Issue 30 · Jul 20–26, 2026
Every candidate
All 2,728 papers were scored from their abstracts by gpt-5.6-luna; 992 were not skipped. Shown here: the top 150 of those, in score order. Picks are marked.
-
The Intruder Threshold: A Spectral Law for LoRA Fine-Tuning
The paper proposes a parameter-free spectral law that predicts when a LoRA update creates new leading singular directions—so-called intruder dimensions—that are associated with catastrophic forgetting. Across 9,840 layer scans spanning seven model families and external adapters, it predicts threshold crossings and perplexity degradation reasonably well, and an SVD-based spike-budget intervention reduces forgetting by 62% without hurting the target task.
This offers a genuinely new, mechanistic and deployment-time explanation for LoRA’s apparent forgetting asymmetry, backed by broad layer-level validation and an intervention that improves the tradeoff without validation sweeps.
-
Programming protein shape as an explicit design layer via CAD blueprint-guided diffusion
The authors combine CAD-style geometric blueprints with diffusion models to directly specify global protein features such as curvature, twist, chirality, and superhelical form. They validated 31 of 59 generated designs structurally and used flow-optimized helical proteins to build an F1-ATPase-powered molecular swimmer, demonstrating that designed shape can produce an emergent mechanical function.
This is a genuinely new design interface for protein generation, backed by substantial structural validation and a striking functional demonstration rather than only computational metrics.
-
ABot-World-0: Infinite Interactive World Rollout on a Single Desktop GPU
ABot-World-0 is an action-conditioned video world model designed for interactive, long-horizon rollouts from keyboard input. It combines large-scale data collection across games, simulators, and videos with teacher-to-student distillation, a long-rollout training method called LongForcing, and an optimized inference stack; the reported system generates 720p video at up to 16 FPS on one RTX 5090 with 1.2-second latency. The main novelty is packaging long-horizon controllable generation and practical single-GPU streaming into one system, though the abstract gives few quantitative comparisons.
Real-time, long-horizon interactive video generation on a single desktop GPU could be an important capability and efficiency result, but the vague claim of competitive performance and lack of benchmark numbers make it premature for a strong recommendation.
-
StateAct: Program State, before Pixels, for Long-Horizon Computer-Use Agents
StateAct lets computer-use agents inspect and modify underlying program state—such as files, application backends, and DOMs—directly with code, using screenshot-based GUI interaction only when necessary. On OSWorld 2.0, it improves Claude Opus 4.8 from 20.6% to 26.9% binary success and reduces cost by about 9×, while independent state-based verification catches structural failures. The results suggest that direct state access and verification may be more useful for long-horizon tasks than relying primarily on better screenshot perception.
The state-first framing, large reported cost reduction, and evidence that code-only access is insufficient without selective GUI interaction are genuinely interesting, but the evaluation is concentrated on one benchmark and agent/model setup, so it is not yet a clear field-level breakthrough.
-
Generative World Renderer at the Speed of Play
The paper distills a slow generative renderer that converts physics-engine world states and prompts into video frames into a streaming model running at 31.54 FPS, up from 0.56 FPS. It also adds lightweight latent codecs and supports unbounded streams, demonstrating a playable physics-driven generative world at roughly real-time speed while retaining the teacher’s interfaces.
Real-time, physics-consistent generative rendering is a meaningful capability jump and an interesting alternative to prompt-driven world models, but the abstract gives few quantitative quality results beyond the large speedup, so it falls short of a strong recommendation.
-
Motor Learning and Transfer Are Symmetric Across Hands
A meta-analysis of 114 datasets and preregistered experiments with 526 participants tested whether motor adaptation favors the dominant hand or transfers asymmetrically between limbs. Both adaptation and transfer were symmetric, suggesting that lateralized motor control and motor learning rely on partially distinct neural organizations.
The paper directly challenges the expectation that lateralized motor control should imply lateralized motor learning, using unusually broad meta-analytic and preregistered experimental evidence.
-
Sample-Efficient Learning from Agent Experience
The paper proposes “Experience Distillation”: first let an agent improve through in-context trial-and-error histories, then distill those histories into the model’s weights without collecting more environment data. Across 749 software-engineering tasks and six text-adventure games, this retains at least 64.8% of the in-context-learning gains, while ordinary supervised fine-tuning recovers only 3.8%; against classical RL, it reaches comparable performance with at least 9.6× fewer environment samples.
The combination of in-context learning and no-additional-interaction distillation addresses a real bottleneck in agent training, with unusually strong sample-efficiency results across two domains, though the abstract does not establish robustness beyond these settings or clarify how much depends on the specific distillation implementation.
-
Barzilai-Borwein Fails Superlinear Convergence on an Open Set of Quadratics for Every Dimension $n\geq 4$
This paper proves that the long-step Barzilai–Borwein (BB1) method is not root-superlinearly convergent on a nonempty open family of strictly convex quadratic problems in every dimension n≥4. The result comes from a computer-assisted proof of an attracting seven-cycle in the projective BB dynamics, yielding two-sided geometric decay bounds for gradients and errors rather than superlinear decay.
It gives a concrete, assumption-challenging negative answer to a longstanding convergence question, but its impact is concentrated in the theory of a specific classical optimizer rather than broadly changing practical optimization.
-
Self Gradient Forcing: Native Long Video Extrapolation
The paper identifies a limitation of Self Forcing for autoregressive video diffusion: later-frame losses cannot train earlier generated frames to produce better key-value memory for future prediction. Self Gradient Forcing uses a no-gradient rollout followed by a parallel reconstruction pass that backpropagates future losses into how past context is encoded, and claims substantially better long-horizon consistency, including extrapolation from a 5-second training window to videos lasting several minutes.
The memory-writing supervision mechanism is a meaningful and potentially general training idea for long-context video generation, but the abstract gives no quantitative results, baselines, or scale details to substantiate the unusually strong multi-minute extrapolation claim.
-
Three-Body Scattering for Generative Modeling
The paper proposes Three-Body Scattering Modeling, a one-step generative-training method in which each generated sample is attracted to one real target and repelled from an independently generated target. This provides constant-cost, sample-level supervision derived from an energy-distance/Wasserstein gradient-flow view, avoiding minibatch-wide pairwise interactions. On ImageNet-256 it reports one-step FID scores of 2.23 in pixel space and 1.63 in latent space.
The combination of distributional-energy geometry and constant-cost scattering supervision is a genuinely interesting alternative route to one-step generation, with unusually strong reported ImageNet FIDs, but the abstract does not provide enough baseline, ablation, or reproducibility detail to merit a strong verdict.
-
Text Template Tokens Are Implicit Semantic Registers in Diffusion Transformers
The paper studies how chat-template tokens in text-to-image diffusion transformers contribute during generation, using causal interventions rather than treating them as irrelevant formatting. It finds that these tokens become attention sinks that help preserve object identity by reading from image latents, and uses the analysis to prune some heads, reducing joint-attention FLOPs by 20% at a 1.4-point GenEval drop.
The causal finding that template tokens act as latent semantic registers is genuinely non-obvious, and the associated compute reduction is useful, but the abstract gives limited evidence about how broadly the mechanism and pruning rule generalize across models and tasks.
-
Panache: One-Pass Motif Discovery at Every Window Length
Panache is a streaming algorithm for finding recurring, z-normalized motifs across many possible time-series window lengths in one pass, rather than running a quadratic matrix-profile computation separately for each length. It maintains sliding Fourier statistics, hashes candidate subsequences, and uses Parseval-based lower bounds to discard most comparisons; on 17 UCR configurations it recovered the exact top-20 motifs, and on a five-million-sample Wafer series it was substantially faster than CPU and GPU baselines.
The one-pass, all-window-length formulation appears genuinely novel and could change the cost of a useful time-series primitive, but the evidence is still limited to benchmark configurations and the abstract leaves important scalability and approximation/resource-budget details unresolved.
-
Robostral Navigate
Robostral Navigate is an 8B vision-language navigation policy that uses only monocular RGB video and predicts the next waypoint directly in image space, allowing the same policy to transfer across wheeled, legged, and aerial robots without recalibration. It combines large-scale synthetic trajectory generation with prefix-cached episode training, reportedly reducing training tokens 22-fold, and achieves 77.4% success on R2R-CE and 75.1% on RxR-CE, exceeding prior monocular and even depth/multi-camera systems on R2R-CE.
The combination of sensor-minimal, image-space control, cross-embodiment deployment, and a substantial training-efficiency gain is notably interesting, but the abstract provides benchmark results rather than convincing real-robot transfer or broad deployment evidence.
-
Microglia extract neuronal proteolytic organelles via skoupocytosis
The paper reports that microglia remove oversized proteolytic organelles directly from neuronal processes by pinching off small pieces of membrane, rather than relying on retrograde transport to the soma. It names this process “skoupocytosis” and implicates the PS lipase ABHD16a and lyso-PS signaling in initiating the interaction, with observations in vitro and in vivo.
This proposes a genuinely new neuron–microglia housekeeping mechanism with a plausible molecular trigger and in vivo support, but the abstract provides no quantitative evidence about prevalence, functional necessity, or consequences.
-
Transcription Policy as a Latent Variable: Activating Controllable Verbatim ASR with Word-Level Timing
The paper treats transcription style—verbatim versus cleaned-up intended text—as an explicit, controllable mode rather than an uncontrolled property of ASR training data. Task tokens trained on parallel transcripts reportedly improve German disfluency F1 from 10% to 79% zero-shot after English-only training, while supervised cross-attention tuning improves word-level timestamps on disfluent speech and enables a scalable “verbatimize” corpus-building workflow.
The combination of controllable transcription policy, strong cross-lingual disfluency transfer, and improved timing addresses a real ASR failure mode, but the abstract gives limited experimental detail and the broad claims need verification across languages, models, and datasets.
-
Multi-Head Attention Residuals
The paper replaces the single depth-wise attention distribution used by attention residuals with multiple independent distributions, one per feature subspace. This adds no parameters and improves validation loss over standard Transformers at 100M–1B parameters, with the best tradeoff at 4–8 heads; an 8B mid-training experiment also reports gains on GSM8K and GPQA. The authors support the mechanism with probes showing that different subspaces prefer different layers, and provide fused kernels that reduce—but do not eliminate—the training overhead.
This is a plausible, conceptually clean extension of attention residuals with evidence across scales and an explanation for why it helps, but it remains an architectural refinement rather than a clearly transformative capability or efficiency breakthrough.
-
From Hodgkin-Huxley to Pretrained Neural Inference AI
The authors pretrain neural networks only on large-scale biophysical simulations of neural population recordings, then use them to infer single-neuron activity and cell-type properties from real high-density probe data without real-data training. They report zero-shot transfer across brain regions, species, and experimental settings, and claim the method reveals many weakly active neurons that change the interpretation of ocular dominance in mouse visual cortex.
Simulation-only pretraining that transfers across species and uncovers a potentially important biological bias is genuinely notable, but the abstract gives no quantitative validation or details supporting its very broad claims.
-
Visual Contrastive Self-Distillation
The paper proposes self-distillation for vision-language models using only input conditioning: an EMA teacher predicts the next token both with the image and with its visual content erased, then uses the log-probability difference to emphasize image-grounded tokens. On ViRL39K, this improves matched on-policy self-distillation across Qwen3-VL and Qwen3.5 model sizes, with aggregate gains of roughly 1.9–4.8 percentage points, without external teachers, answer labels, reasoning traces, or extra inference cost.
The image-versus-content-erased contrast is a clean and potentially reusable way to create the asymmetry needed for self-distillation without privileged signals, and it has consistent multi-model gains, though the evidence is limited to one training setting and does not yet establish broad generality.
-
Masked Visual Actions for Unified World Modeling
The paper represents robot actions as partially revealed pixel trajectories in video, allowing one model to predict scene responses to robot motion or infer robot motion from a desired object trajectory. A single model trained on 15 hours of real and simulated masked examples is used for visual rollout prediction, candidate-future ranking, and inverse manipulation planning across scenes and embodiments.
The pixel-space action interface is a genuinely interesting attempt to align control with pretrained video-model representations and unify forward and inverse manipulation, but the abstract gives no quantitative results or concrete real-world success rates to establish a major advance.
