In briefReinforcement learning for optimizing target polarization in nuclear physics experiments
SummaryDynamically polarized targets need continuous microwave frequency tuning to offset radiation damage and material changes, a task traditionally done by expert trial and error. Using operational data from the APOLLO cryogenic target, the authors combine surrogate models (MLPs and Gaussian process regression) with reinforcement learning in a data-driven control framework. They report that their RL agent achieved almost a 2x improvement over operators' actions.
Armen Kasparian, Torri Jeske, Monibor Rahman et al.
In briefSimulation-Based Agentic VLM Framework for Wildfire Monitoring and Reporting
SummaryThe paper proposes a vision-language model framework that converts 2D wildfire simulations into labeled video episodes, using low-detail Blender 3D proxies and controllable video generation to build multimodal memory. A training-free multi-agent VLM retrieves reference episodes and produces structured wildfire reports. On held-out generated episodes, video memory reaches 51.5% exact four-tag accuracy, and the full system reaches 77.3% on six simulator-derived report fields.
In briefCounterexample generation with per-theorem verifiers: when imitation hurts and reinforcement learning repairs it
SummaryThe paper observes that large language models often prove a theorem forward but fail to disprove a closely related false one. The authors release SymCE, a corpus of 4,707 false undergraduate algebra and real analysis conjectures, each paired with an executable Python verifier that also serves as the reward function. Supervised fine-tuning on counterexamples alone drops true-theorem recognition from 0.27 to 0.00, while reinforcement learning with a sparse outcome-only reward repairs this and reaches 0.66.
WhyFrames counterexample generation with deterministic verifiers as RL reward, revealing an imitation trap; multi-seed and multi-benchmark experiments give solid evidence, with impact mostly in a subfiel
Omar Farouk Zouak, Houssam Eddine Boukhalfa, Soumaya Lakehal et al.
In briefAI-driven planning of data center capacity that accounts for heat
SummaryLarge language model workloads create localized GPU hotspots and sudden thermal spikes that make cooling hard to control. This paper presents an AI framework that performs thermal-aware capacity planning for a real-world data center in seconds. The model learns from parameters such as rack power, server power, server placement, and HVAC settings, and predicts temperatures within milliseconds. Tested against high-fidelity CFD simulations on unseen designs, it reports high accuracy with a 10000X speedup.
In briefContextual Bandits Using Additive Residual Quantization Representations
SummaryThe paper proposes Residual Quantization, an offline-trained codebook that maps continuous contexts to discrete multi-level centroid assignments, enabling additive bandit algorithms with nonlinear expressivity. The method is designed to keep memory strictly bounded while allowing efficient online updates. Across 13 datasets, the RQ variants outperform their non-RQ counterparts on 11.
In briefA generative graph model combining interpretable graphon parameters with neural inverse operators for inference on unseen graph sizes
SummaryThe paper proposes a generative model for complex networks based on multifractal step graphons, which compactly parameterize networks through a recursive construction. A neural inverse operator recovers the parameters, allowing inference on graph sizes not seen during training. In a zero-shot graph-generation benchmark, the method achieves the best average performance on three of four metrics, and on single-observation networks it performs comparably to an instance-specific optimization method. In a multi-subject EEG case study, the inferred parameters track a reversible change in brain state more sensitively than traditional network statistics.
WhyCombines multifractal step graphons with neural inverse operators to unify mechanistic and amortized inference; zero-shot and EEG experiments support it, impact mainly within network science.
Wooseong Choi, Italo'Ivo Lima Dias Pinto, Chen Sun et al.
In briefDo LLMs Synthesize or Recall Algorithm Code? A Benchmark for Algorithmic Code Retrieval
SummaryThis paper argues that for well-known algorithms, LLM code generation is better understood as parametric code retrieval, meaning reproducing an algorithm from memorized knowledge rather than creating a new one. The authors introduce AlgoREval, a benchmark of 599 problems covering 77 classical algorithms across 7 programming languages and 4 graph-input representations. They evaluate 15 models with 7B to 34B parameters in a zero-shot setting, finding large variation in accuracy across languages and input formats. The results also compare prompt augmentation, SFT, and GRPO, and caution against deploying AI-generated algorithmic code without systematic validation.
In briefLearning Style, Forgetting Semantics: SFT vs. RFT on Classification Tasks
SummaryThis paper studies why supervised fine-tuning (SFT) causes more forgetting than reinforcement fine-tuning (RFT) even when teacher demonstrations are semantically correct. It uses classification tasks where same-class tokens express one answer in different styles, and a linear-softmax policy to split updates into semantic and style components. The authors show that SFT with a nonuniform teacher develops style drift, giving a positive forgetting lower bound under explicit conditions, while RFT with exact policy gradients keeps zero semantic error. Simulations over task sequences support these predictions.
Haodong Liang, Yanhao Jin, Krishnakumar Balasubramanian et al.
In briefToken utility for on-policy distillation via local trust-region optimization (paper)
SummaryThis paper addresses which tokens to prioritize in on-policy distillation (OPD), deriving token importance directly from the distillation objective as the loss decrease under an equal local KL budget. The resulting utility, computed from the logit gradient and the categorical Fisher matrix, is used to build Fisher Utility Weighting (FUW), a continuous, mean-normalized token weighting. Across seven math reasoning benchmarks and three Qwen3 student-teacher pairs, FUW improves over uniform OPD and is competitive with existing selective methods.
In briefCRISP: Rebuilding Binary Neural Networks' Final Hidden Layer as Readable Logic Clauses
SummaryThe paper presents CRISP, a framework that reconstructs the last-layer activation vector of binary neural teachers using Tsetlin Machine clauses, so each hidden bit traces back to named input thresholds. It is evaluated on MNIST, KMNIST, FashionMNIST, SVHN, and CIFAR10 with BinaryConnect CNN and fully binary network teachers. The results report that sign-binarization did not reduce teacher-head accuracy in the tested BNN setting, and give reconstruction fidelity across several Booleanization methods.
Alex Chan, Shafi Muhtasim Chowdhury, Ekin Can Erku\c{s} et al.
In briefGeometry-aware time reparameterization to make flow-map distillation easier to learn
SummaryThis paper studies whether changing a teacher model's time parameterization makes flow-map distillation, which enables one- and few-step generation, easier to learn. The authors propose a geometry-aware reparameterization that allocates more student time to trajectory regions with large normal acceleration while preserving the teacher's geometric paths and terminal distribution. The clock is estimated once before distillation and adds no student parameters or inference-time evaluations. Experiments on synthetic data, CIFAR-10, and CelebA-64 show improved sample quality over identity-time distillation at matched inference budgets, especially for one-step generation.
In briefFinding the Move Is Not Winning: XiangqiBench for Closed-Loop LLM Agent Evaluation
SummaryThe paper introduces XiangqiBench, an executable Chinese chess benchmark where LLM agents must deliver checkmate against an engine across 119 tactical endgames. The authors recorded 8,568 multi-turn trajectories from 12 frontier models and found that naming a correct first move often does not lead to a win. Reliability also drops sharply across repeated trials, and simulated rollouts frequently differ from the real opponent's replies. The authors argue that agent evaluations should score closed-loop outcomes and report reliability alongside coverage.
