Issue 31 · Jul 27 – Aug 2, 2026
Every candidate
All 3,172 papers were scored from their abstracts by gpt-5.6-luna; 1,223 were not skipped. Shown here: the top 150 of those, in score order. Picks are marked.
-
DiffusionGemma Technical Report
DiffusionGemma adapts a Gemma 4 mixture-of-experts model into a discrete-diffusion language model that refines 256-token blocks in parallel rather than decoding token by token. It reports about 1,500 output tokens per second on one H100, using fewer than 10% of the original model’s training-token budget, while largely retaining capability, multimodal input, long-context support, and even autoregressive generation.
Parallel blockwise diffusion decoding combined with high claimed speed, low conversion cost, and compatibility with existing AR capabilities represents a potentially major alternative to the autoregressive bottleneck, though the abstract gives limited quality and benchmark detail.
-
Keep It InMind: Benchmarking the Implicit-Association Blind Spot in Agent Memory
The paper introduces InMind, a 125-task benchmark for testing whether agent memory can use facts that are only indirectly related to a query—for example, applying a stored tree-nut allergy when answering a macaron request. Across six memory systems, indirect-query performance reaches only 14.4% despite near-perfect recall when the relevant memory is explicitly requested, while keeping the memory visible raises backbone accuracy to 84.0%. The results suggest the main failure is query-conditioned routing and retrieval, not storage or missing world knowledge, and that simply increasing embedding dimensionality does not solve it.
It identifies and cleanly isolates a substantial, plausibly widespread failure mode in agent memory: systems can store and recall facts yet fail to surface them when the needed connection is implicit, with strong cross-system evidence that common scaling does not fix it.
-
The Kinetics of Training: A Driven-Nucleation Rate Law for Emergence, Plasticity Loss, and Circuit Control in Language Models
The paper models language-model capability formation as a rare, all-at-once alignment event: circuits receive no partial credit, and the difficulty is determined mainly by how many components remain missing. Across Pythia models and seven capabilities, it reports abrupt emergence, predictable ignition timing, and a distinct form of plasticity loss that can be reversed by reinitializing query-key weights but not value weights. It proposes a rate law for emergence and tests it with controlled attention models and scheduled circuit reinitialization.
The combination of a quantitative nucleation theory, evidence against partial-credit learning, predictive onset estimates, and a targeted intervention restoring learnability would substantially change how capability emergence and catastrophic plasticity loss are understood if the results hold up.
-
Judging Is Not Enumerating: Silent Omissions in LLM-Authored Acceptable Sets
This paper studies whether LLMs can author complete acceptable sets—such as test suites, answer keys, or reward criteria—rather than merely judge individual candidates. Across algorithmic, code, and lexical settings, models judge membership much better than they enumerate the full set: they omit many valid items, while a predicate-format control reaches about 0.99 F1, suggesting the bottleneck is materializing the induced set rather than specifying it. The paper also shows that omissions are hard to audit, can degrade RLVR substantially, and that execution-based repair improves suite yield by 3.3–10.6×.
The judging-versus-enumerating distinction exposes a consequential and non-obvious failure mode in LLM-generated evaluators, supported by results across multiple reference types and practical RLVR experiments.
-
ReToken: One Token to Improve Vision-Language Models for Visual Retrieval
ReToken adds one learned query token that retrieves a sparse, question-relevant subset of visual tokens from an already computed visual KV cache, rather than processing all visual context equally. Trained on a small image-QA dataset, it improves long-context image retrieval substantially and transfers zero-shot to long-video understanding, while fitting training and inference on a single H100.
The combination of a very lightweight retrieval bottleneck, large reported gains on visual haystack tasks, and zero-shot transfer to long video is a non-obvious and potentially broadly useful result.
-
Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-design
The paper trains a diffusion transformer over a unified token representation of robot bodies, states, and actions, allowing it to jointly generate robot designs and controllers conditioned on target motions. Its dynamics-based self-guidance can optimize rewards not seen during training, and experiments across several embodiment spaces outperform evolutionary search; a fabricated ALOHA redesign reduced tracking error by more than 70%.
This is a potentially important shift from optimizing controllers for fixed robots to using one learned dynamics model to co-design embodiments and control across robot types, with a substantial hardware result rather than only simulation evidence.
-
HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone
HiFi-UMI is a robot-free teleoperation data-capture system designed to make demonstrations accurate enough for direct real-robot deployment. It combines stereo-inertial tracking, directly measured relative gripper pose, microsecond synchronization, and wide-FOV cameras, achieving about 3 mm local end-effector accuracy. Policies trained or post-trained only on this data reportedly match real-robot teleoperation baselines across three model families, while 4,000 hours of pretraining data improve action error and task success substantially.
The potentially important result is that sufficiently high-fidelity robot-free demonstrations can eliminate the usual real-robot anchor and transfer directly to hardware, supported by multiple backbones and real manipulation experiments, though the abstract does not establish how broadly this holds beyond the reported tasks.
-
DarwinX: Evolving Agent Harnesses Through Natural Selection
DarwinX evolves an LLM agent’s prompts, tools, skills, and control flow as a population rather than optimizing a single lineage, while keeping the underlying model frozen. Its preserve-and-extend selection and archive-based recombination reportedly improve performance across Terminal-Bench, TerminalWorld, WebArena-Infinity, and SWE-bench transfer, including a rise from 43.5% to 93.0% on audited WebArena-Infinity tasks and about 17 points on average across benchmarks.
The combination of population-based harness evolution, verifier-derived fitness, and strong cross-task transfer suggests a potentially important alternative to scaling or fine-tuning model weights, though the unusually large claims merit careful verification.
-
Climbing fibers encode the gradient of a loss function for the cerebellum
In marmosets learning rewarded saccades, the authors varied the value of different targets and measured climbing-fiber signals in the cerebellum. They find that these signals combine reward prediction errors with movement errors and, on average, encode the dot product between a reward-weighted error and each Purkinje cell’s motor influence, resembling the gradient of a loss function rather than a generic error signal.
This offers a concrete physiological and causal account of how cerebellar teaching signals could solve credit assignment by transmitting neuron-specific loss gradients, supported by reward manipulation and perturbation-based estimates of Purkinje-cell motor effects.
-
Pictura: Perspective-View Self-Play at Scale for Driving
Pictura is a GPU-accelerated simulator that renders egocentric camera views for every agent during large-scale multi-agent self-play, avoiding the usual privileged-state training followed by camera-policy distillation. Using plain PPO and 50 billion agent steps, the resulting Alberti policy nearly matches a privileged-observation policy and transfers zero-shot to re-rendered Waymo layouts, where it reportedly outperforms privileged agents.
Directly training a driving policy from perspective images through massive self-play is a meaningful change in regime, and the claimed scale plus zero-shot transfer challenge the assumption that privileged-state self-play is necessary, although the abstract provides few quantitative details.
-
Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering
The paper presents an open framework for studying AI systems that improve machine-learning engineering through executable feedback and long-horizon program evolution. A 35B model trained with four operators—draft, improve, debug, and crossover—reportedly raises MLE-Bench Lite performance from 39.39% to 60.61%, or 71.21% with an expanded search procedure, and transfers to NatureBench Lite. The notable idea is to combine post-training and evolutionary search in one execution-grounded loop rather than treating model improvement and inference-time search separately.
This is a potentially important, executable formulation of recursive self-improvement with large reported gains and cross-benchmark transfer, but the claims depend heavily on benchmark/search comparisons and the abstract does not provide enough detail to establish that the improvement is genuinely general rather than search engineering or evaluation-specific.
-
Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation
The paper proposes Explorative Modeling, which samples multiple candidate outputs and trains against whichever best matches the data, factoring the training loop rather than decomposing generation into stages. Across image, video, and language experiments, increasing the number of candidates reportedly improves scaling and efficiency, while an end-to-end version matches diffusion on control tasks with 16–256× fewer inference steps. The claimed gains are substantial, but the abstract does not provide enough methodological or experimental detail to assess how broadly they hold.
The best-of-K training formulation and its claim that exploration is a third scaling axis are genuinely notable, but the unusually broad and large efficiency claims need careful verification before this merits a strong recommendation.
-
TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM
TurboVLA replaces the usual large-language-model-centered vision-to-language-to-action pipeline with separate visual and language encoders, lightweight bidirectional interaction, and a compact action decoder. On LIBERO, it reports 97.7% average success using 0.2B parameters, 31.2 ms inference latency, and 0.9 GB VRAM on an RTX 4090, matching or exceeding much larger policies. The main contribution is an efficient VLA design aimed at real-time, low-memory robot control rather than another LLM-based policy.
The unusually low latency and memory use, together with a credible non-LLM-centric VLA formulation, are worth attention, but the evidence is limited to a single simulated benchmark and does not yet establish real-world generality or robustness.
-
Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents
Qwen-UI-Agent is a broad GUI-agent system designed to operate across mobile devices, desktops, browsers, and search environments. It combines GUI actions with CLI commands, batches actions, trains on long-horizon trajectories using online RL and large-scale parallel rollouts, and uses an automated data-generation and failure-analysis loop. The report claims strong results on mobile-use benchmarks, including real-device evaluation, while remaining competitive on computer and browser tasks.
The combination of real-device training, unified GUI/CLI actions, 100+ turn online-RL trajectories, and automated environment/data iteration is a substantial systems direction, but the abstract provides limited ablation or independent evidence beyond benchmark scores and state-of-the-art claims.
-
A Change of Frame Makes Balance Observable: Distillation-Free Humanoid Single-Leg Stance
The paper shows that expressing center-of-mass velocity relative to the support foot makes the capture-point signal recoverable from a humanoid’s encoders and IMU, without adding sensors or using distillation. Adding this dynamic-CoM observation and human-inspired balance rewards lets a policy maintain clean single-leg stance on 89/90 held-out simulated motions, versus 0/90 for eight existing general policies, and transfer to a real Unitree G1; removing the observation reduces performance by 43 points.
This combines a simple but non-obvious frame-change that makes a key balance variable deployable with unusually large gains over existing humanoid policies, a strong ablation, cross-simulator evaluation, and real-robot transfer.
-
Bicycle Acrobatics with Reinforcement Learning
The authors use several reinforcement-learning objectives plus a state-triggered policy orchestrator to teach a custom bicycle robot a repertoire of dynamic stunts, including jumps, wheelies, front flips, kip-ups, and kip-downs. The system performs these behaviors repeatedly on hardware, composes them into longer sequences, follows waypoints across previously unseen multi-table layouts, and completes repeated steerable maneuver repertoires.
This is a substantial real-world capability result: an underactuated wheeled robot performs and composes acrobatics usually associated with legged robots, with repeated hardware demonstrations rather than simulation-only claims.
-
Neuropil aggrecan, not perineuronal nets, closes the critical period for visual plasticity
Using cell-type-specific Acan deletions in mice, the authors find that removing aggrecan from PV interneurons eliminates perineuronal nets but does not reopen visual critical-period plasticity. Plasticity remains in adulthood only when aggrecan is removed from excitatory neurons or broadly across neurons, implicating neuropil aggrecan from excitatory cells rather than PNNs themselves as the limiting mechanism.
The study directly challenges a long-standing PNN-based explanation for critical-period closure with targeted genetic manipulations and identifies a different cellular source and mechanism for restricting plasticity.
-
VideoCoCo: Code-as-CoT for Physically-Consistent Video Generation via an Agentic Dual-Engine System
VideoCoCo uses an LLM-generated Blender program as an executable chain-of-thought: the simulator produces a temporally explicit draft, and a video model converts that draft into a photorealistic result. It reports large gains over OmniWeaving on PhyGenBench (0.475→0.558) and VBench-2.0 (52.18→77.88), suggesting that separating physical process specification from visual rendering can improve consistency and controllability.
The executable-simulation intermediate representation is a genuinely interesting direction for physically grounded video generation, but the abstract provides only benchmark-level evidence and limited information about robustness, comparisons, or whether the gains generalize beyond the curated VideoCoCo-3K setup.
-
A candidate electromagnetic channel for coordination between physicallyseparated Drosophila
The authors report that individually housed fruit flies show weak but reproducible movement coordination despite opaque barriers, interrupted mechanical paths, and a shared light–dark stimulus. Material-dependent effects and RF coupling measurements are interpreted as consistent with near-field magnetic radiofrequency communication, but the carrier frequency, detector, and biological mechanism remain unidentified.
The proposed inter-animal magnetic-RF channel would be highly paradigm-shifting, but the abstract provides limited quantitative evidence and leaves substantial alternative explanations and the actual signal mechanism unresolved.
