main contribution is treating the agent harness and cost, rather than just the language model, as central variables in coding-agent evaluation.
The benchmark itself is not inherently compelling, but the large same-model gap from adapter design and the finding that harness choice rivals model choice are non-obvious and potentially important for evaluating coding agents.
-
VIA-SD: Verification via Intra-Model Routing for Speculative Decoding
VIA-SD adds an intermediate verification tier to speculative decoding: high-confidence draft tokens are accepted directly, medium-confidence tokens are regenerated by a routed slim version of the verifier, and only uncertain tokens use the full model. Across four tasks and multiple model families, it reports 10–20% speedups over strong speculative-decoding baselines and 2.5–3× over ordinary decoding, without changing draft-model training.
The intra-model slim-verifier idea is a meaningful and potentially general extension of speculative decoding, but the reported gains are moderate and the abstract does not establish how much routing overhead or hardware dependence affects the result.
-
Squeeze-Release: Iterative Pruning with Exact Structural Minimization
The paper proposes repeatedly pruning a network, physically removing zeroed structures, and then adding small noise back into the freed positions so later training can discover further redundancy. It also introduces a function-preserving LayerNorm replacement to enable channel reduction in residual networks. The authors report 39× compression for a fully connected model and 14.8× for ConvNeXt-Tiny at comparable accuracy, while only proving—not demonstrating—the extension to transformers.
The iterative prune–compact–release loop and claimed order-of-magnitude dense-model compression are potentially substantial, but the abstract gives limited experimental detail and does not show transformer results, so the claims need verification.
-
Next Forcing: Causal World Modeling with Multi-Chunk Prediction
Next Forcing trains causal video/world models to predict several future video chunks simultaneously, using lightweight auxiliary modules whose predictions form a causal chain. The authors report faster convergence, improved high-frame-rate accuracy, roughly 2x inference speedup, better physical-law adherence, and strong RoboTwin results, including a 50%+ FVD reduction in general video pretraining.
The multi-horizon, multi-chunk objective is a meaningful adaptation of multi-token prediction to causal video modeling and could improve both training and inference, but the abstract provides limited detail and mostly reports claimed benchmark gains rather than enough evidence for a strong verdict.
-
Interpreting and Steering a Text-to-Speech Language Model with Sparse Autoencoders
The paper applies BatchTopK sparse autoencoders to CosyVoice3’s shared text-and-speech-token language-model backbone, and labels features according to whether they respond to text, speech, or both. It finds interpretable features for phonemes, laughter, accents, and speaker attributes, then uses latent interventions to control laughter, perceived gender, and speaking rate while retaining content. The notable contribution is showing causal, usable control directions in a TTS model rather than only post-hoc feature descriptions.
The combination of sparse-feature interpretability and causal steering in a multimodal TTS residual stream is genuinely interesting, especially the large laughter intervention, but the abstract gives limited evidence about breadth, robustness, perceptual evaluation, and whether the approach generalizes beyond one model.
-
Dexterous Point Policy: Learning Point-based Dexterous Hand Policies from Human Demonstrations
Dexterous Point Policy learns a transformer policy over 3D object, wrist, and fingertip keypoints extracted from human videos, then transfers it directly to a multi-fingered robot without robot demonstrations. On real-robot pick-and-place and tool-use tasks, it reports 75% success versus 1% for a VLA baseline, with generalization to multiple objects and unseen categories.
The striking result is that a shared wrist-and-fingertip 3D representation may largely overcome the human-to-dexterous-robot embodiment gap without costly robot data, but the abstract gives too few details about task count, comparisons, and evaluation breadth to warrant a strong verdict.
-
Conformal calibration and look-elsewhere effect in anomaly detection for new-physics searches
The paper adds a conformal-calibration layer to machine-learned anomaly detectors, aiming to produce finite-sample-valid local p-values and account for scanning many regions through a Gross–Vitells look-elsewhere correction. On LHC Olympics data and background-only pseudoexperiments, it reports that mass sculpting and standard asymptotic procedures can create apparently huge false excesses, while weighted conformal calibration suppresses these false alarms without retraining the detector.
The combination of distribution-free calibration, sideband-aware weighting, and trials-factor correction addresses a serious but specialized failure mode, with striking claimed false-discovery results; however, the evidence is limited to the abstract and one public physics benchmark, and the underlying conformal components are established techniques.
-
Microchimerism in the human brain, quantitative assessment and single nuclei profiling establish cell types and diversity
The study finds maternal cells in the brains of 70% of epilepsy-surgery patients, often across multiple brain regions and in substantial quantities. Single-nucleus profiling suggests these cells can adopt several neuronal and glial identities, while analysis of public datasets indicates they may persist from gestation into old age, especially as L2/3 neurons or microglia-like cells.
The widespread, long-lived presence of maternal-derived cells across diverse human brain cell types is highly unexpected and potentially important for models of brain development and self, but the evidence is based largely on epilepsy tissue and retrospective public datasets, with functional consequences still unknown.
-
Intact learning and memory in mice incapable of de novo myelination
Using inducible genetic manipulations in adult mice, the authors blocked compact myelin formation or oligodendrocyte metabolic support in newly generated oligodendrocytes, while separately eliminating those cells. Mice could still learn and retain motor skills and contextual fear memories when new oligodendrocytes lacked functional myelin or MCT1, but learning was impaired when the newly formed oligodendrocytes were eliminated, suggesting they support learning through a non-canonical function rather than faster saltatory conduction.
The dissociation between the presence of newly generated oligodendrocytes and their canonical myelination function challenges a standard account of myelin-dependent learning, though the abstract does not yet identify the alternative mechanism or establish how broadly the result generalizes.
-
Rethinking Psychometric Evaluation of LLMs: When and Why Self-Reports Predict Behavior
The paper tests whether LLM self-reports can predict their behavior, comparing broad Big Five personality measures with behavior-specific Theory of Planned Behavior measures across four tasks and 11 frontier models. Behavior-specific self-reports match behavior at roughly human levels within a shared conversation, but this coherence often disappears across conversations—especially when immediate context strongly primes behavior; persona prompts improve report consistency without improving behavioral alignment.
It offers a non-obvious, empirically tested account of when LLM self-reports are meaningful, distinguishing behavior-specific coherence from broad personality consistency across contexts, though the abstract does not quantify the effects or establish broad deployment value.
-
LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies
LaWAM predicts how candidate robot actions will change a scene in a compact vision-feature space, rather than generating full future video, and feeds those predicted visual subgoals into a policy. It reports 98.6% success on LIBERO, 91.22% on RoboTwin, real-world manipulation results, and up to 24× lower wall-clock latency than pixel-space world-action models.
The latent-subgoal formulation addresses a real computational bottleneck in dynamics-aware robot policies and claims a large latency reduction, but the abstract does not establish how broad or robust the gains are beyond standard manipulation benchmarks.
-
Avatar V: Scaling Video-Reference Avatar Video Generation
Avatar V conditions video generation on an entire reference video rather than a single image, aiming to preserve both a person’s appearance and their characteristic timing, gestures, and expressions. It combines long-context reference attention, a motion stream, identity-aware super-resolution, and large-scale multi-stage training, but the abstract gives no quantitative results beyond broad claims of outperforming other systems.
Video-reference conditioning for behavioral identity transfer could be an important advance for avatar generation, but the evidence is weakly specified: there are no actual metric values, dataset details, or independent evaluation, and the production-scale claims are easy to overstate.
-
Unstable Features, Reproducible Subspaces: Understanding Seed Dependence in Sparse Autoencoders
This paper studies why sparse autoencoder features differ across random seeds. Across models, layers, dictionary sizes, and SAE variants, it finds that individually stable features carry most functional signal, while unstable features often reflect different basis choices within reproducible low-dimensional subspaces; a synthetic experiment supports this explanation. Pooling features across seeds produces more stable SAEs without losing explained variance in their experiments.
The subspace-versus-individual-feature framing is a meaningful and somewhat surprising contribution to SAE interpretability, but the abstract provides no quantitative results and the practical impact beyond this setting is not yet established.
-
Decentralized Multi-Agent Systems with Shared Context
The paper proposes DeLM, a decentralized multi-agent framework in which agents asynchronously claim tasks and write compact, verified progress to a shared context instead of relying on a central coordinator. It reports up to 10.5-point gains on SWE-bench Verified and 5.7 points on LongBench-v2, while roughly halving task cost, across multiple frontier model families.
The combination of asynchronous task allocation, shared verified memory, and substantially lower reported cost is a meaningful alternative to centralized LLM orchestration, but the underlying blackboard-style coordination idea is established and the abstract does not provide enough experimental detail to justify a strong verdict.
-
WEAVER, Better, Faster, Longer: An Effective World Model for Robotic Manipulation
WEAVER is a multi-view world model for robotic manipulation that predicts future latent states and rewards using flow matching, with architectural and memory choices aimed at improving visual fidelity, long-horizon consistency, and simulation speed. On real robotic hardware, it reports strong policy-evaluation correlation (0.87), 38% relative improvement for policy improvement, and 14% improvement in test-time planning while running 5–10× faster than prior world models, including better out-of-distribution performance.