In briefAI Multi-Stage Method for Assessing Androgenetic Alopecia From Low-Magnification Scalp Images
SummaryThis paper presents an automated scalp-analysis and decision-support framework for assessing androgenetic alopecia (AGA). It combines follicular unit localization, visible-shaft counting, calibrated shaft-width estimation, regional aggregation, and an interpretable rule layer. Under patient-disjoint evaluation, YOLOv8m reached a test mAP@0.5 of 0.920 and recall of 0.860, while EfficientNet-B5 with a support-map channel reached 87.0% expert-box count accuracy. The system is intended to assist, not replace, dermatologist interpretation.
In briefAI Theorist uncovers excitonic structure in α-RuCl3
SummaryThis paper introduces AI Theorist, a system of AI agents that autonomously develops physical models through hypothesis generation, first-principles calculations, and evidence-driven refinement. The authors apply it to α-RuCl3, a candidate material for a Kitaev quantum spin liquid, analyzing its optical spectra and photocurrent observations. The system proposes a new interpretation identifying distinct excitonic states with different optical selection rules and real-space distributions. The authors state this is the first demonstration of an AI system autonomously building a physical model to explain previously unpublished experimental data in a quantum material.
In briefPrompted to Discriminate: Generalizing Malicious-Input Probes in the Wild
SummaryThis paper studies whether appending a short classification instruction after user input improves activation probes that detect harmful inputs such as prompt injections and jailbreaks in LLM agents. Across 13 safety benchmarks and three open-weight model families, using leave-one-dataset-out evaluation, classification-style suffixes improved detection of unseen attack types by up to about 4 AUC points. The gain comes mainly from the classification format rather than the named criterion, and it also appears in multi-position probes, though the best suffix depends on the model and readout.
In briefFaster neural field training through region-based adaptive sampling (ACES)
SummaryImplicit neural representations model high-dimensional continuous fields, but uniform sampling during training wastes effort on uneven spatial regions. This paper proposes ACES, which builds adaptive spatial partitions for domain coverage and applies region-level importance weighting to focus training on informative regions. The authors provide a theoretical analysis of gradient variance and bias, and report on scientific field learning tasks that ACES converges faster and reaches lower error than uniform and pointwise adaptive baselines, with the largest gains on highly localized fields.
In briefWhy terminal-agent training stalls: data generation and verification issues
SummaryThis paper examines using frontier models such as Claude Opus as meta-agents to generate terminal tasks and verifiers for RL training. It argues that a runnable Docker image and executable test suite do not guarantee a reliable end-to-end pipeline. The authors diagnose three failure classes: benchmark invalidity, harness brittleness, and reward misalignment. They report that prompt redesign and context extension raised baseline solvability 5.6 times, but adding hard tasks lowered mean pass@2 to 20.6% under the same training configuration.
In briefTrained agentic context management lets a small model handle long documents with short context
SummaryThe paper studies long-context language models using a minimal harness: a tool that lets the model call itself with any prompt, and a tool that reads a specified token range from the input. The authors finetune Qwen3.6-35B-A3B on a diverse synthetic dataset using this setup. They report that, with only 8,000 tokens of context, the small model matches GPT-5.4 with 1M tokens of context on the OOLONG-synth benchmark when documents exceed 40K tokens.
In briefEnergy Saving in 5G Networks Using Quantum Reinforcement Learning
SummaryThis paper addresses rising energy consumption in 5G and beyond networks, focusing on base station power use. It proposes a Quantum Reinforcement Learning (QRL) algorithm built on parameterized quantum circuits, which the authors say converges faster than Deep Reinforcement Learning. Simulations reportedly show substantial energy reduction while maintaining Quality of Service, even with highly dynamic users, and better convergence speed and learning complexity than DRL and Q-Learning.
Muhammad Usman, Nguyen Van Huynh, Marianna Lezzi et al.
In briefVisAudit: Benchmark for AI Agents Diagnosing and Fixing Chart Errors
SummaryThis paper introduces VisAudit, a benchmark that tests whether multimodal AI agents can diagnose, repair, and verify errors in data visualizations. It includes 1,900 flawed charts across 21 chart types and 10 flaw categories, plus 300 initially correct charts, built by controlled perturbation of validated visualizations. Experiments show a substantial gap from reliable autonomous review, as the strongest evaluated model fully repaired only 47.4% of flawed charts in the autonomous-repair setting.
In briefValidated Data Onboarding for AI Demand Forecasting on U.S. Building Meter Data
SummaryThis paper presents a pipeline that detects and repairs defects in building meter data before model training, using only information available at forecast time. Tested on hourly electricity data from twelve U.S. buildings in the public Building Data Genome 2 dataset, injected defects raised a gradient-boosting forecaster's error by 86%, and after detection and repair the error returned to clean-data levels while retaining 93% of training targets. An earlier version that over-cleaned natural data made forecasts 25% worse; this negative result is retained and corrected through a dated amendment.
In briefInherit-MAS: Improving multi-agent AI systems during use by reusing workflows and results
SummaryThis paper introduces Inherit-MAS, a method for refining multi-agent systems built from large language models while they run. Each refinement starts from the latest completed workflow, and stored results are reused only when the full request and execution context match. On WorkBench and HotpotQA FullWiki, it outperforms EvoAgent, EvoMAS, and TacoMAS, and it reduces token usage compared with rerunning without execution inheritance.
In briefFlashSinkhorn 2: A Faster Solver for Large-Scale Optimal Transport
SummaryThis paper presents FlashSinkhorn 2 (FS2), a GPU solver for large discrete entropic optimal transport problems on low-dimensional point clouds with squared-Euclidean cost. It couples a coarse stage that solves on cell centroids with a block-sparse fine stage that corrects the centroid error. On one A100 GPU, the authors report solving discrete EOT between two measures of 1.34×10^8 particles from a cosmological N-body simulation to a marginal residual below 0.01 in under 2.5 hours. An open-source implementation is released.
WhyA coarse-to-fine block-sparse solver substantively improves streaming Sinkhorn, validated by large-scale experiments on 134 million particles.
Felix X.-F. Ye, Yu Chin Fabian Lim, Naigang Wang et al.
In briefHesitation Has a Geometry: Entropy-Trained Hyperbolic Probes for Sparse Activation Steering
SummaryThis paper proposes Hyperbolic Entropy Steering (HEST), a method for steering language model activations that embeds hidden states in the Poincaré ball using a lightweight probe trained on the model's own next-token entropy. Steering is applied only where entropy exceeds a threshold, moving states along the geodesic of steepest descent of the probe's readout. Tested on three instruction-tuned models, HEST with the Busemann readout improves greedy accuracy on MATH-500 and GSM8K in five of six settings, while a contrastive steering vector added at every token lowers accuracy.
Zeyong Zhang, Tung Sum Thomas Kwok, Tengfei Ma et al.
In briefOctrees as an Explicit 3D Language: OctLLM adds 3D ability to LLMs using sparse octree tokens and separate branches
SummaryThe paper presents OctLLM, which represents 3D geometry as an explicit sequence of octree occupancy tokens, pruned into a shorter Sparse Octree (S-Octree). Mesh tokens pass through independent trainable branches in some blocks, while text and image tokens keep a frozen vision-language pathway, interacting via shared self-attention. The authors report training far fewer parameters than full fine-tuning, achieving state-of-the-art results among unified multimodal LLMs and matching the backbone on general language benchmarks.