-
MindForge: Teaching Small Language Models Whole-Life-Cycle Software Engineering via Source-Free Program Synthesis
MindForge automatically turns open-source command-line programs into training environments that reveal only a compiled executable and documentation, allowing agents to practice the full from-scratch software lifecycle without source-code supervision. Fine-tuning Qwen3.6-27B on trajectories generated by a larger teacher raises ProgramBench pass rate from 37.98% to 49.51% and produces gains across seven unseen benchmarks covering repository generation, translation, bug fixing, and feature work.
The source-free, whole-lifecycle training environment is a substantive idea, and the broad cross-benchmark gains are unusually large, but the abstract provides too little detail to establish robustness or rule out benchmark and teacher-data artifacts.
-
The Phase Transition in Online PCA Depends on $n/d\log(d)$, not $n/d$
This paper shows that Oja’s streaming PCA algorithm needs about d log d samples—not merely d samples—to obtain nonzero asymptotic alignment with the top population eigenvector under a spiked Gaussian model. It derives an explicit phase transition and a critical-window distribution, contrasting sharply with ordinary batch PCA, where a constant sample-to-dimension ratio can suffice.
The logarithmic sample-complexity threshold for online PCA is a genuinely non-obvious correction to the usual constant-aspect-ratio picture, with a precise theorem and critical scaling, though the result is currently limited to Oja’s algorithm and an idealized covariance model.
-
Kimi K3: Open Frontier Intelligence
Kimi K3 is presented as a 2.8T-parameter mixture-of-experts model with 104B active parameters, native vision, and a million-token context window. It combines new attention and residual mechanisms with a highly sparse expert router, and claims roughly 2.5× better scaling efficiency than Kimi K2, plus strong long-horizon coding and agentic performance; however, the abstract provides no concrete benchmark results or detailed evidence for these claims.
The combination of open frontier-scale weights, million-token agentic training, and claimed 2.5× scaling efficiency could be important, but the abstract is too promotional and underspecified to justify a strong recommendation without inspecting the technical report and results.
-
Stable FP4 Training via Transposition-Invariant Block Quantization
The paper argues that FP4 training becomes unstable because conventional 1D block quantization uses different scales before and after matrix transposition, making forward and backward computations inconsistent. It introduces 2D block quantization with transposition-invariant scales, plus truncation-free scaling and stochastic rounding, and reports stable training of dense models up to 7B parameters and a 30B-parameter MoE model over 100B tokens, with under 1.3% degradation versus BF16; attention uses MXFP8 rather than FP4.
This provides a specific explanation for FP4 training instability and a relatively simple fix demonstrated at unusually large model and token scales, potentially enabling substantially cheaper LLM training beyond FP8.
-
PhiZero: A World Model Built Around Physical Language
PhiZero represents predicted physical state transitions as a compact discrete “physical language,” then renders those transitions as video rather than predicting pixels directly. It is trained self-supervised on in-the-wild video and is reported to support physically coherent generation, action-conditioned simulation, interactive modeling, and zero-shot motion transfer.
The explicit reason-then-render world-modeling approach is a potentially important departure from pixel-space prediction, but the abstract gives no quantitative results or details sufficient to establish that it materially improves capability or generality.
-
HumanCLAW: Can Vision-Language Models Act Through a Body?
HumanCLAW evaluates whether vision-language models can choose the right next action for a physical body without conflating decision errors with failures of low-level control. It uses a harnessed body that executes atomic commands as short continuous motion segments under gravity and collisions, and tests nine VLMs on 1,218 long-horizon indoor navigation and interaction episodes. The best model succeeds on only 16.8% of episodes; the reported failure is not mainly object recognition, but losing track of the body’s location, progress, and collisions.
The decision-versus-execution decoupling and the finding that current VLMs fail mainly through weak embodied self-awareness are genuinely interesting, but this is primarily an evaluation framework and benchmark, with limited evidence in the abstract that it enables a major capability advance.
-
SpatialCLI: Learning to Reason With Spatial Tools, Then Without Them
SpatialCLI trains a vision-language model to use specialist spatial vision tools for localization, segmentation, depth, and pose, then distills successful tool-use behavior so the model can perform much of the reasoning without those tools. On the 516-example MindCube benchmark, Qwen3-VL-8B improves from 29.3% to 84.6% with tools and reaches 73.8% after internalization, though the abstract gives little detail about evaluation breadth or tool costs.
The combination of tool-augmented spatial reasoning and subsequent capability internalization produces a striking reported gain, but the evidence is currently limited to one small benchmark and lacks enough detail to establish broad generality.
-
Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers
Chimera combines linear-time Kimi Delta Attention, occasional global latent attention, local convolutions, and sparse MoE layers in a single text/image/video diffusion backbone. Its proposed HeteroP scaling method tunes the heterogeneous modules and produces compute-scaling laws; an 11B-parameter model with 2B active parameters reportedly achieves 7.3× system-level compute efficiency over a matched full-attention baseline and extrapolates from 5- to 30-second videos without length-specific fine-tuning.
The combination of efficient long-context attention with modality-aware diffusion scaling and striking zero-shot video-length extrapolation is genuinely interesting, but the abstract gives limited evidence beyond loss efficiency and one video metric, so it falls short of a strong recommendation.
-
Meshy T2: Fast Native Mesh Generation with Flow Matching
Meshy T2 generates polygonal 3D meshes from images using parallel flow-matching models rather than slow autoregressive token decoding. A vertex-level mesh VAE and coarse-to-fine voxel/mesh cascade preserve geometry and artist-style connectivity, while a requested vertex budget controls complexity; the authors report median end-to-end generation time of 6 seconds, over 10× faster than autoregressive baselines.
The combination of native vertex-set mesh representation and flow-based parallel generation appears to offer a meaningful speed and controllability advance, but the abstract provides too few benchmark details to establish whether the claimed fidelity and topology advantages are broadly credible.
-
RestoreKV: Recovering Full-Cache Behavior Under Aggressive Query-Agnostic KV Cache Eviction
RestoreKV augments query-agnostic KV-cache eviction with a small learned, context-conditioned “restore” cache generated during prefill, while keeping the same total cache budget and leaving later queries unchanged. Across four backbones, four long-context benchmarks, and five eviction methods, it substantially recovers lost accuracy—for example, improving KVzip from 38.2 to 73.2 on RULER-4K at a 5% budget—using adapters covering only 0.4% of parameters and under 0.5% extra cache-construction overhead.
The selection-plus-learned-restoration formulation is a genuinely useful new direction for aggressive KV compression, and the broad, budget-matched results show unusually large gains, though the evidence is still benchmark-level rather than a demonstrated change in real deployment capability.
-
X-NavDP: Generalizing Navigation Diffusion Policy to Novel Behavior and Embodiments with Group Q-score Reweighted Matching
The paper post-trains a pretrained navigation diffusion policy with online reinforcement learning, using behavior perturbations for exploration and group-normalized Q-scores to reweight diffusion score matching. It reports much better generalization across robot embodiments and difficult local-navigation behaviors: simulation success rises from 61.2% to 84.3%, while real-world hard-case success rises from 10% to 65%.
The combination of data-efficient RL with diffusion-policy post-training appears to unlock substantial cross-embodiment and real-world navigation gains, including behaviors poorly represented by oracle-planner demonstrations, though the abstract does not establish how broad or fairly measured the comparisons are.
-
Millisecond-scale, single-neuron credit assignment in a songbird
The authors used neuron-specific, activity-contingent auditory perturbations in songbirds to test how precisely vocal learning can assign credit. Perturbing one premotor neuron at a particular moment caused adaptive changes in that neuron’s firing with about 3.2 ms temporal precision, without comparable changes in neighboring neurons. The result suggests that reinforcement-driven learning in the brain can be far more temporally and spatially precise than the delayed, diffuse dopamine signal alone would imply.
This is an unusually striking, potentially assumption-changing result about single-neuron and millisecond-scale credit assignment, but the abstract provides limited information about sample size, replication, and robustness, and it is currently a bioRxiv preprint.
-
Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model
Mage-VL is a streaming vision-language model that uses video-codec signals—motion vectors, residuals, and sparse keyframes—to encode only visually changing regions instead of uniformly sampling frames. Its dual-system design combines a lightweight event detector with a causal reasoning decoder, reportedly cutting visual tokens by over 75% and achieving up to 3.5× faster inference while matching or exceeding larger models on video and spatial tasks. The paper also reports training-data and supervised-fine-tuning findings, but the abstract provides limited detail on evaluation breadth and absolute results.
Codec-native selective visual tokenization plus proactive streaming is a meaningful direction for real-time multimodal models, with potentially substantial efficiency gains, but the broad superiority claims and many auxiliary findings need closer inspection of benchmarks and comparisons.
-
Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory
This paper scales a previously proposed parametric long-term memory module to 6.9B parameters and 300B training tokens, using distributed indexing and sparse retrieval to make the kNN-based system feasible. Allocating parameters to the memory module appears more effective than scaling the base language model alone: a 410M Pythia model plus 6.9B memory reportedly matches or slightly exceeds Pythia-12B on 17 benchmarks with 39% fewer total parameters, while domain memories improve Qwen3 models by over 9 average points.
The independently scalable memory component and strong parameter-efficiency results are genuinely noteworthy, but this is primarily a large-scale validation and systems extension of an existing Memory Decoder idea rather than an entirely new direction.
-
Weak-to-Strong On-Policy Distillation
The paper proposes distilling from weaker models by forming a proxy teacher from the logit difference between a positive and negative model, then adding that capability direction to the strong student’s base distribution. Across math and code benchmarks, this reportedly improves the student beyond both standard on-policy distillation and the available domain teacher, including when all supervision models are weaker; different contrast pairs provide different reasoning signals.
The weak-to-strong logit-contrast formulation is a non-obvious direction that challenges the usual requirement for a stronger teacher, but the abstract gives no quantitative results or details about model scale and robustness, so it does not yet merit a strong recommendation.
-
Pyramidal Width Can Increase Under Vertex Insertion
The paper gives an exact three-dimensional counterexample to a 2015 conjecture that adding a vertex to a polytope cannot increase its pyramidal width, assuming the original vertices remain vertices. It provides integer coordinates, certifies the relevant face structures and distances exactly, and shows a roughly 1.41× increase, with a verifier included.
This cleanly overturns a published geometric assumption underlying analyses of Frank–Wolfe-type methods, with unusually strong exact verification, but its direct algorithmic implications and broader practical significance are not established in the abstract.
-
Harness-G: A Graph-Structured Harness for Search Agents
Harness-G changes RL search agents from generating arbitrary natural-language queries to choosing among explicit evidence sentences, entities, or an answer action in a graph-structured environment. It also adds a non-myopic credit assignment method that evaluates alternative actions and credits earlier choices for later gains; across six QA benchmarks, it reports large F1 improvements over Graph-R1, especially at 1.5B parameters.
The explicit reformulation of retrieval as finite action selection addresses a real policy–environment aliasing problem and shows sizeable gains across six benchmarks, but the evidence is still confined to benchmark QA results and the abstract does not establish broad generality or robustness.
-
WCM: A World Critic Model for Vision-Language-Action Reinforcement Learning
The paper augments a value critic for vision-language-action policies with a lightweight world-model objective: it predicts future latent states while estimating returns, rather than learning values from isolated frames. Across 149 simulated tasks and seven real-world manipulation tasks, the authors report better in-distribution and especially out-of-distribution performance, working with several VLA backbones and both on- and off-policy RL.
The explicit combination of latent future prediction and value estimation addresses a real partial-observability weakness in VLA critics and is tested broadly, but the abstract gives no quantitative gains or ablations to establish how much it improves over strong alternatives.
-
Reinforcement Learning for Code Optimization
This paper develops an RL setup for optimizing already-correct code, where execution speed is a noisy and sparse reward. It combines a calibrated timing benchmark, an offline simulator for selecting promising configurations, and modifications to GRPO and evaluation; on Qwen 2.5 7B and CWM 32B, optimization-aware training substantially improves speed-sensitive pass rates while preserving correctness, though it reaches only about half the rate of human complexity improvements.
The combination of learnable timing rewards, offline configuration selection, and RL adaptations addresses a real failure mode in code optimization and reports sizeable gains, but it is still a specialized extension of RL-for-code rather than an obviously general new paradigm.