-
ATSplat: Compact Feed-forward 3D Gaussian Splatting with Adaptive Token Expansion
ATSplat is a feed-forward 3D Gaussian Splatting method that starts with sparse 3D patch tokens and selectively expands tokens in regions predicted to be difficult to reconstruct. On RealEstate10K and DL3DV, it reportedly achieves state-of-the-art rendering quality with over 5.7× fewer Gaussians than dense feed-forward methods, reconstructing in under a second and rendering at 1136 FPS with 311K Gaussians.
The adaptive token-based allocation addresses a real weakness of feed-forward 3DGS and reports a substantial representation and rendering-efficiency improvement, but the evidence is limited to two datasets and lacks quantitative comparisons in the abstract beyond the compression claim.
-
FlashRT: Agent Harness for Guiding Agents to Deploy Real-Time Multimodal Applications
FlashRT uses a coding agent to turn simple reference implementations of real-time multimodal pipelines into optimized multi-GPU deployments. It introduces an intermediate representation, static analysis, and an iterative implement-verify-benchmark loop to explore placement, streaming, and parallelism choices; across video world models and multimodal LLMs, it reports latency reductions up to about 70x and throughput gains up to 3.6x, including a 65% latency reduction over vLLM-Omni on AMD hardware.
The agent-guided, measurement-gated optimization workflow for heterogeneous real-time model pipelines is a meaningful systems direction, but the very large speedups may partly reflect weak reference baselines and the abstract lacks enough detail for a strong verdict.
-
Where Should Optimizer State Live? Tiered State Allocation for Memory-Efficient Mixture-of-Experts Training
SkewAdam assigns different optimizer-state representations to MoE components: momentum plus factored second moments for the dense backbone, only factored second moments for experts, and exact second moments for the router. On a 6.78B-parameter MoE, it reduces optimizer state from 50.6 GB to 1.29 GB and peak memory from 81.4 GB to 31.3 GB, while matching or exceeding the reported baselines in a short 82M-token run; ablations indicate the accuracy comes mainly from momentum and update clipping rather than the memory-saving allocation itself.
The extreme, component-aware optimizer-state reduction could materially change the hardware requirements for MoE training, but the evidence is limited to one model and a short training comparison, and the perplexity advantage is not primarily attributable to the novel tier allocation.
-
Train the Model, Not the Reader: Decodability Supervision for Verifiable Activation Explanations
The paper shows that activation explanations can achieve high reconstruction scores while making mostly unsupported or false specific claims, because the explainer and decoder can learn private co-adapted codes. It proposes RECAP, which trains auxiliary probes to make designated activation content independently decodable, and reports that fresh verbalizers then rank true claims above false ones, including under adversarially edited explanations, though the evidence is mainly from synthetic sandbox models and a small Pythia-160M replication.
The specific demonstration that reconstruction-based explanation faithfulness can fail catastrophically, plus independent probe-based auditing and RECAP, is a meaningful interpretability contribution, but the limited model scale and partial transfer from probes to natural-language explanations make it short of a strong recommendation.
-
Persistent Computational State: A Session-Centric Runtime for Generative World Models
The paper argues that many world-model failures under branching, backtracking, and resumed simulation come not from the models themselves but from stateless request-oriented serving. It identifies the minimal persistent runtime state—such as RNG state, memory, or KV context—and shows that checkpointing and restoring it can reproduce the original continuation exactly across three model types, while enabling cheap session management and relevance-based eviction.
The claim that a major apparent world-model limitation is primarily a serving-runtime failure, plus a general PCS abstraction and measured session runtime, is genuinely non-obvious, but the abstract lacks enough experimental detail and breadth to establish how broadly this changes world-model practice.
-
ReViV: Reconstructing the Viewer and the View in 4D from Monocular Egocentric Video
ReViV presents a single feed-forward model that jointly reconstructs the wearer and the surrounding scene from monocular egocentric RGB video. It predicts camera motion, gaze, body and hand motion, depth, and temporally consistent 4D representations together, rather than relying on separate pipelines or precomputed camera trajectories. The authors report state-of-the-art accuracy and faster inference across several egocentric benchmarks, though the abstract gives no quantitative margins.
The unified reconstruction of viewer and view from only monocular video is a meaningful and potentially adoptable direction, but the lack of numerical results and detailed evidence makes the broad claims insufficient for a strong recommendation.
-
Optimizing the Preconditioner: A Black-box Online-to-Nonconvex Conversion with Static Regret Minimization Oracles
The paper gives a black-box reduction from stochastic nonconvex optimization to ordinary static-regret online convex optimization. An online learner selects preconditioners while a separate gradient tracker supplies update information, yielding the usual $O(T^{-1/2})$ stationary-point rate for smooth objectives and the optimal $O(T^{-2/7})$ Goldstein-stationarity rate for merely Lipschitz objectives. The latter extension uses only standard static regret and resolves an open problem from Chen and Hazan (2024).
This is a non-obvious theoretical conversion that removes the need for stronger online-regret notions and resolves a stated open problem, but its practical impact and breadth beyond the proved oracle framework are not established in the abstract.
-
A Controlled Study of Attention-Only Transformers
This paper systematically tests whether transformer feed-forward layers are necessary by training attention-only decoder models under matched parameter, compute, and depth budgets. Removing FFNs directly hurts, but reallocating their budget to additional attention nearly closes the loss gap—within 0.006 nats at matched parameters—and the remaining deficit appears concentrated in retrieving knowledge from model weights rather than using context. The authors link this behavior to how rank accumulates in attention projections and show that QK normalization is critical for training deep attention-only models.
The controlled result that attention-only transformers can nearly match conventional transformers, together with a specific account of the remaining parametric-recall deficit, challenges the assumed necessity of FFNs, but the evidence is limited to relatively small models and tested training regimes.
-
Patch Policy: Efficient Embodied Control via Dense Visual Representations
Patch Policy lets robot-control transformers consume dense pretrained ViT patch features directly, using a block-causal attention mask to preserve temporal causality without the cost of a full vision-language-action model. Across four simulation and three real-world suites, it reports a 40% relative gain over global-token policies and an 18% gain over fine-tuned OpenVLA-OFT, with about 0.7% as many parameters.
The combination of dense visual representations, lightweight causal policy design, and large claimed gains across simulated and real robots is substantially more than a routine policy tweak, though the abstract lacks absolute results and implementation details needed to establish how broadly the comparison holds.
-
Style over Substance: A Shortcut Audit of Emotion-Description Preference Evaluation
The paper audits EmoPrefer, a multimodal benchmark that ranks emotion descriptions, and finds that a logistic regression using only description length and generator identity nearly matches 7B text and audio-visual judges (65.8 vs. 66.8 WAF). Generator style is highly identifiable, and human preferences often follow generator-specific win-rate priors; after controlling for length and style, adding video and audio provides no significant gain. The authors argue that the benchmark can be scored without checking whether descriptions are grounded in the media, and propose source balancing and stricter controls.
This is a concrete, well-supported shortcut audit suggesting a widely used multimodal preference metric may measure generator style rather than emotion understanding, though the evidence is confined to one benchmark and does not establish that the issue generalizes broadly.
-
Mixing-Free and Signal-Optimal Learning of Gaussian Graphical Models from Glauber Dynamics
This paper learns the exact graph of a Gaussian graphical model from a single, potentially non-stationary trajectory of random-scan Glauber dynamics. Its two algorithms avoid mixing-time assumptions entirely by extracting fresh Gaussian innovations from the update sequence, achieving the information-theoretically optimal inverse-quadratic dependence on minimum edge strength, with recovery horizons of roughly p d^2/κ^2 and p d^4/κ^2.
The mixing-free treatment of dependent, non-stationary samples together with optimal κ^{-2} scaling is a substantive theoretical advance, but the result is specialized and the abstract provides no broad empirical or practical validation.
-
Stronger Memory-Query Tradeoffs for Convex Optimization: The Limitations of Subquadratic Memory
This paper proves substantially stronger lower bounds on the number of first-order oracle queries needed to optimize a Lipschitz convex function when the algorithm has limited memory. The deterministic bound reveals a sharp phase transition near quadratic memory, while the randomized bound shows that roughly d² bits of memory are necessary to approach unconstrained query complexity even when the target accuracy is only polynomially small in d.
It gives the first sharp memory–query tradeoff around m≈d² and closes an important gap in understanding whether subquadratic-memory optimization can retain optimal query efficiency, but its impact is primarily foundational rather than an immediate ML capability advance.
-
ATLAS: A Foundation Neural Sampler for Amorphous Materials
ATLAS learns an equivariant diffusion process that samples amorphous material structures from a target energy function, aiming to bypass the rare barrier crossings that make low-temperature molecular simulation slow. Across model glass and metallic-glass systems, it reportedly matches parallel-tempering distributions and thermodynamics with under 0.2% free-energy error using over 500× fewer energy evaluations, while also supporting composition transfer, steering, and inverse design.
The combination of energy-based diffusion sampling, amortization across compositions and temperatures, and demonstrated large reductions in expensive energy evaluations is a genuinely notable direction, but the abstract alone does not establish how broadly the method beats strong existing samplers or whether the ambitious “foundation model” framing generalizes.
-
Eversion-based robots can enable safe access,steering and endoscopic imaging within the spinal subarachnoid space
The authors build a 2 mm pressure-driven eversion robot that grows forward at its tip instead of being pushed through tissue, reducing friction and shear in the spinal subarachnoid space. In phantoms it reduced mean and peak interaction forces by 65.2% and 48.0%, and in an intact human cadaver it achieved 150 mm of controlled extension with endoscopic and fluoroscopic imaging without visible macroscopic injury.
This is a genuinely new and clinically relevant navigation mechanism demonstrated over multiple vertebral levels in intact human anatomy, but the evidence is still limited to phantoms, modeling, and cadaver testing rather than physiological or in vivo studies.
-
Fast Wave-optics Rendering of Multiplane Images for 3D Holographic Displays
The paper introduces a wave-optics hologram renderer that operates on multiplane images, converting modern 3D scene representations into content for holographic displays. It reports up to 250,000× faster rendering than primitive-based computer-generated holography at comparable quality, while improving image quality over conventional layer-based methods, with validation in simulation and on captured hardware.
The combination of MPI scene representations with efficient wave-optics hologram synthesis—and especially the very large reported speedup plus experimental validation—looks substantially more than an incremental CGH improvement, though the abstract does not establish how fair or broadly applicable the comparisons are.
-
Beyond Fail-to-Pass: Iterative Hardening of Co-Generated Bug Reproduction Tests and Fixes
The paper argues that a bug-reproduction test should do more than fail on the buggy code and pass on the known fix: it should also reject plausible incorrect patches. It identifies lax tests and test–fix error coupling in co-generation, then proposes CoHarden, which iteratively hardens generated tests against mutation patches; on SWE-bench Verified it reports 69.4% resolved and 78.9% fail-to-pass, with sizable gains over baselines.
The distinction between symptom-reproducing and repair-discriminating tests, plus the reported roughly 8–10 point gains, is a non-obvious and potentially useful direction for LLM-based program repair, though the evidence is limited to benchmark results and a specialized APR setting.
-
Synthetic minority data is redundant or invalid: a data-dependent validity theory and a de-biased test
This paper argues that synthetic minority examples should be evaluated against withheld real minority data rather than the same data used to generate them. It defines validity as population-level class membership probability, proves that class overlap imposes an unavoidable invalidity floor, and reports that across 91 oversampling methods synthetic data rarely improves F1 beyond trivial baselines while often worsening calibration. The authors release a test that requires synthetic data to show both validity and information gain.
The data-dependent validity framing and de-biased evaluation could challenge a widely used practice, with unusually broad empirical claims, but the abstract alone does not establish whether the estimator and conclusions generalize beyond the tested datasets and classifier choices.
-
The Dark Room in the Reward Channel: Dense Prediction Rewards Collapse GRPO-Trained LLM Agents -- and The Channel, Not the Content, Decides What Works
This paper studies why dense prediction rewards can destabilize GRPO-trained LLM agents, using 74 controlled experiments across several environments, model sizes, and reward-delivery mechanisms. It finds that difference-form rewards combined with standard within-group normalization can drive agents into an absorbing “dark room” behavior, while the same prediction signal as an auxiliary loss is often harmless; experimentally, reward-channel variance and hackability matter more than the semantic content of the signal.
The controlled separation of reward content from delivery channel, and the proposed within-group-variance explanation for catastrophic collapse, are potentially important for training LLM agents, but the claims remain concentrated in a narrow GRPO/ALFWorld-style setting and lack enough quantitative detail for a stronger verdict.