In briefSocial Bot Detection in the ChatGPT Era: Challenges and Opportunities
SummaryThis paper reviews the challenges and opportunities in detecting social bots as sophisticated AI chatbots become more common. It examines current detection techniques and real-world applications, identifying gaps and trends with a focus on AI-generated conversations and behaviors. The authors propose four research directions: using generative agents for synthetic data and evaluation, multimodal and cross-platform detection based on coordination signatures, extending detection to non-English and low-resource languages, and collaborative federated learning models that preserve user privacy.
In briefShared Memory Surprisingly Improves Looped Transformers
SummaryLooped Transformers reuse the same layers several times per token to add compute without adding parameters, but each loop writes its own key-value cache, so memory grows with compute. The authors pretrain looped language models so that only the first recursion writes a cache, while later recursions read it and keep a short local window. They find that sharing memory does not reduce quality and in fact improves it. At 150M to 1B parameters, the hybrid variant with five recursions lowers FineWeb-Edu validation perplexity by 1.12 to 1.82 versus a same-size standard Transformer while using 76-79% less context memory.
Giovanni Monea, Keshav Ramji, Yousef El-Kurdi et al.
In briefFactorSplat: Appearance-Controllable Gaussian Proxies for Medical Volume Rendering
SummaryThis paper presents FactorSplat, a per-scene Gaussian splatting proxy for medical volume rendering that accepts region-specific intensity-to-RGBA transfer function curves at inference time. It keeps geometry and directional appearance shared across presets while preserving visibility control. Tested on seven CT and MR scans, it improves mean PSNR and changed-region error over the region-aware VEG baseline, and supports unseen edits without retraining.
In briefLatent-MOPD: Multi-teacher on-policy distillation using hidden representations
SummaryThis paper introduces Latent-MOPD, a method for training a language model student on its own generated responses while learning from several specialist teacher models. It transfers both the teachers' output predictions and their internal hidden states, without training any additional teachers. Across nine math, code and logic benchmarks, it outperforms token-only, representation-only and uniform-averaging baselines, and the student surpasses the best per-benchmark teacher on most of them.
WhyProposes representation-level multi-teacher on-policy distillation for LLMs, beating single-channel baselines on nine benchmarks with fairly solid evidence.
In briefTHPL: A Vision-to-Language Framework for Rainbow Trout Feeding Decisions in Recirculating Aquaculture
SummaryThis paper presents THPL, a framework that supports feeding decisions for rainbow trout in recirculating aquaculture systems. It extracts fish movement trajectories to compute an activity coefficient for feeding intensity, encodes individual and group behavior into physical and implicit tokens, and fine-tunes a large language model with these tokens, environmental data, and expert rules. Reported results show decision accuracy rising from 33.33% with a text-only baseline to 93.33% with the dual-evidence tokens, and 96.67% with counterfactual preference optimization.
In briefFlow Matching for Fast Posterior Sampling in Bayesian Inverse Problems
SummaryThis paper studies conditional flow matching, a generative method that transforms distributions step by step, as an alternative to Markov chain Monte Carlo for posterior sampling in PDE-based Bayesian inverse problems. After one-time training, the model produces approximate posterior samples quickly for any observation. The authors provide computable accuracy estimates for the approximate posterior in total-variation distance and Kullback-Leibler divergence, and propose a hybrid sampler that is asymptotically exact through Metropolization. Numerical examples, including electrical impedance tomography and a likelihood-free Lotka-Volterra model, are used to validate these results.
Jan Blechschmidt, Oliver G. Ernst, Moritz Poguntke et al.
In briefCoco: An Agentic Copilot for Hardware-Software Co-Design
SummaryThis paper presents Coco, an agentic platform developed and deployed with TPU architects to speed up hardware-software co-design, covering experiment setup, simulator sweeps, and insight generation. Coco has four layers: a datastore that automatically registers simulation results in a relational schema, typed tool APIs, agents that encode recurring analyses such as iso-execution comparison, and a platform interface whose navigation state serves as agent context. The authors report early deployment experience showing reduced time to simulation and time to insight.
Samuel Kushnir, Kavya Sreedhar, Yeshwanth Reddy Pogula et al.
In briefEviDent-CBCT: Generating Dental CBCT Reports From Incomplete Report Labels
SummaryThe paper addresses the difficulty of generating dental cone-beam CT reports when training reports do not fully document every finding. It proposes EviDent-CBCT, which first converts each scan into a discrete record of tooth-level, global, and nerve-canal evidence, then generates the report using only that record. The method outperformed a controlled direct baseline on evidence and logical F1 metrics and placed second and third in the ODIN 2026 challenge.
In briefBalancing reward and diversity in text-to-image generation via satisficing
SummaryThis paper addresses text-to-image generation where every image must meet a reward floor and the batch must meet a diversity cutoff, rather than combining both into one score. The authors propose SatisDive, a training-free inference-time method that traverses the resulting satisfaction-diversity Pareto frontier. On Pick-a-Pic, it improves worst-candidate reward over FK steering for two base models and Pareto-dominates FK steering's curve across overlapping DreamSim ranges.
In briefSimulating 5G networks in the training loop for thousands of parallel robots on GPU
SummaryThis paper presents Isaac-Net, a GPU-batched 5G New Radio uplink module that runs in lockstep with Isaac Lab physics across thousands of environments. It decides, every 0.5 ms slot, which robots transmit for all environments at once. Experiments show its median delay closely matches ns-3 5G-LENA and that it reproduces the Age of Information tail, which an independent-delay model underestimates by about three times. The code is open source.
WhyEmbeds GPU-batched 5G simulation in massively parallel robot training, validated against ns-3 on delay and AoI; novelty and evidence are solid, impact is confined to robot learning on private 5G.
In briefAutomating HCI Principles: Skills for On-Demand UI Construction and a Human-AI Thinking Space
SummaryThe paper argues that large language models can generate user interfaces on demand from natural-language task descriptions, and proposes treating the user-AI dialogue as a shared Space to Think in which task decomposition yields the interface. It suggests encoding classic HCI knowledge, such as Nielsen's heuristics and WCAG criteria, as machine-readable skill.md files that a generating agent loads at runtime. The authors outline a research agenda for making HCI design knowledge executable, inspectable, and open.
In briefCo-design Gym: A Benchmark for Jointly Optimizing Agent Body Design and Control Policy
SummaryThis paper introduces Co-Design Gym, a suite of benchmark environments for jointly optimizing an agent's embodiment (physical design) and its control policy. The authors note that most existing benchmarks fix the embodiment and study only policy learning, which can be highly suboptimal. The suite spans 20 environment families, including robotic manipulation and locomotion, multi-robot cooperation, electricity grids, and wireless networks, with over 85 co-design presets. The paper also systematically evaluates representative co-design algorithms.
Aviraj Newatia, Yordan Tsvetkov, Leonard Pleiss et al.