The combination of long-horizon multi-view prediction, fast planning, and real-hardware validation addresses a central bottleneck for useful robot world models, but the abstract lacks enough task, baseline, and absolute-success details to justify a strong verdict.
-
AutoMegaKernel: A Statically-Checked Agent Harness for Self-Retargeting Megakernel Synthesis
AutoMegaKernel uses an agent-driven compiler to generate single-launch persistent CUDA megakernels for Llama-family inference, with a statically checked schedule IR that rejects schedules likely to deadlock or race before execution. It retargets one source across three GPU architectures and reports token-exact correctness, automatic support for 10 models, and 1.25–1.72× self-improvement over its baseline. Its W8A16 kernels beat CUDA-graphed bf16 cuBLAS at batch-1 decode on several inference-oriented GPUs, though they lose on A100/H100 and the largest tested model is only 1.1B parameters.
The combination of agentic megakernel synthesis, static schedule rejection, and cross-GPU retargeting is a genuinely interesting systems direction, but the headline speedups rely on an asymmetric int8-versus-bf16 comparison and relatively small-scale validation, limiting confidence in broad impact.
-
Act on What You See: Unlocking Safe Social Navigation in Vision-Language-Action Models
The paper proposes SALSA, a two-stage post-training method for making vision-language-action policies navigate around people more safely. It connects existing social cues in intermediate VLA representations to the action head, uses counterfactual human/object scenes to reduce visual-saliency shortcuts, and adds automatically generated supervision for anticipating future collision risk. On SCAND and in real-world tests, it reports an 86.4% reduction in near-collisions and an increase in social counterfactual accuracy from 53% to 93%.
The combination of latent-feature-to-action alignment and automatically supervised anticipation addresses a meaningful failure mode in VLA robot navigation, with unusually large reported gains and real-world evaluation, but the abstract does not establish how broadly the method generalizes or how strong the baselines are.
-
Planar-Sector LOS Guidance for Interception of Agile Targets with Lifting-Wing Quadcopters
The paper replaces the usual symmetric image-centered line-of-sight constraint with a planar-sector constraint: it tightly limits lateral tracking error but allows more longitudinal image error, giving a lifting-wing quadcopter more thrust for pursuit. Combined with delay-compensated estimation and nonlinear control, the system intercepts agile, unpredictably moving aerial targets using only a strapdown monocular camera, including outdoor trials at ranges up to 138 m with continuous tracking.
The asymmetric visibility constraint is a genuinely useful guidance idea backed by substantial real-world interception experiments and a reported near-50% thrust increase, though it remains a specialized system-level advance rather than a broadly transformative result.
-
Efficient AI-Inspired Reduction of Feynman Integrals via Tube Seeding
The paper uses machine learning to discover a sparse seed-selection strategy for integration-by-parts reduction of Feynman integrals. Restricting seeds to a thin, zigzagging “tube” makes their number grow linearly with numerator power, and enables reductions of difficult non-planar two-loop five-point integrals that conventional Laporta seeding reportedly cannot handle, with lower time and memory use.
The tube-seeding idea appears to deliver a substantial scaling and memory improvement on a genuinely hard symbolic-computation problem, though its impact is mainly in specialized theoretical-physics workflows rather than general AI/ML.
-
Efficiently Learning Drifting Halfspaces with Massart Noise
This paper studies efficient online learning of margin-separable halfspaces when the target drifts over time and labels have Massart noise. It gives an algorithm with error scaling as η + Õ(Δ^{1/3}/γ), improves prior results in the noiseless case, and proves that Δ^{1/3} dependence is unavoidable for low-degree polynomial methods even though the information-theoretic optimum is Δ^{1/2}.
The combination of an efficient drifting-halfspace learner with evidence of a computational-statistical tradeoff is a substantive theoretical result, but its practical relevance and the strength of the lower bound depend on technical details absent from the abstract.
-
$μ$VLA: On Recurrent Memory for Partially Observable Manipulation in VLA Models
The paper isolates the effect of recurrence in vision-language-action models by adding only a small set of persistent memory tokens, with no auxiliary losses or external memory machinery. On MIKASA-Robo, this raises success from 0.42 to 0.84 on training tasks and from 0.07 to 0.23 on held-out tasks, while preserving performance on fully observable LIBERO; however, it does not help when the task requires a different memory structure.
The controlled finding that minimal in-backbone recurrence can substantially improve partially observable manipulation is useful and somewhat surprising, but the gains appear benchmark- and memory-structure-dependent rather than a broadly demonstrated capability jump.
-
Natural-Language Temporal Grounding in Hour-Long Videos is a Search Problem: A Benchmark and Empirical Decomposition
The paper studies temporal grounding in unusually long videos—up to nine hours—using 2,273 queries across 194 videos. It finds that current Video-LLMs mainly fail because they cannot search the full video effectively, while frame retrieval followed by local grounding improves performance by 6.7× over a monolithic Video-LLM. The benchmark and decomposition suggest that long-video grounding may be better treated as retrieval plus grounding rather than a single end-to-end recognition task.
The hour-scale setting and strong evidence that search—not local recognition—is the dominant bottleneck are genuinely useful, but this is primarily a benchmark and empirical decomposition rather than a demonstrated major new model capability.
-
Ambient Diffusion Policy: Imitation Learning from Suboptimal Data in Robotics
Ambient Diffusion Policy trains diffusion-based robot policies on low-quality demonstrations while limiting when those demonstrations influence learning: mainly at very early and very late diffusion times. The authors motivate this with a spectral power-law structure in robot actions and a global-to-local hierarchy, then evaluate across noisy data, sim-to-real shifts, task mismatch, and heterogeneous Open X-Embodiment mixtures. It reportedly beats existing co-training methods by up to 33% on six tasks, though the abstract does not specify the metric or absolute performance.
Selective diffusion-time use of suboptimal demonstrations is a distinctive, theoretically motivated approach to a central robotics bottleneck, with unusually broad shift and data-mixture experiments, but the headline improvement is underspecified and the theory is based on a simplified model.
-
Modality Forcing for Scalable Spatial Generation
The paper introduces Modality Forcing, a way to train one diffusion transformer to generate or predict images and depth in arbitrary conditioning orders by assigning separate noise levels to each modality. It uses sparse real-world depth data and shows that scaling the underlying text-to-image model and image pretraining improves depth estimation, with the largest model reportedly competitive with specialized monocular depth estimators and substantially better than prior joint generative approaches.
The notable result is evidence that large-scale image-generation pretraining can transfer into spatial perception using sparse depth supervision, but the abstract does not establish a broad enough capability or benchmark lead for a strong verdict.
-
Nonlinear Two-Time-Scale Stochastic Approximation: A Sharp Phase Transition and How to Beat It
This paper identifies a regularity-dependent phase transition in nonlinear two-time-scale stochastic approximation: the usual slow iterate reaches the ideal $k^{-1}$ mean-square rate only when the nonlinear remainder satisfies $a(1+\rho)\ge 1$, with a matching lower bound showing the slower rate is unavoidable for the naive update. It then introduces an online estimator of the nonlinear bias and subtracts it, proving that the corrected recursion achieves $O(k^{-1})$ across all regularity regimes, with extensions to general nonlinear TTSA via fast-manifold coordinates.
The sharp lower-bound phase transition plus a bias-correction mechanism that removes it is a genuinely interesting theoretical advance for stochastic approximation, but its practical impact and breadth beyond the stated normal forms are not yet demonstrated.
-
Representation Costs in Data Science: Foundations and the Quasi-Banach Spaces of Deep Neural Networks
The paper develops a general way to translate parameter regularization into a function-space representation cost, recovering familiar cases such as RKHSs, Besov spaces, and variation spaces. Its main result is that weight decay in depth- L ReLU networks induces a depth-dependent quasi-seminorm; for depth greater than two, the corresponding native function space has a nonconvex quasi-Banach geometry.
This offers a genuinely new theoretical lens on what weight decay means in deep ReLU networks and links depth to nonconvex function-space geometry, but the abstract gives no direct evidence yet that the framework changes practical training or explains major empirical phenomena.
-
Not All Skills Help: Measuring and Repairing Agent Knowledge
The paper treats accumulated natural-language skills as hypotheses that should be tested empirically, rather than trusting an LLM to decide which skills are useful. ASSAY estimates each skill’s causal effect by masking it on a development set, then selects or suppresses skills differently for each task; across seven models on AppWorld and tau-bench, it reports substantial gains without weight updates, including 69.3% task completion for DeepSeek-V3 on AppWorld’s hardest split and an 8.7% relative gain for GPT-4.1 on tau-bench retail.
Per-task causal skill selection is a meaningful and non-obvious alternative to global skill curation, with broad model coverage and large reported gains, but the evidence is limited to two agent benchmarks and the claimed gains may depend on potentially costly development-set masking.
