In briefBeyond Overall Accuracy: Per-Question Changes in Automatic Prompt Tuning
SummaryThis paper introduces RFDPO, a method that tracks, for each question, whether answers change from wrong to right (recoveries), right to wrong (regressions), or change at all (churn) during automatic prompt optimization. Tested on five benchmarks covering law, knowledge, instruction following, math, and privacy, it shows that overall accuracy can hide many item-level changes. In a controlled ablation on LegalBench-Hearsay, item-level feedback reduced churn by 40% while keeping similar net gains.
In briefBackpropagating verifier rewards through continuous flow language models, a reinforcement learning method for models that generate continuous token embeddings.
SummaryThe paper proposes a reinforcement learning post-training framework for continuous flow language models, which denoise token embeddings over several steps and decode to discrete tokens only at the end. A critic estimates the value of intermediate embeddings, bypassing the non-differentiable decoding step so that low-variance pathwise gradients can be backpropagated. Across four reasoning tasks, the method reportedly improves the model and outperforms baselines of the same type.
In briefImproved Hierarchical Bayesian Deep Learning for Uncertainty Estimates
SummaryThis paper introduces MLPMMH, a multilevel sampling method that jointly infers neural network parameters and prior hyperparameters to quantify uncertainty. The approach combines pseudo-marginal inference for hyperparameters with coupled sequential Monte Carlo across a hierarchy of network-width truncations. The authors prove a mean-square error and computational cost rate under standard multilevel coupling assumptions and verify it on a Bayesian regression problem. They also propose S-MLPMMH, a practical approximation using a PCA-based hierarchy, which shows competitive results on MNIST.
Neil K. Chada, Simon Cotter, Kody J. H. Law et al.
In briefMonoMIP: Multi-height inverse perspective mapping for monocular 3D object detection
SummaryThis paper proposes MonoMIP, a dual-decoder framework for detecting 3D objects from a single camera image. It adds bird's-eye view features built with multi-height inverse perspective mapping to complement perspective-space depth-guided predictions. The authors report 25.84% Moderate AP3D|R40 for Car on the KITTI validation set and 21.66% on the official test set, without extra training data, and state it outperforms recent published monocular methods on KITTI and Waymo.
In briefPlan or Retrieve First? Scale Study of Tool Discovery Architectures
SummaryThis paper compares two ways LLM agents find tools in large API registries: retrieving first then planning (RtP), and planning abstract tool stubs first then retrieving (PtR). On ToolBench, RtP performs better for single-tool queries while PtR performs better for multi-tool queries, a crossover that holds across retrieval budgets and registry sizes. The authors also analyze PtR's roughly doubled inference cost and suggest complexity-adaptive routing driven by RtP's top-1 retrieval score.
In briefGenerative vs. dense retrieval: a static value prior reverses the gap
SummaryThis paper compares generative retrieval (GR) and dense retrieval (DR) for candidate generation in online retail, using two Amazon Toys datasets and an industrial e-commerce dataset. GR leads on the value of the best relevant item, while DR leads on exact-target recall and relevant value across the ranked list. The authors find that popularity drives most of GR's measured lead, and adding the same static item value to DR reverses the advantage.
In briefHow inner optimizers affect scaling across model replicas
SummaryThis paper studies why increasing the number of model replicas, while keeping the total training budget fixed, tends to hurt performance when training large language models across distributed compute. The authors identify two factors, the covariance structure of pseudo-gradients and how batch size affects their noise scale. They theoretically and empirically show that whitening-based preconditioning works well as an inner optimizer, achieving strong replica scaling from 8 to 96 replicas.
Eugene Belilovsky, Stefan Horoi, Edouard Oyallon et al.
In briefWorldMemBench: A Benchmark for Testing Memory in Video World Models
SummaryThis paper introduces WorldMemBench, a benchmark for systematically evaluating how well video world models remember what they have shown. It uses 260 test cases covering static memory of humans, objects, and environments, and dynamic memory of event progression and state persistence. Tests of nine models found that strong static retention does not guarantee reliable dynamic state evolution, and memory degrades after repeated revisits. Human studies indicate that the automatic evaluations align with human judgments.