-
Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale
Echoverse builds stateful, login-gated web applications from specifications and uses their databases to grade computer-use tasks. A co-evolution loop turns rollout failures into repairs to the environments, tasks, and verifiers while also training the agent; across 14 evaluation splits, a 9B model improves from 36.5% to 67.1%, with additional gains from deep environments, targeted interaction practice, and dense RL rewards. Four environments are released as a benchmark.
The combination of behaviorally deep, self-verifying environments with rollout-driven co-evolution addresses a central bottleneck in computer-use-agent training, and the reported gains are substantial, though the evidence is still limited to 12 environments and four released benchmarks.
-
ST-WAM: Semantic-Temporal World Action Model for Robust Manipulation under Visual Distribution Shifts
This paper argues that pixel-level future prediction in world action models can cause models to hallucinate familiar training-domain visuals when the current scene shifts. ST-WAM combines DINOv3 semantic features with VAE-based visual dynamics and retrieves relevant recent history; it reports large gains under visual shifts, including a rise in real-world success from 25.8% to 61.5% and a 21.3-point zero-shot improvement on LIBERO-Plus.
The diagnosed training-distribution hallucination and the substantial reported robustness gains make this more than an incremental WAM variant, but the abstract does not establish how broad or independently validated the real-world and benchmark results are.
-
To Add Is Machine, To Delete Is Human: Measuring and Mitigating Deletion Avoidance in LLM Code Editing
This paper studies a specific failure mode in LLM code editing: models often preserve code that should be deleted, instead wrapping it in guards or fallbacks so existing tests still pass. Across SWE-bench Verified and a new 200-task deletion-only benchmark, the authors find substantial deletion failures, show that tests requiring actual removal sharply reduce apparent success, and report that post-training focused on deletion can improve the behavior.
The concrete deletion-avoidance phenomenon, Guard-and-Go analysis, and deletion-only benchmark expose a non-obvious gap in current code-agent evaluations, with reasonably broad empirical support, but the demonstrated mitigation and overall capability impact are not yet large enough for a strong verdict.
-
Metis: Memory Foundation Model
Metis proposes making long-term memory an intrinsic, dynamically evolving state inside a foundation model rather than an external retrieval or memory module. It trains the model to compress, maintain, and use historical information through memory attention, with memory updates performed by ordinary forward passes while model weights stay frozen. The paper presents a prototype, specialized training data, and experiments analyzing its memory behavior, but the abstract gives no quantitative results.
The native, gradient-free memory-state architecture is a meaningful research direction, but the lack of reported results or comparisons makes it too early to judge whether it delivers a substantial practical advantage.
-
LeapTalk: Breaking the Latency-Quality Trade-off in Talking Head Generation
LeapTalk is a talking-head generator designed for arbitrarily long, streaming videos using a single forward pass rather than multi-step diffusion or autoregressive rollout. It combines Brownian-bridge data-to-data distillation, a persistent identity reference, SNR-aligned teacher–student matching, and audio-guided classifier-free guidance; the authors report up to 200 FPS with improved long-term stability and lip synchronization.
Single-step generation at claimed 200 FPS while avoiding autoregressive identity drift would be a meaningful advance for real-time talking heads, but the abstract provides no quantitative comparisons or details sufficient to establish that the broad latency and quality claims hold.
-
$π\mathbf{R}^2$: Reactive Real-time Flow Policies
This paper adapts action-chunking flow policies for closed-loop manipulation by separating rapidly refreshed proprioceptive inputs from slower vision-language features, and by using a latency-adaptive one-step denoising schedule that accounts for actions already in flight. On a real xArm6+XHand system, the method reportedly replans at about 25 Hz—roughly four times faster than the base policy—and improves success by up to 30% in real tasks and 23% in simulation.
The combination of asynchronous conditioning and latency-aware flow inference directly addresses a major practical weakness of large action-chunking policies, with credible simulation and real-robot gains, but the abstract lacks enough task breadth and detailed comparisons to justify a strong verdict.
-
Scaling Properties of Text Conditioning in Visual Generation
The paper investigates how the amount of structured information in text prompts affects visual-generation training, using likelihood- and attribute-based measures rather than token count. It reports simple scaling relationships between prompt structure and converged diffusion loss, then uses image-derived annotations and a trained prompter to improve generation and compositional reasoning across many benchmarks.
The finding that diffusion loss tracks structured semantic content, plus the proposed diffusability/promptability framing, is potentially important, but the abstract gives no quantitative results or enough experimental detail to substantiate its broad model-comparison claims.
-
Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability
This paper studies filesystem-based long-term memory for LLM agents, where agents maintain and search a growing tree of files while also storing distilled skills. Across conversation and embodied-task settings, organized hierarchies roughly halve retrieval cost on large stores, but organization degrades as memories accumulate and does not yet improve answer quality; tool choice affects the resulting store nearly as much as model choice.
It provides a useful, non-obvious evaluation of a widespread but underexamined memory design, especially the finding that organization improves search economy without improving answers and is strongly shaped by tooling, though it appears primarily to be a systematic study rather than a major new capability or method.
-
Mental World Modeling
The paper proposes Mental World Modeling, which augments physical world models with agents’ beliefs, goals, intentions, emotions, and social constraints. Its inspectable, training-free MENTIS baseline parses multimodal scenarios, generates target-specific observations, simulates coupled physical and mental state changes, and evaluates possible actions; tests across eight LLM-based models suggest that explicit mental-state tracking improves human-action prediction. The evidence comes from a manually constructed dataset of text, images, and sounding videos, with further analysis of current model limitations.
The explicit coupling of physical dynamics with target-dependent mental states is a potentially important world-modeling direction, but the abstract provides little quantitative evidence and relies on a manually constructed dataset and a training-free baseline.
-
Wonder: Video World Model Done Better
Wonder is an interactive video world model that turns an image or video into a navigable scene. It combines a coordinate-field-based camera control signal, sparse-attention memory for retrieving relevant past context, and modified distillation training to maintain control fidelity and consistency over long rollouts. The authors claim real-time generation at 16 FPS for minute-scale exploration, including revisiting regions and re-shooting existing video scenes.
The combination of controllable camera navigation, long-horizon memory, and real-time generation addresses a meaningful limitation of current video models, but the abstract provides no quantitative evidence for consistency, latency, or comparisons against competing world models.
-
CoTinyVLA: Chain-of-Thought Distillation for a Sub-Billion-Parameter Vision-Language-Action Model
CoTinyVLA distills a large teacher’s hierarchical reasoning into a 0.9B vision-language-action model, using episode-level plans, chunk-level subaction reasoning, long temporal visual context, and command paraphrases. On 10,030 LIBERO-Plus perturbed tasks, it reportedly outperforms larger 7B baselines across all four suites while using only 2.25 GiB of GPU memory; ablations indicate that the plan representation and temporal frame allocation are especially important.
The combination of strong robustness from a sub-billion-parameter model and evidence that distilled episode plans materially drive performance is notably interesting, but the claims are demonstrated on one benchmark and use several bundled components rather than establishing a broadly validated new paradigm.
-
Collusion with Competitive Marginals: Price-Level Audits Are Blind by Construction
The paper argues that agents can coordinate through correlations in their unexplained bid components while leaving each individual agent’s price distribution looking perfectly competitive. It proves that audits based on one agent’s price or bid history cannot detect this behavior, then reports correlated behavior among language-model bidding agents and substantial bidder dependence in Ethereum auctions; it suggests identity resolution and operator counting rather than price-level tests as the practical response.
The combination of an impossibility result with evidence that the proposed hidden coupling appears in LLM agents is genuinely non-obvious, though the practical significance depends on whether the empirical correlations establish strategic collusion rather than common causes or benign dependence.
-
Parallel Decoding Distillation for Fast Image and Video Generation
The paper proposes Parallel Decoding Distillation, a trajectory-distillation method that trains diffusion or flow models to predict several denoising steps per network evaluation. It supports variable inference budgets and reports 4–8 function evaluations for image and video generation, including text-to-video/audio, while claiming better diversity than methods based on adversarial or variational score distillation.
The multi-step prediction formulation and claimed preservation of video diversity address real weaknesses in fast generative-model distillation, but the abstract provides no quantitative comparisons or details sufficient to distinguish a major advance from another distillation variant.
-
OSReward: Instituting Standardized Evaluation for Cross-Platform Computer-Use Reward Models
OSReward benchmarks vision-language models used to judge whether computer-use agents actually completed tasks across different platforms. It finds that even strong judges are systematically too lenient, while affordable open models perform poorly; the authors release hard and fine-grained test sets plus OS-Shepherd models trained on 100K annotated judgments that reportedly approach commercial judge quality at 30–60x lower cost.
This is more than a routine benchmark because it exposes a potentially important failure mode in agent evaluation and offers low-cost reward models, but the abstract gives few quantitative details to substantiate the broad performance and cost claims.
-
CAST: Game Solvers as Turn-Level Teachers for LLM Agents
CAST uses a game solver’s change in estimated state value after each action to provide dense, turn-level credit during reinforcement learning, rather than relying only on a final success reward. The authors show that these scalar “solver advantages” can approximate on-policy distillation without requiring teacher logits, and report improvements across three games plus zero-shot gains on ALFWorld and WebShop.
The scalar-value formulation of solver-guided credit assignment is a useful and plausibly general alternative to sparse RLVR or logit-level distillation, but the abstract gives no quantitative gains and the evidence is concentrated in game environments.
-
MerchantBench: Benchmarking LLM Agents for Long-Term Coherence in E-Commerce Operations
MerchantBench evaluates LLM agents in a 365-day simulated e-commerce business where sourcing, pricing, cash management, and delayed feedback interact over time. Across eight models and two agent frameworks, the best system reached only 27.3% of the final net assets achieved by human participants, exposing substantial weaknesses in maintaining coherent strategies over long horizons.
This is more than a standard agent benchmark because it measures persistent, financially coupled decision-making with delayed consequences, and the large human–LLM gap is a useful warning about current long-horizon agent capability, though the evidence is limited to a simulated domain and 48 runs.
-
Lost in Compression: A Controlled Cross-Lingual Audit of Extractive Prompt Compressors
This paper evaluates four learned and four deterministic prompt compressors across ten languages, five scripts, and eleven target LLMs using parallel data and tokenizer-matched budgets. English-trained learned compressors lose much more useful context in non-English languages—especially at aggressive rates—while multilingual training avoids the transfer gap; in some languages, aggressive compression performs no better than providing no context. Translate-then-compress can match or outperform native compression at about half the token cost in several tested languages.
The broad controlled audit reveals a practically important, supervision-dependent cross-lingual failure mode in prompt compression, supported by 250,000 calls, though it is primarily an empirical audit rather than a new compression method.
-
Instruction-Tuned Language Models Cannot Sample from Distributions They Can Describe
The paper argues that instruction-tuned language models often fail to produce genuine independent samples of a persona’s response distribution: repeated calls collapse onto the same answer, despite the model being able to describe the distribution accurately. Across three model families, the authors link this behavior to concentrated logits and alignment training, and show that directly asking for the distribution is more accurate than aggregating persona samples; their prompt-perturbation method improves sampling-based estimates by 21% at no extra call cost.
The claimed KNOWS/DOES split and evidence that instruction tuning destroys sampling diversity challenge a common assumption behind silicon sampling, but the abstract provides limited quantitative and cross-task detail beyond one benchmark and three model families.
-
IRIS: Reusable Identity Representations from Frozen LLMs for Entity Alignment
IRIS uses a frozen LLM to elicit a reusable, identity-focused representation for each knowledge-graph entity. Entities can then be aligned by direct similarity in a shared space, without constructing representations for each KG pair or rerunning the LLM for candidate comparisons; it reports near-perfect Hits@1 on four benchmarks.
The reusable, candidate-independent identity-space formulation and unusually high reported accuracy are potentially important, but the abstract gives too little detail about prompts, baselines, leakage, and robustness to trust the near-perfect results as a major advance.
-
Entropy-Smooth Convex Optimization Cannot Be Accelerated
The paper proves that no first-order method can generally achieve accelerated convergence for convex functions that are smooth relative to negative entropy on the probability simplex: the rate is at best Ω(L/T) when the dimension is sufficiently large. This contrasts with ordinary ℓ1-smoothness, where acceleration is available, and extends the lower bound to quantum states using von Neumann entropy.
It gives a non-pathological, structure-specific impossibility result showing that entropy geometry itself can block acceleration, but the result is primarily a specialized theoretical lower bound with a dimension condition rather than a broad practical advance.
