In briefGeoGuesser Bench: Testing AI Agents on Interactive Location Finding
SummaryGeoGuesser Bench is a benchmark of 282 tasks in an offline Street View environment, where AI agents rotate, zoom, and move between real panoramas to report their starting coordinates, name the surrounding place, or find and count nearby objects. The study evaluates 23 vision-language models, with the best reaching 62.5% accuracy on place naming and 35.0% on object finding. It finds that world knowledge and exploration ability are only weakly correlated across models.
In brief3L-D: A low-light dataset built with a sensor-agnostic, alignment-preserving capture framework
SummaryThis paper introduces 3L-Dataset, a physically calibrated low-light image and video dataset created with a modular, sensor-agnostic acquisition framework. The test bench reproduces five standardized Midavaine night levels, from a full moon (about 1000 mlx) down to an overcast moonless night (about 0.1 mlx), across visible to near-infrared light and several scale-model environments. Using high-precision programmable mechanical actuation, each synchronized sextuplet of one reference image and five images at the five night levels achieves sub-pixel spatial alignment without post-processing warping.
L\'eo Rueff, Maurg\`ene Ghezal, Valentin Gorse et al.
In briefMathematical foundations for estimating and comparing information in robot teams, and how far they reach in the cognition loop
SummaryThis book builds the Generalized Singular Value Decomposition (GSVD) from the ground up. The GSVD compares two linear operators acting on the same unknown, showing along which state directions one dominates the other and by how much. It re-derives the needed linear algebra and estimation theory, gives two proofs of the GSVD, and works a small matrix pair through five lenses. It separates information attribution, subspace geometry, statistical dependence, and decision-theoretic discrimination for collective systems, and evaluates the GSVD's role and limits across the cognition loop.
In briefLLMs Cannot Be Neutral: How Candidate and Evaluator Gender Affects AI Decisions
SummaryThis paper examines whether large language models used to evaluate job candidates change their decisions based on gender. Eight models judged 240 synthetic professional narratives in a hypothetical city council vote, with candidate and evaluator gender varied independently, producing 30,720 judgments. Explicit candidate gender labels raised Yes-vote probability by 1.05 to 1.73 percentage points compared with unspecified gender, and 2.5 to 3.3 percent of decisions reversed relative to the baseline.
In briefClosing the dynamics gap via self-directed robotic interaction
SummaryThis paper presents CUE, an agentic system that closes the gap between simulated and real physics by having a robot interact with its environment directly. Given only a high-level instruction and no resets, the agent builds its own perception, identifies physical parameters such as friction and restitution, and targets the interactions where its simulation is least accurate. On billiards in simulation across 25 hidden-physics conditions, it reduces held-out ball-position error by 72% relative to the strongest baseline, and it is also deployed on a real pool table.
In briefHuman-Oriented World Model: A Conceptual Framework for Tracking and Forecasting Human States
SummaryThe paper proposes the Human-Oriented World Model (HOWM), a conceptual framework centered on task-relevant human states and how they evolve. It links current observations and individual history with environmental and social context to support continuing state estimation, personalized dynamics prediction, and action support. The framework is illustrated through hypothetical training and learning scenarios, and the authors outline research questions on whether accumulated interaction experience improves estimation and prediction and whether the model helps people pursue their goals.
In briefUsing idle serving compute and zeroth-order curvature to speed up LLM pretraining
SummaryThe paper proposes ZO-Curv, which uses forward-only inference hardware that sits idle off-peak to estimate curvature and adapt first-order training updates. The authors argue that zeroth-order estimates, though too noisy to serve as gradients, still carry useful curvature information. Experiments on models from 125M to 1.5B parameters show faster pretraining; at 770M parameters, ZO-Curv reaches AdamW's 20K-step validation loss in 3.9 hours instead of 7.5 and ends with a lower loss.
In briefFinding Meaningful Image Regions at Different Scales by Clustering Pretrained Visual Features
SummaryThis paper proposes a method that clusters pretrained visual representations without semantic supervision to discover image regions at multiple levels of detail. Pixel-wise principal representation maps are derived via SVD-based projection, and a semantic continuity rate with signal-to-noise analysis assesses clustering quality. Evaluated on COCO-2017 and VOC-2012 validation sets, the approach shows competitive segmentation performance.
In briefDynamic hypergraph network with contrastive refinement for skeleton-based action recognition
SummaryThis paper addresses skeleton-based action recognition, where standard graph networks model only pairwise joint relationships and existing hypergraph methods use static structures. The authors propose DyGCN, whose core DyGCT block combines geometry-constrained dynamic hypergraph construction (GeoHC) with contrastive score refinement (CSR) based on prototype learning. CSR operates only during training, adding no inference-time overhead. The paper reports state-of-the-art results on NTU RGB+D, NTU RGB+D 120, and Northwestern-UCLA.
In briefHistoTx: Early fusion of H&E tissue images and spatial gene expression data using self-supervised learning
SummaryThe paper proposes HistoTx, an early-fusion vision transformer that builds transcript tokens at arbitrary spatial resolution, merges them with image patch tokens, and processes both in a shared transformer. It starts from a pretrained pathology foundation model and continues self-supervised pretraining on paired image and spatial transcriptomics data. With image input alone, performance stays comparable to the baseline on most tasks and improves on tasks such as gene expression prediction. When both image and transcript inputs are provided, the model outperforms either modality alone.
Alice Driessen, Emil Malmsten, Steffen Wolf et al.
In briefVioTech: Frame-Level Violin Technique Detection Using Multi-Scale Mixture-of-Experts
SummaryThe paper proposes VioTech, a framework that detects violin playing techniques directly at 10 ms frame resolution without explicit note segmentation. It combines a frozen MERT encoder with a multi-scale Mixture-of-Experts architecture and task-specific routing for tonal, articulation, and legato predictions. The authors also introduce VioTech120, a manually annotated dataset of violin techniques built from real solo performance recordings.
Ting-Kang Wang, Vincent K.M. Cheung, Nan-Tien Lai et al.
In briefReap: A lightweight model that reveals how Monte Carlo proof search works
SummaryThe paper presents Reap, an open framework for learning and analyzing formal proof search in Lean4, implementing policy-value-guided Monte Carlo Tree Search directly within Lean. Using it, the authors train Reaper-1, a 1.7B-parameter model, which solves 79.9% of miniF2F-test at pass@32. The reported result is state-of-the-art among open provers of comparable size.
In briefAuditing Fraud Detection in Representation Space: Bayesian Inverse RL for Calibrated Uncertainty and Its Limits
SummaryThis paper examines shilling-attack detection in recommender systems and introduces VEIL, a Bayesian framework that infers rewards from pairwise comparisons over frozen encoder embeddings, with a base posterior, a hierarchical ARD extension with learned confidence, and an organic/attacker mixture. The authors prove that a detector built on a representation invariant to a user-profile transformation cannot beat chance against attacks indistinguishable from organic profiles under that invariance. Detection is near-perfect against synthetic-structure attacks but falls to chance on MovieLens-1M under donor-mimicry, while a non-learned baseline using raw item-set Jaccard similarity still achieves AUROC of 0.92-0.97 across the donor-mimicry configurations.