In briefRLPE: Training LLMs to Write Prompts via Reinforcement Learning
SummaryThis paper introduces Reinforcement Learned Prompt Engineering (RLPE), which trains a model to analyze the execution traces of a separate task model and write prompts targeting its observed failures. The trained model is rewarded only on prompt performance. Prompts are generated once and reused, so the trained model is not needed at inference time, and only query access to the target model is required.
In briefConformal physics-informed weak regression for estimating parameters of epidemic models
SummaryThis paper proposes physics-informed weak regression (PIWR), which fits epidemic compartment model parameters by integrating the model in time, avoiding derivatives of the data. It wraps the estimators in split conformal prediction to produce simultaneous intervals with finite-sample coverage guarantees. Reported results on eleven benchmark models and real 1978 boarding-school influenza data show higher accuracy and much faster fitting than the compared methods.
In briefDeep generative models for more accurate phylogenetic tree inference
SummaryThis paper introduces a flexible family of conditional generative models that infer posterior distributions over phylogenetic trees from sequence data, avoiding the limits of likelihood-based methods. The models build trees step by step using top-down, middle-out, or bottom-up paradigms and are trained with a discrete diffusion framework. On real and synthetic datasets of up to 65 taxa, they recover MCMC posteriors more accurately than Phyloformer 2. Compared with IQ-Tree, they match on simple data and are more accurate on more realistic data.
In briefWhy VLMs struggle at measuring vital signs from video, and a new method to fix it
SummaryThis paper examines why Vision-Language Models (VLMs) perform poorly at remote physiological measurement from video, finding that generic visual features miss fine physiological cues and that discrete token prediction misaligns with continuous numerical values. The authors propose VITAL, which keeps the original visual pathway and adds an independent physiological pathway, along with a structured numerical alignment method for more reliable heart rate estimation. Experiments on multiple rPPG datasets show improved cross-domain estimation while retaining the VLM's general visual understanding ability.
In briefSmart Alone, Blind Together: Why Accuracy Alone Cannot Certify Theory of Mind in Agent Teams
SummaryThe paper argues that accuracy is an inadequate certificate for modules that infer teammates' intentions in multi-agent teams, because the required accuracy depends on the cost of wrong actions and errors across modules are often correlated. Tests in task assignment, Level-Based Foraging and Hanabi show accuracy can wrongly approve or reject deployments. The authors propose a Blind-Spot Gate that certifies the deployment by pricing error costs, pairwise interactions and measured dependence, and report better results than accuracy on held-out and preregistered tests.
WhyProposes a deployment-level gate pricing error costs and dependence instead of accuracy, validated across tasks and human judgments.
In briefTime-series-aware keyframe diffusion to synthesize cardiac dynamics from sparse observations (CardioFAD)
SummaryClinicians often see only very sparse cardiac measurements, making full heart motion hard to recover. The paper proposes CardioFAD, which learns motion-salient keyframes that both preserve the visual trajectory and align with ECG signals, then uses diffusion to generate keyframes and interpolate the remaining frames. The authors report improved visual quality, temporal coherence, and functional consistency on cardiac MRI and echocardiography videos under an extreme one-shot setting.
In briefPre-trained language models as no-training tools for social science tasks like record linkage and name-based ethnicity inference
SummaryThe paper examines whether pre-trained language models can perform record linkage and name-based demographic inference without task-specific labeled data, which such tasks usually require. It introduces EnsembleLink, which combines pre-trained embeddings with string-based scorers to match records. For ethnicity inference, large language models reached up to 83.8% accuracy on U.S. voter files, compared with 78.0% for name-supplemented BISG.
In briefWhat Information Buys in Active Inference: When Calibrated Information Gain Beats Confidence Widths
SummaryThis paper compares two exploration methods, Active Inference (AIF) and Bayes-UCBVI, on tabular MDPs. It finds that AIF's information gain decays too quickly, causing greedy agents to miss optimal actions, and that square-root calibration yields regret of order √T for an exact abstraction. With fixed rewards, calibrated AIF offers no better guarantee than Bayes-UCBVI, but when an adversary chooses rewards from the agent's data, calibrated information gain acts as a reward-agnostic ℓ1 confidence set and is preferable.