-
Learning What Matters: Supervising Global Context Pruning with Causal Evidence Sets
This paper argues that attention is a poor supervisory signal for long-context pruning because it often highlights information the answer does not causally depend on, including outdated facts. Using answer-preserving masking to recover causal evidence sets, the authors train routers that reach 99% routing accuracy on a two-step reference task across training runs, compared with attention-distilled routers ranging from 36% to 98%; similar conflicts appear in pretrained Qwen and Gemma models. The work also finds that a lightweight cross-layer trace can recover causal chains, suggesting the limitation is in the readout rather than necessarily in the model’s internal computation.
It presents a non-obvious and practically relevant failure of attention-based context pruning, backed by strong intervention-based results, but the evidence appears concentrated on retrieval-style tasks and a small set of models rather than broad real-world long-context workloads.
-
Optimal use of a black-box learner in semiparametric estimation
This paper studies estimation of a linear effect when the nuisance functions are learned by potentially misspecified black-box models. It proposes an estimator—and the TAME procedure for editing debiasing weights—that removes a mixed error term present in standard double machine learning, achieving a rate of 1/√n plus the product of approximation errors and the squared estimation error; a matching lower bound shows this is optimal. The ideas are also presented as extending beyond the partial linear model and to settings with multiple learners.
The optimal rate and adversarial moment-calibration construction appear to offer a genuinely nontrivial improvement over DML under nuisance-model misspecification, backed by a matching lower bound, though the contribution is mainly theoretical and its practical impact is not demonstrated in the abstract.
-
Similarity Is Not Logic: Factored Inference for Dual-Encoder Vision-Language Models
The paper argues that dual-encoder VLMs often represent concepts adequately but fail because their single similarity score combines evidence like a bag-of-concepts, ignoring logical operators such as negation and conjunction. It introduces LCSE, a training-free inference procedure that extracts concept scores from frozen encoders and executes constraints externally; on the proposed benchmark it substantially outperforms fine-tuned baselines and improves negation accuracy while preserving retrieval performance.
The interface-level diagnosis and training-free separation of concept extraction from constraint execution are meaningful, with unusually large gains on compositional and negation tests, but the evidence is still centered on author-created benchmarks and a relatively narrow dual-encoder failure mode.
-
Track-Leakage-Free Hold-Out Self-Validation for Photogrammetric Reconstruction: Protocol, Sensitivity, and Limits
The paper proposes a leakage-free hold-out test for judging photogrammetric reconstructions without ground truth: withheld images are re-localized using only 3D points built from the retained images. Across real captures and benchmarks, the test detects fragmented or failed reconstructions but remains maximally confident when the entire reconstruction is coherently and substantially distorted, including errors of tens of meters. It therefore measures internal geometric consistency, not metric accuracy.
This is a well-supported negative result that directly challenges the assumption that self-consistency can serve as ground-truth-free accuracy validation, though its impact is concentrated in photogrammetric inspection.
-
Predict before you train: Scaling Laws for particle physics foundation models
The paper fits compute-and-data scaling laws using only relatively small transformer models trained on collider jets, then tests whether those laws predict much larger models trained later. The forecast predicts held-out models trained with over 100 times more compute to within 1% loss, and lower pretraining loss tracks better downstream tagging performance; the released models and code also enable reproduction and further study.
It provides unusually strong evidence that scaling laws can forecast unseen, much larger particle-physics models and translate training compute into downstream physics performance, though the result is limited to one model family, corpus, and task domain.
-
A neural circuit for olfactory motion detection
This study identifies a Drosophila antennal-lobe circuit that detects the direction of odor movement by comparing bilateral odor signals with a roughly 40 ms delay. Connectomic, physiological, behavioral, and modeling results point to a GABAergic inhibitory neuron implementing delayed contralateral inhibition alongside ipsilateral excitation, producing direction-selective responses across many odor types. The notable advance is showing that an early olfactory circuit performs a computation analogous to visual motion detection.
It provides a concrete, biologically grounded mechanism for a previously unexplained sensory computation—olfactory motion detection—but the abstract does not yet establish how broadly this principle generalizes beyond flies or how much it improves navigation.
-
Immunoproteasome Deficiency Impairs Microglial Clearance and Worsens Tau and Amyloid Pathology
The study tests whether immunoproteasome induction in Alzheimer’s disease is protective or inflammatory by removing two immunoproteasome catalytic subunits in tauopathy and amyloid/tau mouse models. Deficiency worsened tau and amyloid pathology because microglia could engulf aggregates but failed to degrade the internalized material, accompanied by altered phagolysosomal and mitochondrial programs; related transcriptional patterns appeared in human datasets.
It provides a specific mechanistic resolution to the debate over immunoproteasomes by identifying a post-engulfment microglial degradation failure, supported across mouse genetics, longitudinal imaging, single-cell data, and human transcriptomic reanalysis, though the evidence remains largely preclinical and associative in humans.
-
Serotonin induces DOWN states across the anesthetized mouse forebrain
The study optogenetically activates serotonin neurons in the dorsal raphe while recording simultaneously from many forebrain regions in lightly anesthetized mice. Serotonin reliably triggered DOWN states across cortex and striatum, but not much in the midbrain; a connectivity-constrained model suggests that some effects arise through synchronized network dynamics rather than direct serotonin projections.
It offers a potentially important resolution to conflicting reports about serotonin’s global effect on brain state, supported by broad Neuropixel recordings and a mechanistic network model, but the result is still limited to optogenetic stimulation under anesthesia and lacks enough quantitative detail for a strong verdict.
-
Peripheral nerve injury reallocates primary afferent input through spinal parvalbumin microcircuits
This study shows that nerve injury changes how low-threshold touch afferents recruit spinal parvalbumin interneurons: signaling shifts away from inhibitory PV neurons and toward excitatory PV neurons, even though synapse numbers and connectivity remain intact. The authors identify a previously unrecognized Complexin-1 inhibitory circuit contributing to this switch and show that the two PV populations have opposing effects on mechanical hypersensitivity.
It provides a non-structural, functionally causal mechanism for touch-to-pain conversion and identifies a new inhibitory microcircuit, but the findings remain focused on a specific spinal pain circuit and are reported from a preprint abstract.
-
A predictive theory of experimental design for inferring neural population geometry in large-scale recordings
The paper develops a theory for predicting how reliably PCA-based neural population geometry can be estimated as the number of recorded neurons and trials changes. It argues, counterintuitively, that recording more neurons can reduce the number of trials needed for accurate geometric estimates, and derives related scaling laws for neural prediction, validated across species and recording modalities.
The proposed “blessing of dimensionality” and experimental-design scaling laws could materially change how large-scale neural recordings are planned, but the abstract provides no quantitative effect sizes and the theory is limited initially to PCA.
-
Developmental tuning of functional manifold dimensionality across the human brain
Using five naturalistic fMRI datasets spanning infancy to adulthood (781 people), the authors estimated the intrinsic dimensionality of activity across brain regions and tasks. They find that task-related regions become relatively high-dimensional because task-irrelevant regions undergo increasingly selective dimensionality compression, changing from broad and nonspecific in infants to localized and precise in adults. The result frames developmental specialization as mostly pruning irrelevant representational variation rather than adding task-specific complexity.
The large developmental sample and cross-modal result support a non-obvious selective-compression account of cortical specialization, though the correlational, cross-sectional fMRI analysis and dependence on a manifold estimator warrant closer scrutiny before treating it as a general developmental mechanism.
-
Reading and Steering Representations of Materials-Science Mechanisms in an Open-Weight Language Model
This paper probes an open-weight language model to distinguish whether materials-science knowledge is merely encoded in vocabulary-like features or is represented through relationships between internal states. Using hidden-state readouts, Jacobian-based probes, counterfactual prompts, and causal interventions, it finds that directional physical laws are more reliably reflected in state transitions than in static representations: 39 of 40 directional relations were oriented correctly, and interventions changed answer probabilities in the expected directions. It also shows that some apparent organization of mechanisms in hidden-state space can be explained by numerical similarity rather than physical structure.
The separation of readable concepts from causally meaningful, law-consistent state transformations is a genuinely interesting mechanistic-interpretability result, but the evidence is still narrow—one model, 60 relations, and limited intervention cases—so it is worth a look rather than a must-read.
-
OpenForgeRL: Train Harness-native Agents in Any Environment
OpenForgeRL is an infrastructure framework for training agents inside existing stateful, multi-process inference harnesses rather than simplified custom loops. It uses a proxy to record harness model calls for standard RL pipelines and isolated Kubernetes environments for rollouts; experiments on coding/tool-use and GUI agents show gains over similarly sized open baselines, with some results matching larger models, while error recovery remains weak.
The practically important contribution is making end-to-end RL over real agent harnesses and environments feasible, supported by diverse benchmark results, though the core advance is largely systems/infrastructure and the abstract does not establish a major capability breakthrough.
-
SWE-Pruner Pro: The Coder LLM Already Knows What to Prune
SWE-Pruner Pro uses a small prediction head on a coding agent’s own hidden representations to decide which lines of tool output to retain, rather than relying on a separate code-relevance classifier. Across two open-weight models and four multi-turn benchmarks, it reportedly cuts prompt and completion tokens by up to 39% without reducing task quality, and even improves results on SWE-Bench Verified and Oolong for one model.
The agent-internal pruning mechanism is a meaningful efficiency idea, and the reported token savings with maintained or improved performance are credible enough to warrant a look, though it appears specialized to coding-agent context management rather than a major new direction.
-
Scaling Laws for Hypernetwork-Based Knowledge Injection in Large Language Models
The paper trains a hypernetwork to convert factual examples into a fixed LoRA adapter for an LLM, then studies how injection quality changes with hypernetwork depth, width, and target-model size. Using a tens-of-millions-example multi-hop QA corpus, it reports power-law scaling and improving out-of-distribution factual reasoning as models grow, apparently outperforming ordinary LoRA and full fine-tuning on OOD evaluations.
The systematic scaling-law treatment of hypernetworks for train-time knowledge injection is a genuinely interesting direction, but the abstract gives no quantitative gains, model scales, or details sufficient to assess whether the claimed advantage is robust rather than benchmark-specific.
-
Oxygen-TryOn: Fashion-Native Foundation Model for Any-item Virtual Try-On
Oxygen-TryOn is a fashion-specific image generation model for virtual try-on, accepting multiple garment references—including product images and worn photos—and a target person image. It uses a large try-on data pipeline plus continued pretraining, supervised fine-tuning, and reward-based training, and claims broad support for different item types, body views, and multi-item outfits while preserving identity and garment details.
The unified, multi-item, non-mask-based formulation could represent a meaningful capability expansion over garment-specific try-on systems, but the abstract provides no quantitative results or independent evidence behind broad proprietary-model comparisons.
-
HPD-Parsing: Hierarchical Parallel Document Parsing
HPD-Parsing separates document parsing into a global layout stage and concurrent block-level content decoders, rather than generating the whole page in one autoregressive sequence. Combined with progressive multi-token prediction, it reports 4,752 tokens/s—2.62× faster than the prior fastest parser and 3.06× faster than a vanilla autoregressive baseline—while retaining competitive accuracy.
The hierarchical parallel decoding design is a meaningful efficiency direction for unified document VLMs, but the abstract provides limited detail on accuracy tradeoffs, hardware, and breadth of evaluation, so the claimed speedup is not enough for a strong recommendation.
-
Federated Lightweight Fine-Tuning
FLITE represents a model update as a small latent vector passed through a shared affine, low-rank generator, so clients can average latents while exactly averaging the corresponding weight updates. On CIFAR-100, it reduces communication to about 5 KB per client per round—an 8,718× reduction—while staying within roughly 0.5 percentage points of full-weight FedAvg, including under non-IID data; int4 latents reduce this further to 648 bytes without accuracy loss.
The exact latent-averaging construction combined with an additive pretrained-base formulation offers an unusually large communication reduction, but the evidence is mainly from one CIFAR-100/ResNet-18 setting, so its scalability and generality remain uncertain.
-
FlowBlock: Wavefront-Parallel Decoding for Self-Correcting Diffusion Language Models
FlowBlock accelerates block-wise diffusion language model decoding by allowing later blocks to begin from provisional upstream predictions, then correcting them as context improves. It uses gated wavefront scheduling and packs asynchronous requests efficiently, reporting up to 2.95–4.01× higher throughput, 53.6–77.1% lower latency, and a 1.3-point accuracy improvement over serial baselines, with much higher throughput than a training-based parallel method.