In briefChain of Faith: Rewards That Make Models Rely on Their Reasoning
SummaryThis paper studies how to make a model's final answer depend more on its generated chain-of-thought reasoning, rather than on explanations that only sound plausible. It proposes four rewards built from the policy's own signals and adds each to group relative policy optimization (GRPO). Experiments on GSM8K and StrategyQA with two small instruction-tuned models, using reasoning edits, paraphrases, and prompt hints, show that some configurations raise reliance on reasoning edits, while effects on accuracy and paraphrase robustness vary.
In briefTME-ContextNet: Joint Cell Nucleus Segmentation and Classification with Class-Based Spatial Supervision
SummaryThe paper proposes TME-ContextNet, a multi-task framework for segmenting and classifying individual cell nuclei in pathology images. It turns existing nuclear annotations into dense spatial supervision using Euclidean distance transforms, without needing extra tissue-region labels, and splits decoding into geometry and semantic task groups. Reported results on PanMix, CoNSeP, and PanNuke include Dice and mPQ+ scores, with 36.40 million parameters compared with 54.74 million for HoVer-Net.
In briefUPSHIFT: Steering Assistant Responses to User Preferences from Feedback Traces
SummaryThis paper proposes UPSHIFT, a training-free method that compares hidden representations of initial and feedback-revised responses within each session and averages these shifts across a user's sessions. The resulting preference direction is injected when answering new queries to personalize responses. In tests with 40 simulated users on PrefEval, the best configuration raised macro-averaged Elo from 1364.0 to 1605.2 for Qwen3-8B and from 1341.9 to 1591.8 for Qwen3-32B, outperforming the unsteered baseline.
In briefFine-tuning a few-step masked diffusion language model on its own sampling likelihood
SummaryThe paper observes that a masked diffusion language model sampled in a few parallel steps in random order effectively runs a mixture over reveal plans, which standard cross-entropy training does not fit. The authors propose mixture-likelihood fine-tuning (Mix-FT), which optimizes an importance-weighted bound on the model's own sampler likelihood without a teacher or architectural changes. Experiments on LLaDA-8B and a 110M block diffusion model report that Mix-FT produces better continuations than continued cross-entropy and other baselines at fewer sampling steps.
WhyThe authors propose mixture-likelihood fine-tuning for few-step masked diffusion language models, needing no teacher or architecture change. The abstract reports gains over continued cross-entropy and
In briefPriCBMs: Multi-granular Concept Bottleneck Vision-Language Models for Visual Privacy Recognition
SummaryThis paper proposes PriCBMs, a model for recognizing privacy-sensitive information in shared images, such as details accidentally exposed on social media. It uses privacy-related concepts as intermediate representations, first matching visual cues to these concepts and then predicting privacy categories from the concept responses, which makes predictions more interpretable. The authors add a concept selection strategy based on category relevance, cross-category discriminability, and semantic diversity, along with a module that models both global and regional evidence. Experiments on VISPR and FineViP show that PriCBMs outperforms the compared methods on all four multi-label metrics.
In briefParametric CAD Bench v3: Testing AI Agents at Building 3D Parts from Text and Drawings
SummaryThis paper introduces Parametric CAD Bench v3, a benchmark of 100 tasks covering text-to-CAD creation, text-based create-and-edit, and reconstructing mechanical parts from engineering drawings. Agents work in a pinned FreeCAD environment, and a separate programmatic verifier scores their native CAD outputs using a continuous reward that blends geometry similarity and specification consistency. In a September 24, 2026 snapshot, Claude Code with Claude Opus 5.5 achieved the highest overall score, 61.03%.
In briefOctreeSplat: Octree-guided feed-forward 3D Gaussian splatting for scene reconstruction
SummaryOctreeSplat is a feed-forward 3D Gaussian splatting method guided by a hierarchical octree, used for novel view synthesis and 3D reconstruction. It extracts multi-view features and geometric priors from a frozen visual foundation model, builds the octree, and predicts densification scores so that Gaussian primitives are allocated at densities that vary with local complexity. Experiments on Replica, TUM, Tanks and Temples, and DTU show competitive RGB view synthesis, depth synthesis, and surface reconstruction, with fewer Gaussian primitives while maintaining high reconstruction quality.
In briefMapping Artworld Influencers Through Social Media Data Mining
SummaryThe paper presents a method for building a network map of the artworld from social media. Starting with manually selected exemplar artworld actors, it uses their tweets and natural language processing to identify significant actors, including those without a Twitter account. The identified actors are grouped into communities using the Louvain method and their authority is scored with HITS centrality. The author reports that the network can discover new artworld actors from a small initial set and yields data that would be hard to gather otherwise.
In briefVerdictAD: A training-free method that refines anomaly detection using few normal images
SummaryVerdictAD is a training-free framework that improves existing anomaly detectors. It uses a typed decision model to compare a query image and its local regions against a few normal references, producing structured decision probabilities. These decisions refine the overall anomaly score and the anomaly map. The authors report consistent gains for existing detectors on MVTec AD, VisA, and MVTec LOCO AD.
WhyThe authors report a training-free plug-and-play refinement using structured visual decisions, improving existing detectors on MVTec AD, VisA, and MVTec LOCO AD.
In briefProbe-guided gradient balancing to help weaker modalities in multimodal learning
SummaryIn jointly trained multimodal networks, faster-learning modalities can dominate the shared objective and suppress gradients to slower ones. The paper proposes Probe-Guided Gradient Boosting (PGGB), which attaches lightweight linear classifiers to stop-gradient features to estimate each modality's representation quality. The resulting utilization gap adaptively and boundedly scales weaker-modality gradients without changing the loss or fusion architecture. Experiments on eight benchmarks report gains on highly imbalanced data and competitive results under low imbalance.
WhyThe authors propose independent linear probes to estimate modality representation quality and boost weaker-modality gradients, evaluated on eight benchmarks across four domains, with bounded-scaling a
In briefComputer vision for art history: reflections and future research directions
SummaryThe paper argues that art historians' ongoing debate over computational methods partly stems from such methods not addressing the core questions of current art history research, and from weak collaboration between the two research communities. The authors propose a methodology that studies articles from art history journals from a computer science perspective, to identify the image features art historians use and describe them in terms meaningful to the computer vision community. The paper concludes by offering several paradigms as a new starting point for how computer vision applications can address the central questions of contemporary art history.
In briefReinSeg: A Method for 3-D Road Anomaly Segmentation Using Bounded Multimodal Refinement
SummaryThis paper presents ReinSeg, a framework for point-wise 3-D road anomaly segmentation in autonomous driving. It retrieves scene-specific visual context from surround-view images for LiDAR bird's-eye-view cells and feeds it back through a relative-norm-bounded residual while keeping the frozen LIDO scoring interface. On the Spotting the Unexpected benchmark, validation average precision rises from 27.54% to 35.67%, and the false-positive rate at 95% anomaly recall falls from 34.85% to 10.61%.