-
Free Energy Heuristics: Fast-And-Frugal Cognition as Active Inference Under Uncertain Precision
This paper proposes that chain-of-thought hurts when a model is highly uncertain about whether its own evidence is reliable: additional reasoning then accumulates false confidence rather than useful signal. It formalizes this with uncertain-precision free-energy models, showing conditions under which optimal inference stops after a few reliable cues and becomes equivalent to a take-the-best heuristic, then reports a preregistered study across seven LLMs where accuracy fell 17.3 points on specially constructed high-meta-uncertainty items but not on matched items with definite answers.
The combination of a principled account of when CoT fails, a formal link between active inference and fast-and-frugal heuristics, and a sizable preregistered cross-model effect is genuinely interesting, but the empirical benchmark and strong theoretical conclusions need close scrutiny before treating it as a broadly established result.
-
Intensity-dependent topographical expansion of sensory representations
Across four fMRI datasets, the authors find that stronger sensory stimulation expands the spatial extent of cortical and cerebellar activation rather than merely increasing response magnitude in fixed regions. This effect replicated for painful, tactile, auditory, and visual stimuli, and the amount of expansion tracked perceptual discriminability, involving both larger active regions and recruitment of new ones.
The replicated, cross-modal evidence that stimulus intensity is encoded partly through expanding neural topographies challenges a common fixed-map assumption, though the finding is descriptive and based on the spatial resolution limits of fMRI rather than a demonstrated computational mechanism.
-
Mutant SOD1 expressed by oligodendrocytes aggregates in myelinic nanochannels and accelerates disease progression in familial ALS mice
In two familial ALS mouse models, mutant SOD1 in oligodendrocytes was found to aggregate in narrow cytosolic channels within compact myelin, including paranodal loops and the inner periaxonal tongue. Removing mutant SOD1 before myelin compaction slowed disease, whereas later silencing did not; experimentally collapsing these channels accelerated disease, supporting a causal role for disrupted oligodendrocyte-to-axon support.
The paper offers a specific, anatomically localized mechanism linking oligodendrocyte dysfunction to ALS progression and supports causality with complementary genetic interventions, but the evidence is confined to mouse SOD1 models.
-
Isolation of postnatal human neural stem cells
The authors prospectively isolate two distinct populations of neural stem cells from postnatal human brains using surface markers and index sorting, then characterize their lineage behavior with clonal barcoding and xenotransplantation. One population is biased toward interneuron and oligodendrocyte production, while the other favors astrocytes; both decline sharply during the first two decades but remain detectable through age 90. The main advance is a functional, experimentally accessible framework for studying human postnatal neural stem cells rather than inferring them from mature neurons alone.
Prospective isolation and functional subdivision of human postnatal neural stem cells—including persistence into very old age—would materially advance a contentious area, though the abstract lacks enough quantitative and replication detail for a strong verdict.
-
Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning
NVIDIA introduces a 550B-parameter MoE language model with 55B active parameters, combining Mamba and attention, trained on 20T tokens and extended to 1M-token context. It combines several efficiency and post-training techniques and is claimed to match leading public models with up to 6× higher inference throughput, while releasing checkpoints, data, and training recipes.
The open, very large hybrid Mamba–Transformer system and claimed throughput/context advances are potentially important, but the abstract gives no benchmark details or conditions supporting the 6× claim, so it does not merit a strong verdict yet.
-
EvoArena: Tracking Memory Evolution for Robust LLM Agents in Dynamic Environments
EvoArena evaluates LLM agents in environments that change through sequences of updates across terminal, software, and social-preference tasks. The authors also introduce EvoMem, which stores memory as structured patches and histories; it improves performance modestly on the new benchmark and more substantially on GAIA and LoCoMo, while analysis suggests better retention of changing state.
The focus on evaluating and remembering environment evolution is a useful, nonstandard framing for agent reliability, but the reported 1.5% gain on the main benchmark is modest and the abstract does not establish a major capability or generality jump.
-
Evoflux: Inference-Time Evolution of Executable Tool Workflows for Compact Agents
Evoflux uses inference-time evolutionary search to generate and repair typed tool-use workflow graphs, using execution failures, schema checks, and dependency errors as feedback. On MCP-Bench with 250 tools, it increases workflow execution feasibility for small language-model planners from about 3% to 17–24%, while supervised distillation is unreliable and ReAct costs more tokens.
The execution-grounded evolutionary repair framing is a credible and somewhat non-obvious alternative to distilling tool traces, but the reported feasibility remains low and the evidence is limited to one benchmark setting.
-
Getting Better at Working With You: Compiling User Corrections into Runtime Enforcement for Coding Agents
TRACE turns corrections a user gives during one coding-agent interaction into explicit rules and runtime checks that are enforced in later tasks, rather than merely storing them as memory. In simulated user-in-the-loop evaluations, it substantially lowers preference violations on two task suites, though the strongest results are benchmark-specific and the experiments are not reported with real users or deployed agents.
The compelling idea is compiling user corrections into executable completion checks—a materially different approach from passive memory—but the evidence is limited to simulated evaluations and the striking baseline/results setup needs closer scrutiny.
-
Demystifying Hidden-State Recurrence: Switchable Latent Reasoning with On-Policy Reinforcement Learning
The paper introduces SWITCH, a latent-reasoning method that uses explicit entry and exit tokens to delimit recurrent hidden-state computation. These discrete boundaries make on-policy GRPO training easier and provide handles for probing and causal interventions; the authors report better performance than prior hidden-state recurrence methods and find that computation is concentrated at the transition into latent mode.
The unified use of switch tokens for both RL compatibility and mechanistic analysis is a meaningful design idea, but the abstract gives no quantitative results, scale details, or comparisons strong enough to justify a higher verdict.
-
AlloSpatial: Agentic Harness Framework for Spatial Reasoning in Foundation Models
AlloSpatial gives multimodal models an external allocentric world model: it converts egocentric observations into structured spatial trees and route maps, then uses tools and consistency checks to answer spatial queries. On VSI-Bench and MindCube, the framework reportedly improves models by 5–18% without training, while a Qwen3-VL version trained with reinforcement learning outperforms larger general-purpose models; notably, the spatial trees can support reasoning after the visual input is removed.
The explicit conversion from egocentric vision to queryable allocentric representations, especially the claim that the representation preserves useful reasoning without the original images, is a non-obvious idea, but the abstract provides limited detail and benchmark evidence for a framework that may partly combine familiar mapping, tool-use, and RL techniques.
-
Optical Reasoning: Rethinking Images as an Expressive Reasoning Medium Beyond Text
The paper proposes using images—not text—as the intermediate reasoning representation for language and multimodal models. It implements this with compact typographic layouts and structured graphics, reporting comparable or better performance on math, science, and multimodal benchmarks while reducing the number of reasoning tokens by 16–28.57%.
The genuinely unusual idea is treating visual compositions as standalone chain-of-thought, but the abstract gives limited detail about model design, token accounting, and breadth of validation, so the claimed efficiency and capability gains need scrutiny.
-
Quickest Detection of Hallucination Onset: Delay Bounds and Learned CUSUM Statistics
This paper reframes hallucination monitoring as a sequential change-detection problem, measuring how many tokens elapse after hallucination onset rather than relying on token-level AUC. On RAGTruth, a causal recurrent detector catches some onsets in 11–13 tokens versus 31 for a linear baseline, but catches fewer than a third at the stated false-alarm rate; the authors also derive delay bounds and show that the main limitation is the quality of per-token evidence, not temporal accumulation.
The sequential-analysis framing, recall-honest delay metric, and information-rate diagnosis are genuinely useful, but the practical results remain limited by low recall and a narrow RAGTruth-based evaluation, so this is not yet a major capability advance.
-
Echo-Memory: A Controlled Study of Memory in Action World Models
Echo-Memory compares several memory mechanisms for action-conditioned video world models while holding the backbone, training setup, action interface, and evaluation fixed. It finds that replay fidelity does not reliably measure whether a model remembers objects after leaving and revisiting a scene: raw context helps open-domain returns, while aggressive compression loses salient details, and block-wise state-space recurrence performs best in that setting.
The controlled comparison and finding that replay quality diverges from genuine revisit memory are useful and somewhat surprising, but the abstract gives no quantitative effect sizes or evidence of a major capability advance.
-
Order Matters: Unveiling the Hidden Impact of Macro Placement Sequences via Proxy-Guided LLM Evolution
The paper treats macro-placement order as a learned optimization variable rather than a fixed heuristic. Its OrderPlace system uses an LLM to evolve code-level ordering policies and a cheap greedy proxy to screen them, reporting 34.04% lower wirelength than WireMask-EA and 14.08% lower than EGPlace on ISPD 2005 benchmarks.
The combination of sequence-sensitive placement, proxy-guided search, and LLM-generated ordering policies is a non-obvious direction with large reported gains, but the evidence is limited to one benchmark family and the abstract does not establish generalization or fair evaluation against stronger baselines.