The paper presents a genuinely interesting training-free way to trade block finality for self-correction, with substantial claimed serving speedups, but the abstract gives only peak results and limited detail on workload breadth and reproducibility.
-
Autonomous Discovery of Wireless Communications Algorithms
The paper presents AITE, an LLM-guided evolutionary search system that automatically designs wireless communication algorithms while considering both accuracy and computational cost. On OTFS equalization it reportedly beats prior methods with 3.6× lower latency, and on pilot-free OFDM with a custom constellation it finds explicit algorithms matching neural receivers, suggesting that search can produce interpretable alternatives to learned physical-layer systems.
The combination of LLM-driven algorithm discovery with strong, lower-latency and interpretable wireless receivers is genuinely interesting, but the abstract lacks enough detail about search cost, baselines, generalization, and independent validation for a strong recommendation.
-
FF-ProCams: Feed-Forward Gaussian Splatting for Projector-Camera System
FF-ProCams predicts a relightable 3D Gaussian representation of a scene from just eight projector-camera views, rather than optimizing a separate scene representation at test time. It combines a Mamba/Transformer encoder with a projector-aware renderer and reports comparable or better rendering than optimization-based methods using 297 views, while reducing reconstruction time to 0.13 seconds; however, the training data is synthetic.
The feed-forward, projector-aware inverse-rendering setup and reported three-to-five-order-of-magnitude speedup are genuinely notable, but confidence is limited because the abstract emphasizes synthetic data and provides little evidence about real-world scenes or robustness.
-
Physical Self-Supervised Learning: IMU Sensing without Manual Labels
The paper proposes learning IMU sensing without manual labels by replacing the usual neural decoder with an adaptive decoder based on kinematic equations. It combines this with structured latent reconstruction, motion-disentangling frequency/spatial constraints, multi-view kinematic trees, and uncertainty modeling, reporting up to 5× lower tracking error and 4× lower motion-capture error in difficult cross-device, cross-placement, and cross-user settings.
The explicit use of learnable physical kinematics as the self-supervision mechanism, together with strong claimed label-free generalization gains, is genuinely interesting, but the abstract gives few quantitative details and bundles several familiar components, so it does not yet justify a strong verdict.
-
SAGE: Subgoal-Conditioned Action Generation for Latent World Model Planning
SAGE improves latent-world-model planning by predicting reachable latent subgoals at multiple time scales and using them to guide candidate action-sequence generation, rather than starting proposals randomly. On long-horizon PushT and OGBench Cube tasks, it reports large success-rate gains at a target offset of 150: 12.7% to 64.7% and 26.7% to 67.3%, respectively, while retaining short-horizon performance.
The subgoal-conditioned proposal mechanism addresses a real scaling bottleneck in latent planning and shows unusually large gains, but evidence is limited to two benchmark environments and the abstract does not establish broad generality or comparisons beyond the stated baselines.
-
The Price of Hidden Curvature: Improved Lower Bounds for Bandit Convex Optimization
This paper proves stronger minimax regret lower bounds for stochastic bandit convex optimization: $\Omega(d^{4/3}\sqrt n)$ when the horizon is large, exceeding the $d\sqrt n$ dependence for linear bandits, and $\Omega(\sqrt d\,n^{3/4})$ in an intermediate regime. Its construction hides the useful quadratic curvature behind an unknown geometric “tube,” forcing the learner to spend samples discovering where curvature can be observed; it also shows the classic Flaxman et al. rate is optimal in that regime.
The result gives a meaningful separation between general bandit convex optimization and linear bandits, plus an optimality result for a longstanding algorithmic rate, but the abstract alone provides only theorem-level claims without enough detail to assess the proof or robustness.
-
DocAtlas: Long-Document Understanding as Mutable-State Interaction
DocAtlas reframes long-document question answering as interaction with a mutable external workspace rather than one-shot retrieval from a fixed index. Its harness lets an agent search, selectively read, organize evidence in a hierarchy, take notes, and review them under a context budget; it improves GPT-5.4 to 71.4% on MMLongBench-Doc and raises a Qwen3.5-4B VLM from 54.4% to 63.7% after end-to-end RL.
The mutable-state document environment and demonstrated gains for a compact VLM are substantially more interesting than ordinary RAG tweaks, but the evidence is limited to one benchmark and the abstract does not establish broad generality or disentangle the individual components.
-
Milo, a Fully Autonomous Indoor/Outdoor Robotic Guide Dog
The authors build Milo, a roughly $2,000 modified Unitree Go2 that autonomously guides a person indoors and outdoors without prior maps, external computation, or infrastructure. It combines onboard perception and voxel mapping with a learned obstacle-avoidance policy, and in indoor/outdoor obstacle courses it navigates more smoothly and causes fewer handler collisions than a costmap baseline; the hardware and software are open source.
This demonstrates a potentially important real-world capability at unusually low cost and with strong autonomy claims, but the evidence is limited to obstacle-course comparisons rather than sustained deployment with blind users or demanding natural environments.
-
Nuclear Quantum Effects as a Denoising Problem
The paper reframes quantum nuclear sampling as a denoising problem: a denoiser learned only from classical Boltzmann statistics is combined at sampling time with an analytic Gaussian term encoding quantum effects. The authors claim this gives exact transfer across temperature, isotope mass, dissipation, and imaginary-time boundary conditions without retraining, including recovery of displacement and momentum distributions from open paths.
The potentially important idea is that one learned denoiser can be reused across many quantum contexts because the context enters through a known quadratic noise structure, but the abstract provides limited numerical evidence and no indication yet of scale or practical advantage over existing path-integral methods.
-
OSVE: One Step Video Editing with One Step Diffusion Models
OSVE adapts one-step text-to-image diffusion models to text-guided video editing. It uses a learned per-frame noise encoder, a structure-aware training loss, cross-frame latent attention, and anchor-based sliding windows to preserve geometry and temporal consistency. The authors report quality comparable to or better than multi-step editors while being roughly 155–171 times faster.
The combination of one-step inversion and temporally consistent video editing could be a substantial efficiency advance, but the abstract gives no benchmark details or independent evidence to validate the very large speed and quality claims.
-
HOST:Robots Acquire Manipulation Skills in Seconds from a Single Human Video
HOST lets a robot learn a new manipulation skill at inference time from one human demonstration video, without conventional robot-data fine-tuning. It predicts task progress, converts the demonstrated future into expected robot observations, and uses those predictions to generate actions; across its evaluation it reports 62% average success, with skill acquisition taking about 29 seconds and prior skills retained. The claimed advantage is substantial compared with zero-shot and 50-demonstration fine-tuning baselines, though the abstract does not specify the task breadth or evaluation conditions.
One-shot, seconds-scale transfer from human video with retention of existing skills is a genuinely interesting direction, but the moderate success rate and unspecified experimental scope make the headline comparisons insufficient for a strong verdict.
-
Generative AI floods and dilutes the market for books
The authors track 14,419 self-published genre-fiction books on Amazon, using full-text AI detection and daily sales data, and find that books with substantial detected AI content are gaining market share despite selling less well on average. The number of books generating sales grew much faster than total revenue, reducing revenue per selling book; AI-heavy books also increasingly occupy top ranks and show more overlap with existing books among bestsellers. The paper argues that generative AI can reshape publishing through sheer production scale rather than superior quality.
This is a potentially important, non-obvious empirical account of generative AI’s market impact, but the conclusions depend heavily on the validity of AI-text detection and observational sales comparisons, which the abstract does not establish.
-
Nova3D: Code-Native Generation of Programmable 3D Assets
Nova3D generates 3D assets as executable Blender programs rather than only producing opaque meshes. On a 54-item benchmark, it reliably produced named part hierarchies, numeric constraints, editable local components, and articulated joints, while retaining competitive visual quality; the main tradeoff was weaker texture realism than baked-PBR systems. The key contribution is making semantic structure and programmability native to generation instead of recovering them afterward.
The code-native representation appears to unlock several capabilities that mesh generators and post-hoc segmentation do not, but the evidence comes primarily from a small, likely author-defined benchmark, so it is promising rather than yet a must-read result.
-
Spectral theory for population density dynamics of spiking neurons with refractoriness
The paper develops a mathematically rigorous population-density model for spiking neurons with a finite absolute refractory period, treating refractory history as part of the state. It derives the generator’s spectrum and an exact linear-response transfer function, including threshold-noise terms omitted by earlier heuristic treatments, and shows in a mean-field model that refractoriness can promote stable firing-rate oscillations.
This addresses a known gap in population-density theory with a genuinely new operator framework and a non-obvious prediction that refractoriness can facilitate, rather than merely suppress, collective oscillations, though the abstract provides little quantitative validation beyond the theoretical derivations.
-
GaugeQuant: Online Learning of Quantization-Optimal Bases from LLM Symmetries
GaugeQuant uses the continuous basis-change symmetries of transformer layers to learn an internal representation that is easier to quantize, without changing the language-modeling objective. It adds a LogSumExp regularizer and updates only the symmetry transformations; on LLaMA-2 7B, it reports large perplexity improvements under W4A4 and especially W4A16 quantization, without calibration data or quantization simulation.
The use of transformer gauge symmetries as an online, training-time route to quantization robustness is a genuinely interesting direction, but the abstract gives limited experimental context and the unusually large W4A16 gain needs careful verification.
-
Sparse Concept Channels in Frozen 3D CT Vision Encoders
The authors probe frozen 3D medical vision encoders and find that individual radiological findings are concentrated in roughly 10 channels, whose ablation selectively removes the corresponding prediction while preserving unrelated findings. The same pattern appears in two architecturally different CT vision-language models, and their training-free probe improves report-generation metrics over CT-CHAT while running 22× faster.
The sparse, selectively causal concept-channel finding replicated across unrelated 3D medical encoders is genuinely interesting and potentially useful for interpretability and efficient clinical inference, but the abstract lacks enough dataset, statistical, and comparison detail to justify a strong recommendation.
-
Out of Sight, Still in Mind: Token Compression for Omni-LLMs
ReMo is a training-free method for reducing visual-token overhead in omni-modal language models. It removes visual information that is redundant with audio or other visual tokens, and replaces some object-level visual tokens with short text descriptions; on Qwen2.5-Omni, it cuts 54% of input tokens across two model sizes without reducing average accuracy on five audio-visual benchmarks.
The cross-modal redundancy criterion and text proxies offer a nontrivial route to substantially cheaper omni-modal inference, but the abstract provides limited detail and benchmark evidence, so the claimed slight improvement over full tokens needs verification.
-
ARGON: A GNN-Empowered Compilation Framework for Scalable Neutral Atom Computing
ARGON separates neutral-atom compilation into offline geometric conflict resolution and online temporal routing. It precomputes hardware-valid layouts, uses a GNN to select layouts with awareness of future routing bottlenecks, and then performs collision-free atom transport; the authors report compilation speedups up to 10^4× and fidelity improvements up to 100× on dense circuits.
The spatiotemporal decoupling strategy and long-horizon learned layout selection could address a real scaling bottleneck, but the abstract gives only headline ‘up to’ improvements and insufficient detail about circuit families, baselines, and tradeoffs to justify a strong recommendation.
-
How Many Bits Can an Adapter Write? Measuring the Capacity and Memorization of Parameter-Efficient Fine-Tuning
The paper measures how many bits of training information LoRA adapters actually encode, rather than inferring capacity from their parameter count. It finds that adapters store only a few bits per trainable parameter, with capacity depending strongly on layer placement—MLP adapters reportedly store nearly twice as much as attention adapters—and that supervised fine-tuning can memorize examples while verifiable-reward training does not show the same verbatim recording.
The bit-level measurement framework and the finding that adapter location, rather than parameter count alone, controls memorization are genuinely interesting, but the abstract gives too few quantitative details and validation results to justify a strong verdict.
-
Mean-to-Score Discrete Diffusion: Posterior-Mean Denoisers for Score Entropy
The paper argues that SEDD’s unconstrained score ratios can be mathematically inconsistent with any valid clean-token posterior, causing invalid intermediate sampling weights. It proposes mean-to-score (M2S), which predicts a posterior mean and maps it through the known forward CTMC kernel to obtain realizable scores; this improves perplexity and FID on CIFAR-10 and gives substantially lower generative perplexity than several discrete-diffusion baselines at matched sampling budgets.
The posterior-realizability diagnosis and exact constrained parameterization are a meaningful advance, with large reported language-model gains, but the abstract has incomplete CIFAR FID results and limited detail on the breadth and comparability of the evaluations.