In briefCoati: Host-Grounded Execution Framework for Installable Mobile GUI Agents
SummaryThis paper presents Coati, an open-source agent that users install directly as an Android app, rather than driving a phone from a separate computer over ADB. Its on-device host records action effects and cross-app results, checks completion claims against this evidence, and uses per-app memory and a fast executor to speed up tasks. On all 116 AndroidWorld tasks, Coati succeeds on 86.5% over three trials, and the authors argue host evidence matters more than model self-assessment.
In briefSearchInject: Controlling Retrieval Exposure to Indirect Prompt Injection in Language-Model Search
SummaryThe paper proposes SearchInject, a joint scoring method that decides whether retrieved pages should be quarantined before they reach a search-augmented language model's context. On 300 held-out SciFact queries across four construction families, top-10 target exposure dropped from about 0.94, 0.92, and 0.85 to about 0.16, 0.15, and 0.14 for BM25, a normalized hybrid, and latent semantic analysis, respectively. The frozen threshold produced clean test false-quarantine rates of 0.59-0.65%, above the calibration constraint. A separate generation follow-up did not allow estimating any reduction in attack success.
In briefInterFuse: Adaptive Fusion of Multiple Payment Risk Models for Transaction Ranking
SummaryInterFuse is a framework that combines several payment risk models, each trained on a different view of transaction data, to rank transactions under limited review budgets. It builds a shared fusion state from interactions among the models and uses it to adaptively aggregate their representations, predictions, and input coverage. Tested on an industrial payment replay and three public financial benchmarks, it achieves the highest top-ranked recall and AUC on all four datasets, improving Recall@1% by 1.11 to 8.09 percentage points over the strongest baselines.
In briefFinding Where and for Whom a Prediction Model Is Miscalibrated
SummaryThis paper proposes a local audit that identifies sets of similar observations where a prediction model systematically over- or under-predicts. It also checks whether average prediction errors differ between demographic groups, such as race or sex, within those sets. Observations are grouped into pockets using random-forest similarity and ridge-regression grouping, learned on one sample and tested on another. The method is demonstrated on simulations, public synthetic data, and several real models.
Bernardo Modenesi, Taoyi Chen, Sima Najafzadehkhoei et al.
In briefVariable names as priors: how meaningful names help LLMs predict numbers from little data
SummaryThis paper studies how large language models use variable names and in-context examples to make numerical predictions on tabular data. When variable names semantically match the values they describe, five LLMs without task-specific training matched or outperformed seven supervised machine learning models across 20 datasets from nine scientific domains, especially when examples were scarce. Controlled experiments based on physical formulas indicate that the models combine domain knowledge with examples in a manner consistent with Bayesian updating.
WhyThe authors show variable-name semantics act as priors letting LLMs match or beat supervised models on 20 datasets, with physics-formula tasks showing Bayesian-like updating.
In briefSYNC: A training-free inference method that dynamically picks layers to steer multiple personas in language models
SummaryThe paper introduces SYNC, a training-free inference control framework that lets a model express several personas at once without changing its weights. It estimates each layer's sensitivity to each persona from cosine similarity trajectories and assigns every persona to its best injection layer at each decoding step, avoiding offline layer search. Across two models, two persona profiles, and steering vectors from unsupervised and supervised extraction, SYNC leads on 21 of 28 persona configurations and is significantly better on 15.
Yosef Nuraga Wicaksana, Muhammad Reza Qorib, Fadhil Muhammad et al.
In briefEvolving Explanations: Refining Faithful Explanations for Symbolic Regression by Complexity
SummaryThe paper notes that natural-language explanations from large language models for machine learning models can sound fluent yet may not faithfully reflect the model, and formalizes this as a faithfulness gap. It proposes a complexity-adaptive framework in which an explainer LLM writes increasingly detailed descriptions and an independent validator LLM extracts a symbolic expression to compare against the original model until numerical agreement is reached. Across 250 runs on real-world and synthetic datasets, the refinement loop achieves an overall pass rate of 84.0%, with most explanations converging at low complexity.
In briefMeasuring AI agents' teamwork ability with a collaborative intelligence benchmark
SummaryThe paper introduces the Collaborative Intelligence Benchmark, which evaluates how well AI agents work as teammates across 240 tasks covering 8 dimensions of teamwork. Tasks are drawn from real records or existing benchmarks, such as aviation accident reports, clinical cases, coding, and negotiations, and have programmatically verifiable outcomes. Evaluating 18 models, the authors find that every model has clear teamwork weaknesses, with the best scoring 82.2 out of 100.
In briefNeural network susceptibilities when learning from physical data
SummaryThis paper studies how neural networks trained on data from physical systems learn regularities imposed by physical laws. It uses susceptibilities, which measure a network's response to perturbations of a data distribution parameter, to identify which input features are most informative for learning and to guide data selection, with a prediction validated experimentally. For distributions without a natural parameter, the authors introduce one via a physically meaningful scalar function of the physical state and show that susceptibilities track the emergence of capabilities such as collision detection during training.
Rohan Hitchcock, Gary W. Delaney, Jonathan H. Manton et al.
In briefUniCross3D: Unified Cross-View and Cross-Domain Diffusion for Consistent 3D Generation from One Image
SummaryThis paper introduces UniCross3D, a diffusion framework that generates 3D content from a single image with consistent geometry and appearance. It adds a cross-view latent regularization that penalizes latent variance to keep synthesized viewpoints geometrically consistent, and a cross-domain mutual information objective that aligns generated color and normal maps. Experiments report better view consistency, semantic alignment, and reconstruction fidelity than state-of-the-art methods, especially with challenging textures and ambiguous viewpoints.
In briefICJ-Graph: A Citation Graph and Layered Corpus of International Court of Justice Documents
SummaryThis paper presents ICJ-Graph, a machine-readable resource linking International Court of Justice documents to cited cases, legal provisions and UN resolutions. The collection separates court decisions, written pleadings and oral records into distinct layers and includes 6,396 ICJ documents, 505 documents from other legal bodies, 12,375 provisions and 20,291 UN resolutions, with text, metadata and embeddings. The authors evaluate five models on citation prediction for 199 decisions, finding that tool-assisted models outperform non-tool baselines in F1 across case, provision and resolution citations.
In briefRecovering missing variability in simulation-trained cosmological field models
SummaryThe paper addresses simulation-trained generative models for cosmological field-level inference, where finite training data causes variance shrinkage that signals missing physical variability. The authors add physics-informed latent corrections to a continuous normalizing flow, recovering a substantial share of the missing variability without extra training simulations. In a weak-lensing field-level inference application, posterior analysis shows these latent directions are used during inference and improve robustness to out-of-distribution shifts.