-
INFUSER: Influence-Guided Self-Evolution Improves Reasoning
INFUSER co-trains a question Generator and reasoning Solver from automatically collected documents. Instead of rewarding the Generator merely for producing difficult problems, it estimates each problem’s influence on future Solver performance and uses that signal with a dual-normalized GRPO variant; on Qwen3-8B-Base, it reports over 20% relative gains on Olympiad and SuperGPQA, and an 8B co-evolving Generator outperforming a frozen 32B reasoning model on math and coding. The central contribution is an optimizer-aware, adaptive curriculum for self-evolution rather than another fixed synthetic-data pipeline.
The influence-based feedback loop and co-evolutionary curriculum are a meaningful departure from difficulty-driven self-training, with promising cross-task and model-size results, but the abstract lacks absolute scores and enough detail to establish how robust the reported gains are.
-
REFINE: Super-efficient 3D Gaussian Splatting Pruning via Rendering-Free Primitive Importance
REFINE prunes 3D Gaussian Splatting scenes using an analytically approximated, rendering-aware Hessian-based importance score instead of repeatedly rendering the scene to estimate which primitives can be removed. The authors report comparable rendering quality with 3,000× lower pruning computation and about 20× lower device latency than prior pruning methods, though the abstract gives no quantitative quality or scale details.
The rendering-free, geometry- and visibility-aware importance formulation and very large claimed speedup are genuinely interesting, but the abstract lacks enough numerical quality tradeoffs and broad experimental detail to justify a strong recommendation.
-
MotionWAM: Towards Foundation World Action Models for Real-Time Humanoid Loco-Manipulation
MotionWAM adapts video-based World Action Models for real-time humanoid loco-manipulation from a single egocentric camera. It uses intermediate video-model features and a unified whole-body motion representation, rather than separate upper-body manipulation and lower-body locomotion policies. On nine real Unitree G1 tasks, it reportedly outperforms similarly fine-tuned VLA baselines by over 30% in overall success and enables task-driven foot interactions.
The unified, real-time whole-body action space and demonstrated foot-mediated manipulation are meaningful advances, but the abstract gives limited detail about the tasks, compute/latency, and the size and nature of the reported 30% improvement.
-
Training-Free Generalized Few-Shot Segmentation through Open-Vocabulary Semantic Arbitration
Open-V performs generalized few-shot semantic segmentation without fine-tuning by combining frozen SAM3 segmentation proposals with a CLIP centroid built from the support examples, using per-pixel semantic arbitration. It reports strong results across PASCAL-5i, COCO-20i, and ADE-OW, including a 17.7-point harmonic mIoU improvement over the strongest trained baseline on 1-shot PASCAL-5i, and highlights preprocessing/evaluation mismatches as a reproducibility issue.
The inference-only coordination of foundation-model priors and the large reported GFSS gain are genuinely interesting, but the abstract provides limited experimental detail and the approach may depend heavily on the particular SAM3/CLIP implementation and evaluation corrections.
-
Self-Harness: Harnesses That Improve Themselves
Self-Harness lets an LLM agent iteratively modify its own tool-use and interaction harness. It mines failures from execution traces, proposes small harness changes, and keeps only edits that pass regression tests; across three models and three agent benchmarks, the authors report improvements on both held-in and held-out tasks, with relative gains up to 132%.
The model-specific, self-improving harness loop is a genuinely interesting direction with broad evaluation, but the abstract gives no absolute pass rates, comparison details, or safeguards against benchmark-specific overfitting, so the large gains are not yet strong enough for a strong verdict.
-
Next-Token Prediction Learns Generalisable Representations of Sleep Physiology
Hypnos is a multimodal sleep foundation model trained autoregressively to predict discretized next tokens across eight physiological signal types, including EEG, ECG, and respiration. Trained on more than 20,000 overnight recordings, it reportedly matches supervised sleep-stage classifiers with 100 times less labeled data and transfers to daytime atrial-fibrillation detection, outperforming a dedicated ECG foundation model.
The combination of joint next-token prediction across multimodal physiological streams and strong low-label, cross-domain transfer is a genuinely interesting alternative to masked reconstruction and contrastive learning, but the abstract gives too few quantitative details to justify a strong verdict.
-
Integral Field Unit Spectroscopy with One Fiber
The paper trains a probabilistic model to infer spatially resolved galaxy spectra from broadband images and the location of a virtual fiber. It uses millions of images paired with single-fiber DESI spectra, exploiting varied fiber placements and repeated galaxy morphology rather than any direct IFU training data; its predicted emission-line maps reportedly agree with independent MaNGA observations and approach an IFU-supervised baseline.
The potentially important idea is recovering IFU-like spatial spectroscopy from cheap single-fiber observations and images, but the abstract gives no quantitative comparisons, uncertainty calibration results, or detail sufficient to establish how reliable the reconstruction is.
-
Harnessing the Collective Intelligence of AI Agents in the Wild for New Discoveries
EinsteinArena is an open platform where AI agents tackle mathematical problems, submit verifiable solutions, discuss partial results, and build on one another’s work. The authors report 12 new best-known results, including improving the dimension-11 kissing-number lower bound from 593 to 604, with progress emerging across multiple agents rather than from one isolated run.
The reported multi-agent discovery process and substantial kissing-number improvement are genuinely notable, but the abstract provides too little detail about the agents, verification, attribution, and breadth of the 12 results to justify a strong verdict.
-
A prior-free blind detection of information leakage from model predictions
This paper studies whether data leakage can be detected using only a model’s predictions and observed outcomes, without training code, external data, or domain expertise. It proves that leakage can be fundamentally indistinguishable from an honestly better, recalibrated model, but shows that near-deterministic leaked subgroups create a detectable high-purity region; experiments on UK Biobank illustrate a detection threshold around Δc ≈ 0.007 for one endpoint and a sub-second test.
The impossibility result and matching of leakage types to detectable or undetectable regimes are genuinely useful, but the empirical support is limited to one cohort and leakage setting, so this is not yet a strong recommendation.
-
UniDexTok: A Unified Dexterous Hand Tokenizer from Real Data
The paper introduces a shared 22-DoF semantic representation for human and robot dexterous hands, then learns embodiment-conditioned discrete tokens directly from real joint-state data without simulation or explicit retargeting. It reports dramatic reconstruction improvements over UniHM, reducing errors from centimeter-scale to sub-millimeter scale, and finds that data from other hand embodiments improves reconstruction and enables zero-/few-shot adaptation to new hands.
A unified, retargeting-free representation that transfers across heterogeneous dexterous hands could materially improve multi-embodiment manipulation learning, but the exceptionally large gains are difficult to assess from the abstract without knowing the evaluation setup, baselines, and whether the shared representation preserves task-relevant kinematics.
-
Bilinear gating of motor primitives: a principle linking dendritic computation to rapid goal-directed adaptation
The paper reports that in macaque motor cortex, the fraction of spikes emitted in bursts carries reach-direction information more selectively than overall firing rate, consistently across 12 sessions. It proposes that dendritic coincidence detection implements a multiplicative interaction between goal-related and movement-state inputs, and shows in a spiking model and reinforcement-learning agent that this mechanism supports zero-shot goal generalization and rapid adaptation.
The combination of a burst-specific motor code, a concrete dendritic bilinear-gating mechanism, and an adaptation benefit is genuinely interesting, but the abstract gives limited quantitative evidence and the computational advantage is demonstrated mainly in a minimal model.
-
WorldKernel: A World Model is the Coupling Kernel of Admissible Possible Worlds
The paper argues that ordinary prediction can recover marginal or identified causal quantities but cannot determine how different counterfactual worlds are coupled, even with unlimited data. It represents these couplings with a positive-semidefinite kernel over admissible worlds, showing that the kernel can bound unidentified counterfactuals, benefit from logical constraints and targeted interventions, and is computationally hard to reconstruct exactly.
The off-diagonal coupling-kernel formulation is a genuinely interesting reframing of counterfactual uncertainty, with potentially useful computational consequences, but the abstract provides limited detail on the experiments and the practical scope of the claimed bounds.
-
Bellman-sufficient Information Complexity
The paper proposes a general information-theoretic framework for sequential decision problems in which a Bellman-sufficient state and a decision-relevant information index yield matching upper and lower risk bounds. Applying it to Gaussian-process bandits, it constructs a bounded continuous kernel where minimax regret is Θ(T^{1−α}) along infinitely many horizons, while standard anytime maximal-information GP-UCB and the original RKHS exploration schedule suffer linear regret; a different Bellman policy achieves the minimax order. The result argues that uniform optimism can waste information through cloud-like exploration, making localization within the Bellman recursion essential.
The framework and the claimed counterexample to canonical GP-UCB optimality are genuinely non-obvious, but the contribution is highly theoretical and the abstract provides limited detail about how broad or practically consequential the separation is beyond the constructed kernel.