In briefRingNav: Task-Shaped Circular Memory for Visual Heading
SummaryThis paper presents RingNav, a visual heading estimator with about 10.6k parameters that restricts its memory to the circle S¹. It learns only the turn between consecutive images and composes heading directly on the circle. On the TartanAir Visual SLAM Challenge, it cuts mean absolute turn error by 27% and 4-second accumulated heading error by 15% compared with a similarly sized learned-memory model. It transfers to real drone experiments without fine-tuning and runs on Raspberry Pi Zero 2W and GAP8 processors.
Naitri Rajyaguru, Aryan Tajne, Stephen Jardim et al.
In briefGraph coarsening helps escape local optima in compute-limited nonconvex MAX-CUT
SummaryThe paper studies MAX-CUT, a problem whose binary optimization landscape has exponentially many local optima, making it sensitive to initialization. The authors warm-start Particle Polynomial Relaxation, a nonconvex continuous reformulation, with Combinatorial Multigrid, a spectral graph-coarsening method that needs no training data. The reported results include 97.5% of optimum on average for the t2g20 instance and matching published optima on all sixty Biq Mac instances when discrete local search is added. A sparse representation lets the approach scale to G-set instances with up to 20,000 vertices.
M. Harper Langston, Pierre-David Letourneau, Dalton Jones et al.
In briefLoss curvature as a cheap proxy for memorization in diffusion models
SummaryThis paper studies why diffusion models memorize training data, looking at both when memorization begins during training and how likely each individual sample is to be memorized. The authors propose input loss curvature, the Frobenius norm of the Hessian of the training loss, as a metric for both views. They extend a counterfactual memorization definition from supervised learning and show it is governed by the curvature of a leave-one-out model. Experiments indicate that input-dependent curvature serves as a proxy for memorization.
In briefTracing How Language Models Learn Morphology, Meaning, and Divergence During Training
SummaryThis paper examines how Large Language Models acquire language by analyzing intermediate training checkpoints from open-source models such as Pythia and OLMo under teacher-forcing evaluation. It studies four aspects: morphological structure acquisition, embedding drift, repetition dominance, and co-occurrence effects. The authors find that models acquire morphological structure before semantic embeddings stabilize. They also report that larger training datasets benefit divergence recognition, while pre-training co-occurrence and inference-time repetition hurt divergence recognition and associative link generation.
Forough Shirin Abkenar, Haoshuo Fu, Liang Feng et al.
In briefLag: A Mean-Field Optimal Policy for Reusable-Resource Online Assignment
SummaryThis paper studies online task assignment where servers are temporarily busy and can be reused, as in peer review, call routing, and ride matching. The authors model it as a contextual restless multi-armed bandit and use an occupancy-measure linear program to upper-bound achievable reward. From its dual they derive Lag, a closed-form index policy. Under a scaling regime where server copies and arrivals grow together, Lag is shown to be asymptotically optimal. On peer review and call center data it reaches 83-99% of the LP bound and beats greedy assignment by up to 12.3% and 36%.
WhyProposes an occupancy-measure LP and closed-form Lagrangian index policy, with a first deterministic fluid limit, validated on real and synthetic datasets.
In briefWho Should We Listen to More? A Welfare-Aware Method for Choosing Whose Human Feedback to Collect
SummaryThe paper studies how to allocate limited human feedback across people to build training data for pluralistic AI alignment, since AI simulators label preferences less accurately for some individuals. It proposes Welfare-Aware Preference Acquisition (WAPA), which selects queries by estimated marginal social welfare gain and proves the algorithm is exactly optimal for its estimated objective under a fixed budget. In experiments, acquiring human labels for 20% of comparisons recovered 75-86% of the average predictive utility gap between AI-only and ground-truth-only training.
In briefHow Many Ranks a Token Needs: A Noise-Floor Gate for Rank-One LoRA Experts
SummaryThis paper addresses the fact that standard LoRA applies every rank-one component to every token. It proposes SQUELCH, a router-free method that soft-thresholds each token's responses against its own noise level, so the active rank varies per token, with the threshold chosen by Stein's unbiased risk estimate. A redundancy-suppression step further refines the selection. On commonsense reasoning, knowledge, math, code, and GLUE benchmarks, it improves over LoRA and the strongest mixture baseline while activating on average 8.6 of 32 components per token.