In briefAuditing Pedestrian Detection Explanations Under Domain Shift, Controlling for Confidence
SummaryThis paper audits post-hoc explanations of a fixed YOLOv8s pedestrian detector on the PIE and JAAD datasets, measuring faithfulness with ROI-based D-Deletion. Deletion-based faithfulness is strongly correlated with detection confidence at explanation time (Spearman 0.70 to 0.82 for D-RISE), so naive confidence-stratified comparisons are unreliable. After controlling for confidence, D-RISE faithfulness still differs by dataset in the central confidence range, with PIE higher than JAAD.
In briefArrivalBench: Agent-built data pipelines pass once but fail under out-of-order arrivals
SummaryThe paper introduces ArrivalBench, a benchmark that re-executes agent-generated data pipelines under replayable schedules with late, duplicated, out-of-order and retried records, requiring the final state to match a batch recomputation of the full log. On 40 tasks, single-execution grading certified 86-100% of pipelines from eleven models, while re-execution found 7.0-79.2% of those certified pipelines silently wrong. The authors also report that idempotency hazards fail more often than ordering hazards in every model.
In briefSEDIMA: Persistent Memory That Lets Evolutionary Code-Search Agents Learn Across Runs
SummaryLLM-driven evolutionary search can automatically discover programs and algorithms, but most existing systems have no memory and restart from scratch each run. This paper introduces SEDIMA, a persistent hierarchical memory that turns raw search traces into natural-language insights, clusters them by semantic similarity, and retrieves relevant guidance to shape future mutations. Under a fixed budget of 100 evaluated candidates, it improves average final performance by 5.5% on AlgoTune and 6.6% on ALE-Bench LITE. Under OpenEvolve, it needs 32.3% fewer iterations on average to reach baseline-best performance across five backbones.
In briefLexicographic Multi-Objective On-Policy Distillation: Merging Reward Specialists with Explicit Priorities
SummaryThis paper proposes Lexicographic Multi-Objective On-Policy Distillation (LMOPD), a multi-teacher method that integrates reward-specialized language-model policies under an explicit priority order. For each rollout, it selects the specialist for the first objective whose gate detects a deficiency, and projects out components that conflict with higher-priority specialists. Evaluated on 30B-A3B mixture-of-experts models across three math benchmarks with two and four experts, LMOPD better retains top-priority capabilities than the baselines compared.[]
In briefConformal prediction for time series using deep sequence models
SummaryThis paper systematically studies how deep sequence models, such as recurrent neural networks and Transformers, can be used in conformal prediction for time series. It examines three approaches: conditional quantile regression, conditional quantile function estimation, and localized conformal prediction. The authors provide a theoretical analysis establishing asymptotic conditional coverage guarantees for all three under suitable assumptions. Experiments on real-world datasets demonstrate the approaches' effectiveness.
In briefWhy adaptive batching helps LLM pretraining: an unbounded-variance view
SummaryThis paper studies why increasing batch size during large language model pretraining works, using a theoretical view based on growing gradient variance. It introduces a generalized BG-a noise model in which variance grows with distance from initialization, and derives under L-smoothness a lower bound and a matching upper bound on oracle complexity. The authors propose an adaptive batch scheduler and report that, when pretraining OLMo2 models up to 1B parameters on C4, it achieves lower validation loss than small and large batch training under matched token budgets, using under 10% of the iterations of small batch training.
In briefDo Every User Need a Private LoRA? Separating Personalization from Per-User Adaptation
SummaryThe paper argues that personalized large language models that store a full adaptation state for each user scale poorly as user populations grow. It proposes LINEUP, which learns a bank of reusable low-rank personalization factors composed through user-conditioned recall and query-dependent calibration, so each target user optimizes only eight scalars. Across six personalized classification, prediction, and generation tasks, the authors report that LINEUP leads on all 12 metrics and keeps its advantage under limited user history, compared with a private-LoRA configuration using 4.19 million per-user parameters.
WhyDecouples personalization into shared low-rank factors and tiny user codes; leads on 12 metrics across six tasks with theoretical bounds, suggesting practical impact for scalable personalized LLMs.
In briefDeReAct: Decomposed Reasoning and Acting for More Reliable AI Agents
SummaryThe paper proposes DeReAct, a modular agent design that separates action approval and task-completion decisions from a single language model. A Critic checks proposed actions before they run, and a Context Manager rebuilds the environment-supported state and certifies completion. On GAIA and SWE-bench Verified, the gains in Pass@1 are largest for weaker Brain models, such as 6.5 to 7.0 points for Qwen3-Coder-480B. With Claude Opus 4.5, Pass@1 stays comparable to ReAct, but the trajectories are more evidence-complete and constraint-satisfying.
In briefFrom behavior to provenance: tracing a tabular foundation model's behavior back to its synthetic pretraining data
SummaryThe paper uses O'PRIOR, a synthetic task generator that records the provenance of each task, to build a testbed for attributing a tabular foundation model's behavior to its pretraining data. Each task has explicit lineage over mechanisms such as missingness, confounding, shortcuts, and distribution shift. Removing the top-attributed 5% of synthetic tasks lowers mean ROC-AUC on held-out real tasks by 0.013, compared with 0.002 under random removal. Provenance discrimination is weaker, with ranking AUROC between 0.55 and 0.62.</summary>
Mohamed Bouadi, Nassim Bouarour, Shivam Dubey et al.
In briefReducing convergence collapse in fixed-target anomaly detectors with kernel-anchored locality regularization
SummaryThis paper studies tabular anomaly detectors that train a neural network toward a fixed target and score anomalies by test-time residual. The authors show that better convergence makes these detectors worse, because the learned map also matches the target on anomalies, so the residual signal vanishes. They propose Kernel Contraction Matching (KCM), a closed-form training-free method, and the Kernel-Anchored Regularizer (KAR), which improves AUROC under prolonged training on ADBench datasets.
Jos\'e Lucas De Melo Costa, Fabrice Popineau, Arpad Rimmel et al.
In briefDrive vs. Decay: Training Dynamics of Joint-Embedding Predictive Architectures
SummaryThis paper studies representation collapse in Joint-Embedding Predictive Architectures (JEPAs) during early training. By linearising the coupled gradient flow, it identifies a driving force and a decay effect whose per-mode ratio determines stability. The authors confirm the predicted phase boundary across more than 800 tabular JEPA configurations. They also introduce ResidualPred, a predictor that improves effective rank and accuracy on tabular benchmarks and I-JEPA pretraining on CIFAR-10, CIFAR-100, STL-10, and ImageNet.
WhyUnifies JEPA collapse-avoidance heuristics into a stability theory with a predicted phase boundary, validated across 800+ configurations and CIFAR and ImageNet pretraining.
Jos\'e Lucas De Melo Costa, Seong Woo Ahn, Fabrice Popineau et al.
In briefSCOPE-4D: An Endoscopic 4D Geometry Foundation Model
SummarySCOPE-4D is a model that jointly estimates camera parameters, dense geometry, and 3D tissue trajectories from monocular RGB endoscopy video in one forward pass. It is trained on SCOPE-5K, a collection of about 5,000 real and synthetic gastrointestinal endoscopy and laparoscopy clips. A Common-Residual Motion constraint separates tissue deformation from shared tissue movement. The paper reports improved camera and depth estimation, superior 3D tracking, and more stable long-sequence colon reconstruction on public and newly collected benchmarks.