-
SwiftCTS: Fast Cross-Design Prediction and Pareto Optimization of Clock Tree Metrics via Few-Shot Calibration
SwiftCTS uses physics-inspired features and gradient-boosted surrogate models to predict clock-tree power, wirelength, and timing, then searches configurations with an evolutionary optimizer instead of repeatedly invoking the EDA flow. Its main contribution is few-shot multiplicative calibration: one or two physical runs reportedly reduce large out-of-distribution prediction errors to roughly 1–3%, enabling 100,000 configuration evaluations in under ten seconds and producing physically validated Pareto frontiers.
The combination of cross-design few-shot calibration and very fast surrogate-driven CTS optimization is a genuinely useful direction with striking reported gains, but the abstract gives limited information about the number and diversity of designs, baselines, and independent validation needed to trust the broad OOD claims.
-
Overcoming State Inertia in Full-Duplex Spoken Language Models via Activation Steering
The paper identifies “state inertia” in full-duplex spoken language models: after the model is speaking, its internal representations remain biased toward generation briefly after a user interruption, causing it to miss the beginning of the user’s speech. It introduces a zero-buffer interruption benchmark and uses a training-free activation-steering vector to shift the model toward perception, improving PersonaPlex interruption correctness from 28% to 45% and initial-word capture from 40% to 72%.
The state-inertia diagnosis and zero-buffer framing target a real limitation of interactive speech models, and the sizable cross-model, training-free gains make this worth a closer look, though the abstract does not establish broad robustness beyond the reported evaluations.
-
Mirror Descent Beyond Euclidean Stability: An Exponential Separation in Initialization Sensitivity
The paper shows that, unlike gradient descent with quadratic regularization, mirror descent with a smooth, strongly convex, well-conditioned but nonquadratic regularizer can amplify tiny initialization differences exponentially quickly, even on convex smooth objectives. It constructs this separation in three dimensions, shows analogous instability for KL-regularized updates on the simplex, and proposes anchoring the Bregman regularizer at a fixed point as a stabilizing modification.
This is a potentially important, assumption-challenging stability result for mirror descent and KL-regularized policy optimization, but the evidence in the abstract is primarily theoretical constructions rather than broad validation of practical impact.
-
DAM-VLA: Decoupled Asynchronous Multimodal Vision Language Action model
DAM-VLA separates the update rates of vision, language, and action-related modalities instead of forcing them onto one synchronous clock. It uses modality-specific latent buffers and gated cross-attention, and reports 95.2% average success across seven contact-rich real-world manipulation tasks versus 40.95% for a synchronous baseline, with 100 Hz control.
The per-modality asynchronous design and very large reported real-world gain are genuinely interesting, but the abstract does not establish how strong the baseline is or provide enough ablation and generalization evidence for a strong recommendation.
-
CHORUS: Decentralized Multi-Embodiment Collaboration with One VLA Policy
CHORUS adapts one pretrained vision-language-action model to control different robots independently, using only each robot’s local camera observations plus a robot-specific prompt. In real-world tasks such as tape measurement, book handovers, and lifting a laundry basket, it reports substantially better decentralized coordination than from-scratch and centralized baselines, without inference-time communication or explicit alignment.
A shared VLA enabling reactive multi-robot collaboration from local observations alone is a meaningful and somewhat surprising direction, with real-robot demonstrations, but the abstract does not provide enough experimental detail to establish how broad or robust the gains are.
-
FACTR 2: Learning External Force Sensing for Commodity Robot Arms Improves Policy Learning
The paper introduces NEXT, which estimates external joint torques on inexpensive robot arms using only a minute of training from ten minutes of free-motion data, avoiding dedicated force sensors. It uses these estimates both for force-feedback teleoperation and for FIRST, a behavior-cloning strategy that emphasizes pre-contact and contact segments; across five long-horizon tasks, FIRST reportedly improves task progress by over 17%.
It could make contact-aware manipulation substantially cheaper and more widely deployable, but the abstract provides limited detail on generalization, hardware diversity, and how the torque estimates compare quantitatively with real sensors.
-
FTP-1: A Generalist Foundation Tactile Policy Across Tactile Sensors for Contact-Rich Manipulation
FTP-1 trains one tactile manipulation policy across image-, array-, and state-based sensors by encoding them into shared morphology-aware tokens. Using roughly 3,000 hours from 26 sources and 21 sensors, it improves fine-tuned performance on known hardware and reports substantial transfer gains on two tactile setups not seen during pretraining.
A cross-sensor, cross-embodiment tactile foundation policy is a meaningful new direction, and the reported gains on unseen sensors are notably non-obvious, but the abstract lacks absolute success rates and broader evidence needed for a strong verdict.
-
Proprioceptive-visual correspondence enables self-other distinction in humanoid robots
The paper shows a humanoid robot learning to identify its own body among humans or identical robots using only synchronized proprioceptive and visual signals, without identity labels or an explicit kinematic model. This self/other distinction is then used to learn a 3D predictive body-occupancy model and support reaching, collision-aware planning, and motion retargeting.
The label-free route from sensor correspondence to self-identification and a usable body model is a genuinely interesting embodied-robotics direction, but the abstract gives no quantitative scale, baselines, or evidence that the downstream capabilities substantially exceed existing self-modeling methods.
-
Scalable Dynamic Tactile Sensing Enabled by Passive and Flexible Acoustic Waveguides
The paper presents a flexible tactile skin that routes mechanical signals through a passive network of miniature Helmholtz resonators and reinforced acoustic waveguides, allowing sparse microphones to localize touch over many sensing nodes. A 4-microphone, 64-node array reaches up to 4 mm spatial resolution and over 99% localization accuracy, with waveform reconstruction in under 5.5 ms, and demonstrations include gloves, fingertip arrays, pulse sensing, and large-area skins.
The passive acoustic-waveguide architecture is a distinctive approach to avoiding wiring and flexibility trade-offs in scalable tactile skins, with promising array-scale results, but the abstract does not establish durability, comparisons to leading tactile technologies, or performance under realistic robotic manipulation.
-
Learning with Simulators: No Regret in a Computationally Bounded World
The paper studies learning from arbitrarily dependent data when the learner has access to a simulator approximating the data-generating process. It claims that simulator access restores VC-dimension-style generalization guarantees, gives conditional sampling statistical and computational benefits, and yields one algorithm that learns every VC class over all polynomial-time-samplable processes with regret tied to the process’s time-bounded Kolmogorov complexity. The main contribution is a computationally grounded extension of the classical PAC framework beyond independent examples.
The simulator-based formulation and universal algorithm for dependent, computationally bounded processes are conceptually substantial and somewhat surprising, but the abstract provides no concrete bounds or experimental/theoretical details sufficient for a strong recommendation.
-
Beyond the Commitment Boundary: Probing Epiphenomenal Chain-of-Thought in Large Reasoning Models
The paper studies when a reasoning model’s chain of thought actually determines its final answer, using early-exit interventions to estimate the causal importance of each step. Across several model families and tasks, it reports a sharp “commitment boundary”: the answer often stabilizes well before the trace ends, with later steps having little effect; detecting this boundary enables up to 55% shorter reasoning traces with negligible performance loss.
The combination of a potentially useful causal account of when answers form and a large inference-efficiency gain is notable, but the abstract gives too few details about models, tasks, baselines, and robustness to justify a strong recommendation.
-
HyperTool: Beyond Step-Wise Tool Calls for Tool-Augmented Agents
HyperTool lets an LLM package multiple tool calls, intermediate data transformations, and value passing into one executable code block, rather than exposing every atomic call in the main reasoning trace. Training on verified compositional tool-use trajectories yields large gains on MCP-Universe accuracy for Qwen3-8B and 32B, roughly tripling performance, though the evaluation appears limited to one benchmark and the underlying idea resembles existing programmatic tool-use approaches.
The execution-granularity framing and large reported gains make this worth tracking, but the concept is adjacent to established code-mediated tool use and the abstract provides only single-benchmark evidence.
-
Temporal Backtracking Search for Test-time Generative Video Reasoning
The paper treats video generation as a search problem over time rather than repeatedly sampling complete videos. Its Temporal Backtracking Search verifies partial trajectories, identifies where they fail, and restarts from valid prefixes; across algorithmic, navigation, and robotics tasks it reportedly beats equal-budget Best-of-N, improving an OOD success rate from 0.7% to 22.7%.
Temporal prefix backtracking is a genuinely non-obvious test-time scaling direction, and the large OOD improvement is promising, but the abstract provides limited detail about task scale, baselines, and verification reliability, so it falls short of a strong recommendation.
-
BayLing-Duplex: Native Full-Duplex Speech Dialogue with a Single Autoregressive LLM
BayLing-Duplex trains a single autoregressive speech LLM to decide when to listen, speak, and stop, using special tokens rather than a separate VAD or turn-taking controller. Fine-tuning GLM-4-Voice on 400K full-duplex examples plus DPO reportedly gives 92% turn-taking and 100% interruption success, while improving speech-response quality over Moshi without hurting task performance.