In briefLearning Safety Filters for Agile Collision Avoidance from Data
SummaryThis paper proposes learning the dynamics, nominal controller, and barrier function of an integral control barrier function from data. Dynamics are modeled with neural ODEs, the nominal controller by imitation, and the barrier by supervised regression, with the filter constraint built analytically from these models. The barrier is trained on viability, so a state is labeled safe only if an emergency maneuver under actuator limits can still avoid the constraint. On agile racing with a 6-D dynamic bicycle, the filter removes all 12 unsafe episodes from 200 held-out layouts while keeping the success rate unchanged.
Dasari Raj Vamsi, Vrushabh Zinage, Efstathios Bakolas et al.
In briefLatent Diffusion-GAN: Adversarial Learning in a Compressed Latent Space
SummaryThis paper proposes Latent Diffusion Generative Adversarial Networks (LD-GAN), which adds adversarial learning to diffusion models without changing their original pipeline. It uses the pretrained variational autoencoder from latent diffusion models as an energy-based discriminator, so no extra parameters are needed. A structural consistency energy is also introduced to align encoder and decoder feature representations. The authors report improved sample fidelity, sharpness, and diversity over baseline methods across various generation tasks, with efficient training.
In briefA Judge That Inspects Environment State to Widen Harm Evaluation
SummaryThe paper proposes detecting harm by agents in stateful environments using an LLM judge with read-only access to environment state and a harm taxonomy. It is tested in two DecodingTrust-Arena environments and one OSWorld environment, comparing trajectory-only, state-only, and combined judges. Combining verdicts from the trajectory and state-only judges recalls 94 percent of confirmed harms in the DecodingTrust-Arena environments versus 53 percent for the trajectory judge alone, and 77 percent versus 67 percent on OS-Harm.
In briefAdaptive method for localizing knowledge negation in continual LLM unlearning
SummaryThis paper addresses the problem of utility degradation when large language models continually unlearn harmful or unlawful content in real-world settings. It proposes ALKN, a method built on the task vector framework that uses dynamic masking to sparsify training gradients and adaptively adjusts unlearning intensity based on inter-task relationships. Experiments on three established LLM unlearning datasets show that ALKN outperforms baseline methods in both unlearning effectiveness and utility retention under continual unlearning settings.
Abudukelimu Wuerkaixi, Qizhou Wang, Sen Cui et al.
In briefScalaris: A scalable memory system for enterprise chat assistants
SummaryThis paper presents Scalaris, an enterprise memory architecture for conversational AI assistants that unifies long-term and short-term memory. It handles user interactions and assistant responses separately, each with specialized representations and ingestion and retrieval pipelines. Using Qwen3-32B, Scalaris reaches 92.2% on LongMemEval-S and 90.77% on LoCoMo, and supplies about 10 times fewer retrieved context tokens than Hindsight. When integrated with proprietary assistants, it improved overall response quality by 36%.
Devarsh Gandhi, Sarthak Pawar, Hemprasad Badgujar et al.
In briefWhitened Advantage Summation Policy Optimization: a correlation-aware way to combine multiple rewards in reinforcement learning
SummaryThe paper addresses a problem in reinforcement learning for language models, where correlated reward signals are summed with fixed weights and their shared variation is counted multiple times. The proposed method, WAS-PO, whitens normalized advantages before summing them so that the chosen weights apply to decorrelated components. On tool calling with Qwen2.5-1.5B-Instruct, WAS-PO achieved higher training reward than GDPO with three and four rewards, but held-out accuracy did not improve correspondingly, and with two rewards training reward was lower than GDPO's.
In briefBoundary-guided network for more accurate polyp outlines in colonoscopy images
SummaryThis paper presents BGD-SF PolySegNet, an encoder-decoder network that uses boundary information actively during polyp segmentation in colonoscopy images. It applies boundary guidance to multi-scale context aggregation, skip-feature fusion, and mask refinement. On the internal test set it reaches a Dice score of 0.887, and it improves boundary F1, HD95, and ASSD over the strongest CNN baselines by 15.7%, 27.5%, and 34.6%.
Nishi Kanta Paul, Shamia Maherin, Morsheda Akter et al.
In briefA method for on-device multimodal agents that prunes models to fit hardware cores and manages memory by data lifetime.
SummaryThe paper presents PARC, which combines structured pruning aligned to the core count of edge accelerators with a KV cache that separates data by lifetime. Pruning uses first-order Taylor importance and grouped-query-attention-aware selection so attention dimensions divide evenly across cores. At runtime, historical visual tokens leave the persistent cache, allowing more interaction rounds within a fixed memory budget.
In briefGazeTags: camera-free object discovery and interaction using head gaze
SummaryThis paper presents GazeTags, a way for wearable devices to know which objects a user is looking at without cameras or computer vision. The wearable emits a narrow infrared signal aligned with head direction, and only the low-cost tags inside that beam respond with their identity and optional data such as object state or sensor readings. The authors describe the hardware, characterize system performance and tradeoffs, and demonstrate gaze-aware environments.
In briefAVID: A 5T fMRI Dataset for Testing Brain Decoding of Sound-Triggered Visual Imagery
SummaryThis paper introduces AVID, a large-scale 5.0 Tesla fMRI dataset for decoding visual mental imagery induced by auditory descriptions. It contains about 67 hours of recordings from nine participants, split into a complex scene-level set using naturalistic Mandarin descriptions and a simple set with explicitly specified visual attributes. The authors also propose a standardized evaluation protocol that separates source-image consistency from semantic alignment. Baseline results show that current reconstruction methods remain limited on AVID, though the best reconstructions preserve recognizable scene semantics.
In briefTesting whether open AI models can accurately recall Hadith texts
SummaryThis study builds a bilingual English and Arabic benchmark to test five open-weight language models on reproducing exact Hadith texts from their bibliographic references, across all six canonical Sunni collections. Models received no retrieval support, and their 30,000 responses were scored for lexical and semantic fidelity. None of the responses matched the source text on both measures, and Arabic results were consistently worse than English. The authors conclude that such models are unreliable for exact religious-text recall without retrieval over a verified corpus.
Dayam Nadeem, Safdar Mustafa, Ameen Ur Rehman et al.
In briefA multilingual check of stereotype bias in AI language models
SummaryThis paper tests whether large language models show gender, racial, and religious stereotype bias, using 12,000 Reddit posts written locally in English, Spanish, French, and German. Six models, three open-weight and three proprietary, were evaluated. Religion showed the largest choice imbalance across every model and language, and the API models scored below the open-weight models. The authors note that these differences measure how concentrated the model's choices are, not endorsement of stereotypes or causal country effects.
In briefBasisKV: Value Latents for Communication and Key Routing
SummaryThis paper proposes BasisKV, a low-rank KV-cache compression method for long-context LLM inference on multiple GPUs with tensor parallelism. Its low-rank Value representation reduces both Value-cache width and tensor-parallel communication, and the compressed Values help select which Key pages to fetch for exact attention scores. Across several LLM families and system settings, it reports up to 1.42x end-to-end speedup under 8-GPU tensor-parallel inference.