Chaoyi Zhou, Zhongpai Gao, Anwesa Choudhuri et al.
In briefSelecting Key Network Graph Metrics to Detect Cyber Attacks More Efficiently
SummaryThis paper evaluates 21 topological metrics derived from natural visibility graphs to see whether a smaller subset can still detect cyber attacks. It combines four importance methods (SHAP, grouped permutation importance, Boruta, and RFE) through consensus ranking, then tests reduced metric sets on the CICIDS2018 dataset with a CNN classifier. The top three metrics alone reached 97.148% mean accuracy, above the 95.999% of the full set, while total runtime dropped from 14,961.39 s to 589.22 s.
In briefNEEDLEWORK: Rewriting Robot Data Offline with Verified Local Action Bridges
SummaryThe paper introduces NEEDLE, an offline dataset-augmentation algorithm that adds short, verified action bridges between recorded observations in high-dimensional robot demonstrations. It bypasses suboptimal detours, broadens action coverage, and can incorporate failed trajectories, using only RGB images, proprioception, and episode outcomes. A sampling technique lets policies use accepted bridges without synthesizing intermediate images or discarding original demonstrations. On real-robot tasks, the authors report an average success-rate improvement of 21 percentage points over the strongest baseline per task.
In briefSoTa: Soft Tactile Skins for Skilled Robot Hand Manipulation
SummaryThe paper presents SoTa, a low-cost capacitive tactile skin that covers full human and robot hands with a shared layout of 202 sensing points, allowing human demonstration data to be used alongside robot data. In three contact-rich manipulation tasks, tactile sensing improved success over vision-only policies. With a fixed robot demonstration budget, adding human demonstrations raised mean success across eight evaluation conditions from 22.8% to 45.9%.
In briefJoint Movement, Compression, and Power Design for Mobile Embodied AI Networks
SummaryThis paper studies an uplink mobile embodied AI network (MEAN) in which transmit power, movement distance, and semantic compression ratio are jointly optimized to maximize the minimum energy efficiency. The problem is non-convex due to coupled SINR, mobility-dependent channel gains, and a fractional objective. The authors propose an alternating optimization algorithm combining Dinkelbach transformation, successive convex approximation, and grid search. Simulations show it outperforms no-mobility and no-compression baselines.
In briefWorld Editing: Deeper Interventions on Executable Game Worlds
SummaryThe paper defines world editing as intervening on an existing executable world while keeping unchanged properties preserved, and introduces intervention depth to measure how strongly an edit couples entities, dynamics, and systems. The authors present IGMWorld and IGMBench, a benchmark of 110 game-modding tasks with over 1.1K criteria across Minecraft and Terraria. The strongest configuration solves 78.2% of tasks under a strict task-level criterion, and reliability generally decreases as intervention depth increases. Visual consistency is a separate weakness, with all configurations below 50% joint visual pass rate.
In briefComparing Tool Options Before Acting for Long-Horizon AI Agents
SummaryThe paper argues that tool-use agents should estimate the long-term value of a possible next tool call before executing it. It proposes CITA, which trains a Comparative Inference Model using observed tool behavior, supervision from a Bayesian tool-graph simulator, and LLM-based semantic comparisons. Across three tool-use benchmarks and several backbone LLMs, CITA improves Tool F1 and task success.
In briefLeast-Squares Objectives for Time-Series Forecasting: A Comparison
SummaryThe paper contrasts two least-squares objectives: one minimizes error on the forecast target coordinate, the other fits the whole next latent vector. In a four-dimensional time-series experiment, the first concentrates weight on the true signal and matches the innovation variance, while the rank-1 vector fit puts nearly all mass on a repeated coordinate and forecasts poorly. Iterated forecasts from the fitted coefficient also show error growing from 0.992 at one step to 2.700 at eight steps.
In briefSlow-Fast Multi-Teacher On-Policy Distillation for Preserving Capabilities
SummaryThis paper proposes SF-MOPD, a method for merging domain-specific expertise from multiple teacher models into one student model while limiting capability interference. A fast model is updated directly by each teacher, while a slow model, an exponential moving average of the student, gradually absorbs the learning and serves as a capability reference. Experiments across multiple model scales show reduced interference, better specialized multimodal ability, and smaller average drops on general benchmarks, outperforming vanilla MOPD.
In briefWorld-Calibrated Proposal-to-Action Flow for Vision-Language-Action Models
SummaryThis paper studies flow-based vision-language-action (VLA) robot policies, which generate action chunks from a task-agnostic isotropic Gaussian source. The authors argue this source ignores recent motion and predicted future evolution. They propose ProAct, which uses a Proposal Expert to produce a scene-aware hypothesis near demonstrated actions, and a World Expert to predict a task-consistent latent future that calibrates an anisotropic, proposal-centered source. Compared with π0.5, the authors report better performance on simulation and real-world tasks, with 50% fewer denoising steps, up to 25.8% lower inference latency, and up to 34.8% higher throughput.
In briefSCION: Scene composition using reusable neural building blocks
SummarySCION is a hierarchical scene representation that replaces many independent Gaussians with a compact set of reusable primitives and lightweight instances placed throughout the scene. It is fitted to multi-view captures through joint optimization of discrete and continuous parameters. According to the paper, it stays high quality at 1.2 MB, outperforms existing Gaussian compression methods on rate-distortion, and supports instance-level editing and animation without retraining.
In briefDeskForge: a controllable desktop environment that generates dense training data for computer-use agents.
SummaryThe paper introduces DeskForge, a controllable desktop environment that composes real applications to produce large-scale, densely annotated screenshots for training computer-use agents. The authors build DeskForge-1M, containing 1.2M annotated desktop observations with 159.7M element instances, and fine-tune four vision-language models on 200K grounding examples from it. The fine-tuned models improve on held-out conditions, five external GUI grounding benchmarks, and long-horizon WebArena-Infinity and OpenApps task completion.
WhyControllable composition of real desktops yields dense supervision, with consistent gains across five benchmarks and long-horizon tasks; solid evidence and practical value.
A. Said Gurbuz, Ahmed Nassar, Sunghwan Hong et al.
In briefMoore, Escher, Penrose: A Conformal Golden Braid
SummaryInspired by M.C. Escher's 1956 lithograph Print Gallery, the paper builds on a conformal power map z↦z^α to create self-referential scenes. It uses a frozen text-to-image diffusion model to generate new compositions. The authors introduce a generalized inverse T† of the transformation and braid denoising steps between source and transformed spaces so the scene and its distortion develop together.
In briefSphere Encoder 2: an improved sphere-latent autoencoder for image generation
SummarySphere Encoder is an autoencoder that generates images by decoding random points from a high-dimensional latent sphere. The paper identifies two limitations of the original method: a gap between random sampling points and the encoded latent distribution, and pixel-wise reconstruction loss that makes outputs blurry. The authors present Sphere Encoder 2 to address both issues, reporting substantially better image generation quality while keeping the speed and simplicity of an autoencoder.