Mei Okonkwo, Owen Castellanos, Ingrid Solberg et al.
In briefCrossing the Gender Binary: How Science Fiction and General Fiction Use Gendered Words
SummaryThis paper uses the masked language model RoBERTa to test how often man/men, woman/women, and person/people are replaced by neutral or other social-category terms in science fiction and general fiction corpora. Both corpora cross the gender binary, but the replacements are unequal, with masked woman more often drawing masculine substitutes. Science fiction shows greater gender permeability and a stronger tendency toward degendering than general fiction. The method could be applied to other category vocabularies such as race and class.
In briefOptimize Once, Check Feasibility Broadly: Separating Optimization from Feasibility in Chance-Constrained Programs
SummaryThis paper addresses chance-constrained stochastic programs, which require solutions to satisfy uncertain constraints with high probability. It proposes ACCESS, a framework that first generates candidate solutions by solving many cheap single-scenario subproblems, then assesses feasibility separately through large-scale batched simulation. A variant, ACCESS(δ), provides certified feasibility with probability 1-δ. Experiments report that ACCESS matches or exceeds the best objective in nearly every setting, conservative reformulations are several percent worse, and SAA is often infeasible and much slower.
Justin Dumouchelle, Shunyu Yao, Beste Basciftci et al.
In briefJoint Subspace Sparse Attention for Faster Long-Context LLM Decoding
SummaryThis paper proposes Joint Subspace Sparse Attention (JSSA), a method that speeds up long-context LLM decoding by reducing KV-cache memory traffic. It learns a shared low-dimensional query-key subspace for better token selection and uses a gather-free indexed-attention kernel for efficient execution. The authors report accuracy close to full-KV attention across several architectures and up to 6.43× decoding throughput on an NVIDIA B200 at 128K context.
In briefHow Far Can Coding Agents Go in Visual Generation? A study of whether programming AI agents can create images by writing drawing code.
SummaryThis paper studies general-purpose coding agents that build images by writing executable programs, tested zero-shot on AgentDrawBench, a 200-prompt benchmark. The agents perform well at preserving structure and meeting precise visual constraints but are limited in perceptual richness. The authors also show that reconstructed drawing code retains substantial information for visual question answering and explore whether agents share an internal scene representation across drawing languages.
In briefCan LLMs use their world knowledge to predict events?
SummaryThis paper examines why large language models often underperform simpler models at temporal event prediction, arguing that the problem lies in how events are presented rather than in inaccessible knowledge. It proposes an interface that grounds each event in a natural-language description and separates event semantics from the prediction label space. The authors introduce DescribeEvent, a benchmark of about 150K events across seven datasets and six domains. Capable LLMs benefit substantially from descriptions and symbolic labels, while weaker models gain little.
Zijun Wu, Zhitian Zhang, Hossein Hajimirsadeghi et al.
In briefDiscrete diffusion models reach minimax-optimal rates under sparsity
SummaryThis paper studies the statistical convergence rates of discrete diffusion models, which generate text and other categorical data, under sparsity assumptions. The authors analyze a nonparametric estimator derived from a regularized score-entropy objective. They show that it achieves minimax-optimal rates in the sparse regime. The regularization adapts label smoothing and does not require knowing the sparsity parameter.
In briefIR-PRM+ with Previous-Utility Prediction: Constant-Regret and Best-Iterate Convergence Guarantees
SummaryThis paper studies regret matching variants and counterfactual regret minimization, algorithms widely used for solving normal-form and extensive-form games. It proposes using previous-utility prediction for IR-PRM+, covering both simultaneous and alternating updates. For finite two-player zero-sum matrix games, the method recovers constant-regret and convergence-rate guarantees previously obtained with extra-gradient prediction. The analysis extends to finitely many simplex learners and to the CFR variant IR-PCFR+ for perfect-recall extensive-form games.
WhyPrevious-utility prediction recovers constant-regret and convergence guarantees for alternating updates and extensive-form games, a clear theoretical extension.