In briefSMILE: A structured multi-agent framework for liver ultrasound report generation
SummaryThe paper proposes SMILE, a framework that generates liver ultrasound reports using three specialist agents for image interpretation, clinical-stage reasoning, and visual question answering. A judge vision-language model consolidates their outputs through consensus filtering into one clinically coherent report. SMILE outperforms state-of-the-art vision-language models on a public and an internal liver ultrasound dataset, and the authors suggest it is promising for data-scarce clinical settings.
In briefTesting Local Conditional Fit for Detecting Diffusion Model Training Members
SummaryThis paper studies membership inference, which determines whether a candidate example was used to train a text-to-image diffusion model, a useful tool for auditing training-data use. The authors propose LoFIT (Local Conditional Fit), which measures how the denoising loss initially responds to shifts in predicted noise under alternative text conditions. LoFIT needs only forward queries to the denoiser and no target-model gradients. Across four datasets, it improves membership discrimination for both fine-tuned and pretrained models.
In briefInterpretable non-uniform uncertainty sets for testing how robust trained reinforcement learning policies are to state perturbations.
SummaryDeep reinforcement learning agents can be disturbed by changes to their observed state inputs. This paper focuses on attacks applied at evaluation time to pretrained policies and proposes PDNUUS, an interpretable, state-dependent, dimension-wise non-uniform uncertainty set that can be synthesized online during evaluation. The authors first formulate an ideal constrained optimization problem and then introduce a surrogate constraint to make it tractable. Experiments across multiple continuous-control environments, policy algorithms, and state attacks show that the method reveals differences in policy sensitivity across states and dimensions, while performance degradation is moderately controlled.
In briefRLVR Sharpens the Pencil, Not the Mind: How Verifiable-Reward RL Affects Reasoning
SummaryThis paper examines how Reinforcement Learning with Verifiable Rewards (RLVR) improves reasoning in language models. The authors tracked OLMo-2-1B across 13 training checkpoints on GSM8K, finding first-try accuracy rose from 58.0% to 61.3% while solution coverage over 256 samples rose from 94.5% to 96.9%, suggesting the knowledge was already present. Probes for detecting planted reasoning errors performed about the same across checkpoints, and the main internal change appeared at the final answer token.</summary>
In briefOrganizational learning in materials discovery: which databases, calculations, and agent teams are worth using
SummaryThis paper examines how combining databases, calculations, and agent teams affects materials discovery campaigns. Using archived measured band gaps, the authors rank information sources and find that pooling all sources ranks worse than pooling the best three. They also report that flat language-model agent teams trailed a random-pick baseline, while an adaptive team found more of the lowest-formation-energy compounds within its initial compute budget.
In briefSeMo3D: One Model for 3D Generation, Editing, and Completion
SummaryThis paper proposes Selective Modeling, which synthesizes only the designated region of a 3D asset (the region to generate, edit, or complete) while keeping the rest unchanged. The resulting SeMo3D framework unifies the three tasks and can be trained without paired editing or completion data, using region-selective denoising on standard 3D assets. It also introduces Omni3DiT, a multi-condition 3D diffusion transformer that fuses image, text, and point guidance. The authors report that SeMo3D outperforms state-of-the-art methods on all tasks across three benchmarks.
In briefCR-Bench-v2: Testing How Useful AI Code Review Agents Really Are
SummaryThis paper introduces CR-Bench, a benchmark that frames code review as a task of finding known defects in pull requests with full repository context. It contains 584 instances, including a 174-instance subset derived from SWE-Bench Verified, each labeled with its defect, ground-truth fix, and a taxonomy of category, impact, and severity. The authors propose two metrics, Known-Defect Detection Rate and Usefulness Rate, and evaluate eight code review agents plus Gemini Code Assist. Even the strongest systems detect only about one-third of known defects, while Gemini Code Assist detects 8.1%.
Kristen Pereira, Neelabh Sinha, Rajat Ghosh et al.
In briefEEFA: A framework for checking the ethics of AI agents
SummaryThe paper argues that current AI agent benchmarks measure what agents can do but not how they behave ethically. It reviews twelve leading benchmarks to show which ethical dimensions they cover, then introduces the Ethical Evaluation Framework for AI Agents (EEFA). EEFA has five dimensions: Autonomy Preservation, Harm Avoidance, Transparency and Non-Deception, Privacy Compliance, and Accountability and Traceability, and comes with a reproducible validation protocol. The authors say it maps onto the EU AI Act and the NIST AI Risk Management Framework.
In briefSequential recovery updates for learned robot controllers that keep successful trajectories intact
SummaryThe paper addresses improving an already competent learned controller by adding feedback skills at measured failures, with an entry rule fitted to exclude every state on an accumulated set of successful trajectories. A skill is admitted only if the composed controller reproduces the new recovery and replays all protected trajectories exactly. Tested on four streams covering two motions on each of two humanoid systems, two updates per stream added 62 recoveries over 192 withheld perturbations without losing any base-controller successes.
Zongtan Li (University of Georgia, Athens, GA et al.
In briefFlat Features, Detailed Images? A Simple Detail Control Method for Diffusion Transformers
SummaryThe paper finds that prompts encouraging detail raise the entropy of feature token energy distributions in diffusion transformers while largely preserving their ordering, so flatter distributions correspond to more detail. Building on this, the authors propose a training-free method that continuously adjusts detail with a scalar, using a time-aware scheduler to focus the intervention on intermediate sampling steps. Experiments across several diffusion transformers show continuous control over detail metrics while largely preserving image quality and semantic structure.
In briefRe³Rank: Listwise Reranking Using Reasoning-Conditioned Retrieval Heads
SummaryThe paper proposes Re³Rank, a listwise reranker built on reasoning-conditioned retrieval heads (RRHeads). A generator first writes a reasoning trace over the question and candidates, then a scorer uses the trace tokens as attention queries to produce continuous relevance scores. A straight-through estimator lets the ranking loss pass through the discrete trace so both components are trained jointly. Averaged over 9 benchmarks, the method outperforms 11 baseline rerankers, with the 4B variant leading the strongest baseline by 3.11 NDCG@10 points and the 0.6B variant surpassing rerankers up to 32B parameters.
In briefEnergy-Based Policy Orchestration: Choosing Feasible Robot Policy-Goal Pairs at Runtime
SummaryThis paper addresses hierarchical robot control, where a fixed planner-to-policy interface can fail when contacts, balance, or disturbances change faster than high-level updates. It formulates policy orchestration as selecting, at runtime, a policy-goal pair that stays feasible while tracking the planner's target. The authors propose EPO, which learns a policy-specific energy over low-level goals, where lower energy means higher feasibility, and reports improved precision-robustness trade-offs in simulation. They also demonstrate sim-to-real deployment of the same frozen policies and energy models on real humanoid loco-manipulation tasks.