In briefOne Shared Basis Drives All Avatars: Gaussian Blendshape Distillation for Real-Time Animation
SummaryThe paper introduces GALA, a distillation method that replaces costly per-frame neural decoding in 3D Gaussian avatars with a shallow MLP that predicts coefficients for a linear blend of identity-independent blendshapes. Applied to three avatar models for facial expressions and full-body clothing dynamics, it cuts CPU animation cost by up to three orders of magnitude while preserving most rendering quality. The method generalizes to held-out identities and reaches up to 60 fps on mobile devices.
Ramazan Fazylov, Stamatis Lefkimmiatis, Ivan Laptev
In briefPROWBench: Do Video Models Show What the Program Specifies?
SummaryThis paper introduces PROWBench, a benchmark for checking whether videos generated by programmable world models faithfully show events specified by a program. It includes 170 programmatically built episodes and 600 proxy videos, and records entity states and timestamped events as replayable records. The benchmark evaluates entity control and long-horizon memory, plus two VLM-based metrics for timeline adherence and event rendering.
In briefSelf-improving robot agents without retraining: the RPG framework
SummaryThis paper presents RPG (Reconstruct, Practice, Go Real), a framework that lets robot execution systems improve autonomously without updating model weights. It builds practice tasks in simulation from an offline dataset, diagnoses failures, and revises reusable skills and the system prompt. On held-out initializations of 22 manipulation tasks, success rose from 28.6% after the first round to 95.0% after 15 rounds, outperforming the evaluated baselines. The frozen system also succeeded in all 30 physical trials across three tasks.
In briefEmbedding prediction for image generation: predicted embeddings replace fixed class or text conditions to guide diffusion transformers.
SummaryDiffusion transformers usually embed a class label or text prompt once and reuse that condition at every denoising step. This paper introduces NEPA (Next-Embedding Predictive Autoregression), which predicts the next continuous embedding, and uses Multi-Embedding Prediction to predict the embeddings of the clean image at once. A DiT generator is conditioned on these predictions, recomputed at each denoising step to match the current noisy state. On class-conditional ImageNet 256×256, the final model NEPA-DiT-XL, combined with REPA, reaches an FID of 1.32 using about a third of REPA's training compute.
In briefScholarCatalyst: A Benchmark for Finding Papers That Spark New Research
SummaryThe authors built ScholarCatalyst, a benchmark in which 184 lead authors labeled which earlier papers advanced 207 of their recent computer science projects, each with a rationale. The retrieval task asks systems to find these papers from only the literature available when each project began, given an initial research question. Agentic search (0.42 Recall@20) did no better than embedding retrieval (0.48), and an agent built on Claude Fable 5.1 reached only 0.51. The results suggest models need new training methods to match expert judgment in searching broad literature.
In briefSILSA: Sliding-window slice latents for high-resolution 3D shape generation that keeps topology intact
SummaryThis paper presents SILSA, a 3D generation framework that represents shapes with compact sliding-window slice latents along three axes instead of expensive voxel tokens. It uses a Slice VAE, a Volumetric Anchor Lattice, and slice-level topology supervision based on persistence diagrams and Betti transitions. The authors report better structural fidelity and lower generation cost, with fewer tokens, reduced training memory, and faster inference than the baselines.
In briefVISTA: A Visual Harness for Reasoning in Interactive Environments
SummaryThis paper introduces VISTA, a visual harness that gives general-purpose multimodal models long-horizon vision in interactive environments. It lets the model perceive visual observations directly and keeps a lossless visual memory of past observations that it can retrieve and reorganize during reasoning. On ARC-AGI-3, VISTA raised Claude Opus 5.0's Relative Human Action Efficiency score from 40.68 to 100.00, completing all 25 public games with 57.4% fewer actions than first-time human participants. It also substantially outperformed minimal-harness baselines on three additional visual game and puzzle benchmarks.
In briefTACO: A Memory-Light Optimizer for Fine-Tuning Large Language Models
SummaryFull-parameter fine-tuning of large language models requires large optimizer state memory. This paper proposes TACO, which for each column of weight matrices keeps only the sign of the largest-magnitude gradient entry, retaining first-order gradients while making optimizer state nearly negligible. On OPT-13B, the authors report that persistent optimizer state falls from 27.7 GB to 0.16 GB and peak training memory from 80.6 GB to 27.5 GB, with comparable accuracy and runtime. They also report full-parameter fine-tuning of 30-32B-parameter models on a single 80 GB H100 GPU.
Jichao Jiang, Cristian McGee, El Houcine Bergou et al.
In briefFERPO: A reinforcement learning method that improves policies from critic values without action gradients.
SummaryThis paper proposes FERPO, an on-policy maximum entropy reinforcement learning algorithm for continuous control. It improves the policy using critic values without differentiating the critic with respect to actions, deriving a target action distribution regularized by entropy and KL divergence. The actor is then fit to this target by minimizing a forward-KL objective estimated with self-normalized importance sampling. Experiments on MuJoCo Playground and ManiSkill report competitive performance and sample-efficiency gains, along with faster actor updates than REPPO.
Sebastian Sanokowski, Alireza Sarmadi, Majid Khadiv
In briefHiPhy: Hierarchical Alignment for Physically Plausible Multi-Principle Video Generation
SummaryThis paper addresses the problem that video generation models, despite high visual fidelity, often fail to follow physical laws, especially when several physical principles must act together in one video. The authors propose HiPhy, a reinforcement learning framework that enforces the temporal dynamics of individual physical principles locally and the physical and semantic coherence of the whole scene globally. They also build a 50K-prompt dataset and the MultiPhyBench benchmark. Experiments report that HiPhy outperforms prior methods, with the largest gains in scenes involving multiple concurrent physical principles.
Tahira Kazimi, Shubhankar Borse, Munawar Hayat et al.
In briefInterEvolve: Evolving Reward Programs at Test Time for Humanoid Movement and Manipulation
SummaryThis paper studies how a humanoid robot controller can solve tasks it was never trained for by reusing its existing skills, improving from its own attempts, and retaining what it learns, without retraining. The method combines an object-aware forward-backward behavioral foundation model with staged reward programs that a large language model agent revises and a numerical optimizer tunes. Experiments show that the evolved programs unlock competence that human-designed rewards leave untapped, and they produce behaviors for diverse tasks in simulation. The evolved skills also run autonomously on a physical Unitree G1 robot using onboard perception.
In briefExact cost-augmented Schrödinger bridges on graphs via a Feynman-Kac tilt
SummaryThis paper studies the generalized Schrödinger bridge on graphs, which moves mass between two distributions while charging for visited states. The authors fold the state cost into the reference process as a Feynman-Kac tilt, so the cost-augmented bridge becomes a plain bridge against the tilted reference and no temporal-difference penalty is needed. The bridge is computed exactly by alternating two endpoint rescalings, each a sparse matrix-exponential application, without time discretization or learning.
In briefHierarchical Continuous Diffusion Language Models: a text generation method linking discrete tokens with a continuous latent trajectory
SummaryThis paper addresses two limitations of diffusion language models: discrete models lose dependencies among tokens decoded in parallel, while continuous models' denoisers are not tied to a valid token configuration. It proposes Hierarchical Continuous Diffusion Language Models (HC-DLM), which couple discrete token generation with a continuous latent trajectory in a single denoising process trained with a variational bound on token likelihood. The authors report improved puzzle accuracy on Sudoku and Countdown and better generative perplexity on LM1B over discrete and continuous diffusion baselines at matched model size.