Native full-duplex control inside one standard autoregressive model is a meaningful and somewhat assumption-challenging direction, but the abstract lacks latency, robustness, and broader comparative details needed to establish a major practical breakthrough.
-
Towards Direct Latent-Space Synthesis for Parallel Branches in LLM-Agent Workflows
The paper proposes letting a synthesizer consume and combine the KV caches from parallel worker agents directly, rather than serializing their outputs into one long text prompt. A cache-mapping layer and fine-tuned adapter make this interface workable; across nine tasks it matches or beats text-based synthesis on seven while reducing time-to-first-token by 2.5–11×.
Direct synthesis over parallel KV caches is a genuinely interesting systems/interface idea with potentially large latency benefits, but the abstract gives limited detail about quality tradeoffs, hardware/setup, and how broadly the approach transfers beyond the reported datasets.
-
MemoryVAM: Integrating Memory into Video Action Model for Robot Manipulation
MemoryVAM adds episodic memory to video-world-model robot policies so actions can depend on events outside the short observation window. Its Recap-Cue module compresses visual history into memory tokens and uses language-conditioned cues for episode progress, improving LIBERO-Mem success from 5% to 42.5% and reaching roughly 75–80% success on several real-robot long-horizon tasks.
The combination of learned episodic recall and progress cues addresses a real limitation of short-context video policies, with large reported gains, but the evidence appears confined to a specialized benchmark and a few task categories, so it is not yet a strong verdict.
-
Exploring Starts Are Not Enough: Counterexamples and a Fix for Monte Carlo Exploring Starts
This paper shows that tabular Monte Carlo Exploring Starts can converge to suboptimal policies, for both initial-visit and first-visit variants, despite the usual intuition that randomizing starting states should ensure sufficient exploration. It also gives a learning-rate modification—scaling updates by state-action update frequencies—that restores convergence for initial-visit MCES and is intended to remain usable with function approximation.
It addresses a fundamental open convergence question with explicit counterexamples and a potentially scalable fix, but the abstract provides too few details to justify the rare strong verdict.
-
Accelerating GPU Inference of Large Language Models with Moderately Unstructured Sparse Weight Matrices
This paper introduces a storage format and GPU SpMM kernel for LLM weight matrices with roughly 50% unstructured sparsity, combining sparse tensor cores, CUDA cores, compressed storage, and a residual path. The authors report up to 1.64× kernel speedup over SpInfer and 1.41× end-to-end speedup over FlashLLM, claiming to be the first to beat dense multiplication on modern HBM-equipped GPUs at this sparsity level.
It addresses a real bottleneck with a fairly distinctive hardware-software design and reports meaningful speedups where moderate unstructured sparsity is usually slower than dense computation, but the abstract lacks breadth on models, GPUs, sparsity-quality tradeoffs, and end-to-end workloads needed for a stronger verdict.
-
FlashNav: Ultra-Fast Policy Training for Robot Navigation within 20 Seconds
FlashNav trains range-based reinforcement-learning navigation policies in under 20 seconds by replacing detailed simulation with a GPU-resident bitmap simulator that retains only the components needed for velocity-level navigation. The authors report 100% success in their tested setups and transfer to TurtleBot2 and Unitree Go2 robots in static and dynamic indoor environments.
The claimed seconds-level, deployable policy training and sim-to-real transfer would be a substantial practical improvement, but the abstract gives limited evidence about task breadth, baselines, and how robust the 100% success result is.
-
LoComposition: Terrain-Adaptive Energy-Efficient Quadruped Locomotion without Gait Priors
The paper trains quadruped locomotion policies by separating task rewards, safety constraints, energy minimization, and terrain perception, rather than encoding a preferred gait through hand-designed reward terms. On simulated terrain, this reportedly maintains traversal performance while cutting cost of transport by 56% and operational-limit violations by 96% versus a complex-reward baseline, with zero-shot transfer to a Unitree Go2 using LiDAR elevation maps.
The combination of gait-prior-free learning, explicit constraints, and terrain-adaptive energy use is a meaningful robotics direction with strong reported efficiency and hardware-transfer results, but the abstract gives limited detail on terrain breadth, baselines, and experimental scale.
-
Learning the generating functional for variance reduction in lattice QCD
The paper trains normalizing flows to represent the generating functional of a lattice gauge theory, allowing many correlation functions to be estimated with substantially lower variance rather than designing a separate estimator for each observable. In demonstrations on glueball correlators and Wilson loops in QCD and Yang–Mills theory, it reports variance reductions of up to three orders of magnitude and a systematic approach toward nearly noiseless estimates.
The combination of learned generative models with the generating-functional formalism could provide a broadly reusable route to major Monte Carlo efficiency gains, but the abstract gives limited detail on the range of lattice sizes, computational cost, and how consistently the striking three-order-of-magnitude result holds.
-
A modular neural circuit for computing the motion of objects
Using high-density recordings in macaque visual area MT, the authors identify distinct neuron types for component motion and whole-pattern motion. They report that pattern-motion neurons are organized in anatomically separate modules and integrate inputs from multiple component-motion neurons, revealing a hierarchical circuit for recovering an object’s motion direction.
This provides a concrete circuit-level account of a long-standing visual-computation problem, but the abstract does not yet establish how comprehensive or causally validated the proposed architecture is.
-
Robust learning-driven structural and functional plasticity of spines in the mature mouse cortex
In adult mice learning a visual association task, cortical dendritic spines showed more complex nanostructure, increased formation, and smaller average size rather than remaining largely static. Spine responses became more selective for task-relevant orientations, and pre-training response discriminability predicted later task performance, linking structural remodeling to functional learning signals.
The combination of persistent spine formation, nanostructural remodeling, and predictive functional selectivity challenges the view that mature-cortex spines are mostly stable, but the abstract provides limited quantitative and behavioral detail.
-
The Critical Period Microbiota Shape Brain Plasticity
In juvenile mice, disrupting the gut microbiota with antibiotics impaired ocular-dominance plasticity in visual cortex and altered cortical gene expression related to extracellular matrix, the blood-brain barrier, and myelination. Transplanting microbiota from juveniles into adults restored this form of plasticity, suggesting that gut microbes or their signals help regulate developmental critical periods and may influence plasticity later in life.
The adult rescue of visual-cortex plasticity by juvenile microbiota is a genuinely interesting, assumption-challenging result, but the abstract provides no effect sizes, mechanistic identification, or detail on controls and remains preprint-level evidence.
-
The mouse superior colliculus promotes competing actions independently of sensory inputs
In a mouse audiovisual left-versus-right choice task, the authors find that superior colliculus neurons separately encode sensory and action-related information rather than combining them within the same cells. Inactivating one side reduced actions toward the opposite side without changing sensory sensitivity, while inactivating both sides restored choice balance. The results support a model in which the two colliculi provide stimulus-independent drive promoting competing actions, while sensory and prefrontal circuits determine which action is favored.
The causal dissociation between sensory sensitivity and stimulus-independent action promotion challenges the standard view of the superior colliculus as primarily integrating sensory evidence, though the abstract provides limited quantitative detail and the result is from one behavioral paradigm.
-
The blueprint of human functional architecture shifts from cognition to anatomy during perturbations of consciousness
The study compares fMRI functional connectivity with eight anatomical, molecular, electrophysiological, metabolic, and cognitive representations of the brain. It finds that, in awake people, cognitive co-activation is the strongest predictor of inter-region synchrony, but across disorders of consciousness and several anesthetics this influence largely disappears and anatomical and molecular constraints become more predictive.
The cross-condition result offers a non-obvious account of how consciousness reshapes functional brain organization, but the abstract provides no effect sizes or details sufficient to establish how robustly the multimodal comparisons support the claim.
-
A Two-Dimensional Grid-Cell Code for Three-Dimensional Navigation in Freely Flying Bats
The authors recorded medial entorhinal grid cells wirelessly from freely flying bats and report that grid-cell ensembles retain a two-dimensional toroidal organization. They argue that bats’ flight paths are themselves largely organized along transient two-dimensional planes, allowing a conventional 2D hexagonal grid code to support navigation in 3D without requiring a fully three-dimensional spatial code.
This offers a non-obvious resolution to how grid cells could support 3D navigation—by matching a 2D neural code to behaviorally relevant planes—but the abstract gives limited quantitative detail and the preprint evidence is not yet enough for a stronger recommendation.
-
Cross-cue reconstruction of perceived 3D object structure from human visual cortex
The authors decode fMRI responses into latent features of a pretrained 3D point-cloud autoencoder and reconstruct explicit 3D objects. A decoder trained on ordinary 2D images generalized to novel categories and to disparity-defined random-dot stereograms, including stimuli with identical 2D outlines but different perceived slants, especially in higher dorsal-stream visual areas. This suggests the decoded signal reflects a cue-invariant representation of perceived 3D structure rather than simply image appearance or object category.