In briefFindLink: Learning to Assess Connections Between Scientific Findings From Source-Grounded Premises
SummaryThe paper introduces GI-TraceBench, a benchmark of 10,282 gastrointestinal oncology questions covering provenance question answering, knowledge discovery, and hypothesis evaluation. It also proposes FindLink, a two-stage framework that builds source-bound premises with study context and support status, then organizes them into support plans to generate evidence-conditioned answers. On a 1,545-question fixed-evidence test set, FindLink raises the balanced family macro score from 67.95 to 87.25 over direct supervised fine-tuning of the same backbone.
In briefPixelBench: Can LLMs Generate Images Pixel by Pixel?
SummaryThis paper examines whether large language models can create images directly as pixel grids, rather than through symbolic formats like SVG or CAD. The authors introduce PixelBench, which tests text-to-pixel and image-to-pixel generation at several resolutions, using an RGB palette and a grid of color indices, along with a pixel-understanding evaluation. They also propose Pixel Reconstruction Tuning, which post-trains Molmo2-4B on caption-palette-grid data and reports improved results across 16 visual benchmarks.
In briefAgent-based model of knowledge arbitrage in science communities
SummaryThe paper introduces Scientific Liquidity Agents (SLA), an agent-based model in which domain specialists and intermediary arbitrageurs move ideas, methods and results between research contexts. Two exploratory simulations with 10 and 18 agents produced 68 and 274 papers. The authors report that no recorded replication succeeded, suggesting verification rather than production is the bottleneck.
In briefFORGE: Generating Executable Repair Code for Interlocking Assembly Faults
SummaryThis paper presents FORGE, a framework that helps large language models generate executable rework programs to restore interlocking assemblies to constrained states after a fault. The authors also introduce FORGE-Bench, which contains 4,000 tasks across 26 interlocking assemblies, with five held-out assemblies used to test generalization to unseen structures. Experiments show that locally reasonable repair code is insufficient for reliable recovery, and that linking structural recovery reasoning with executable action realization is essential.
In briefTune the Baseline First: Most of Muon's Scaled Weight Decay Gain Comes From Tuning, Yet Its Small Real Edge Grows With Model Size
SummaryThis paper compares Muon with scaled weight decay (Muon-SW) against constant decoupled weight decay for training language models. The authors tune learning rate and weight decay jointly for each method across a mixture-of-experts scaling ladder. They find that about 90% of the apparent advantage disappears once both methods are tuned, while the remaining small advantage grows with model size.
In briefINTERACT: Learning how traffic and the ego car interact from driving videos to plan safer paths
SummaryThis paper proposes INTERACT, a three-stage framework that learns an object-centric world model from unlabeled driving videos and transfers it to trajectory planning. It groups dense visual features into slots representing scene entities, infers latent ego actions, and uses them to predict how entities evolve under the vehicle's motion. The learned representations are distilled into a planner to guide trajectory generation and selection. Experiments on NAVSIM and nuScenes show state-of-the-art results on NAVSIM with improved collision metrics.
In briefA study of how alternating updates make regret matching algorithms converge
SummaryThis paper analyzes the convergence of regret matching algorithms when players are updated in turn rather than simultaneously. The authors introduce an alternating version of the RVU property and use it to show optimal T^{-1} average-iterate convergence for two predictive regret matching+ variants. They find that half prediction is the unique prediction multiplier satisfying this property, and report comparable performance to state-of-the-art solvers on 15 game benchmarks.
WhyThe authors introduce an alternating RVU property, obtaining optimal T^-1 average-iterate convergence and showing half-prediction is the unique satisfying multiplier.
Anh Duc Nguyen, Ioannis Anagnostides, Brian Hu Zhang et al.
In briefMedHEB: A benchmark testing medical embeddings across mixed clinical evidence types
SummaryThis paper introduces MedHEB, a unified benchmark of 107 retrieval tasks drawn from 54 datasets, covering 2D images, 3D volumes, and text. The tasks span image-image, image-text, text-text, and modality-bridging visual matching, at global, question-conditioned, and region-level granularity. Results show that models strong on conventional image-text or text-text retrieval are less reliable on modality-bridging and region-level tasks. The authors also release MedEmb, an MLLM-based reference embedding model, along with evaluation code and data access manifests.
In briefEvolving, audited guardrail tools to keep LLM agents safe
SummaryThis paper addresses the problem that LLM agents operating in open-world settings face safety risks that cannot all be predefined, while most existing guardrails are static. It presents EVOLVE, a self-evolving defense framework that turns security knowledge into executable guardrail tools that are inherited, independently audited, and refined through lifelong interaction. The authors report that EVOLVE outperforms existing defenses by an average of 22% in F1 score across multiple benchmarks, backbone models, and embodied environments. They also validate the framework in real-world robotic and autonomous agent systems.
In briefLatent spectral prompting for tracking objects in hyperspectral video
SummarySProTrack is a framework for tracking objects in hyperspectral video, which offers rich spectral detail useful under camouflage or low light. It transfers temporal knowledge from pretrained RGB trackers to the hyperspectral domain using latent spectral prompts. An autoencoder learns a global scene-level spectral representation from full hyperspectral frames, guiding prompt extraction while the RGB tracker handles temporal modeling. Experiments on HOT2020 and HOT2023 show the best overall performance across spectral domains.
In briefWhen Predictions Shape the Future: A Research Agenda for Reflexive Time-Series Analysis
SummaryThe paper notes that time-series models in areas such as inventory management and clinical care trigger interventions whose actions change later observations, creating what the authors call AI-shaped trajectories. Reusing such records as passive data can lead to misreadings, such as treating a stockout prevented by a prediction as stable demand. The authors propose analyzing the prediction-action-observation loop rather than intervention-agnostic sequences, and call for closed-loop provenance, sequence identification, and closed-loop evaluation.
In briefNeuro-Inspired Robotics: Lessons from the Sensorimotor Brain for Robot Learning
SummaryThis survey argues that robot policies still fail at everyday manipulation despite large-scale pretraining, and that the gap stems from design rather than data alone. It draws on the nervous system, including body compliance, a fast visual pathway for action, cerebellar prediction, and motor cortex composition and verification of actions. The authors propose these designs for next-generation robot models and offer each as a hypothesis to be tested.
In briefBalloon effect: blocking one channel moves a hidden trait to another
SummaryThis paper studies how language models trained on other models' outputs can acquire traits never stated in those outputs, a phenomenon called subliminal learning. The authors close each trait's output channels one by one across two stress tests, two models, six output formats and 16 traits. Banning surface tokens did not remove the trait; it shifted to wherever the format still allowed, which they call the balloon effect. They trace this transfer to individual attention heads and neurons and show that patching the circuit that writes the trait into the output halves the trait passed to a student model.
In briefAttentionIndex: Faster Learned Sparse Attention Using a Causal Inverted-File Index
SummaryThis paper addresses the cost of learned sparse attention, where a lightweight indexer scores every prior key and this selection takes over half of prefill time at one million tokens. It proposes AttentionIndex, which routes each query through an inverted-file index that groups keys by similarity and refines candidates with the native indexer score. On GLM-5.3 and DeepSeek-V4-Flash, it achieves 1.3 to 1.8 times the prefill throughput of exhaustive selection and improves long-context benchmark accuracy by up to 18.4 percentage points over HiSA.