In briefPaper: Diagnosing and Repairing Mathematical Reasoning in Large Language Models
SummaryThis paper examines whether large language models truly understand the mathematical structure behind their solutions. It introduces the notion of Mathematical Primitive and a benchmark that evaluates reasoning across four dimensions: Discovery, Generation, Digestion, and Execution. The authors find that solution accuracy hides distinct capability profiles, that Discovery is the main bottleneck, and that discovery-related failures are easier to repair. They also propose a primitive-guided self-distillation framework that reportedly improves reasoning over baselines across model scales and benchmarks.
In briefChoosing Step Sizes for LLM Fine-Tuning with a Mix of Zero- and First-Order Methods
SummaryThe paper proposes ZFO, a lightweight framework that uses a first-order optimizer to pick the search direction and then uses two extra objective evaluations along that direction to select a curvature-aware step size. The authors provide theoretical guarantees on the finite-difference curvature estimates and convergence to a neighborhood of a stationary point. In the evaluated language model and dataset settings, ZFO frequently improves optimization and final performance over fixed-step first-order baselines, with gains depending on the objective. Code is publicly available.
In briefGenerating intrinsically disordered protein regions by reinforcing sparse autoencoder features
SummaryThis paper introduces IDiom, an autoregressive protein language model trained on 54 million predicted intrinsically disordered regions (IDRs) from the AlphaFold Database, which generates sequences resembling natural IDRs. The authors also propose RL-SAE, a post-training method that rewards sequences activating specified sparse autoencoder features. Across eight design tasks, RL-SAE activates on average 90% of 30 targeted features, versus 24% for activation steering.
WhyProposes IDR-focused language model and RL with sparse autoencoder features, outperforming steering across eight design tasks with solid comparative evidence.
Jason X. Liu, Sebastian Ibarraran, Frank Hu et al.
In briefDMAD: Distribution Matching as Adversarial Distillation for Fast Image and Video Generation
SummaryThis paper proposes DMAD, which replaces the separately fitted auxiliary diffusion model in Distribution Matching Distillation with discriminator heads that learn log-density ratios directly. The authors prove that at the discriminator optimum these losses recover the DMD distribution-matching gradient. They report the best results among compared few-step methods and multi-step teachers on ImageNet-64x64, SDXL, and Wan2.1-T2V-14B benchmarks.
WhyRecasts distribution matching as adversarial classification, dropping auxiliary score models; validated across ImageNet, SDXL, and Wan2.1 with broad impact.
In briefGrammar-based molecular representation for generation and transfer learning in chemistry
SummaryThis paper introduces Higher-order Grammar Representation (HGR), which lifts molecules to combinatorial complexes and parses them into compact production-rule sequences under a context-free higher-order grammar, so standard sequence models can capture ring systems and recurring motifs. The authors also build RingDiv, a ring-enriched benchmark of 1.18 million molecules, and propose a ring diversity index (RDI) to measure ring-system coverage. HGR-based generative models achieve 100% validity by construction and rank first in FCD on all five generation benchmarks. HGR-FM attains the highest mean AUC across seven MoleculeNet benchmarks under both probing and full fine-tuning.
Yiming Huang, Yujie Zeng, Vijay Prakash Dwivedi et al.
In briefDecoding Looped Transformers Better for Almost Free
SummaryLooped Transformers reuse a shared block across recurrent loops to save parameters, but standard decoding ignores the intermediate states from earlier loops. The paper introduces LoopCD, a training-free contrastive decoding method that compares the final prediction with an earlier loop's output, in either logit or hidden-state space. Across four model families, it gives consistent gains at full depth, and halving the loop count still matches or exceeds full-depth baselines while cutting forward FLOPs by 22.5% to 48.2%.
In briefSoftServe: A Scalable Quasi-Newton Optimizer for Deep Learning
SummaryThis paper introduces SoftServe, a family of quasi-Newton optimization methods for training deep neural networks, designed to handle non-convexity and very large parameter counts without line searches or ad hoc curvature fixes. It presents diagonal and Kronecker-factored variants and uses the stable coupled Newton-Schulz iteration for matrix operations. The authors report that SoftServe performs well on severely ill-conditioned problems, including recurrent networks, deep autoencoders, physics-informed networks, and a 136M-parameter diffusion model, often reaching lower losses than Adam, Muon, and SOAP.
In briefOmniSeek: Native Tool Integration for Multi-turn Audio-Visual Reasoning
SummaryThis paper presents OmniSeek, a framework that turns an omni-modal large language model into a multi-turn agent that actively retrieves audio or visual evidence while reasoning. The model decides when and over which time window to look or listen, and appends the retrieved raw segments back into its context. The authors build the OmniTraj-170K dataset of multi-hop trajectories with interleaved audio-visual evidence, train with supervised learning and two-stage reinforcement learning, and add an Audio-Visual Necessity objective to discourage single-modality shortcuts. Experiments across benchmarks report consistent gains in audio-visual reasoning.
In briefGenerative Cinematographer: jointly controlling camera and object motion in 3D video generation
SummaryThe paper presents GenCine, which lifts a single image into an editable 3D scene where users specify a camera path and move foreground regions with local 3D motion handles. These controls are projected into guidance maps that inform a pretrained Wan video model via a lightweight branch and LoRA adapters. Experiments report consistent camera-relative motion and improved geometric consistency under viewpoint changes.
Jiahan Zhang, Chaohao Yang, Namitha Guruprasad et al.
In briefSimuVerity: A benchmark testing whether AI agents can generate engineering-ready Simulink models
SummaryExisting Simulink benchmarks mainly check whether generated models compile, run, or resemble a reference, which does not show if they meet engineering requirements. SimuVerity introduces 101 text-to-executable model generation tasks across ten engineering domains, with a hierarchical evaluator scoring qualified models on six dimensions. Six agent systems were tested, and the best reached only 42.86 overall. The results indicate structural similarity is a poor proxy for engineering performance.
In briefFrom Gradients to Capabilities: Multi-Teacher On-Policy Distillation
SummaryThis paper studies multi-teacher on-policy distillation (MOPD), which aims to merge the strengths of several RL-trained domain teachers into one student model. Using Qwen3-1.7B with four domain teachers, the authors compare gradients, optimizer updates, and task learning curves, with SmolLM3-3B as an additional diagnostic. They report that loss averaging, Adam's first moment, BF16 rounding, and top-64 intersection KL gradients all affect teacher signals, with mathematics accuracy differing depending on the averaging method.
In briefPowerBench: Measuring Language Model Bias in Power-Shifting Requests
SummaryThis paper introduces PowerBench, an evaluation that tests how language models respond to requests that shift power, distinguishing self-empowerment, disempowerment, and power grabbing, with a control set of requests that shift no power. The authors release an open-source dataset and evaluate 24 models from US and Chinese developers under conditions including nationality, AI-agent users, and eight request languages. They find models refuse power grabbing more than disempowerment, and disempowerment more than self-empowerment, with biases varying by nationality, user type, and language.
Nicolas Martorell, Wendy Brau, Gonzalo A. Heredia et al.