-
Language Models Agree With Each Other, Not With Readers
The paper compares sentence selections from 18 language-model arms with 2,523 people’s unsolicited reading highlights across 120 web documents, using a calibrated overlap measure that controls for selection depth and length. Models agree with one another substantially more than readers agree with one another, while none matches readers better than another reader does; the pattern persists across vendors, model sizes, prompts, procedures, and four later models tested out of sample.
This is a carefully controlled and potentially important challenge to the idea that model agreement reflects human-like consensus, but the result is primarily a measurement and behavioral characterization rather than a demonstrated capability or method advance.
-
BM25 Wins at Scale: A Scaling Study of Retrieval-Augmented Generation Paradigms
This study compares lexical, dense, graph-based, and agentic RAG across 28 nested corpus sizes, while keeping the questions and relevant/adversarial documents fixed. Agentic search works best on small corpora, but its sequential search cost grows sharply; above about 10 million corpus tokens, BM25 becomes more accurate and remains the strongest low-cost option, with nearly a 20-point advantage at the largest scale. The main finding is that corpus growth favors global lexical ranking, while agentic reasoning is better used after retrieval than instead of it.
The scale-dependent crossover and strong large-corpus performance of BM25 are a useful counterpoint to current enthusiasm for agentic and dense RAG, though this is still a controlled comparative study using one reader model and judging protocol rather than a new method.
-
SkillRise: Agentic Reinforcement Learning for Cross-Task Skill Evolution
SkillRise trains an LLM agent to solve a sequence of related tasks while maintaining and refining a shared skill document. It uses separate credit assignment for task success and downstream skill curation, and reports 2.3–8.5 point Pass@1 gains over the strongest baselines on ALFWorld, WebShop, and ScienceWorld, with improved performance as the task sequence grows. The main novelty is treating cross-task skill evolution as a single reinforcement-learning process rather than a separate extraction, retrieval, and execution pipeline.
The cross-task, test-time skill accumulation setup is a meaningful and potentially reusable direction, but the abstract provides limited detail about task-scale, generalization, and whether the gains exceed strong memory or prompting baselines.
-
One Future, Every Robot: Label-Efficient Collective-State Prediction with Decentralized JEPA
The paper trains decentralized robots to predict the same latent future of their collective using only local histories and bounded neighbor messages, without explicitly enforcing agreement or pooling predictions at deployment. In an independent replication, this approach reportedly improves both inter-robot agreement and prediction accuracy, especially under changes in swarm size and communication topology, while action-conditioned tests suggest the latent remains useful for local decision-making.
The potentially important result is that shared latent-future prediction can produce aligned, useful decentralized representations without consensus loss, with substantial robustness under topology and swarm-size shifts, but the abstract gives too little detail about tasks, robot scale, and baselines to justify a stronger verdict.
-
Beyond Geometric Complementarity: Coherent Overlap in Sparse Mixture-of-Experts Routing
This paper argues that MoE experts do not need to occupy separate representation directions to be functionally complementary. Across six architectures and a 39-cell controlled analysis, routed experts overlap geometrically but still explain token-specific residuals better than alternatives, and adding later experts often improves prediction; a Top-2 training study also beats Top-1 in all three seeds. It introduces the term “coherent overlap” and separates route quality, geometric separation, contextual interaction, and functional redundancy.
The useful contribution is a careful, intervention-backed challenge to the assumption that MoE gains require geometrically disjoint experts, though the results mainly refine understanding rather than establish a major new capability or training method.
-
GPTQ-2D: Cubic-Time Two-Sided Adaptive Rounding
GPTQ-2D extends GPTQ-style adaptive rounding from one-sided to two-sided matrix objectives, where the quadratic metric has a Kronecker-product structure. By rounding entries anti-diagonal by anti-diagonal, it produces the same result as naive vectorized GPTQ while reducing runtime from quartic to cubic in the matrix dimension and enabling parallel rounding within each anti-diagonal.
The exact-output-preserving cubic-time algorithm for two-sided adaptive rounding is a meaningful structural improvement, though its practical importance and breadth are unclear from the abstract.
-
Context Is King: How In-Context Specification Shapes the Geometry of Concepts
The paper argues that language models construct the conceptual geometry they use from the current in-context specification, rather than simply retrieving a fixed pretrained representation. Across Gemma and Qwen models, prompts can make the same tokens form either cycles or branching trees—even arbitrary meaningless tokens—and activation patching suggests this imposed structure causally guides successor answers. Larger models use the structure more cleanly and can override strong pretrained priors, while smaller models form weaker maps but do not consistently rely on them.
The combination of context-dependent topology, arbitrary-token constructions, cross-model scaling effects, and causal activation-patching evidence is a genuinely non-obvious contribution to understanding in-context learning, but the abstract does not establish how broad or durable the effect is beyond the tested model families and tasks.
-
LOCKS: Page-Local Compact Key Summaries for Efficient Long-Context Decoding
LOCKS compresses each page of the KV cache with its own low-rank spectral summary, using the summary to estimate attention mass and select only the most relevant pages without reading their full keys or values. On several long-context QA, retrieval, and reasoning tests, it reportedly stays close to full-cache quality while reading 10–25× less KV data; at million-token context on an H200, it halves decode latency and attends to about 2% of tokens.
The page-local summaries and candidate-free page selection are a substantive alternative to global KV compression or eviction, with unusually strong long-context and systems results, but the claims need verification beyond the abstract and reported benchmarks.
-
What Can Latent World Models Know? Physical Parameter Identifiability in Multimodal Predictive Representations
The paper studies which physical properties latent world-model representations actually encode, rather than assuming that prediction automatically produces a complete physical model. Controlled experiments show that observability is necessary but insufficient: prediction targets determine what is retained—contact stiffness appears when touch is predicted, while drag remains poorly represented despite being recoverable from observations; similar effects replicate on real-robot data across scaling curves.
The certificate-gated identifiability analysis and the finding that more data cannot recover parameters omitted by the objective are genuinely useful for world-model design, but the evidence is still limited to a synthetic environment and one robotics dataset, so it falls short of a strong recommendation.
-
Bunraku: Turning a Single Illustration into an Editable Live2D Character
Bunraku converts one 2D character illustration into an editable Live2D asset: ordered RGBA layers with completed occluded regions, meshes, and parameterized deformation poses. It combines layered diffusion with a joint transformer over all layer vertices, finding that joint prediction improves coherent motion while simply scaling the model does not; it also enables text-based retexturing while preserving the animation structure. The paper introduces a dataset and benchmark, but reports its main deformation result on only 50 held-out characters.
The end-to-end generation of a production-oriented, editable animation representation—and the finding that cross-layer joint prediction matters more than 112× scaling—looks genuinely novel, but the limited quantitative evidence makes this worth a closer look rather than a strong recommendation.
-
Thinking Once Is Enough: Intermediate-Layer Evidence Routing for High-Resolution VQA
Thinking-Once argues that high-resolution VQA models often encode the needed fine-grained visual evidence already, but lose it in later layers rather than needing repeated image inspection. Its training-free method identifies and routes question-relevant intermediate-layer evidence forward in a single visual pass, improving three HR-VQA benchmarks by roughly 2.7–3.1 points on average while reducing peak memory by about 4 GB; on Qwen2.5-VL-7B it reports larger gains and a 97.2% inference-time reduction versus DeepScan.
The combination of a single-pass intermediate-layer evidence-routing mechanism with substantial reported speed, memory, and accuracy gains is genuinely interesting, though the abstract does not establish how robust the gains are across settings or whether the routing is more than a model-specific intervention.
-
WitCert: Sound Runtime Risk Observability and Gating for KV-Cache Quantization
WitCert adds a runtime monitor for KV-cache quantization that estimates, per layer, attention head, and decoding step, an upper bound on the change from exact attention. It uses a deterministic bound that applies broadly and a tighter probabilistic certificate for dithered INT8, then uses the estimates to selectively gate risky compression; on hard RULER tasks this reportedly raises raw FP8 quality from 22.8 to 79.7 while staying within 0.8 points of uncompressed quality, and the certified INT8 setup serves 1.88× more KV tokens at equal memory.
The combination of sound per-request risk observability, adaptive-safe certificates, runtime repair, and machine-checked proofs is substantially more than another KV-quantization tweak, though the strongest guarantees are restricted and the large gains need verification from the released artifacts.
-
Understanding Is Done Early: A Depth Division of Labor in Large Language Models and Its Use for Unbounded-Context Memory
The paper finds that lower and middle Transformer layers mainly build context representations, while upper layers make them query-specific predictions. CoMem exploits this by caching intermediate-layer residuals rather than full KV states, retrieving a fixed-size set for each query, and rerunning only the upper layers; this makes read-time compute and memory independent of the amount of stored context. On long-context benchmarks it reports large gains over full-context KV retrieval, while an 8B-model control reduces 128k memory from 89.36 GB to 18.26 GB and speeds prefill by 7.83×, though bounded retrieval introduces some compression loss.
Caching along the layer axis instead of only compressing tokens is a genuinely interesting memory design with substantial reported efficiency and long-context gains, but the abstract does not establish how broadly it beats strong alternatives or how much of the benchmark advantage comes from training and retrieval choices.
-
TokTier: Exact Stateful CPU+GPU Tokenization for Agentic LLM Serving
TokTier avoids re-tokenizing entire prompts in agentic LLM sessions by incrementally repairing tokenization near appended text, while falling back to exact full tokenization when boundaries are uncertain; it also provides GPU-based tokenization for requests without reusable state. Across 17 tokenizer families and large-scale differential testing, it reports zero divergences, sub-millisecond processing for very long continuations, and 16–34% median TTFT reductions in vLLM under production-like bursts.
The combination of exact stateful incremental tokenization, GPU acceleration, and production-scale validation addresses a real serving bottleneck with unusually large reported latency and throughput gains, though the evidence is still primarily from the authors’ system and benchmarks rather than independent reproduction.
-
The Gate, Not the Cache: Gate Provenance Bounds the Closed-Loop Reliability of Training-Free VLA Token Skipping
This paper shows that training-free token skipping in vision-language-action models can fail in closed-loop control because the next gate is computed from an already-skipped forward, causing errors to compound over time. It proposes an unconditional dense “actuation-slack refresh” during execution to produce a clean gate and KV cache for the next step; across two VLA policies, four LIBERO suites, four SIMPLER tasks, and a physical robot, this restores near-dense performance while reducing serving latency by 18–22%.
The useful new insight is that gate provenance, rather than the particular skipping or caching mechanism, determines closed-loop reliability, supported by cross-method experiments and physical-robot latency results, though the evaluation scope is still limited.
-
Grasp Execution Without a Planner: Configuration-Space Grasp Distance Fields with Certified Safety & Guaranteed Quality
The paper introduces Grasp Distance Fields, which represent many feasible arm-hand grasp configurations as a smooth field and use its gradient directly as a feedback controller, avoiding explicit trajectory planning and replanning. A safety-filtering QP enforces collision and workspace constraints, while contact-mode logic and a wrench-quality constraint preserve grasp quality; experiments on a fixed-base arm and a Unitree G1 achieved 46/50 successful cluttered-object grasps with moving obstacles, 94% median retained quality margin, and 0.09 ms QP solves.
This is a substantial and unusually integrated planner-free grasp-execution approach with formal safety and quality guarantees plus real-robot evidence, but the evaluation is still limited to 50 objects and one underactuated hand, so its generality is not yet established.
-
Round-Trip Consistency: Bidirectional Diffusion Models Can Predict Their Own Rollout Errors
The paper trains one latent diffusion model to predict dynamics in either time direction, then uses a forward rollout followed by a backward rollout as a ground-truth-free estimate of accumulated prediction error. The round-trip discrepancy strongly correlates with actual error on MHD data, detects an out-of-distribution vortex, supports calibrated uncertainty-based intervention, and transfers to video and Navier–Stokes settings; bidirectionality also improves performance and provides an inverse solver.
The combination of bidirectional generative dynamics and round-trip error monitoring is a genuinely useful, broadly relevant idea with unusually strong quantitative validation, but its reliability may depend on approximate reversibility and the evidence is still concentrated in a few modeled domains.
-
When May a Model Replace the Experiment? Audits, Licenses, and the Price of Trust in Surrogate-Driven Design
This paper studies when a surrogate model can safely stand in for expensive experiments or full evaluations during design and search. It argues that predictive accuracy such as R² is not enough: selecting the best of many candidates amplifies overprediction, while certified claims must ultimately rely on true evaluations; selection-aware audits can estimate whether the surrogate preserves rankings and reduce certified evaluation costs. The theory is tested across 432 surrogate fits and three fully ground-truthed tasks, with audits correlating strongly with deployed performance and a reported 25× cost reduction.