Cross-cue reconstruction of disparity-defined 3D geometry is a genuinely interesting test of shared perceptual representations, but the abstract gives too little information about sample size, reconstruction accuracy, and robustness to justify a strong recommendation.
-
Mechanistic simulation identifies predictive dose-dependent biomarkers of propofol anesthesia
The authors build a multiscale thalamocortical model in which propofol acts only through GABA-A modulation, without fitting to anesthesia data. The same model reproduces auditory oddball effects in macaques and functional-connectivity changes in anesthetized humans, and predicts a dose-dependent increase in residual cortical activity between stimuli that is subsequently observed in macaque recordings.
The simulation-first discovery and cross-species validation of a previously unnoticed dose-related biomarker are genuinely interesting, but the abstract provides no quantitative effect sizes or robustness details, so it falls short of a strong recommendation.
-
Interplay of sleep neural oscillations enhances coordinated memory reactivation between cortex and hippocampus
Using mice trained on two visually distinct, lateralized virtual-reality memories, the authors tracked replay in visual cortex and hippocampus during sleep. They find that hippocampal ripple replay is preferentially aligned with the matching cortical memory when ripples coincide with cortical spindles and appropriately phased slow oscillations, suggesting these rhythms coordinate memory content rather than merely increasing overall communication.
The content-specific, three-way coordination of hippocampal and cortical replay is a meaningful mechanistic advance, but the abstract provides correlational evidence and does not establish that this coordination is causally necessary for consolidation.
-
$μ_0$: A Scalable 3D Interaction-Trace World Model
The paper proposes learning a robot world model that predicts smooth 3D trajectories of salient interaction points—objects, hands, tools, and contacts—instead of reconstructing pixels or predicting embodiment-specific actions. It introduces TraceExtract to obtain these traces and language associations from diverse videos, then uses the frozen model as a reusable interface for training policies on different robot embodiments; the resulting policies reportedly match action-supervised VLA baselines such as π0.
The 3D-trace interface is a plausible and meaningfully different route to cross-embodiment robot learning, and matching action-supervised models without action-labeled pretraining is notable, but the abstract gives no quantitative results or real-world breadth sufficient for a strong verdict.
-
Memento: Reconstruct to Remember for Consistent Long Video Generation
Memento trains long-video generation not only to continue the next shot, but also to reconstruct recurring subjects from the accumulated memory, making identity preservation an explicit objective. It uses separate memory queries for long-range identity evidence and short-range continuation context, supported by subject-focused training data; the abstract claims improved consistency and video quality, but gives no quantitative results.
The reconstruction-based formulation and dual-query memory are a meaningful approach to a central long-video-generation problem, but the abstract provides only unsupported state-of-the-art claims and no scale or numerical evidence.
-
RepFusion: Leveraging Multimodal Priors for Denoising in Representation Space
RepFusion uses a pretrained multimodal language model to repeatedly encode the current noisy visual representation, then conditions a diffusion transformer on those features to generate images. The paper argues that semantically structured representation-autoencoder latents make MLLM priors useful for denoising, and reports better results than similarly budgeted newly initialized denoisers, though no quantitative results are given in the abstract.
Repurposing an MLLM as a noise-aware, repeatedly queried denoising prior is a non-obvious architectural direction, but the abstract provides insufficient quantitative or broad evidence to justify a stronger recommendation.
-
SwiftVR: Real-Time One-Step Generative Video Restoration
SwiftVR is a causal, chunk-wise one-step generative video-restoration system designed for live high-resolution streaming. It replaces masked or sparse attention with deterministic shifted-window packing that uses standard dense GPU attention, and adds a lightweight restoration-oriented autoencoder; the authors report 26 FPS at 1080p on an RTX 5090 and 14 FPS at 4K on an H100, where compared diffusion baselines exceed memory limits.
Real-time 1080p generative restoration on a consumer GPU is a meaningful deployment result, but the abstract gives limited restoration-quality numbers and does not establish how much of the gain comes from genuinely new modeling versus careful systems optimization.
-
MVEB: Massive Video Embedding Benchmark
MVEB evaluates 33 video-embedding models across 23 tasks, including classification, retrieval, clustering, zero-shot recognition, and video QA. It finds that different model families excel in different settings, and that adding audio helps or hurts depending on whether the task labels were created using audio as well as visual information. The benchmark is integrated with MTEB and released with a larger 184-task pool.
This is primarily a benchmark paper, but the consistent finding that audio can hurt when annotations are visual-only is a useful and somewhat surprising warning for multimodal embedding evaluation.
-
World Model Self-Distillation: Training World Models to Solve General Tasks
The paper trains a video world model to solve tasks from only an image and short instruction, rather than requiring detailed textual plans. A vision-language model proposes tasks and evaluates generated outcomes; the system distills behavior from a caption-conditioned “Demonstrator” into an “Executor,” then improves it with reinforcement learning. It reports better task-solving scores than the Demonstrator on a new benchmark and competitive transfer to robotic tasks, but the evidence relies heavily on VLM-based evaluation.
The combination of self-distillation and VLM-feedback RL addresses a real scalability bottleneck in using video generators for planning, but the abstract gives no quantitative results and the central evaluation may be vulnerable to judging-model artifacts.
-
iMaC: Translating Actions into Motion and Contact Images for Embodied World Models
iMaC represents robot actions as visual images rather than joint commands or end-effector vectors. An image-action encoder maps target-driven images into control embeddings, and a world model predicts future states conditioned on them; the authors report better prediction, task success, and cross-scene generalization on manipulation benchmarks and real robots. The main novelty is treating visual action representations as a general control interface across embodiments, though the abstract gives no quantitative results or details about how images are generated and executed.
The visual-image action interface is a potentially useful and nonstandard alternative to hand-designed kinematic actions, but the abstract provides insufficient quantitative evidence to establish that it is a substantial advance rather than a repackaging of goal-image or visual-foresight control.
-
You Don't Need Strong Assumptions: Visual Representation Learning via Temporal Differences
The paper proposes Temporal Difference in Vision (TDV), a video self-supervised method that learns an image representation and a motion representation whose sum predicts the representation of the next frame. Unlike common methods using crops, masks, or augmentations, it relies mainly on the assumption that the past causes the future, and reportedly matches state-of-the-art approaches on dense vision tasks while using fewer engineered biases.
The additive image-plus-motion formulation and attempt to remove standard self-supervised biases are genuinely interesting, but the abstract gives no quantitative results or breadth of evaluation to establish that it is more than a promising alternative.
-
Improving Human Diving Endurance with a Field-Deployable, Untethered Exoskeleton
The authors built DiveMate, an untethered underwater exoskeleton that adaptively assists a diver’s kicking in naturalistic conditions. They report 42.9% more travel per unit of breathing-gas energy, 54.9% longer dive duration, and a 47.0% lower gas-consumption rate, alongside reduced muscle activation and more regular motion.
This is a notable real-world extension of exoskeletons to underwater endurance with large reported gains, but the abstract gives too little information about participant numbers, controls, and experimental breadth to justify a strong recommendation.
-
EgoEngine: From Egocentric Human Videos to High-Fidelity Dexterous Robot Demonstrations
EgoEngine converts egocentric human manipulation videos into robot-like observation videos and executable dexterous action trajectories, attempting to address both the human-to-robot visual and action gaps. In simulation and real-robot experiments, it reportedly enables zero-shot visuomotor dexterous policy learning without collecting real-robot demonstrations.
The claimed ability to learn dexterous robot policies from human videos with no robot demonstrations is a genuinely important capability, but the abstract gives no quantitative results, task breadth, or details sufficient to establish how robust or general the conversion is.
-
EmbodiSteer: Steering Embodiment-Agnostic Visuomotor Policies with Joint-Space Guidance for Zero-Shot Cross-Embodiment Deployment
EmbodiSteer adds embodiment awareness at inference time to Cartesian visuomotor diffusion policies. It uses forward kinematics, Jacobian-based joint-space updates, and collision-aware guidance during denoising, without retraining the policy; across nine simulated robots and two physical robots, it substantially reduces collisions and improves task success in constrained scenes.
The training-free joint-space steering mechanism is a meaningful approach to zero-shot cross-embodiment deployment, supported by results across multiple simulated and physical robots, though the abstract does not establish broad robustness beyond constrained collision-avoidance scenarios.
-
Improving Robotic Generalist Policies via Flow Reversal Steering
The paper introduces Flow Reversal Steering, which inverts a flow-matching robot policy to map coarse, suboptimal action suggestions from humans or VLMs onto nearby actions already supported by the policy’s behavioral prior. This improves zero-shot manipulation, can be distilled into an auxiliary policy in under a minute, and helps bootstrap reinforcement learning on tasks where standard RL fails; experiments cover simulated and real robots, with reported task-success gains of up to 95 percentage points.