In briefCellPlex: A Benchmark for Predicting Single-Cell Spatial Proteomics from H&E Slides
SummarySpatial proteomics measures dozens of proteins in each cell of a tissue section but is costly, so it is used mainly in small studies. This paper introduces CellPlex, a benchmark that evaluates methods for predicting per-cell protein levels from routine H&E-stained slides across seven cohorts, 717 samples, 44.8 million cells and 149 proteins. Adapted pathology foundation models were the most accurate on every cohort, but kept less than half their accuracy on cohorts they were not trained on.
WhyThe authors build a cell-level benchmark across seven cohorts (717 samples, 44.8M cells, 149 proteins) comparing pathology foundation models against specialist virtual staining models.
In briefQFacePlacement: Where Quantum-Inspired Tensor Models Fit in Face Verification
SummaryThis paper examines where quantum-inspired tensor-network models should be placed in face verification pipelines. It compares two interfaces: aligning AdaFace embeddings to ArcFace space, and building a local correspondence tensor before global pooling. Dense mappings remain stronger than tensor compression, which is credible only when the interface offers task-aligned, compressible structure. The authors also argue that aggregate AUC should be assessed alongside low-false-acceptance metrics.
Krishna Bhatia, Kavita Choudhary, Pruthwik Mishra et al.
In briefAligning LLMs to multiple values by locating value-sensitive attention heads
SummaryThis paper studies multi-value alignment of large language models from a mechanistic interpretability perspective, using a probing framework to locate attention heads sensitive to values such as harmlessness and honesty, as well as style preferences like coherence. It proposes an inference-time activation steering method that controls behavior toward single or combined values without updating model parameters. Experiments across alignment benchmarks show effective control while largely preserving performance on non-target behaviors.
In briefSpecPilot: Automatically Discovering Faster Ways to Run Speculative Decoding for Language Models
SummaryThis paper introduces SpecPilot, an autonomous agent loop that finds algorithm and systems optimizations for speculative decoding in each deployment setting. A cost model identifies where the current system falls furthest short of its ceiling, and parallel agents explore and integrate successful changes. Evaluated across four target models, three drafters, two GPU platforms and five benchmarks, it raises the speedup over autoregressive decoding from 4.46x to 10.87x with one request and from 1.94x to 3.93x with batching, without training or quality loss.
In briefEvoOcc: A Fast-Slow Framework Balancing Accuracy and Latency in 3D Occupancy Prediction for Autonomous Driving
SummaryThis paper addresses 3D semantic occupancy prediction for autonomous driving, where accurate and low-latency scene representations are both needed. The authors propose EvoOcc, a framework that combines the complementary strengths of slow, accurate models and fast, deployment-friendly models without changing their architectures. It models occupancy evolution as a continuous-time controlled dynamical system and splits the evolution into ego state and scene state components. Experiments on three fast-slow system pairs and irregular prediction arrivals show consistent gains over fast baselines with much lower latency than the slow system.
In briefSemantic caches for LLM services leave error budgets unused
SummaryThe paper presents Surplus, a semantic caching method for LLM-based web services that ranks queries by a shared risk score and opens risk bins as far as an audited error estimate allows. This spends unused error budget where reuse is most likely correct, while free labels and a ledger of realized errors keep the error rate within budget under drift. Across five query streams Surplus stays within the error budget in every test cell, serves 40% more hits than vCache on shuffled streams, and cuts mean latency by 31.59% at the highest load in a vLLM testbed.
In briefRoboColosseum: Large-scale real-world testing of robot physical intelligence
SummaryThis paper introduces RoboColosseum, a distributed framework that runs standardized robot tests across many locally operated robots through shared policy serving and common protocols. It evaluates five vision-language-action policies and three frontier AI agents on four robot embodiments, using blinded pairwise trials and a fine-tuning track with 100 demonstrations per task. Frontier agents matched specialized policies, but all systems struggled to complete open-world tasks reliably. The authors also find that open-track rankings poorly predict performance after fine-tuning, and that measuring partial progress reveals differences missed by task success alone.
Tan Sang Nguyen, Yi Ru Wang, Shailesh Shailesh et al.
In briefSame sentence, different construct: testing whether LLM hate speech classifiers judge groups equally
SummaryThe paper tests large language model hate speech classifiers using 421 templates that differ only in the identity term for seven protected groups. Four models each exceed 0.95 balanced accuracy, yet they miss hateful statements about women at notably higher rates, and their stated rationales shift for the same statement depending on the group. The authors argue that validating LLM measurement instruments requires identity-controlled tests of measurement invariance that examine both decisions and the criteria models give for them.
In briefPretraining one encoder on unlabeled time series from multiple industrial domains
SummaryThis paper proposes JEPA-HITS, a self-supervised joint-embedding pretraining method that learns a single encoder from heterogeneous industrial time series in battery, semiconductor CMP, CNC machining, and bearing domains while preserving each dataset's temporal scales. With leave-one-domain-out pretraining, compared with random initialization, battery remaining useful life RMSE drops from 10.72 to 7.95, CMP from 4.34 to 3.89, and tool wear regression from 13.53 to 11.72, while bearing fault diagnosis macro-F1 rises from 0.547 to 0.737. The results suggest unlabeled measurements from heterogeneous industrial processes can yield reusable representations for downstream learning with limited labels.
Jonathan Adam Rico, Nagarajan Raghavan, Senthilnath Jayavelu
In briefStability of Online Active Learning in Large Neural Networks
SummaryThis paper studies whether online active learning, where deep neural networks keep adapting during real-time interaction, remains stable as networks grow to multiple layers. The authors use prediction error as feedback and link the update process to Lyapunov-based energy minimization. They show that output error convergence alone is not enough for parameter convergence, and use persistency of excitation and active damping together with a cascaded Lyapunov analysis to prove asymptotic stability of the overall learning process.
In briefRecRSI: A framework that finds better recommendation algorithms through mechanism-based hypotheses
SummaryThis paper presents RecRSI, an LLM-based research framework that organizes the discovery of recommendation algorithms around mechanism hypotheses. Its components derive design primitives, develop and test hypotheses, and feed experimental outcomes back into later research. Across four initial recommenders and four datasets, it achieved average relative gains of 8.48% in NDCG@5 and 6.88% in HR@5 over the initial algorithms.
In briefTHReD benchmark: testing an LLM agent and its tools for turning health records into AI-ready data
SummaryThis paper benchmarks THReD, a web system that converts electronic health record extracts into AI-ready datasets through a chat agent whose proposed steps run only after user confirmation. On 141 tasks, the agent completed 92.7% on its first answer, abstained on every advice request, recovered from every injected fault, and never acted on instructions planted in the data. Tool experiments show high accuracy in inferring patient classes and vital-sign labels, byte-identical outputs when saved pipelines are replayed, and processing time growing roughly linearly with record count.