The selection-aware trust framework and proposed audit criterion address a real failure mode in surrogate-driven optimization with substantial theoretical and empirical support, though the core concerns overlap with established ideas about selection bias, Goodhart effects, and model-based optimization.
-
Cryoaerosolization Enables Scalable Vitrification-Based Cell Cryopreservation
The authors develop a cryopreservation system that aerosolizes cell-containing droplets into liquid nitrogen, combining extremely rapid cooling and warming with much higher throughput than prior vitrification methods. Across human iPSCs, fibroblasts, and pig red blood cells, it reports over 90% post-thaw viability or recovery at relatively low cryoprotectant concentrations, with preserved colony-forming ability in iPSCs and processing rates above 100 mL/hour.
This is a genuinely interesting scale-up of vitrification that appears to address the key throughput-versus-viability tradeoff in cell preservation, though the abstract does not establish long-term storage, product-scale reproducibility, or clinical compatibility.
-
Combinatorial Control of Corticospinal Axon Growth by Retinoic Acid Receptors
This study identifies the retinoic acid receptors RARA and RARG as cooperative regulators of axon regeneration in the adult mammalian CNS. Expressing both, but not either alone, promoted neurite growth in cultured neurons and induced sprouting, long-tract regeneration, and improved motor behavior after spinal or corticospinal injury; single-cell and binding analyses suggest that the receptors control complementary enhancer and promoter programs. The cooperation appears specific to CNS neurons, since RARG alone was sufficient in peripheral sensory neurons.
The combination-specific, compartmentalized mechanism for CNS axon regeneration is genuinely interesting and supported by multiple in vitro, in vivo, transcriptomic, and chromatin assays, but the abstract lacks quantitative effect sizes and independent validation needed for a strong verdict.
-
Carbonyl stress primes the metastable aging brain for Alzheimer's disease
Using spatial transcriptomics, proteomics, lipidomics, and single-nucleus data, the authors argue that Alzheimer’s inflammation is strongest in white matter rather than directly around amyloid plaques, and is associated with lipid storage, myelin loss, carbonyl stress, and iron-handling changes. An oligodendrocyte lipid-droplet program tracked cognitive decline independently of amyloid and tau, leading them to propose that amyloid and tau catalyze an already primed, metastable glial state rather than fully initiating the inflammatory response.
The cross-modal, spatially resolved evidence supports a non-obvious shift from plaque-centered inflammation toward a white-matter lipid/carbonyl-stress model, but the causal claim remains largely inferential and observational.
-
Memory encoding reprograms neuronal transcriptional responses via durable chromatin remodeling
Using genetic tagging and single-cell multiomic sequencing, the authors tracked mouse medial prefrontal cortex neurons activated during fear-memory formation over four weeks. When those same neurons were later reactivated by memory recall, they showed altered transcriptional responses compared with otherwise similar neurons, including reduced responses to excitation and changes in regulatory programs. The result suggests that memory-related chromatin remodeling does more than preserve a molecular trace: it changes how engram neurons respond to future experiences, potentially reducing interference between memories.
The compelling result is evidence that memory storage produces a slow, cell-specific change in the future transcriptional response of engram neurons, but the abstract provides correlational multiomic evidence without a causal test linking the remodeling to reduced memory interference.
-
Striatal acetylcholine enables latent-state creation during reversal learning
The study argues that dorsomedial striatal acetylcholine helps animals create new latent states when outcomes are poorly predicted during reversal learning. In mice, suppressing phasic acetylcholine impaired reversal, while enhancing it on unrewarded trials improved performance; importantly, an explicit cue signaling reward omission reduced the deficit from acetylcholine knockdown, as predicted by a circuit-constrained reinforcement-learning model.
This provides a relatively specific computational and causal account of striatal acetylcholine—state construction under ambiguous reward omission—with an out-of-sample behavioral prediction, though the evidence is still limited to one reversal-learning paradigm in mice.
-
INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models
INTACT trains a latent world model to map a desired visual state change directly to an action distribution, avoiding test-time search for control. It uses shared, structurally matched representations for observed transitions and future goals, and reports high success on four LeWM tasks with 2.9–5.5 ms direct inference; optional sampling-based refinement uses far fewer candidates than standard CEM. The main novelty is integrating intent representation learning and action prediction without assuming pointwise latent matching or globally linear dynamics.
The search-free intent-to-action formulation and shared action-law structure are plausibly meaningful, but the evidence is limited to four benchmark tasks and the abstract does not establish broad real-world generality or a decisive advantage over strong alternatives.
-
A Frozen Pixel-Space Diffusion Model Can Guide Itself with Its Own Samples
The paper adds a small prediction head to an already-trained pixel-space diffusion transformer and uses the difference between an intermediate-layer prediction and the final prediction to guide sampling. The head can be trained entirely on the model’s own generated images, with less than 1% of the compute needed for full retraining; on ImageNet pixel-diffusion models, this reportedly cuts FID by more than 50% in some unguided settings and improves CFG baselines as well.
The self-guidance mechanism and finding that synthetic samples outperform real images for training the adapter are non-obvious and potentially useful, but the evidence is limited to ImageNet and a narrow family of pixel-space diffusion models, so the broader impact is unclear.
-
DreamTraj: Generating 6-DoF Object Trajectories by Reading Unrendered Video Diffusion Latents
DreamTraj predicts an object’s future 6-DoF manipulation trajectory from one RGB image and a natural-language instruction. Instead of generating video and extracting motion afterward, it decodes motion-related information from intermediate features of a frozen image-to-video diffusion model, trained with a new dataset of 5,038 language-labeled trajectories. The approach reportedly outperforms predictors with multi-frame or privileged inputs and is 4.6× faster than generate-then-extract pipelines, though the abstract gives no quantitative accuracy results.
The direct use of intermediate video-diffusion representations as a trajectory predictor is a genuinely interesting robotics direction, but the claimed superiority is difficult to assess without numerical results, baselines, or evidence that the latent representations generalize beyond the dataset.
-
Interpretable MEG Decoding of Perceived Speech: Cortical Sources and the Stimulus Features That Drive Retrieval
The authors redesign a MEG-to-speech retrieval model so its spatial operations align with the 3D sensor geometry and its branches map more directly to putative cortical sources. On a 1,005-way speech retrieval task, it retains strong performance while using about 20 times fewer decoder parameters, and interventions suggest that narrative structure, temporal detail, silence, intensity, vowels, and acoustic onsets drive much of the recoverable signal. The model’s source-aligned weights also show plausible speech-network localization and a left-hemisphere bias for higher-frequency rhythms.
This is more than a benchmark gain because it combines parameter-efficient MEG decoding with source-aligned interpretability and feature-level interventions, but the core retrieval setup and architectural changes remain incremental rather than a clearly new decoding capability.
-
MPIE-Bench: Benchmarking Anatomically Plausible Multi-Person Interaction Editing
MPIE-Bench evaluates whether image editors can place multiple named people into physical interactions without fused limbs, missing body parts, or body interpenetration. It combines 2,500 video-mined editing examples with mesh-based scores for anatomical completeness and contact geometry; across ten editors, these scores remain moderate while VLM checklist judges exceed 0.95, and human ratings favor the proposed metrics.
This is primarily a benchmark, but the substantial gap between highly favorable VLM judgments and visibly poor mesh-based interaction quality is a non-obvious finding with practical implications for evaluating multi-person image editing.
-
Visual prompt engineering for video models
The paper proposes modifying a task image or video—such as converting a schematic scene into a photorealistic one—as a form of prompt engineering for video reasoning models. Across several visual reasoning tasks, these transformations reportedly improve performance and can outperform text prompting and test-time scaling while using relatively little compute.
The idea of changing visual presentation to elicit stronger video reasoning is a genuinely interesting prompting direction, but the abstract gives no quantitative results, task breadth, or evidence that the gains reflect improved reasoning rather than advantageous image transformations.
-
ShadowDancer: Teaching Video World Models Any Action by Learning Unified Dynamics Representations from a Video and Its Shadow
ShadowDancer learns reusable action representations from pairs of videos showing the same motion or dynamics under different appearances. By predicting one appearance-conditioned video from another, it aims to discard scene-specific visual details and transfer demonstrated actions to new environments without action labels or fine-tuning. The authors report better action transfer and longer rollouts than latent-action and interactive-world-model baselines, with an 86% blinded preference in rollout comparisons.
The shadow-pair/cross-shadow formulation is a potentially useful new way to separate dynamics from appearance, but the abstract gives few concrete experimental details and the headline 86% result is not enough to establish broad capability or robustness.
-
DistillAlign: Coordinating Mode Covering and Mode Seeking in Autoregressive Video Distillation
The paper argues that autoregressive video distillation should preserve the teacher’s mode coverage during both initialization and refinement, rather than optimizing only for visual quality or mode-seeking fidelity. It proposes measuring precision and coverage in a shared latent space and combining DMD with a consistency-distillation constraint; the resulting student reportedly improves quality, diversity, and coverage, even beating models distilled from a larger teacher.
The distributional analysis and explicit mode-covering objective are a useful, relevant reframing for video distillation, and the claim of a 1.3B-teacher model outperforming 14B-teacher baselines merits checking, though the abstract provides no quantitative results and the method combines largely established distillation ideas.
-
Can AI agents conduct open-ended AI research? Early evidence from two case studies
The paper proposes “shadow evaluations”: AI agents attempt to solve the central research question from an unpublished paper, and the original authors judge the result. In two six-day NeurIPS case studies, frontier agents completed the coding and experiments without help but failed to make substantial research progress, with recurring problems in judgment, creativity, backtracking, resource use, and instruction following.
The evaluation setup and diagnosis of open-ended research failures are useful and plausibly novel, but evidence from only two unpublished papers is too limited for strong conclusions about AI research automation.
-
Accurate structural modeling of chemically diverse molecular interfaces with Vilya-2
Vilya-2 extends an all-atom diffusion-transformer model from individual molecules to protein–molecule complexes, targeting peptides with non-canonical residues, macrocycles, and other difficult chemistries. It reportedly predicts 59.1% of peptide interfaces below 2 Å backbone RMSD, while also performing strongly on small-molecule docking and generalizing to larger or chemically novel macrocycles and miniproteins; the model can additionally be fine-tuned for hit discovery.
The unified all-atom treatment of chemically diverse peptides and protein interfaces, together with the reported generalization beyond training chemistry, is potentially a substantial advance, but the abstract provides limited experimental detail and several broad state-of-the-art claims that need verification.
-
Specula: Scaling formal specifications for autonomous model checking of system code
Specula uses coding agents to autonomously build TLA+ models and invariants for large system-code projects, then runs model checking to find bugs. In experiments across 48 open-source projects, it reportedly found 249 bugs, including bugs that existing tools have difficulty detecting, while using iterative self-improvement to reduce hallucinated or low-quality specifications.
Autonomously generating useful formal models for real system code at this scale could substantially lower the barrier to formal verification, but the abstract gives too little detail about baseline comparisons, bug validation, and how many findings are genuinely new.
-
Schrödinger's Cat: Probabilistic Representation and Prediction of Potential Scene Kinematics
The paper introduces GARFIELD, a probabilistic latent representation for the distribution of possible future scene motions from partial visual observations and optional constraints. Unlike video predictors or trajectory samplers, it supports joint trajectory generation, localized uncertainty estimates, and direct density evaluation; the authors report competitive planning performance while sampling 97× faster and estimating densities up to 100× faster than Monte Carlo alternatives.
The explicit, efficiently queryable representation of scene-level future-motion distributions is a meaningful direction with potentially large planning benefits, but the abstract gives limited detail about task breadth, baselines, and experimental scale, so it falls short of a strong recommendation.
-
Incast-Free MoE Rate-Based Scheduling
The paper identifies a previously unrecognized incast problem caused by round-robin routing of Mixture-of-Experts traffic, where synchronized expert communication can exponentially overload parts of the network. It proposes a proactive, rate-based fair scheduler, sketches NIC support, and reports from simulations that it avoids incast, keeps links nearly fully utilized, and lowers collective completion time.
The claimed interaction between MoE traffic and exponential incast is potentially important and non-obvious, but the abstract provides only simulation evidence and no quantitative gains, real-cluster deployment, or comparison details to justify a strong recommendation.