In briefScaling cheap synthetic pretraining improves indoor radio-map prediction
SummaryReal measurements are too scarce to train indoor pathloss predictors that generalize across buildings. The authors build a generator for diverse indoor layouts and a fast open-source approximate ray tracer, producing 128M synthetic samples. Pretrained ResNet-101 and Swin-T models, fine-tuned on benchmark data, show a 15% average error reduction for ResNet-101 over the best known methods across five tasks.
Khoren Petrosyan, Artashes Mkrtchyan, Rafayel Mkrtchyan et al.
In briefData Density Scaling Laws for Image Self-Distillation: How Repeating Samples Affects Model Performance
SummaryThis paper tests the common assumption that more unique training samples improve model performance while repeated samples add diminishing value. Across datasets including Maxar, SATLAS, Waymo, nuPlan, Ego4D, ImageNet, and YFCC15M, the authors find that repetition effects are dataset-dependent, and that repeatedly training on a smaller random subset can sometimes outperform a larger set of unique samples. They propose a modified Chinchilla-style scaling law that accounts for both unique sample count and repetitions per sample, which improves prediction in some repetition-heavy regimes, though gains vary across datasets.
Alvard Barseghyan, Ani Vanyan, Hakob Tamazyan et al.
In briefPMO-Dock: A Benchmark for Docking-Based Molecular Design, Specificity, and Generalization
SummaryThis paper introduces PMO-Dock, a benchmark and evaluation protocol for molecular optimization that uses docking, a computational simulation of how small molecules bind to a target. It contains 25 tasks covering hit generation, lead optimization, and a specificity task that rewards strong on-target binding while penalizing off-target interactions. Validation tasks are used for tuning, while held-out test tasks are evaluated under a strict oracle budget. Four methods were benchmarked, and the results show no single method performs best across all tasks, with hyperparameter transfer often limited and method-dependent.
Gor Simonyan, Tatevik Abrahamyan, Narek Abelyan et al.
In briefAction Jets: Learning the local shape of action-conditioned image manifolds for faster planning
SummaryThe paper proposes Action Jets, a network that estimates the local shape of the set of images a world model can produce from a fixed state by varying a low-dimensional action. The shape is represented as a third-order Taylor polynomial, so inference runs once per state and can score many candidate actions. The authors report over 300x speedup over the source generator when scoring sixteen candidates, on benchmarks including BridgeData V2 and DL3DV scenes.
In briefTesting AI Models at Summarizing English and Bengali Text
SummaryThis paper benchmarks eight closed-source and open-weight large language models on abstractive summarization of English and Bengali news articles, using size-matched XL-Sum test sets in a zero-shot setting. The evaluation combines lexical and semantic metrics, reference-free factuality checks, and multidimensional LLM-as-a-Judge assessment. The authors find that strong English performance does not reliably carry over to Bengali, and that different evaluation methods produce different model rankings.
In briefFinding hidden concepts learned by LoRA adapters in image-generation models
SummaryThis paper presents a method that, given only a LoRA adapter for a diffusion model, finds the concept the adapter was trained to generate, to support later investigation. It frames concept discovery as searching for a conditioning input that jointly maximizes output coherence, distinctiveness from the base model, and alignment with the learned LoRA subspace. The authors propose a two-stage framework: render-free Subspace-Guided Evolutionary Search shortlists candidate prompts, then Paired Visual Identification renders and compares LoRA and base-model outputs. The approach aims to discover previously unknown concepts efficiently and reliably.
In briefNon-Abelian quantum Hall states under strong Landau level mixing, using AI to find anyons
SummaryThis paper studies non-Abelian anyons, particles that could underpin fault-tolerant topological quantum computing, in fractional quantum Hall states. For spin-polarized electrons at filling factor 3/2, the authors give a complete solution including all electrons and Landau levels, using a neural-network variational wave function optimized only on the microscopic Hamiltonian. They find anyons with quantum dimension √2 and report that Landau level mixing favors the anti-Pfaffian state over other candidates.
WhyFirst complete neural-network solution including all Landau levels identifies anti-Pfaffian ground states with quantum dimension sqrt(2) anyons, with overlap evidence.