-
AREX: Towards a Recursively Self-Improving Agent for Deep Research
AREX is a deep-research agent that alternates between gathering evidence and auditing its provisional answer against individual constraints, then performs targeted follow-up research on unresolved claims. It also learns a context-compression tool to preserve verified evidence and open issues over long interactions, and trains 4B and 122B-MoE versions with long-horizon reinforcement learning. The abstract reports broad benchmark improvements, but gives no quantitative results or details sufficient to assess how much comes from the recursive procedure versus training and scale.
The explicit discovery–verification recursion and learned improvement-state mechanism are plausible, nontrivial additions to research agents, but the abstract lacks numbers and ablations, so the claimed broad gains are not yet compelling.
-
Surprisal Theory is Tautological (without Rational Grounding)
The paper argues that surprisal theory, stated as “processing difficulty is an affine function of surprisal,” is unfalsifiable because any difficulty pattern can be represented by a suitably chosen language model. It says the theory becomes testable only when the relevant model is independently constrained—for example by memory limits or processing goals—rather than selected for corpus fit or tuned against behavioral data.
This is a potentially important conceptual reframing of a central connection between language models and human processing, but the abstract supports it mainly with a theoretical argument and does not yet establish how the proposed rational grounding would work empirically.
-
Addressing the Orchestration Gap in Generalist Robots via Physical Agency
The paper proposes placing a closed-loop high-level orchestrator above frozen vision-language-action policies and predefined skills, rather than expecting one learned policy to handle planning, monitoring, recovery, and control. Its Pigey system reportedly improves LIBERO-PRO from 12.8% to 53.3% without task-specific fine-tuning and raises a real robot’s success rate from near-zero to over 90% on reasoning-limited tasks.
The decomposition of generalist robot behavior into reusable motor skills plus a physical, failure-aware orchestration loop—and the large reported gains without new training—are genuinely interesting, but the abstract does not provide enough experimental detail to establish how broadly or robustly the results generalize.
-
RED-PIM: Reducing Data Movement for Transformers using Processing-in-Memory
RED-PIM is an algorithm–hardware co-design for running Transformer attention in processing-in-memory systems. It reorganizes attention computations to reduce inter-memory-bank communication from quadratic to linear in sequence length and avoids explicitly materializing the full N×N attention matrix, reporting large inference-time reductions—especially for long sequences—while preserving accuracy.
The claimed reduction in attention communication and intermediate storage could be a substantial advance for long-context PIM inference, but the abstract does not clarify the exact attention approximation/reformulation, hardware assumptions, or breadth of baselines enough to support a stronger verdict.
-
Listen, Do Not Copy: Internalizing Audio-Grounded Scaffold Context for Robust Omni-Model Speech Understanding
The paper identifies a “perception bypass” problem in omni-model speech systems: providing scene descriptions can improve transcript scores by letting the model copy answer-bearing text rather than listen, especially in overlapping or noisy speech. It proposes Audio-Grounded Scaffold Context, which uses non-answer audio-derived clues during training, tests for leakage and audio dependence, and removes the clues at inference; across three omni-models, reported no-clue capped permutation WER falls from 25–71% to 9–15%, with little inference overhead.
The combination of explicit perception-bypass diagnostics with training-time audio-grounded scaffolding that transfers to clue-free inference is a non-obvious contribution, and the reported multi-model WER reductions are large, though the abstract lacks enough experimental detail to establish how robust or general the gains are.
-
Generative Video Compression with Adaptive Score Distillation
The paper presents a video compression diffusion model trained specifically for compression rather than adapted from text-to-video models. It introduces Adaptive Score Distillation, which suppresses misleading distillation updates that otherwise produce motion-stalled one-step reconstructions; the resulting 478M-parameter codec reportedly achieves large bitrate savings at matched perceptual quality and decodes 1080p video at 15.1 fps on an A100.
A compression-native diffusion model plus a concrete diagnosis and fix for motion collapse is a substantial, relevant direction, but the abstract provides limited evidence beyond comparisons to one prior codec and perceptual metrics.
-
One Hand Watches The Other: Dynamic Multi-Agent Cooperation for Sample-Efficient Bimanual Manipulation in Dynamic Environments
DynaMAC treats the other robot arm—or a manipulated moving object—as a dynamic task parameter rather than an external reference frame, allowing a policy-agnostic multi-stream controller to handle interaction and bimanual coordination without a fixed leader. The authors report over 35 percentage points higher performance than probabilistic and generative baselines with 20× fewer samples, including zero-shot transfer from static demonstrations to dynamic environments; they also introduce the DynaBench evaluation suite.
The causal reframing and claimed zero-shot static-to-dynamic transfer could be important for sample-efficient bimanual manipulation, but the abstract gives too few experimental details to trust the unusually large gains at the strongest level.
-
Deconstructing Off-Policy Ratios: Entropy-Scaled Trust Regions for Asynchronous Reinforcement Learning
The paper argues that off-policy importance-ratio errors in asynchronous RL have different meanings at different token entropies: low-entropy tokens amplify sampling noise, while high-entropy tokens reflect useful exploration from in-flight updates. It introduces an entropy-scaled trust-region rule that adjusts correction thresholds per token without extra forward passes, and reports matching synchronous GRPO while achieving a 2.6× speedup on BrowseComp-Plus and multi-turn GSM8K.
The entropy-dependent interpretation of stale-policy ratios is a non-obvious and potentially useful idea for making asynchronous LLM RL stable, but the abstract gives limited detail about baselines, ablations, and breadth of validation, so the strong claims do not yet warrant a strong verdict.
-
Practical advantage beyond the quadratic speedup limit with fully-quantum walks
The paper proposes a fully quantum Metropolis walk whose proposal mechanism uses Hamiltonian simulation rather than quantizing a classical Markov chain. For low-temperature sampling of dense Ising models, it claims a sixth-degree polynomial query advantage over the best classical walk and estimates that fault-tolerant hardware could reduce the conventional quantum-walk crossover from roughly 1,000 years to under one day under matched hardware assumptions.
The genuinely quantum proposal mechanism and claimed beyond-quadratic advantage are notable, but the practical result is an architecture- and resource-estimation claim for a specialized Ising-sampling problem rather than demonstrated hardware advantage.
-
Multi-primitive in-memory computing for Monte Carlo tree search
The paper proposes decomposing Monte Carlo tree search into hardware-native primitives so that selection, expansion, rollout, and backpropagation all execute on-chip using different in-memory-computing components. A 22 nm evaluation using fabricated RRAM parameters estimates about 60 mW for 9×9 Go—roughly 96× more energy-efficient than a CPU and 65–2,059× than an H100—while matching open-source Go engines within sampling uncertainty; the substrate is also demonstrated across eight applications.
The phase-to-primitive approach is a potentially important way to make irregular algorithms compatible with in-memory computing, and the projected efficiency gains are substantial, but the central results appear to be hardware-model estimates rather than a fully demonstrated fabricated system.
-
What CLIP Knows but Cannot Say: Recovering Negation from Frozen Intermediate Features
The paper argues that CLIP’s middle text-encoder layers represent compositional syntax such as negation, but that this information is discarded in the final embedding as visual alignment increases. PeakPatch taps the intermediate “compositional peak” and learns small embedding and score corrections while keeping CLIP frozen; it reports large gains on negation benchmarks, out-of-distribution retrieval, text-to-image generation, and several backbones.
The intermediate-layer diagnosis and post-hoc recovery of negation are genuinely interesting and potentially broadly useful, but the abstract does not establish whether the gains hold beyond the authors’ benchmark and claimed transfer settings.
-
Breaking the Total Variance Barrier: Sharp Sample Complexity for Linear Heteroscedastic Bandits with Fixed Action Set
This paper studies stochastic linear bandits whose observation noise varies substantially over time, focusing on a fixed action set. It introduces variance-adaptive exploration and elimination methods whose simple-regret bounds depend on a near-harmonic mean of the per-round variances rather than only on their cumulative sum, and proves a nearly matching lower bound.
The claimed lower-bound-supported break from the usual square-root cumulative-variance dependence is a genuinely non-obvious theoretical advance, but the abstract gives no quantitative comparisons or experimental evidence to establish how broadly consequential it is.
-
Coexisting but dissociable place and spatial view codes in the primate hippocampus
The authors recorded macaque hippocampal neurons during navigation and used a statistical model to separate coding for the animal’s current position from coding for the visual region it was viewing. They found distinct position, view, and conjunctive cell populations, with robust rodent-like place coding alongside stable spatial-view coding. This suggests that earlier reports of mainly view-based coding in primates may have reflected task and measurement choices rather than a fundamental species difference.
The direct dissociation of place and view codes in primate hippocampus addresses a longstanding cross-species debate and could revise how primate spatial representations are interpreted, but the abstract provides no sample sizes or quantitative effect sizes to establish how robust or general the result is.
-
APOE4 disrupts the central dogma by arresting neuronal proteome dynamics
Using isogenic human iPSC-derived neurons, the authors compared APOE3 and APOE4 across transcription, translation, protein abundance, and protein turnover. They report that APOE4 causes relatively modest RNA changes but disrupts translation and both lysosomal and proteasomal clearance, leading to globally longer-lived and accumulating proteins before a senescence-like state appears. The proposed new mechanism is that defective neuronal proteome renewal is an early consequence of APOE4 and may drive vulnerability in Alzheimer’s disease.
The cross-layer evidence for an APOE4-linked failure of proteome renewal, preceding overt stress, is a potentially important mechanistic reframing, but the abstract gives no quantitative effect sizes or in vivo/clinical validation to justify a strong recommendation.
-
Astrocyte-driven small vessel disease is an early, amyloid-independent feature of PSEN1 E280A familial Alzheimer's disease
The study reports that cerebral small-vessel abnormalities appear early in PSEN1 E280A familial Alzheimer’s disease, before amyloid plaques or tau tangles. Human tissue, PSEN1 knock-in mice, proteomics, and single-nucleus RNA sequencing all point to abnormal astrocytes, blood–brain barrier structures, extracellular matrix, and RNA-associated pathways as possible direct consequences of presenilin dysfunction.
The combination of human and amyloid/tau-free mouse evidence supports a potentially important shift from viewing small-vessel disease as downstream pathology to viewing it as an early, presenilin-linked process, though the abstract provides limited quantitative and causal evidence.
-
Multilevel engineering of cyanobacterial energy metabolism advances photosynthetic hydrogen production while revealing its constraints
The authors systematically identify successive bottlenecks limiting sustained hydrogen production in engineered cyanobacteria. They combine hydrogenase overexpression with synthetic CP12-based regulators that redirect photosynthetic electrons away from carbon fixation, reporting more than a 100-fold improvement over a prior benchmark while also observing metabolic adaptation and genetic instability. The main new insight is that high production requires coordinated control of enzyme capacity, energy/redox balance, and competing electron sinks rather than fixing one bottleneck.
The reported two-order-of-magnitude gain and explicit demonstration of hierarchical, state-dependent constraints are notable, but the abstract provides no absolute production rates, durability details, or breadth of validation, so this is promising rather than clearly a must-read.
-
Sensory experience and mTORC1 interplay orchestrates the maturation of cortical interneuron connectivity and tactile sensitivity
In a mouse model with mTORC1 hyperactivation in parvalbumin interneurons, the authors find reduced PV-cell connectivity and activity alongside stronger but less temporally precise tactile cortical responses and abnormal tactile and social behavior. Brief whisker trimming during the third postnatal week prevented these circuit and behavioral abnormalities, suggesting that sensory experience interacts with mTORC1 during a developmental critical period to shape interneuron maturation.
The non-obvious finding is that a narrowly timed reduction of sensory experience can prevent lasting PV-interneuron, tactile, and social abnormalities caused by mTORC1 dysregulation, though the abstract provides limited detail on model scope and experimental strength.
-
Embodied reinforcement learning in the primate cortico-basal ganglia system
The authors recorded macaque neural activity across eight cortico-basal-ganglia regions while animals learned the same stimulus–reward associations using either saccades or reaches. Value-related population activity occupied largely different dimensions for the two motor systems, even in limbic areas, so value could not be reliably decoded across actions. This argues that neural value representations are embodied and action-dependent rather than a single motor-invariant RL signal.
The circuit-wide evidence for motor-specific value coding challenges a central assumption in biological RL and has a direct connection to embodied AI, but the abstract provides limited quantitative detail and the result does not yet establish how RL models should be revised.