-
HiFloat4 Format for End-To-End Reinforcement Learning Post-Training of Large Language Models
The paper studies fully 4-bit reinforcement-learning post-training for language models, including both rollout and optimization. It finds that rollout activation underflow—not primarily training-side quantization—is the main source of degradation, and proposes Rollout-ResQ, a sparse residual correction for rollout matmuls, combined with a hierarchical HiFloat4 format. On two Qwen models, this reduces the accuracy gap to BF16 from 4.9% to 1.1%, while the same approach with MXFP4 leaves a larger 5.3% gap.
The end-to-end 4-bit RL setting and rollout-specific diagnosis are genuinely interesting, with substantial reported gains, but evidence is limited to two models and the practical hardware and training-cost implications are not established in the abstract.
-
Amortized Moment Matching for Visual Generation
The paper proposes learning distributional moments with neural networks and uses them as an amortized, matrix-free training loss for diffusion and one-step image generators. The resulting Amortized Fréchet Distance is reported to improve stability and one-step ImageNet generation, and to substantially improve instruction following in text-to-image models— reportedly outperforming a multi-step FLUX.2 4B teacher on GenEval while matching it on PickScore.
The conditional, neural formulation of moment matching and the claim that a one-step student can beat its multi-step teacher on instruction following are genuinely interesting, but the abstract gives no quantitative margins or broad evidence, so it falls short of a strong recommendation.
-
SERPO: Self-Evolving Rubric Policy Optimization for Open-Ended Test-Time Reinforcement Learning
SERPO is a test-time reinforcement-learning method for open-ended language tasks, where there is no single correct answer to vote on. It iteratively evolves response archives, task-specific evaluation rubrics, rubric-based reward signals, and the model policy itself, using only the model’s own outputs. Across six benchmarks and two model setups, it reports large gains on HealthBench and ResearchQA and some out-of-domain transfer, though the abstract does not establish how much of the improvement comes from evaluator reliability or benchmark-specific effects.
The closed-loop co-evolution of responses, rubrics, and policy is a meaningful approach to self-supervised open-ended test-time learning, and the reported 8-point macro-average gain is substantial, but the evidence is still abstract-level and relies on self-generated evaluation signals that warrant closer scrutiny.
-
RIPPLE: Generating Multi-Channel Phase, Not Recovering It
RIPPLE treats multi-channel phase as something a generative model must learn, rather than reconstructing it independently after generating magnitudes. It uses Griffin–Lim as a phase prior and refines phase with rectified flow and an explicit inter-channel coherence loss, improving spatial-audio and seismic translation; in the seismic setting, S-wave polarization error falls from about the 57.3° random baseline to 33.8°.
The paper targets a real blind spot in magnitude-focused audio and signal-generation pipelines and reports a substantial cross-channel physical-structure improvement across two unrelated domains, but the abstract provides limited quantitative breadth and does not establish a broadly transformative capability.
-
Deep Learning for Accelerated Long-Horizon Forecasting of Multicomponent Multiphase Microstructure Evolution in High-Entropy Alloys
The paper trains an autoencoder–graph convolution–LSTM surrogate to predict long-term evolution of multicomponent, multiphase alloy microstructures instead of repeatedly solving expensive phase-field equations. Trained on small domains with one nominal AlCrFeNi composition, it reportedly generalizes to larger domains, different compositions, multiple precipitates, and merging or splitting interactions over up to 3 million timesteps, with claimed speedups of roughly 7,200–62,300×.
Zero-shot transfer across domain sizes, compositions, and complex precipitate interactions with very large claimed speedups is notable, but the abstract gives limited quantitative accuracy and the method is still a specialized combination of established surrogate-modeling components.
-
Rescene: band-limited stochastic forcing turns a frozen neural weather operator into a climate emulator
Rescene adds a small wrapper to a frozen neural weather model: a deterministic correction steers long-term behavior toward seasonal climatology, while a band-limited stochastic head restores unresolved variability. In long free-running tests, it reportedly remains stable for 100 years, recovers much of the observed blocking and daily variability, and sustains small-scale weather structure even though noise is injected only at large scales.
The combination of frozen-model stabilization with large-scale stochastic forcing producing realistic small-scale variability is genuinely interesting, but the abstract provides limited information about independent validation, baselines, and whether the impressive climate statistics generalize beyond ERA5-trained tests.
-
Tycho: Active Abstraction with Programmatic World Models for ARC-AGI-3
Tycho is a coding-agent system for interactive ARC-AGI-3 tasks that builds, tests, repairs, and sometimes bypasses executable game models while learning the rules through costly actions. On all 25 public games and 183 levels, frontier-model policies achieved perfect completion and up to 100 RHAE; Opus 5 reportedly used 61% fewer scored actions than aggregate human baselines. The notable finding is that accurate transition models alone are insufficient: agents must also decide when modeling is worth the interaction cost and whether the model captures the task objective.
The active-abstraction framing and distinction between simulating dynamics and discovering objectives are genuinely interesting, but the evidence is limited to one public benchmark run per policy with frontier models and leaves generalization beyond ARC-AGI-3 uncertain.
-
WaiT for the Signal: Simple Frequency-Aware Flow-Matching
WaiT uses a lossless wavelet decomposition to separate coarse and fine image frequencies during flow-matching generation. High-frequency components remain noise until coarse structure has formed, then participate in joint refinement; the authors report up to 50% lower sampling compute, improved native-resolution texture quality, and strong ImageNet, OpenImages, and video results.
The frequency-staged generative process is a genuinely interesting alternative to treating all pixels uniformly, but the abstract provides limited detail and makes unusually broad state-of-the-art claims that need careful verification.
-
Fast Rates for Swap-Agnostic Learning of Proper Losses
This paper introduces an offline and online learner for swap-agnostic learning with proper losses, avoiding a separate learning problem for every prediction value. For finite hypothesis classes, it achieves rates such as $\widetilde{O}((\log |H|/m)^{2/3})$ offline and $\widetilde{O}(T^{1/3}(\log |H|)^{2/3})$ online for smooth losses, improving substantially over rates inherited from swap-omniprediction; the bounds are claimed tight up to logarithmic factors. The key technique reduces the problem to second-order multicalibration using Blackwell approachability with a variance correction.
The improvement from a $T^{2/3}$ to a $T^{1/3}$ online rate, together with a new multicalibration-based reduction and tightness claims, is a meaningful theoretical advance, though its scope is specialized and the abstract provides limited empirical or theorem detail.
-
Mixture-of-Translators: Translating KV Caches Across Heterogeneous Large Language Models
The paper proposes translating a context’s KV cache from one language model into another, allowing heterogeneous models or agents to reuse work without re-prefilling the same context. Its mixture of translator modules and a target-side correction loss reportedly preserves much of the original model’s QA and long-context performance across Qwen2.5, GPT-2, and OPT, including 96.3% of direct-context quality in one cache-augmented generation setting.
Cross-architecture KV-cache reuse is a genuinely interesting systems direction with potentially large benefits for multi-agent and long-context inference, but the abstract gives limited baseline, task, and efficiency detail and the reported QA results are difficult to judge without direct-context comparisons.
-
Curriculum Matters: Data-Efficient Relational PFN Pretraining with Synthetic Data
This paper studies how synthetic-data curricula affect Prior-Data Fitted Network pretraining for tabular and relational prediction. A curriculum that gradually increases schema complexity achieves 0.703 average ROC-AUC with about 45× fewer single-table synthetic datasets, while a relational curriculum recovers 88% of a prior system’s performance using about 220× less relational data; notably, the single-table-trained model nearly matches the relationally trained model without relational adaptation. The main new finding is that training order and synthetic diversity may matter substantially more than the sheer volume or exact relational generator.
The large synthetic-data savings and unexpected transfer of relational reasoning from single-table pretraining are genuinely interesting, but the evidence is limited by abstract-level reporting and comparisons against previously reported baselines rather than clearly controlled, broad experiments.
-
Data Turnstile: A Scalable Open Framework for Function-Calling Data Generation
Data Turnstile is an open framework that generates validated, multi-turn function-calling training conversations from API specifications. Fine-tuning small Qwen3 models on its synthetic data substantially improves tool-use performance: a 0.6B model reaches 75.9% on BFCL, and 0.6B/1.7B models outperform or approach much larger baselines on the reported Telecom split of τ²-bench. The main contribution is a scalable data-generation and validation pipeline rather than a new model architecture.
The large gains and apparent ability to make sub-billion-parameter models competitive with models up to 32B are notable, but the evidence is limited to two benchmarks and one τ²-bench domain, while the underlying stepwise generation and validation approach may be an incremental combination of known techniques.
-
ParaASR: Multi-Token Prediction for Fast and Long-Context LLM-Based Speech Recognition
ParaASR applies multi-token prediction to an audio-encoder/LLM ASR system, training five future-token branches and verifying proposed six-token continuations so that decoding remains exact while requiring fewer decoder passes. It reports about five accepted tokens per step, support for 32K context and 30-minute single-pass transcription, and very low reported real-time factors while maintaining low error rates across Chinese, English, and long-form tests.
The combination of acoustic conditioning with verified multi-token decoding is a meaningful and plausibly broadly useful speedup for LLM-based ASR, but the abstract lacks hardware, baseline, and apples-to-apples details needed to judge how substantial the reported latency gain really is.
-
Parameter-Free Heavy-Tailed Bandits
This paper studies heavy-tailed bandits when the tail exponent and moment bound are both unknown, so the algorithm cannot rely on reliable estimates of rare-event behavior. It proves a lower-bound trade-off for adapting to an unknown moment bound, gives a scheduled-exploration algorithm that matches this frontier up to logarithmic factors, and shows the same method adapts to an unknown tail exponent while establishing that uniform sublinear regret over all positive exponents is impossible.
It appears to resolve a stated open problem with both a sharp impossibility characterization and a matching parameter-free algorithm, though the result is specialized theoretical progress and the abstract provides no quantitative details beyond the headline guarantees.
-
Lost in Compaction: Evaluating Side-Constraint Loss under Context Compaction
This paper studies whether long-context compaction preserves user instructions that should remain active throughout a session, such as not deleting emails without confirmation. Across chat, agent trajectories, and long-horizon research tasks, existing compactors retained only 17% of these constraints on average, while a plug-in constraint extractor reportedly raised retention above 90% without changing the underlying model or compactor.
It identifies a concrete and potentially important failure mode in long-running LLM agents, with unexpectedly poor retention and a simple mitigation, but the abstract provides limited detail about evaluation breadth and whether the proposed fix preserves task quality beyond constraint retention.
-
When Does On-Policy Interaction Help? Representational Tradeoffs in Value-Based Imitation Learning
This paper argues that interactive imitation learning can succeed even when the learner cannot represent the expert’s policy, provided it can represent the expert’s value function. It introduces OVI, an on-policy value-based algorithm with statistical guarantees and a linear optimization implementation, and shows empirically that it can outperform behavior cloning, DAgger, and offline value-based methods when the learner is less expressive than the expert.
The representational tradeoff between policy realizability and value realizability, together with a matching necessity result for offline methods, is a substantive and somewhat counterintuitive theoretical insight, though the abstract gives limited empirical detail.
-
Neural operator learning for collision-aware trajectory planning of spacecraft swarms
The paper proposes a permutation-equivariant neural operator that plans trajectories for an entire spacecraft swarm from the distributions of spacecraft, targets, and debris. It combines self-supervised physics losses and adversarially generated hazards, then uses a batched Gauss–Newton refinement to enforce orbital dynamics; the authors report zero-shot scaling from 10 training spacecraft to 1,000 spacecraft and over 11,000 debris objects while remaining close to per-agent optimal-control solutions.
The combination of swarm-size generalization, collision-aware operator learning, and adversarial self-training is genuinely interesting, but the abstract gives too few quantitative details to establish that the claimed 1,000-agent scaling and optimal-control accuracy are robust rather than a narrow simulation result.
-
Beyond Token-Level Cross-Entropy: Fréchet Distributional Post-Training for Autoregressive Image Generation
This paper post-trains pretrained autoregressive image generators with a representation-space Fréchet-distance objective rather than token-level cross-entropy. It combines generated rollout contexts with a straight-through estimator so image-level distributional gradients can update the generator while retaining hard autoregressive decoding. On eight ImageNet 256×256 configurations across four model families, it reports average FID and feature-distance reductions of 41.4% and 52.0%, with the best FID improving from 2.42 to 1.43 without extra parameters or sampling steps.
The direct optimization of decoded image distributions under realistic rollout contexts is a meaningful and potentially broadly useful departure from teacher-forced cross-entropy, with unusually large reported gains, but the evidence is limited to class-conditional ImageNet and the abstract gives little detail about baselines, statistical reliability, or transfer beyond the tested configurations.