In briefOn-demand evidence chains for long-term conversation memory
SummaryThis paper proposes Chain on Demand (CoD), a framework that organizes evidence at query time over an unchanged turn-level conversation memory. It selects a fixed budget of anchor turns from a candidate pool based on query relevance, lexical diversity, and newly covered query terms, then restores each anchor's same-session neighbors to form a temporary chain used only for the current query. On LoCoMo, CoD reaches 82.2% accuracy on 1,540 non-adversarial questions and 76.3% on all 1,986 questions; on LongMemEval-S it reaches 86.6% across 500 questions.
In briefMetaPOCS: Meta-learned multi-exit network for seismic data filling under varying compute budgets
SummaryThis paper presents MetaPOCS, a framework for interpolating seismic data that balances reconstruction accuracy against computational cost. It builds on a spatial-spectral POCS backbone, supervising intermediate stages and keeping acquired traces at every exit. A Meta-GF optimization learns fusion weights for exit-specific parameter updates, and a latency-calibrated exit policy lets one checkpoint serve several reconstruction budgets.
In briefGigaSpeech 3: Scaling Multi-Dialect Arabic Speech Recognition with Weakly Labeled Data
SummaryThis paper presents GigaSpeech 3, a corpus of 39,000 hours and 10 million weakly transcribed speech segments covering five Arabic dialects: Algerian, Egyptian, Iraqi, Saudi, and Syrian. The corpus was built by training labeling models on 60 hours of manual transcription per dialect and applying them to dialect-matched web audio, followed by noisy-student refinement. Dialect-specific models trained on the corpus outperform their labeling models on all five dialects, reaching a macro-average word error rate of 23.86% on GigaSpeechBench.
In briefGnomon4D: Robot world model that learns where and when to focus using geometric motion
SummaryThis paper presents Gnomon4D, a world-action model that jointly generates future video, robot actions, and GeoDelta, a video-aligned 4D geometric motion field. GeoDelta encodes cross-frame image-plane displacement and depth change for robot and object points, giving the action expert access to their motion relations. Two attention mechanisms use these relations to focus on task-relevant interaction regions and on transitions such as grasp and release. The authors report leading results on RoboTwin 2.0, VLABench, and RoboDojo, and gains of 7.0 and 16.0 percentage points over the strongest real-robot baseline in in-distribution and out-of-distribution settings.
In briefMAGAR: Multimodal AI Agents for 3D Mechanical Assembly via GUI with Minimal Rework
SummaryThis paper introduces MAGAR-Bench, a benchmark of 4,000 multimodal samples from 26 assembly families, which separates assembly decisions, atomic actions, and minimal rework. It also presents MAGAR, a selective correction framework built on a frozen supervised fine-tuned policy, where a learned selector decides whether to apply candidate corrections. On held-out assembly families, the Overall score rises from 37.50 to 41.39, with gains on all three tasks. Of 209 interventions, 70 corrected an error and 22 introduced one.
In briefMulti-view thinking and adaptive generation for generative recommendation (MV-Thinker)
SummaryThis paper proposes MV-Thinker, a framework for semantic-ID-based generative recommendation that predicts the next item autoregressively. Multi-View Temporal Thinking combines overlapping full-history and recent-history views into user-conditioned contexts before SID generation. Hierarchical Residual Adaptation applies nested lightweight adapters to decoder states before a shared vocabulary projection. The authors report that it outperforms strong baselines on four public datasets across multiple ranking metrics.
In briefTraining language models on rounded, bounded, or censored numbers using exact likelihood
SummaryThe paper introduces boundary marginalization, a method that trains autoregressive language models on the exact likelihood of numerical labels that are rounded, bounded, or censored, using their existing digit tokens. It uses a fixed-width, order-preserving numeral representation so that interval probabilities and gradients can be computed from at most two boundary paths. Experiments show that conditional distributions can be recovered from rounded and censored reports alone, and that code latency distributions become more accurate on two backbones.
In briefLearning Entropy-Stabilized Coarse-Grid Turbulence Emulation with Neural Discrete Equilibrium
SummaryThis paper uses the Neural Discrete Equilibrium (NeurDE) method to emulate physical flows on extremely coarse grids. It adds physically grounded control of entropy dynamics to reinforce the system's dissipative mechanism and suppress grid-scale aliasing noise. For turbulent flows at Reynolds number 100,000, the entropically stabilized model produces stable rollouts over thousands of steps on grids as coarse as 16×16. At these scales, it is reported to be more stable than the stabilized numerical method used to generate the training data.
Jose Antonio Lara Benitez, Kareem Hegazy, Anastasis Kratsios et al.
In briefMulti-teacher reasoning distillation with self-exploration for LLM recommendation
SummaryThis paper proposes MUSE, a framework that distills reasoning abilities from large language models into lightweight recommender systems. It builds a reasoning pool from multiple teacher LLMs and expands it through student self-exploration. A Rank-Gain Utility Assessment scores reasoning trajectories by how much they improve the ground-truth item's rank across the full catalog, and dual-level contrastive learning uses these scores for training. The authors report experiments on three real-world datasets showing effectiveness and robustness.
In briefRLiD: reinforcement learning that picks actions directly for language agents and LLMs
SummaryRLiD is a framework that trains language agents by directly optimizing the probabilities of executable action options rather than generating text token by token. Observations, an optional compact memory, and a labeled action menu are fed to a large decision model that selects an action in one forward pass, avoiding invalid action outputs and improving rollout efficiency. The trained policies are also distilled into generative LLMs through supervision on their action trajectories. Experiments on ARC, Countdown, Sokoban and ALFWorld report better performance or lower compute, with ALFWorld reaching 99.3% success in 12.7 H100 GPU-hours.
In briefRecGroupMemory: Feedback-Aware Cohort Memory for Recommendation
SummaryThis paper presents RecGroupMemory, a workflow that gives an LLM-based orchestrator group-level user profiles and advice, separating memory construction from feedback-based revision. It uses the previous day's feedback to revise decision guidance while validation checks keep selected source values unchanged. In online comparisons with Taobao users of low purchasing power, memory with and without feedback increased the number of guided transaction users by 1.58% and 1.78% over a no-memory baseline.
In briefTurbo-GMM: a GPU-accelerated Gaussian mixture model inspired by FlashAttention
SummaryThis paper proposes Turbo-GMM, a fused GPU kernel for the EM algorithm of Gaussian mixture models that completes responsibility normalization and sufficient-statistic accumulation in a single pass over the component dimension. Compared with the two-sweep Flash-GMM, it avoids a second score computation. On an NVIDIA A100 across sixteen configurations, it reports a median 2.15× speedup over Flash-GMM and up to 5.3× and 9.9× over PyTorch and scikit-learn. Its applicability is limited by on-chip resources such as register pressure and occupancy, and larger configurations use a two-pass fallback.