-
Persistent microbial material contributes to Alzheimer disease and is targetable by vaccination
The study reports that the microbial polysaccharide PNAG is present near amyloid plaques in human Alzheimer’s tissue and can trigger TLR2-dependent inflammatory signaling in neuronal and 3D brain models. In APP/PS1 mice, vaccination against PNAG improved cognition and reduced neuroinflammation and amyloid pathology while preserving gut microbiota, suggesting a possible microbial contribution to Alzheimer’s pathology and a vaccine-based intervention.
The proposed link between persistent microbial material, Alzheimer’s pathology, and a vaccine-responsive mechanism is genuinely interesting and potentially important, but the evidence is mainly preclinical and the abstract does not establish that PNAG is causal in human disease.
-
Non-invasive tracking of ripple-like activity during human sleep using MEG
The study uses source-resolved MEG to detect fast, ripple-like events in medial temporal regions during human NREM sleep, and examines them during targeted memory reactivation. These events occur at expected points in slow-oscillation and spindle cycles and are associated with better cue-specific memory reactivation, suggesting a non-invasive way to study hippocampal–cortical consolidation dynamics.
Non-invasive detection of hippocampal ripple-like activity linked to memory reactivation would be a substantial advance, but the abstract provides no quantitative results and the validity of source-resolved MEG for deep, high-frequency hippocampal events needs careful scrutiny.
-
Early-life medial pulvinar disruption drives schizophrenia-relevant prefrontal inhibitory and cognitive deficits in primates
The study lesioned the medial pulvinar in neonatal versus adult marmosets and tracked prefrontal development and behavior. Only early-life lesions caused adult working-memory deficits, along with impaired thalamic input to layer 3 parvalbumin interneurons, weaker gamma activity, and immature inhibitory physiology, suggesting that thalamic disruption can causally derail prefrontal maturation.
This is a plausible and potentially important developmental mechanism linking thalamic circuitry to schizophrenia-relevant cortical dysfunction, but the abstract provides limited detail on sample sizes, lesion specificity, and the breadth of behavioral and physiological evidence.
-
Causal network structure predicts memory organization and neural reinstatement across events
Using fMRI while participants watched and recalled a TV show with five interleaved storylines, the authors separated temporal order from causal relationships between events. Recall order, neural similarity in default-mode regions, and event-boundary reactivation were better explained by causal structure than by time, space, or semantic similarity, suggesting that the brain links causally related events across temporal gaps.
The paper provides a non-obvious challenge to time-centered accounts of episodic-memory organization by showing that causal structure predicts both reinstatement and representational transitions, though the evidence comes from a single naturalistic fMRI paradigm with limited detail on breadth and replication.
-
Learning stabilizes temporal activity but not neuronal selectivity in prefrontal cortex
The authors tracked medial prefrontal neurons in mice over months while the animals learned an association task with changing rules. Learning made the timing of neuronal activity increasingly stable, but individual neurons continued to change which task variables they represented; analysis suggested this was structured recombination of a stable set of task representations rather than random drift.
The dissociation between stable temporal activity and flexible neuronal selectivity, together with a proposed recombination mechanism, is a non-obvious result relevant to how brains support continual learning, but the abstract provides limited quantitative and validation details.
-
CNIH3 is a molecular signature of slow AMPA receptors
The study identifies the auxiliary AMPA-receptor protein CNIH3 as the molecular basis of unusually slow AMPA currents in hippocampal CA1 neurons. Reducing CNIH3 eliminated these currents in ventral CA1, while overexpressing it induced mosaic slow responses in dorsal CA1, suggesting that sparse CNIH3 expression can convert some synapses from rapid signal transmission into stronger, longer-lasting inputs.
This provides a specific molecular explanation for a previously underappreciated form of AMPA-receptor diversity and suggests a potentially important change in how synapses influence spike generation, but the abstract gives limited quantitative and cross-region evidence for the broader claims.
-
Spinal rotational dynamics orchestrate locomotor recovery
The authors recorded motor-unit activity from 16 hindlimb muscles in walking cats, including after complete spinal cord injury, and analyzed the population activity as a low-dimensional dynamical system. They report that activity continuously rotates through the locomotor cycle and persists after injury, supporting an intrinsic spinal cyclic attractor rather than a rhythm generated primarily by sensory feedback or simple flexor–extensor alternation.
The proposed shift from modular flexor–extensor control to an autonomous rotational spinal population dynamic is a potentially important mechanistic reframing, but the abstract gives limited quantitative or cross-condition evidence and the result is not yet enough to warrant a strong recommendation.
-
Distinct forms of dopamine transmission control locomotion and learning
This study argues that striatal dopamine operates through two separable release modes: diffuse extracellular dopamine that supports circuit excitability and locomotion, and spatially restricted, point-to-point signaling that preserves receptor activation, spine density, and learning. By selectively eliminating diffuse dopamine accumulation, the authors dissociate movement from motor and associative learning, suggesting that the spatial geometry of neuromodulator release helps multiplex behavioral functions.
The proposed functional split between diffuse and point-to-point dopamine transmission could substantially change how dopamine multiplexes movement and learning, but the abstract provides no quantitative detail about the selectivity, scale, or robustness of the manipulations.
-
AutoIndex: Learning Representation Programs for Retrieval
AutoIndex searches over executable document-transformation programs—such as slicing, enriching, reweighting, and reorganizing text—rather than tuning the retriever itself. On eight heterogeneous retrieval tasks with BM25 fixed, it improves average Recall@100 by 8.4% and nDCG@10 by 8.3% over indexing full documents, with larger gains on some tasks.
The explicit optimization of document representations through validation-guided program search is a meaningful framing shift, and the reported gains are substantial, but the evidence is limited to eight benchmark tasks and an agent-driven search procedure whose generality and cost are unclear.
-
Streaming Multi-Agent Autoregressive Diffusion Model with World State Registers
WorldWeaver adds persistent, shared “world state registers” to a streaming video diffusion model, rather than relying only on past frames as context. The registers track global scene information and each agent’s status across generated chunks and views, with separate transformer components for state and visual modeling; experiments on two-agent Minecraft generation report better consistency and quality.
Explicit persistent state for multi-agent video generation is a worthwhile architectural direction, but the evidence described is limited to two-agent Minecraft experiments with no quantitative results in the abstract.
-
ISO: An RLVR-Native Optimization Stack
The paper proposes Isospectral Optimization (ISO), which keeps each weight matrix’s singular values fixed while optimizing its singular-vector frames during RL with verifiable rewards. It applies this both to data-free merging of specialist models and to online RLVR training, reporting comparable accuracy in fewer steps—for example, ISO-AdamW matches AdamW’s 0.495 aggregate accuracy in 100 rather than 270 steps and reaches 0.509 after 210 steps on Qwen3-8B-Base.
The fixed-spectrum, frame-only view of RLVR optimization is a substantive and potentially useful new direction, but the abstract provides limited evidence beyond reported runs and does not yet establish broad gains or a clear explanation of why spectral inheritance works.
-
IDEAgent: Agentic Quality-Diversity Search for Research Idea Generation
IDEAgent treats LLM-based research ideation as a quality-diversity search rather than optimizing idea quality or novelty separately. Its multi-agent system evolves idea lineages, using iterative critique and comparisons with prior and rejected ideas, and reports 3.89x higher Yield than the strongest baseline across 32 topics in eight CS domains; the paper also introduces the Yield metric for counting diverse ideas above a quality threshold.
The quality-diversity framing and explicit joint metric are a meaningful conceptual contribution, but the headline gains depend on subjective idea-quality/diversity evaluations and the abstract does not establish that the generated ideas lead to better real research outcomes.
-
Novel Claim or Déjà Vu? Rethinking "Contamination-Free'' Dynamic Evaluation for Multimodal Automated Fact-Checking
This paper tests whether using post-training-cutoff claims really makes multimodal fact-checking evaluations contamination-free. It finds that 17.09%–29.30% of supposedly novel claims may still be answerable from pre-cutoff public knowledge, inflating Macro-F1 by up to 11.34 points and changing system rankings. It proposes stricter contamination controls and guidelines for dynamic evaluation.
The quantified finding that dynamic, post-cutoff benchmarks remain substantially contaminated—and can alter model rankings—is a non-obvious and practically important caveat, though the work is primarily an evaluation audit rather than a new modeling direction.
-
K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs
The paper introduces a curriculum-structured knowledge graph from K–12 textbooks, plus a benchmark and training corpus testing prerequisite relations, concept neighborhoods, evidence links, and visual grounding rather than just exam-question answering. Current models perform poorly on these tasks, while relatively small graph-guided supervision improves both text and vision-language evaluation, with mixed text-and-visual training outperforming either alone.
It is primarily a benchmark and dataset release, but the curriculum-graph framing and reported data-efficient gains across several educational and multimodal evaluations make it more than a routine domain application.
-
AlayaWorld: Interactive Long-Horizon World Modeling -- Full Technical Report
AlayaWorld is a 15B video diffusion-transformer world model that autoregressively generates interactive 540p/720p video at 24 fps while conditioning on camera motion and changing text prompts. It combines persistent visual memory, geometry-aligned history, training on its own corrupted rollouts, and four-step distillation to reduce long-horizon drift and inference cost; the abstract reports best results on iWorld-Bench but gives no quantitative comparisons or latency details.
Interactive long-horizon video generation with persistent memory and four-step inference is directly relevant and potentially important, but the abstract mainly describes a large full-stack system and does not provide enough numbers to establish a major capability or efficiency jump.
-
NVIDIA-labs OO Agents: Native Python Object-Oriented Agents
NOOA proposes representing an LLM agent as an ordinary Python object: methods are callable actions, fields hold state, docstrings and type annotations guide the model, and unfinished methods are implemented through an LLM loop. It combines typed interfaces, live object references, executable code actions, explicit state, and programmable control flow, then reports tests on agent benchmarks including SWE-bench Verified, Terminal-Bench 2.0, and ARC-AGI-3, but the abstract gives no comparative results.
The object-oriented agent interface is a potentially useful and adoptable programming abstraction, but the abstract provides insufficient quantitative evidence that it improves capability, reliability, or efficiency over existing agent frameworks.
-
Token-Level Off-Policy Learning for Faithful Generation Under Distribution Shift
The paper reframes off-policy post-training as token-level classification: instead of training directly on responses from another policy, it predicts which tokens are correct or incorrect and uses that signal to steer generation. On summarization across 11 out-of-distribution datasets and on machine translation, this reportedly generalizes better than sequence- and token-level baselines; the learned LoRA updates also appear interpretable as classifiers or steering vectors.
The token-level correctness-prediction formulation and claimed cross-dataset transfer are a meaningful departure from standard off-policy generation training, but the abstract gives no quantitative gains or details sufficient to establish a major advance.
-
dRAE: Representation Autoencoder with Hyper-Spherical Codes
The paper proposes Hyper-Spherical Quantization (HSQ), which assigns visual representation codes by angular similarity rather than standard Euclidean distance, aiming to avoid codebook collapse caused by variation in feature magnitudes. In a discrete representation autoencoder, it reportedly maintains 100% codebook utilization and scales to a vocabulary of 131,072 codes while retaining reconstruction and semantic performance, but the abstract gives few quantitative comparisons.
The geometric reframing of visual feature quantization and claimed scaling to a very large, fully utilized codebook are plausibly useful, but the abstract lacks enough numerical or broad comparative evidence for a stronger recommendation.
-
RynnBrain 1.1: Towards More Capable and Generalizable Embodied Foundation Model
RynnBrain 1.1 is an embodied multimodal model family ranging from 2B to a 122B mixture-of-experts model. It adds contact-point prediction, native 3D grounding, and a cross-embodiment action interface, then evaluates the resulting models on three robot platforms and several spatial/embodied benchmarks. The abstract reports better benchmark and real-robot results than prior RynnBrain, Qwen-based policies, and other generalist VLAs, but gives no quantitative results or task details.
The unified 3D/contact grounding and cross-embodiment action design is a plausible useful direction with real-robot validation, but this appears primarily to be an incremental model release and the abstract does not substantiate the broad superiority claims with numbers.
-
Closing the Loop in Humanoid VLA: Persistent 3D Object Tokens for Verifiable Loco-Manipulation
The paper adds persistent, role-specific 3D object records to a humanoid vision-language-action policy, using the same object state both to generate whole-body actions and to verify geometric task outcomes. On a Unitree G1, this raises success from 39/80 to 71/80 across eight loco-manipulation task families, with the largest gains in tasks requiring objects to remain tracked through motion, contact, and occlusion; an external comparison also reports 44/50 versus 37/50 successes.