-
Staged Multi-Agent Training (SMAT) for Hip Exoskeletons: Metabolic and Biomechanical Validation of a Simulation-Trained Co-Adaptive Controller
The paper deploys a hip-exoskeleton controller trained entirely in simulation using a staged curriculum in which simulated human and exoskeleton agents adapt together. On eight adults, active assistance reduced metabolic rate by 19.7% versus the passive exoskeleton, with consistently positive hip power and reported generalization across walking speeds and terrains without subject-specific retraining.
The combination of simulation-only co-adaptive training, deployment without user retraining, and physiological validation with a sizable metabolic reduction is unusually compelling, but the evidence is still limited to eight healthy users and the abstract does not provide enough detail to establish how broadly the result transfers.
-
Large language models improve physician accuracy but lead to false reliance
The study evaluates CORA, a retrieval-augmented LLM, as clinical decision support for 46 physicians. Physician accuracy rose from 70.8% unaided to 82.6% with the system, including stronger gains on newer cases, but source-linked citations also made physicians much less likely to reject incorrect model answers when those answers appeared supported. The main finding is that citations improve reliance on correct advice while simultaneously increasing false reliance on incorrect advice.
The measured asymmetry in citation-driven reliance is a genuinely important human-AI interaction result, but the evidence is limited to 46 physicians and a benchmark-style study rather than real clinical deployment.
-
Stress-Relief Annealing: Polynomial-Time Simulation-Free Layout Optimization for Automated Warehouses
The paper introduces Stress-Relief Annealing, which converts warehouse task demand into a per-location traffic “stress field” and uses it to optimize shelf layouts without repeatedly simulating robot traffic. The method reportedly doubles the sustainable robot count of a human-designed layout, matches or beats evolutionary optimization, and reduces computation from 25,000 simulations taking 25 hours to 19 minutes on one CPU core, with results transferring across path-planning algorithms, demand patterns, and larger warehouses.
Simulation-free, theoretically motivated layout optimization with a claimed orders-of-magnitude reduction in search cost and substantial throughput gains is genuinely notable, but the abstract lacks enough experimental detail to establish that the result is broadly reliable rather than specific to the tested warehouse settings.
-
UDT: Reconciling U-Nets and Diffusion Transformers with Data-Adaptive Token Reduction
UDT combines a diffusion-transformer backbone with U-Net-like multiscale processing by adaptively merging and restoring spatial tokens, while keeping the transformer token dimension compatible with cross-attention and representation regularization. On ImageNet 256×256, the authors report convergence to better than SiT’s 7.9 FID in about 40 epochs for an XL model, and final CFG FIDs around 1.35–1.38, substantially faster than the cited baseline.
The data-adaptive token reduction is a meaningful architectural direction with unusually large reported convergence gains, but the abstract does not establish whether the comparisons and training budgets are fully apples-to-apples, so it merits a closer look rather than a strong recommendation.
-
TRPML1 loss drives lysosomal calcium failure and astrocyte dysfunction across Alzheimer's Disease progression
The study identifies loss of astrocytic lysosomal calcium release, linked to reduced TRPML1 expression, as an early event in Alzheimer’s disease models—before amyloid plaques appear. Restoring TRPML1 in astrocytes rescued lysosomal and cytosolic calcium signaling, prevented reactive astrocyte changes, and reduced later plaque burden, suggesting a mechanism that connects early organelle dysfunction to disease progression.
The proposed pre-plaque lysosomal calcium failure and its rescue across both early and late astrocyte dysfunction are genuinely interesting, but the abstract provides no quantitative detail on models, effect sizes, or validation beyond the reported in vivo experiments.
-
A generalizable speech neuroprosthesis
The authors train a transformer speech decoder jointly on intracortical recordings from six people, rather than training a separate model for each user. The shared model improves relative word error rate by over 50% on average and reaches below 7% WER for a new participant after fewer than 200 calibration sentences, suggesting substantially faster deployment of speech BCIs.
Cross-user neural decoding with strong reductions in calibration data is a genuinely important BCI direction, but the abstract gives limited absolute baselines and evidence from only six participants, so it falls short of a strong recommendation.
-
Structure-function dissection of huntingtin exon 1 identifies a PRD-driven modifier of neuronal toxicity in Huntington's disease
The study uses genome-edited neurons and mini-organoids to separate the effects of huntingtin exon 1 domains across human and mouse sequences. It finds that the divergent proline-rich domain (PRD) strongly changes toxicity: the mouse PRD is protective, whereas the human PRD worsens neuronal phenotypes, apparently through actin-cytoskeleton and MKL2/MRTFB dysregulation. Restoring MKL2/MRTFB expression rescued abnormalities in HD models.
The result offers a non-obvious species-specific mechanism showing that the usually overlooked PRD, not just the expanded polyglutamine tract, can materially modify HD toxicity, but the abstract provides limited detail on effect sizes, model breadth, and rescue robustness.
-
Bio-CM{superscript 2}: Distributed computational optics for cortex-widecellular imaging
The paper introduces Bio-CM2, a miniature computational microscope that uses four coordinated optical modules and one sensor to image a large cortical area at cellular resolution. It reports a 7.5 × 10 mm field of view, video-rate imaging of more than 3,000 neurons, and cortex-wide vascular and neuronal connectivity measurements, with additional demonstrations in brain sections and freely behaving worms. The main new idea is distributing optical image formation across modules to avoid the usual field-of-view, resolution, and hardware-complexity trade-offs.
This is a genuinely interesting imaging architecture with potentially important multiscale neuroscience capabilities, but the abstract provides limited quantitative comparison, validation details, and evidence that the platform broadly outperforms existing mesoscopes.
-
A causal spatial role for apical dendrites in the cortex
Using two-photon-guided pruning of individual apical dendrites in awake mouse V1, the authors tested whether these dendrites suppress or drive responses to large visual stimuli. Removing them reduced responses in layer-5 neurons tuned to large stimuli, while imaging showed that their excitatory inputs span an extended region whose size matches the neurons’ visual preference. The work argues that apical dendrites provide spatially organized excitatory drive rather than merely contextual suppression.
It provides a relatively direct causal test of a debated cortical computation and reverses the expected suppressive role, but the abstract gives limited detail on effect sizes, controls, and breadth of validation.
-
Parallel construction of object motion from retina to cortex
Recordings from matched visual stimuli in mouse retina and cortex, combined with modeling, show that cortex does not simply inherit the retina’s object-motion computation. Instead, the retina is tuned to fine jitter and detects moving objects, while cortex favors coarse drift and represents object pattern, suggesting that object-motion perception is built through parallel specialization.
This directly challenges the standard hierarchical view of visual processing with a plausible division-of-labor result, but the abstract provides limited quantitative evidence and the finding is currently only a bioRxiv preprint in mice.
-
Deep conservation of head direction circuits in bees, ants and flies
Using synaptic-resolution maps, the authors compare head-direction circuits in bees, ants, and flies and find that their cell types and large-scale connectivity are remarkably similar across roughly 300 million years of evolution. However, the synaptic wiring differs substantially between species, yet computational models suggest both implementations can produce ring-attractor-like compass functions, revealing how a stable circuit plan can support evolutionary flexibility.
The combination of deep structural conservation with fundamentally different synaptic implementations is a non-obvious result relevant to how neural algorithms evolve, but the abstract provides too little experimental and modeling detail for a stronger recommendation.
-
CaMKII T286 autophosphorylation is not propagated at basal Ca2+ levels and is required only for the induction phase of LTP
This study tests whether CaMKII can keep its T286-phosphorylated state alive through subunit exchange or phosphorylation between holoenzymes after calcium levels return to baseline. The authors find that neither mechanism propagates phosphorylation without renewed calcium stimulation, and that T286 phosphorylation is needed only during the first minutes of LTP induction—not for later expression or maintenance—supporting a transient computational role rather than a self-sustaining molecular memory.
It directly challenges a prominent model in which CaMKII autophosphorylation stores synaptic information, but the abstract provides only moderate evidence detail and the findings await broader replication and experimental scrutiny.
-
A conserved population of genetically defined striatal neurons gates opioid reward
The authors identify a rare striatal neuron subtype marked by Chst9 that has unusually high mu-opioid receptor expression and is conserved across species. Opioids silence these neurons, and removing the receptor specifically from them eliminates fentanyl-conditioned place preference, suggesting that this specialized indirect-pathway population—not the canonical D1/D2 division alone—is a key circuit element for opioid reward.
This offers a specific cellular resolution to the longstanding mismatch between opioid reward and the canonical striatal pathway model, with causal evidence, but the abstract supports mainly a conditioned-place-preference result and does not yet establish broader behavioral or therapeutic relevance.
-
Synaptic engram underlies memory
The authors developed an optical method to selectively weaken synapses between entorhinal-cortex and dentate-gyrus engram cells by controlling pre- and postsynaptic spike timing. Weakening these connections in vivo impaired the associated memory, providing causal evidence that learning-induced connectivity between engram ensembles—not just cellular excitability or other plasticity—supports memory.
The causal, pathway- and engram-specific manipulation is a meaningful advance over correlational engram studies, but the abstract gives evidence from only one circuit and does not establish how general the mechanism is.
-
Thalamocortical orchestration of human theory of mind
Using 7T fMRI and intracranial recordings during a naturalistic movie, the authors find that theory-of-mind events activate the default network while the thalamus shows little local activation. Nevertheless, thalamic activity shares the network’s representational structure and appears to coordinate it through low-frequency phase modulation of high-frequency cortical activity, with stronger coupling associated with better ToM representations.
The proposed role for the thalamus as a coordinating or gating mechanism for cortical mentalizing computations is a genuinely non-obvious neuroscience result, but the abstract provides limited information about sample size, causal evidence, and robustness beyond the reported cross-modal associations.
-
Wearable magnetoencephalography in an open environment
The authors develop a wearable optically pumped magnetometer MEG system that operates in an ordinary office rather than a magnetically shielded room. Room-scale active coils suppress environmental magnetic fields enough for source localization, and the system demonstrates alpha-rhythm recording, sensory-response localization, a nine-target visual speller, and motor-cortex tracking during smartphone use. The key advance is extending millisecond-resolution MEG into relatively natural environments.
Unshielded, source-localizing wearable MEG would be a meaningful enabling capability for naturalistic neuroscience and BCI, but the abstract gives few quantitative performance comparisons or details about robustness outside the demonstrated office setting.
-
MODUS: Decoder-Only Any-to-Any Modeling of Diverse Modalities
MODUS proposes a single decoder-only model that treats diverse modalities symmetrically, allowing arbitrary combinations of modalities as inputs and outputs without separate modality heads or task-specific pipelines. It supports chained cross-modal generation and self-verification, and is reported to be competitive with specialist and multitask systems across several benchmarks, though the abstract gives no quantitative results.
The decoder-only formulation for genuinely any-to-any multimodal generation is a meaningful architectural direction, but the abstract provides too little quantitative evidence to justify a stronger recommendation.
-
Evidence Attribution in Visual Document Understanding without Coordinates or Region Labels
This paper argues that visual evidence attribution often fails partly because models must express evidence as coordinates, even when they know the answer. Replacing coordinate prediction with verbatim evidence quotes plus a multimodal retriever substantially improves evidence recall and roughly halves attribution hallucination across six vision-language models; a GRPO training recipe using retrieved crops raises strict attributed accuracy for an 8B model from 22.4% to 33.8% without region-level labels.
The coordinate-versus-language-interface diagnosis and label-free quote-and-retrieve training scaffold are genuinely interesting, with reasonably concrete multi-model results, but the evidence is limited to one bilingual benchmark and the approach depends on a separate layout parser and retriever.
-
RL$^2$-VLA: Adaptive RL Latent Compositional Steering with Test-Time Scaling for Vision-Language-Action Models
The paper adds a lightweight offline-RL policy to a frozen vision-language-action model and composes its latent action guidance with the base policy at inference time. Unlike uniform test-time steering, it predicts when the base policy is likely to fail and applies extra action diversity only then; it reports up to 17.3% gains on out-of-domain simulation tasks and transfer to real-robot experiments.
The conditional, latent-space RL steering is a meaningful refinement of test-time VLA scaling, and the reported OOD and real-world gains are promising, but the abstract does not establish how broadly it beats competing methods or how large the real-world effect is.
-
See2Think: Do Multimodal Models Really Use Intermediate Visual States?