In briefEffective Resistance and GNN Reliability in Tissue-Specific Interactomes
SummaryThis paper asks whether effective resistance in tissue-specific protein interaction networks can indicate which protein function predictions from graph neural networks to distrust. Across 24 networks, effective resistance is largely driven by inverse node degree. After controlling for several factors, it still explains some additional per-node loss in most held-out networks. The signal is reproducible, but degree degeneration limits its strength.
In briefWhy ablating a model component seems to trigger other components to compensate
SummaryThe paper examines self-repair, where other parts of a language model appear to compensate when one component is ablated. The authors argue the effect comes from a gain already present before ablation, modeled as an affine law for each unit's repair response. They report that 68 of 81 downstream directions across four models follow this law, and that seven of ten reachable heads in GPT-2 Small's IOI circuit act as counterweights.
In briefWatch, Infer, Coordinate: Learning a Robot Partner's Physical Limits for Zero-Shot Teamwork
SummaryThis paper studies whether a helper robot can infer a partner robot's physical limitations, such as hardware degradation or actuator faults, by watching the partner coordinate with another robot in physically coupled manipulation tasks. The authors introduce the Watch, Infer, Coordinate benchmark, which spans three such settings, along with an inference method that scores candidate constraints using observed joint behavior. The method substantially improves constraint inference and zero-shot coordination across all settings, approaching the performance of an oracle that knows the true constraints.
Suyu Ye, Zheyuan Zhang, Vaishnav Tadiparthi et al.
In briefKeep It CALM: Limits of Global Unsafety Removal in Text-to-Image Generation
SummaryThe paper examines training-free safety methods for text-to-image generation that apply one global unsafe direction across all prompts. It finds a trade-off: compact unsafe subspaces miss diverse unsafe content, while broader aggregation distorts safe-adjacent benign prompts. The authors propose CALM, which applies prompt-local counterfactual corrections to only violating token representations and suppresses positively aligned unsafe residuals. They report that CALM improves unsafe content suppression while preserving benign utility.
In briefAutoCompact: Teaching coding agents when to shrink their working memory
SummaryThis paper introduces AutoCompact, a method that trains coding agents to decide when to compact their context, what working state to keep, and how to continue afterward. Training data is collected by having a judge review and correct the base agent's compaction decisions and summaries, followed by supervised fine-tuning and reinforcement learning with task-success rewards. On SWE-bench Verified and SWE-PolyBench Verified, the approach raises pass rates over the base model by an absolute 9.2% and 5.0%, respectively.
In briefWorld Observer: Jointly generating an actor and observers to keep off-screen objects evolving in a world model
SummaryVideo world models usually simulate only what the agent can currently see, so objects that leave the view lose their evolving state. World Observer separates observing from acting by jointly generating a perspective actor view and one or more panoramic observers that watch selected regions. Objects out of the actor's view keep changing in an observer, so their updated states appear when they re-enter, and the paper adds world-space metrics and a benchmark on real and synthetic scenes. The method substantially improves out-of-view dynamics while staying competitive on visual fidelity, camera control, and 3D adherence.
In briefDuoMind: Semantic Communication for Coordinating Multiple Robots
SummaryDuoMind is a distributed framework that lets multiple robots coordinate tasks by exchanging semantic messages. Each robot pairs a vision-language-action model for low-level execution with a vision-language model that handles high-level reasoning and communication with peers. The authors also introduce RoboPoly, a benchmark of long-horizon, coordinated manipulation tasks. Experiments on RoboPoly and RoboTwin show improved multi-robot performance, and ablations confirm the value of hierarchical orchestration and semantic communication.
In brief4Director: Controlling Video World Models with Rigid 3D Geometry
SummaryThis paper introduces 4Director, a video world model that is conditioned on an explicit 4D scene representation for precise camera and object control. Each object is reconstructed as a canonical mesh from the input image and moved by one rigid transformation per frame, then rendered as a depth video. A Motion Adapter turns this geometry into video with view-consistent appearance, illumination, and non-rigid dynamics. The authors report that it outperforms prior methods in visual quality and control.
In briefWhen Do Intrinsic Rewards Lead to Exploration?
SummaryThis paper asks whether intrinsic rewards in reinforcement learning, such as prediction error or count-based and information-gain objectives, actually lead agents to the most informative experience. The authors propose a criterion that compares policies by the counterfactual information their histories provide. They build a simple environment where some intrinsic reward objectives yield Pareto-suboptimal policies, and identify conditions under which existing rewards do drive optimal exploration. They also construct an objective that gives higher value when exploration strictly improves under this criterion.
Scott W. Viteri (Stanford University), Laura Gomezjurado Gonzalez (Stanford University), Clark Barrett (Stanford University)
In briefMuon meets Tamed Langevin: momentum preconditioning for sampling from non-convex, non-Lipschitz potentials
SummaryThis paper studies sampling from Gibbs distributions on matrix spaces whose potential energies are neither convex nor globally gradient-Lipschitz. The authors introduce a family of non-quadratic kinetic energies that yield an underdamped Langevin system with momentum preconditioning, where the kinetic energy gradient smoothly tames the momentum. Under these relaxed potential assumptions, they prove the dynamics leaves the target Gibbs measure invariant and converges exponentially to equilibrium in a weighted total variation distance. They also show the Euler-Maruyama discretization has moment bounds uniform in time, supporting the stability of the sampling algorithm.
In briefΨ-Resilience: model-free feature importance using 1D topological signals
SummaryThis paper proposes Ψ-Resilience, a feature importance method that derives explanations directly from data without a predictive model. It estimates class-conditional densities, takes their pointwise absolute difference to form a 1D signal, and aggregates the 0-dimensional persistent features that survive up to a user-set robustness scale. On synthetic data with known importance, it recovers feature rankings with Spearman correlations up to 0.8, and on real datasets it agrees with SHAP and mutual information with correlations up to 0.9.
In briefMosaiChunk: Spatio-temporal memory for consistent long autoregressive video generation
SummaryAutoregressive video generators have a limited context window, so fine visual details of objects or scenes can be lost and hard to recover when they reappear. MosaiChunk stores a mosaic of selected historical key-value entries across space and time, and keeps the video generator frozen while training only a lightweight router that chooses which history sections to include under a fixed memory budget. The authors also introduce RememBench, a long-horizon revisit benchmark with text-to-video and camera-driven image-to-video splits. Experiments report improved revisit consistency over sliding-window inference and whole-chunk retrieval at matched memory budgets in both settings.
Yiwen Zhang, Haocheng Xi, Michael Tian-Yue Liu et al.
In briefEditHero: A Benchmark for Long-Horizon Part-Level 3D Editing and Vibe Modeling
SummaryThis paper introduces EditHero, which the authors believe is the first benchmark for long-horizon, part-level 3D editing, featuring natural-language instructions and target images for geometry and texture. A deterministic assembly engine produces the exact target after each edit, and every sequence is manually reviewed. The authors compare non-agentic methods, which regenerate whole objects top down, with LLM/VLM agents that edit bottom up through code affecting only instructed parts. Non-agentic methods often miss requested changes and disturb unedited regions, while most LLM agents follow instructions more closely and preserve unedited parts better, though each edit takes minutes.