The substantial real-robot improvement and unified actionable/verifiable object state are notable, but the contribution appears to be a system-level object-centric tracking and checking design tested on a limited set of tasks rather than an evidently broad new VLA paradigm.
-
Sobek: Streaming Equivariant Tensor Product Convolutions
Sobek reorganizes equivariant tensor-product graph convolutions so edge-local intermediate tensors are streamed directly into receiver-side accumulators instead of being materialized. Its generated CUDA implementation supports forward, backward, and double backward, and reports 1.2–49.7× speedups, up to 99% lower peak memory, and workloads up to 100× beyond a prior backend’s capacity.
The execution-schedule reformulation appears to remove a major memory bottleneck in equivariant GNNs with unusually large measured gains, but the contribution is specialized to a relatively narrow operator family and GPU implementation ecosystem.
-
Operational Proto-Introspection in Looped Language Models: Process-Quality Taps, Executable Branching, and the Readout-Control Boundary
The paper tests whether a recurrent language model’s hidden state can reveal the quality of its unfinished computation, and whether that signal can be used to improve decisions. In a frozen 2.6B looped transformer, hidden-state probes predict success better than surface features, replicate across logic tasks and model variants, and improve selective prediction and candidate selection; however, attempts to use the signal for generative steering, adaptive compute, or parameter binding do not work. The authors frame this as decision-useful “proto-introspection” without demonstrated generative control.
The combination of replicated process-quality readouts, executable candidate selection, and clear negative results for causal generative control is a genuinely interesting empirical boundary, but the evidence remains limited to a specialized recurrent 2.6B model and task family.
-
On the Limits of Sampling-Based Reachability: Geometry, Dynamics, and Sample Complexity
This paper analyzes when sampled trajectories can recover a reachable set accurately in Hausdorff distance. It shows that under geometric regularity of the initial set and Lipschitz dynamics, recovery requires roughly \(\tilde{O}((e^{3LT}/r)^n)\) samples, while a matching lower bound shows that exponential dependence on state dimension and time horizon is intrinsic rather than an artifact of the estimator; experiments support the predicted scaling.
The matching minimax lower bound turns a familiar scalability concern into a formal limitation for sampling-based reachability, although it establishes a barrier more than a new practical capability.
-
ROMS-IMLE: A Minimalist Approach to Competitive Single-Step Generative Modelling
ROMS-IMLE revisits implicit maximum likelihood estimation with a relatively simple convolutional network to build a single-step image generator, avoiding adversarial training, transformers, and iterative denoising. The authors report an FID of 2.56 on ImageNet 256 while maintaining good precision and recall, suggesting that gradual diffusion-style transformations may not be essential for high-quality generation.
The potentially important result is a competitive single-step ImageNet generator built from a minimalist IMLE-plus-CNN design, but the abstract gives only one headline metric and insufficient comparisons or scaling evidence to justify a stronger verdict.
-
Diffusion ReRoll: Revisable Denoising for Robotic Sequential Prediction
The paper introduces a diffusion-based sequence predictor that selectively re-noises already-stable portions of a predicted trajectory, letting earlier and later parts revise each other instead of following one irreversible denoising pass. Across planning, diffusion-policy action prediction, and unified video-action modeling, it reports substantial relative success gains over Diffusion Forcing, Diffuser, and Diffusion Policy, including a 56.5% gain on LIBERO-10 and better out-of-distribution performance.
Selective re-noising for cross-horizon revision is a meaningful architectural idea with broad, apparently strong empirical gains, but the abstract lacks absolute results and enough detail to establish whether the improvements generalize beyond the reported settings.
-
PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoning
PRO-LONG gives an LLM agent a complete, structured interaction history and lets a coding-capable model search that history programmatically, rather than relying on selective memory or stuffing context. On the full public ARC-AGI-3 game set, it reportedly improves a base coding agent by 18 percentage points, reaches up to 76.1% pass@1, and uses 4.2–5.8× fewer tokens than specialized harnesses; a best-of-two run reaches 97.4% at $1,750.
The simple shift from lossy memory selection to searchable complete logs appears genuinely useful, with substantial reported gains in both performance and token cost, but the evidence is concentrated on one benchmark and the abstract does not establish broader generality.
-
Co-Evolving LLM Evaluators and Policies via DynamicRubric
The paper argues that policy optimization fails when evaluators assign nearly identical scores to increasingly similar candidate responses, because the score gap is the useful learning signal. DynamicRubric addresses this by generating response-set-specific weighted binary rubric questions and co-evolving the evaluator with the policy; on 8B models it reportedly beats much larger static evaluators and improves reasoning and coding, with deployment at WeChat Search scale.
The response-conditioned, co-evolving evaluator is a meaningful direction backed by theory, comparisons against much larger evaluators, and large-scale deployment, though the abstract gives too few quantitative details to justify a strong verdict.
-
Test-Time Scaling via Error Localization
TTEL uses task or environment feedback to identify where a reasoning trajectory likely first went wrong, then truncates at that point and branches while reusing the valid prefix. On Qwen3 models, it reports substantially better pass-rate versus token-cost tradeoffs than independent sampling and sequential refinement, including 71.0% pass@64 on LiveCodeBench with roughly half the generated tokens.
Token-level credit assignment for inference-time search is a meaningful efficiency improvement over discarding whole failed trajectories, and the reported cost reduction is substantial, though the approach appears dependent on having informative task or environment feedback and the evidence is limited to reasoning benchmarks.
-
Formally Verified Synthesizable Floating-Point Data Types in ARCH HDL
The paper presents a single-source implementation of IEEE-754 FP32 and BF16 operators for the ARCH hardware-description language, emitting synthesizable SystemVerilog, SMT models, and Lean proofs that are structurally linked. It combines exhaustive equivalence checking for tractable operators with Lean proofs of correct rounding for multiplication and FMA, then replaces an impractical 470-bit FMA datapath with a 98-bit pipelined implementation proven equivalent over all inputs. The main novelty is an end-to-end, machine-checked workflow aimed at preventing language-model-generated hardware artifacts from drifting apart.
The unusually strong formal coverage and proof strategy for a synthesizable floating-point implementation are genuinely interesting, especially for reliable LLM-generated hardware, but the contribution is a specialized hardware-verification workflow rather than a broad capability or major AI advance.
-
Memory Efficient Audio Synthesis with Decoupled Temporal Depth Diffusion Transformers
The paper describes an on-device speech detokenizer that converts semantic audio tokens into high-fidelity streaming audio using a shared diffusion-based depth decoder, rather than separate decoders for each residual-quantization level. Causal fixed-window attention keeps memory constant with sequence length; on Apple’s AMX it reportedly runs at about 16× real time with 21 MB peak runtime memory, while improving MOS over the previous on-device system by 0.28 overall and 0.42 for conversational speech.
The combination of decoupled temporal/depth diffusion decoding and constant-memory streaming delivers unusually strong real-device efficiency and quality, though it is primarily a specialized deployment architecture rather than a clearly broad new research direction.
-
The brain-meningeal interface functions as a reservoir and entry site for brain parenchymal macrophages
The study traces how blood-derived monocytes and meningeal border-associated macrophages can enter brain tissue and replace depleted parenchymal macrophages. It argues that entry depends on timing, proximity, and competition with resident-cell self-renewal, and reports peripheral BAM-like macrophages in aged and neurodegenerative human brains, challenging the idea that these cells are exclusively microglial.
The proposed meningeal reservoir and selective replacement of parenchymal macrophages challenge a core assumption about microglial exclusivity, but the abstract provides limited quantitative or mechanistic detail and the work is an unreviewed preprint.
-
On-demand, reversible blood-brain barrier opening via electrical activation of piezoelectric nanoparticles for targeted brain drug delivery
The authors use vascular-targeting piezoelectric nanoparticles activated by transcranial direct-current stimulation to open the blood-brain barrier in a localized and reversible way. In the abstract, this enabled cortical delivery of a small molecule, full-length antibodies, and AAV vectors while preserving neurovascular structure, with the affected area controlled by electrode geometry. The main novelty is combining biodegradable mechanosensitive particles with noninvasive electrical stimulation for programmable regional BBB opening.
Targeted, reversible delivery of biologics and viral vectors across the BBB would be important and non-obvious, but the abstract gives no quantitative permeability, dosing, duration, efficacy, or safety data and does not establish how broadly the method works.
-
ReferTrack: Referring Then Tracking for Embodied Visual Tracking
ReferTrack separates embodied visual tracking into two explicit stages: first identify the language-specified target among image-space bounding boxes, then predict tracking waypoints. It also feeds a history of selected boxes into the model to preserve motion information, and reports strong EVT-Bench results plus sim-to-real tests on legged and humanoid robots using one camera.
The explicit referring-then-tracking decomposition and temporal box history are a useful grounding idea, with credible benchmark and real-robot evidence, but the abstract does not establish a major capability or generality jump beyond a fairly plausible VLA refinement.
-
Uncovering Latent Reasoning Strategies in Language Models
The paper factorizes a pretrained language model’s responses into an input-dependent router over latent reasoning strategies and a strategy-conditioned generator. To prevent the latent variable from being ignored, it weights reconstruction toward tokens that are surprising under the original model, and reports that the resulting codes correspond to distinct reference algorithms while preserving the model’s behavior.
The fractional-information-gain objective is a plausible and fairly original way to address posterior collapse when extracting implicit reasoning strategies, but the abstract gives no quantitative results or evidence that the approach generalizes beyond a purpose-built algorithmic benchmark.
-
Trace: A Taxonomy-Guided Environment for Multidomain Visual Reasoning
Trace is a procedurally generated training environment for visual reasoning that separates scene rendering from executable task and answer generation. It provides 1,000 task types across 277 scene grammars and 11 domains, with exact verifiers and replayable traces; RL training on 64,000 generated instances improves average performance on 24 external benchmarks by 3.51 points for a 3B model and 4.06 points for a 7B model.
The combination of taxonomy-guided procedural visual tasks, exact verification, and evidence of cross-distribution transfer is a meaningful direction for vision-language RL, though the abstract lacks enough detail to establish how broad or durable the gains are.
-
SANA-Video 2.0: Hybrid Linear Attention with Attention Residuals for Efficient Video Generation
The paper proposes a video diffusion transformer that mostly uses linear attention, periodically inserts gated softmax-attention layers, and carries summaries from earlier blocks through attention residuals. In 5B and 14B models, it reports softmax-comparable VBench quality with faster inference, including a 3.2× compiled-forward speedup at 720p/60s and a further 3.58× system-level optimization gain, though the headline comparisons depend heavily on implementation and baseline choices.
The hybrid linear/softmax design targets a real bottleneck in long-video generation and reports substantial speedups, but it is primarily an attention-engineering advance whose strongest system-level claims need careful apples-to-apples validation.
-
Show, Don't Tell: Evaluating Spatial Cognition in Generative Pixels Rather Than LLM Text
The paper proposes ProVisE, a protocol that lets image-generation models answer spatial tasks by drawing or marking directly in pixels, then converts those outputs into standard structured metrics. On a 470-sample benchmark and six external benchmarks, image-generation models perform competitively when pixel-space responses are natural, while text-based VLMs remain better at compositional spatial reasoning.
The pixel-space answer interface is a useful and non-obvious way to evaluate generative vision models, and the reported split between spatial externalization and compositional reasoning is worth checking, though this is primarily a benchmark and evaluation framework rather than a major capability advance.
-
D-NOVA: In-Storage Retrieval Accelerator via Dual-Bound 3D NAND-Optimized Similarity Search with Vector Adaptation
D-NOVA moves IVF-based vector retrieval largely into the 3D NAND array itself, introducing a NAND-compatible dual-bound similarity operation and a lightweight embedding adapter to preserve recall. The authors report up to 41.7× lower latency and 71× better energy efficiency than a CPU baseline, plus substantially higher throughput than prior in-storage RAG accelerators.
The combination of NAND-native similarity sensing and learned vector adaptation is a substantial hardware-software direction for reducing RAG retrieval bottlenecks, but the abstract does not establish how broadly the gains hold or whether they come from measured hardware versus simulation.