See2Think introduces a benchmark and process-tracing framework for testing whether multimodal models genuinely use intermediate visual states during reasoning, rather than merely producing them. Across 1,200 visually dependent problems, the authors find that models often choose relevant visual operations, but rendering those operations faithfully is the main bottleneck; controlled corruption of visual feedback reduces accuracy by more than 10 percentage points, showing some behavioral dependence on intermediate images.
It provides a useful diagnostic beyond final-answer accuracy and reports the non-obvious result that models can select appropriate visual actions yet fail mainly at faithful rendering, though the abstract gives limited detail on model breadth and intervention results.
-
Pass the Baton: Trajectory-Relayed On-Policy Distillation
The paper modifies on-policy distillation by detecting when a student’s reasoning has gone off track, briefly handing generation to the teacher, and then returning control to the student. On eight math-reasoning benchmarks, this improves over standard OPD by 5.73% on average for a 1.7B student and cuts training trajectory length by more than half, with similar gains at 0.6B.
The teacher handoff mechanism is a plausible and fairly concrete response to prefix failure, and the broad benchmark gains plus reduced trajectory cost make it worth a look, though it remains a targeted refinement of distillation rather than a clearly new paradigm.
-
Beacon: Knowing When and How to Perform Agentic Visual Reasoning
This paper analyzes whether multimodal models use visual tools only when needed and whether those tools genuinely help, rather than adding overhead or errors. It finds that current agentic systems often invoke tools unnecessarily, with gains on hard examples offset by mistakes on easy ones, and proposes Beacon, which uses necessity-aware rewards and hint-guided training to improve adaptive tool use and overall performance.
The explicit separation of tool-use necessity from tool effectiveness and the finding that tool benefits can be canceled by harms on easy cases are useful, non-obvious contributions, though the proposed mechanisms appear largely like targeted RL/reward-shaping extensions and the abstract gives no quantitative results.
-
Amortising Trajectory Optimisation for Residual MPC via Implicit Contact Differentiation
The paper uses the Implicit Function Theorem to differentiate converged, regularized contact simulations without unrolling the contact solver or deriving solver-specific KKT sensitivities. In Mujoco MJX, this keeps memory nearly constant as solver iterations increase and substantially reduces memory for larger contact-rich systems. It also distills full-horizon iLQR into a policy that guides short-horizon residual MPC, producing 28–98 percentage-point gains in six-step success on Finger, Franka, and Unitree tasks.
The combination of solver-agnostic implicit contact differentiation and optimizer distillation addresses real scaling bottlenecks in contact-rich MPC, with unusually large reported memory and control-performance gains, though the evidence is limited to a small set of simulated platforms and tasks.
-
Memory Layer: Train the In-Model Cache for Recommendation Models
The paper introduces a trainable in-model key-value cache for recommendation systems, where the item tower writes embeddings during training and serving reads from the same cache. In Instagram Reels production, it achieves full item coverage, reduces embedding staleness from about five minutes to 20 seconds, cuts training/publishing compute by 30%, and improves fresh-content recall and cold-start engagement.
This is a substantial production systems redesign that directly removes a common training-serving mismatch, with unusually concrete deployment results, though its applicability appears concentrated in large-scale recommendation infrastructure.
-
Denoising growth complexity: Data geometry and certified schedules for diffusion sampling
The paper introduces denoising growth complexity (DGC), a geometric quantity tracking how denoising error changes along the Gaussian heat flow. It uses local DGC increments to derive KL guarantees for Euler diffusion samplers, design data-adaptive stepsize schedules, and certify those schedules from samples; the analysis also gives information-theoretic bounds and predicts large gains from adaptation for some Gaussian mixtures.
This is a potentially useful unifying theory for data-adaptive diffusion sampling and claims logarithmic-to-constant schedule improvements, but the abstract provides theoretical results rather than broad empirical validation or evidence that the guarantees materially improve practical samplers.
-
Visko Orbis 1.0: A Live Model for Real-Time Interactive Long Video Generation
Visko Orbis presents a streaming video-generation system that allows users to change prompts during generation, while maintaining subjects, scenes, and style over hour-scale videos. It claims real-time 4K output at 24 FPS across text-to-video, image-to-video, and continuation, with strong results in comparisons against other interactive systems.
Interactive prompt switching with hour-scale consistency and real-time 4K output would be a substantial capability advance, but the abstract gives few quantitative details about quality, hardware cost, latency, or the evaluation and therefore does not support a strong verdict.
-
Dense Soft Weighting for Radar Ego-Velocity Estimation
The paper replaces binary CFAR thresholding in radar ego-velocity estimation with continuous, confidence-weighted use of every range-Doppler cell. A deterministic robust weighted least-squares estimator also produces measurement-derived velocity covariance, avoiding platform-specific training and supporting inertial fusion. On two public datasets and one self-collected dataset, it reports 31–45% lower mean absolute pose error than a strong CFAR baseline while running in real time on embedded hardware.
This is a clean, potentially broadly useful radar front-end that substantially improves robustness and uncertainty estimation without learning or platform-specific training, but the abstract does not establish whether the gains generalize across difficult environments and radar configurations.
-
Parameter-Free Dynamic Regret for Online Convex Optimization under Heavy-Tailed Noise
The paper introduces HT-PAder, an online convex optimization algorithm that handles both changing comparators and heavy-tailed stochastic gradients without knowing the noise, domain, Lipschitz, or nonstationarity parameters in advance. It combines restarted AdaGrad experts with a pathwise meta-algorithm and achieves a dynamic-regret bound depending optimally on comparator path length; the authors also give a matching lower bound, including the first parameter-free minimax result claimed for finite-variance noise.
The combination of parameter-free adaptation, heavy-tailed noise, and dynamic regret with a matching lower bound is a substantive theoretical advance, though its impact is mainly within online-learning theory rather than a broad practical capability jump.
-
InferScale: GPU-Native KV Injection for Personalized LLM Serving
InferScale caches the transformer KV states of individual user-memory facts on the GPU, then injects the relevant states directly into vLLM instead of repeatedly prefilling their text. It introduces position-aware RoPE handling and context-window encoding to reduce the quality loss from encoding facts independently. On three open models and LoCoMo, it reports 3.6–4.8× lower TTFT at 50 retrieved facts and 3.7–4.5× higher concurrent throughput, with modestly lower accuracy than Mem0.
Direct KV reuse for dynamically retrieved, personalized memory is a meaningful serving direction with large reported latency gains, but the evidence is limited to three models and one benchmark and the approach is primarily a systems optimization rather than a broad capability advance.
-
ROCS: Request-Oriented Compute Sharing for Efficient Large-Scale Recommendation
ROCS restructures recommendation inference so that computations depending only on the user/request are performed once and shared across all candidate items, rather than repeated per candidate. It introduces masking and attention mechanisms to preserve this candidate isolation, plus GPU kernel optimizations; experiments report up to 3× higher retrieval throughput and a 50% QPS gain with a 0.5% relative LogLoss improvement in production ranking systems.
The request-oriented factorization is a substantial systems/modeling idea with credible production-scale gains, but it is specialized to recommendation inference and the abstract does not establish a broadly transformative result.
-
Commit Locally, Exit Globally: Coordinating Adaptive Sampling and Early Exit in Diffusion Language Models
The paper introduces C4, a decoding strategy for diffusion language models that jointly accelerates local token/block updates and globally terminates generation when the answer has stabilized. It uses separate mechanisms for these decisions: a confidence-and-stability gate for sequence exit, and a boundary-anchored commit/confirm procedure for safely freezing token positions. Across 12 zero-shot tasks and two models, it reportedly removes 64–95% of decoding steps and achieves 2.6–8.6× measured end-to-end speedups, though the abstract does not quantify any quality tradeoff.
The coordinated treatment of local committing and global early exit is a meaningful, potentially reusable decoding idea with large reported speedups, but it appears mainly an acceleration technique for a still-specialized diffusion-LM setting and the abstract omits detailed quality-retention results.
-
Fully Inductive Cardinality Estimation
FICE is a learned cardinality estimator for SPARQL graph-pattern queries that can transfer to entirely new knowledge graphs, including graphs with unseen relations, without retraining. It combines a local GNN encoder over the graph with a query-topology decoder; across 10 knowledge graphs, it reduces median q-error from 13.54 for the best prior method to 5.34 while supporting sub-millisecond estimation.
The genuinely inductive transfer across unseen graphs and relations addresses a major deployment limitation of learned query optimizers, with a substantial reported error reduction and broad multi-graph evaluation, though the advance is specialized to KG/SPARQL systems.
-
ObjectStream: Latent Objects as Memory Anchors for Streaming Video Understanding
ObjectStream uses latent, spatially coherent objects extracted from frozen Video-LLM representations as persistent memory anchors in streaming video. It tracks object histories, short-lived changes, and recent context without external detectors or model retraining; the authors report a 10-point gain for Qwen2.5-VL-7B on a real-time benchmark, about 50% lower peak memory and time-to-first-token, and better long-video results while dropping 82.5% of visual tokens.
The object-centric memory organization and strong reported efficiency/capability gains are genuinely interesting, but the abstract lacks absolute results, comparisons across more models, and enough detail to establish that the gains are broadly reliable rather than benchmark-specific.
-
ROAD: Reciprocal-Objective Alignment of Discriminative Semantics for 3D Shape Generation
ROAD transfers semantic and geometric knowledge from a pretrained discriminative 3D foundation model into a diffusion transformer during training, using separate objectives for global semantics and fine-grained structural alignment. The foundation model is discarded at inference, and the authors report competitive 3D generation against Step1X-3D using only 1.5% of its training data, with lower training cost.
The potentially important result is that discriminative 3D priors may replace most of the data and compute normally needed for generative training, but the abstract gives only one baseline and no quantitative quality or cost figures, so the strength of the claim is not yet established.
-
AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis
AskChem builds a chemistry literature-search system whose basic unit is an evidence-backed claim rather than an entire paper. It extracts 2.4 million typed claims from 147,000 papers, links them to quotations or evidence locations, and supports faceted search, evidence graphs, and AI-agent interfaces; grounding GPT-5.5 with it improved DOI resolvability from 88.3% to 100%.
The claim-level, provenance-carrying retrieval infrastructure and large live index are useful and plausibly adoptable, but the core approach is an elaboration of retrieval-plus-grounding systems and the abstract gives limited evidence beyond citation/DOI metrics.
-
Agentic Coding in the Wild: Characterizing GitHub Copilot Traces at Production Scale
The paper analyzes 761 million LLM calls from 13 million GitHub Copilot sessions to characterize production agentic-coding workloads. It finds that coding agents generate tightly tool-coupled inference loops with high within-turn KV-cache reuse, but much lower reuse across turns and after context or model changes; users also create long idle periods between turns. A lightweight predictor captures 86–90% of idle time, suggesting that agent-serving systems should schedule and cache work differently from chatbot systems.
The unusually large production trace study exposes workload properties—especially the contrast between high intra-turn cache reuse and long inter-turn idle periods—that challenge standard LLM-serving assumptions, though it is primarily characterization rather than a demonstrated systems breakthrough.
-
TransMem: Transforming Hidden States into Memory for Large Language Models
TransMem adds a lightweight memory module to a frozen LLM by converting selected historical hidden states into reusable representations, avoiding repeated processing of the full interaction history. It trains this module with evidence-conditioned self-distillation, and reports sizable gains on LoCoMo, HotpotQA, and MemoryAgentBench across model sizes and architectures. The main novelty is treating sparse intermediate hidden states—not just text summaries or retrieved tokens—as the substrate for persistent agent memory.
The hidden-state memory mechanism and distillation setup are a meaningful, potentially efficient alternative to standard long-context retrieval, with unusually large reported gains, but the abstract does not establish how much comes from implementation details or how broadly the method generalizes beyond memory benchmarks.
-
EEG-JEPA: Structured Latent Prediction for EEG Foundation Models
EEG-JEPA pretrains EEG models by predicting latent representations of masked electrode–time regions rather than reconstructing raw voltage samples. Using structured masking based on electrode topology and multiple prediction depths, it improves frozen 14-task macro balanced accuracy from 40.49% to 50.42% over masked-waveform reconstruction, reaching 52.94% with additional multi-source training; full fine-tuning gains are smaller, from 68.98% to 70.65%.
The transfer gains are substantial under controlled pretraining conditions, and adapting latent-prediction objectives plus neurotopological masking to EEG is more than a routine loss tweak, but the work remains an application and refinement of the broader JEPA paradigm rather than a clearly new foundation-model direction.