The action-to-noise inversion mechanism is a genuinely interesting way to combine semantic guidance, pretrained generalist policies, and RL, with unusually large reported gains, but the abstract does not establish how broadly it works or whether the improvements persist beyond the evaluated manipulation tasks.
-
QPILOTS: Efficient Test-Time Q-Steering for Flow Policies
QPILOTS steers diffusion or flow-matching action policies at inference time using gradients from a learned Q-function, without retraining or distilling the policy. Its key idea is to first map each noisy denoising intermediate to an estimated clean action before evaluating the critic, avoiding unreliable Q-gradients on noisy actions; experiments report strong results on 50 offline-to-online RL tasks and six simulated manipulation tasks with a frozen vision-language-action model.
Inference-time Q-steering of frozen generative and VLA policies via clean-action projection is a meaningful, potentially reusable idea, supported by broad task coverage, but the abstract lacks detailed comparisons and real-world evidence needed for a strong verdict.
-
Scalable Deep Unfolding of Conic Optimizers
This paper makes deep unfolding practical for larger conic and semidefinite optimization problems. It uses matrix-free implicit differentiation to avoid quadratic memory and a numerically stable derivative for PSD projections with repeated eigenvalues, then learns solver hyperparameters and warm starts. Across SDPs, SOCPs, and covariance-steering subproblems, it reports speedups of up to 50×, including over 30× inside sequential convex programming.
The combination of scalable matrix-free differentiation and stable PSD-cone backpropagation addresses real barriers to learned conic solvers, with substantial reported speedups, but the abstract gives limited detail about problem sizes, baselines, and generalization beyond the tested optimization families.
-
AnyGoal: Vision-Language Guided Multi-Agent Exploration for Training-Free Lifelong Navigation
AnyGoal is a training-free, multi-robot navigation system that uses a vision-language model to score exploration frontiers and a shared Bayesian 2D value map to accumulate goal evidence across subtasks. On 360 unseen GOAT-Bench episodes, two agents reach 52.4% subtask success, 27.5 percentage points above Modular GOAT, while the ablation suggests open-vocabulary perception moves failures from finding goals to verifying them.
The persistent Bayesian evidence map and decentralized multi-agent frontier allocation form a useful new navigation architecture with a large reported benchmark gain, but the evidence is still centered on one benchmark and the claimed improvement may depend on several combined components.
-
Semidefinite Relaxations for Collision-Free Motion Planning
The paper formulates collision-free motion planning around spherical obstacles as a polynomial optimization problem and derives a semidefinite relaxation. It gives a geometric interpretation of the relaxation as exact planning in a potentially higher-dimensional space, characterizes when it is tight, and uses symmetry to keep the SDP cone sizes independent of ambient dimension. Experiments report 10–100× faster solves than SNOPT/IPOPT, lower timing variance, and successful use as a steering function for minimum-snap quadrotor RRT planning.
The combination of a new geometric tightness theory, dimension-efficient SDP reduction, and substantial reported speedups makes this worth a closer look, though the demonstrated setting is still a structured point-robot formulation rather than a broad breakthrough in practical motion planning.
-
Can Neural Networks Achieve Optimal Computational-statistical Tradeoff? An Analysis on Single-Index Model
The paper gives a unified polynomial-time gradient-based procedure for training two-layer networks on Gaussian single-index models. It proves that the method recovers a representation aligned with the hidden signal using nearly the statistically optimal sample complexity, matching known SQ lower bounds up to logarithmic factors across all generative exponents; a new weight-perturbation method extends the result to sparse signals and sparse tensor-PCA-like settings.
Matching the SQ computational-statistical lower bound with a general gradient-based neural-network method, plus a nontrivial sparse extension, is a meaningful theoretical result, though it is confined to structured single-index models and the abstract provides no empirical validation.
-
CausalDrive: Real-time Causal World Models for Autonomous Driving
CausalDrive proposes a driving video world model that predicts other road users instead of receiving their future trajectories as an oracle. It conditions on the initial camera view, the ego vehicle’s planned path, and a broad text prompt, using a distilled flow-matching model to generate reactive scenes at 12 FPS. The authors apply it to closed-loop evaluation, reinforcement learning, and human-in-the-loop simulation, claiming improved interaction behavior and fewer collision artifacts.
The potentially important idea is a real-time, reactive driving simulator without future NPC layouts, but the abstract provides no quantitative comparisons or convincing detail behind the claimed real-world policy gains.
-
FARM: Find Anything using Relational Spatial Memory
FARM gives a robot a compact, continuously updated memory of objects, their appearance, geometry, viewpoints, and spatial relations. It parses queries such as “the lamp below the dartboard and left of the poster,” grounds those relations explicitly, and uses VLMs to rank the matching object; across 44,000 queries in 67 varied scenes it reports large retrieval gains over prior methods while operating at 5–10 Hz, including a closed-loop quadruped deployment.
The combination of real-time relational object memory, explicit spatial grounding, and broad evaluation is a meaningful robotics capability advance, though the abstract gives relative rather than absolute gains and does not establish that it is a broadly new paradigm.
-
Inhibition of protein tyrosine phosphatase PTP1B function ameliorates pathophysiological deficits in Rett Syndrome
The study tests selective allosteric inhibitors of PTP1B in female mouse models of Rett syndrome and reports sustained improvements in motor, muscle, cardiac, and respiratory abnormalities. Genetic removal of PTP1B produces similar long-lasting rescue, supporting an on-target mechanism involving neurotrophic and metabolic signaling rather than merely symptomatic relief.
A potentially important, mechanism-based therapeutic target with unusually broad and durable rescue in a Rett model, but the evidence is still preclinical and the abstract provides few quantitative details while making strong claims about clinical translation.
-
Lineage-selective suicide gene system enables post-engraftment editing of cell therapy composition
The authors develop NeuroGuard, a suicide-gene system that uses NEUROD1-driven recombination to protect mature neurons while making glial, proliferative, and other non-neuronal graft cells vulnerable to inducible ablation. In stem-cell-derived midbrain grafts for Parkinsonian rodents, activation enriched neurons—especially dopaminergic neurons—without reducing dopaminergic cell numbers, innervation, or behavioral recovery, with single-cell data supporting lineage-selective removal.
This is a potentially general and practically important way to adjust graft composition after transplantation rather than choosing between incomplete purification and destroying the whole graft, but the evidence is still limited to in vitro work and rodent transplantation.
-
Curated Residual Decomposition for Increased MU Yield from HD-sEMG
The paper introduces Curated Residual Decomposition, which repeatedly decomposes the residual left after manually cleaning an initial HD-sEMG decomposition to recover motor units that were missed, especially smaller and lower-threshold units. Across validation and public datasets, it increased motor-unit yield by 31–50% in the validation data and 35–142% in public data, while intramuscular recordings showed nearly comparable identification accuracy to the original units.
The large, validated increase in recoverable motor units addresses a known amplitude-selection bias in HD-sEMG, although the evidence is still based on limited datasets and an offline, manually curated workflow rather than a broad change in capability.
-
Ling and Ring 2.6 Technical Report: Efficient and Instant Agentic Intelligence at Trillion-Parameter Scale
This technical report describes Ling-2.6 for fast responses and Ring-2.6 for deeper agentic reasoning, upgrading an earlier model rather than training from scratch. It combines hybrid linear/MLA attention, token-efficiency-focused post-training, and a reinforcement-learning system (KPop) for coding, search, tool use, and workflow environments, reportedly scaling Ring-2.6 to 1T parameters. The abstract gives no quantitative capability, efficiency, or deployment results, though it says the checkpoints are open-sourced.
The trillion-parameter open agent models and unified architecture/training/serving design could be worth examining, but the abstract provides no numbers or credible evidence that these techniques deliver a meaningful advance over existing large agent systems.
-
FORT-Searcher: Synthesizing Shortcut-Resistant Search Tasks for Training Deep Search Agents
The paper proposes FORT, a framework for generating multi-step search questions designed to prevent agents from solving them through cheap shortcuts such as overlapping evidence, highly selective clues, or prior-knowledge cues. It uses trajectory-based diagnostics and adversarial refinement to create harder training data, then reports that SFT on this data produces a strong open-source search agent without specialized RL, though the abstract gives no quantitative results.
The explicit focus on realized search difficulty and shortcut-resistant data synthesis is a useful, potentially general framing, but the abstract lacks numbers and it is unclear whether the gains exceed better data curation or scale effects.
-
Beyond Monolingual Deep Research: Evaluating Agents and Retrievers with Cross-Lingual BrowseComp-Plus
The paper introduces XBCP, a benchmark for testing deep-research agents when questions and supporting evidence are in different languages, including a 12-language mixed corpus. Across four agents and multilingual retrievers, performance falls due both to retrieval failures and difficulty integrating language-mismatched evidence; accuracy remains lower even when the gold evidence is provided directly.
The notable finding is that cross-lingual research fails even with perfect evidence access, pointing to an agent-side reasoning and evidence-integration problem rather than merely inadequate retrieval, though the abstract provides no quantitative results.