-
BMFA: Boundary-Minority Free-Energy Adaptive Screening
The paper analyzes a failure mode of block-mean token summarization in Vision Transformers: a small, high-response boundary region can carry substantial softmax attention mass while contributing almost nothing to the mean. It proposes recursively refining blocks using a lower-bound estimate of local free-energy error, and reports much smaller synthetic and COCO approximation gaps plus ImageNet classification with roughly 56% of tokens retained. The main limitation is that token selection is evaluated after full QK computation, so the work does not yet demonstrate actual sparse-attention speedups.
The free-energy formulation and boundary-minority analysis appear genuinely useful for understanding adaptive token compression, but the practical impact remains uncertain because the prototype does not measure end-to-end computational savings and the abstract gives limited baseline context.
-
C$^2$KV: Compressed and Composable KV Cache Reuse for Efficient LLM Inference
C²KV jointly learns to extract, compress, and concatenate KV-cache representations so that cached context from different sources can be reused without retraining or modifying the base LLM. The authors report large reductions in KV storage and transfer costs, with up to 17× inference speedup on long-context benchmarks while maintaining generation quality across multiple model families.
The composable, position-agnostic compressed KV representation addresses an important limitation of combining cache reuse with compression, but the abstract gives too little detail about quality tradeoffs, workloads, and how broadly the 17× speedup holds to justify a strong recommendation.
-
When to Use Extra Context: Evidence-Grounded Terminology Adaptation for Simultaneous Speech Translation
The paper argues that document context helps simultaneous speech translation mainly by recovering paper-specific terms, rather than by broadly improving each segment’s semantics. Its EGTA method maintains a terminology memory, selects relevant terms for each streaming state, and uses them to adapt ASR and translation decisions without full fine-tuning; experiments show modest BLEU gains but much larger relative improvements in named-entity and acronym recall across several SimulST settings.
The terminology-focused diagnosis and lightweight adaptation across cascaded and end-to-end systems are genuinely useful, with cross-setting evidence, but the headline translation gains are modest and the core retrieval/adaptation approach is not clearly a major new direction.
-
MEVION: Low-Cost Open-Source Data Collection System for Powerful and High-Speed Dual-Arm Manipulation
MEVION is a low-cost, open-source four-arm teleoperation/data-collection platform designed to overcome ALOHA’s limited force and speed. It uses 7 kg, 60 Nm arms with closed-link elbows and costs about $14,000 to build; demonstrations suggest it can collect manipulation data for heavier or faster tasks and support imitation learning.
The combination of an ALOHA-like open data-collection system with substantially stronger, faster arms could broaden the tasks used for robot foundation-model training, but the abstract gives limited quantitative task comparisons or learning results.
-
The Shared Discovery Paradox: How a One-Answer Rule Turns Better Information into Worse Search
This paper studies group search when agents receive noisy private clues but are forced to act on a single shared recommendation. In an exactly solvable 16-box benchmark, pooling improves the best recommendation yet causes everyone to repeat the same action, sharply reducing discovery; coordinating a diverse portfolio restores and slightly improves coverage. It also analyzes self-interested searchers, showing how reward design and correlated clues affect the gap between information quality and group performance.
The useful and non-obvious contribution is an exact separation of information aggregation from action allocation, with incentive results and quantitative limits, though the evidence is confined to a stylized benchmark rather than demonstrated in deployed multi-agent systems.
-
QIRF Quantum-Inspired Non-Orthogonal Function-Space Compression for 3D Gaussian Splatting
QIRF compresses 3D Gaussian Splatting by treating overlapping Gaussians as a non-orthogonal local basis and selecting representatives through generalized eigendecomposition. On 13 scenes, it removes 71.7% of Gaussians, preserves roughly the original reconstruction quality, and increases rendering speed by 34.3%, though the reported PSNR gain is negligible.
The function-space view of Gaussian redundancy is a meaningful departure from ordinary pruning or quantization, and the reported compression/speed gains are substantial, but the evidence is limited to 13 scenes and the core linear-algebraic selection strategy may be an incremental refinement of existing basis-reduction ideas.
-
The Label Complexity of Class-Conditional Coverage under Distribution Shift
This paper studies why conformal prediction can maintain its overall coverage guarantee under distribution shift while badly under-covering particular classes. It proves that when shift affects both features and labels, label-free methods cannot generally achieve efficient class-conditional coverage, and that estimating class-specific thresholds to tolerance requires roughly inverse-quadratic dependence on the tolerance and logarithmic dependence on the number of classes; experiments on several real and synthetic shifts support the theory.
The combination of an impossibility result and matching label-complexity bounds explains a practically important failure mode of conformal prediction, but it does not appear to deliver a major new capability or broadly applicable method.
-
The Calibration Channel Determines the Bayes-Error Proxy: An Exact Law for Temperature-Induced Distortion
This paper derives an exact expression for the soft-label Bayes-error proxy after temperature scaling, showing that it depends on the classifier’s margin distribution rather than just its classification error. The proxy is strictly monotonic in temperature and can be made to take any value between 0 and 1/2 without changing the classifier’s decisions or 0-1 error; experiments on three datasets show distortions of 56x–980x, with a Gaussian-logit approximation closely matching observations.
The exact, model-free characterization and the striking result that temperature alone can induce essentially any reported error proxy are genuinely useful for interpreting calibration-based uncertainty estimates, though the impact is concentrated in a specialized calibration methodology.
-
Automated Discovery Has No Universally Superior Harness
The paper breaks down two LLM-based autonomous-discovery harnesses into their component choices and evaluates 30 combinations across 12 model–problem pairs, using over 3.1 million rollouts and repeated trials. It finds that no fixed harness consistently wins, that OpenEvolve-style variants often lose to simpler alternatives, and that adaptively pruning weak runs and reallocating compute improves performance over fixed or non-adaptive selection.
The large-scale evidence that discovery-harness design does not generalize across problems, together with an empirically validated online allocation strategy, is a non-obvious and potentially useful shift from treating harnesses as universal recipes.
-
Structured Output Collapses Answer Diversity Across 44 Language Models
This study tests 44 language models on prompts with many equally valid answers and finds that merely requesting JSON makes their answers substantially more similar: the modal answer rises from 41% to 64% and answer diversity falls. The effect is specific to formats associated with model training for tool use (JSON and XML), does not increase with decoder-level schema enforcement, and tends to pull distinctive models toward the group’s existing default.
The broad controlled finding that structured-output requests alter model behavior and reduce cross-model diversity is non-obvious and potentially important for evaluating and deploying LLMs, but its practical impact and generality beyond categorical word-choice prompts remain uncertain.
-
Reasoning Fine-Tuning Induces Persistent Latent Policy States
This paper models chain-of-thought activations as a switching dynamical system and discovers discrete, persistent latent “policy states” that organize reasoning over time. Across four benchmarks and models from 1.5B to 32B, reasoning fine-tuning produces more structured and functionally specialized state transitions; interventions and prefix pruning provide evidence that these states are causally relevant, with pruning outperforming self-consistency in 11 of 12 settings.
The potentially important contribution is evidence that reasoning fine-tuning reorganizes global temporal latent dynamics rather than merely improving local token prediction, supported by cross-scale analyses and causal interventions, though the abstract does not establish a major capability or efficiency breakthrough.
-
MAGE: Human-Like Macro Placement via Agentic Multimodal Reasoning
MAGE uses a multimodal, multi-agent workflow to iteratively refine chip macro placements using natural-language floorplanning rules, visual checks, candidate tournaments, and feedback. On nine NanGate45 and GF12nm designs, it reports sizable WNS/TNS improvements over commercial and prior placers, while maintaining comparable wirelength and power; it also proposes metrics for structural qualities such as notches, whitespace, pockets, and alignment.
The combination of agentic multimodal reasoning with explicit expert floorplanning heuristics and reported large gains is plausibly worth a look, but the evidence is limited to nine designs and the abstract does not establish how robust the commercial comparisons or generalization claims are.
-
The Story Shapes the Agent: Narrative Priors in LLM Behavior
The paper tests whether an LLM’s behavior is driven more by a task’s story than by its nominal persona. Using structurally identical investigation games with different narratives, the authors find that narrative framing explains far more behavioral variation than persona, often hurting task success; personas transfer across narratives mainly when they contain concrete action-linked language, which causal ablations support.
The controlled structural-isomorphism setup and causal evidence for narrative-dependent behavior are a non-obvious, potentially useful explanation of why persona prompting transfers poorly, but the scale and model coverage are still limited for a stronger verdict.
-
Attention Without Grounding: Causal Evaluation of Visual Explanations in Medical VLMs
This paper causally tests whether attention and saliency maps from medical VLMs identify the image regions that actually affect their chest X-ray answers. Across several VLM families and datasets, the maps generally fail: some models use the image while their attention is negatively related to occlusion-based importance, while another appears nearly text-only; conventional CXR classifiers pass the same tests. The main contribution is evidence that visually plausible VLM heatmaps are not faithful explanations and should be evaluated with perturbation-based localization tests.
The cross-model causal audit produces a non-obvious negative result—that medical VLM attention can be anti-correlated with image evidence—using multiple datasets, controls, and perturbation tests, though it is primarily an evaluation study rather than a new modeling capability.
-
Measuring the Dependency Gap: Diagnosing Inter-Column Fidelity in Tabular Generative Models
The paper introduces a diagnostic for tabular generators that decomposes a two-sample fidelity score into marginal, dependency, and mixed numerical–categorical components, using factorized data as a controlled reference. Across flow-matching and diffusion models, it finds a persistent dependency deficit that is not fixed by finer sampling, larger capacity, or simple cross-coupling and copula post-processing, suggesting that objectives which do not explicitly score joint structure are the main bottleneck.
The controlled localization of dependency loss and the evidence that capacity and sampling are not the limiting factors are genuinely useful and somewhat surprising, though the contribution is mainly a diagnostic insight in the relatively narrow area of tabular generation.
-
Fabric Pneumatic Artificial Muscles Based on the Drawstring Principle
The paper introduces a pneumatic artificial muscle built around a textile drawstring structure, designed to contract axially without the radial expansion found in conventional pneumatic muscles. The prototype reportedly reaches 44% contraction, supports over 800 times its own weight, and achieves 4.98 kW/kg power density; demonstrations include a bionic robot, an industrial mechanism, and a scalable 2D muscle matrix.
Eliminating radial expansion while retaining high contraction and power density could address a real packaging limitation in soft robots, but the abstract does not provide enough comparative or durability evidence to justify a strong recommendation.
-
Spaghetti Architect: A Contamination-Resistant, By-Construction-Labelled, Multi-Language Code Dataset Generator
Spaghetti Architect generates semantically equivalent but deliberately messy programs from a clean intermediate representation, with controllable intrinsic and incidental difficulty and correctness verified by execution. Across several languages and four open models, it finds that increasing true problem size can collapse output-prediction accuracy while refactoring equivalence remains high, and that model-generated annotations disproportionately help weaker models; reminted held-out variants produce nearly identical scores, supporting contamination resistance.
The controlled, executable generation scheme and separation of semantic difficulty from presentation messiness are genuinely useful, while the contrasting model failures provide a non-obvious result, though this remains primarily a benchmark/tool contribution rather than a major capability advance.
-
How the fly holds a single goal: normalization, not selection, in Drosophila FC2
Using a single-fly connectome and dynamical models, the authors argue that FC2 neurons do not implement winner-take-all selection of a navigational goal. Instead, globally connected FB5A neurons mainly normalize an externally selected activity bump, while hDelta neurons provide only a weaker, potentially competitive interaction; an upstream hDelta network is proposed as the likely site of goal selection. The paper also identifies a direct experiment—silencing FB5A while imaging FC2—to test this account.
It challenges the natural assumption that recurrent inhibition in FC2 selects a single goal, combining connectomic circuit analysis with multiple dynamical models, though the key inhibitory-cell identity and behavioral mechanism remain experimentally unverified.
-
MedDDC-Eval: Diagnosis-Decoupled Evaluation of Multi-Turn Medical Consultation Agents
MedDDC-Eval separates a medical consultation agent’s ability to gather diagnostically useful information from its ability to produce the final diagnosis. It runs every elicited history through the same frozen diagnostic model and reports diagnostic support, coverage, and efficiency; across eight policies, this changed F1 by 2.2–19.0 points and reversed 18–36% of policy rankings. A GRPO experiment also suggests the resulting feedback can improve policy training.
The diagnosis-decoupled evaluation addresses a real confound in multi-turn-agent comparisons and shows that it materially changes conclusions, though it remains primarily an evaluation framework rather than a demonstrated capability breakthrough.