Pierre-Antoine Graham, Ahmed Abouelkomsan, Liang Fu
In briefChunkAct: Action Chunking for Agentic Robotics
SummaryThis paper presents ChunkAct, which brings action chunking to agentic robotics, where an LLM agent controls robot primitives. Online, one call launches a chain of primitives that run autonomously and return control only on failure; offline, memory is compiled into these chunks so the planner handles only decisions needing reasoning. On LIBERO-Pro, it raises task success by 5.0 points over the prior state of the art while using 2.9x fewer LLM calls, 4.4x fewer input tokens, and 2.3x fewer output tokens.*
In briefLoopWAM: Improving general robot policies in a closed loop using mixed-quality deployment rollouts
SummaryLoopWAM is a closed-loop world-action model that learns from robot rollouts of all quality, including successes, suboptimal behavior, and failures, produced by itself and other deployed policies. On the RoboChallenge Table30-V2 benchmark, three deployment-and-training rounds raised success from 23.67% to 43.33%, with a final score of 56.80. The authors also report strong results on RoboTwin 2.0 and RoboCasa365 and show improvement on a shirt folding task after single-task post-training.
WhyNovel closed-loop world-action model learning from mixed-quality rollouts; multi-round gains across three benchmarks give solid evidence, with notable influence on robot learning.
In briefEgoExplorer: A Dataset Capturing Human Walking-and-Manipulating Motion, Vision, and Touch
SummaryThis paper introduces EgoExplorer, a 100-hour dataset of human activities that combine locomotion and object manipulation, recorded with a wearable system. It aligns egocentric video with whole-body motion, finger articulation, tactile signals, and step-level descriptions, focusing on professional tasks. The authors propose a Calibrate-Complete-Constrain pipeline for reconstructing motion from sparse sensors, reporting a 71.9% reduction in body-point error over the strongest baseline. Preliminary experiments explore multimodal generation and whole-body control retargeting for humanoid robots in simulation.
In briefMiniDog: A low-cost dog-breed text-to-image flow matching framework trained on four consumer GPUs
SummaryThe paper presents MiniDog, a framework for pretraining and fine-tuning a diffusion transformer for text-to-image generation on four RTX 3090 GPUs. Pretraining uses 26k ImageNet dog images with captions in about 55 minutes, and fine-tuning uses 2k FLUX.1-schnell generated images in about 15 minutes. The model combines an E2E-INVAE latent space, a 22M-parameter transformer, a small Qwen text encoder, and acceleration via iREPA and an EUPE perception encoder. The authors state they do not claim technical novelty, and the project is open-sourced on GitHub.
Reyhaneh Esmailizadeh, Xingjian Leng, Zhanhao Liang et al.
In briefCepheid: separating scale from shape so one motion model works across different mesh types
SummaryThe paper addresses recovering dense mesh motion from sparse vertex displacements. It finds that residual priors trained on one geometry family often fail to beat parameter-free harmonic extension on other families, and that residual amplitude varies widely across families. The authors learn a dimensionless prior and estimate amplitude from observables with a lightweight correction fitted on training sources, and their unified model outperforms harmonic extension on all evaluated families, including unseen cloth and legged bodies.
In briefQuality-Aware VLA Pretraining: A Benchmark and Empirical Study
SummaryThis paper studies how to use demonstration quality during vision-language-action (VLA) pretraining to improve performance after target-task fine-tuning. The authors compare weighting, filtering, and conditioning across reward definitions and label granularities, using 5,000 naturally suboptimal human demonstrations and 24 target tasks. They also introduce AXIS-Bench, a benchmark built on crowdsourced simulation demonstrations to support these controlled comparisons, and report that simulation pretraining raises real-world success from 9.6% to 31.7%.
In briefTev: Dense progress rewards for training AI agents on long tasks
SummaryThe paper introduces Tev-Gym, a dataset of 1,170 verifiable long-horizon task environments across 7 domains, built by first constructing a hidden specification and then rendering tasks progressively. It proposes progress-potential reward shaping (PPRS) to convert task progress into dense per-step rewards. The authors report that the resulting Tev-27B model improves on ALE-CM and generalizes to Terminal-Bench subsets.
Wenbo Hu, Shraman Pramanick, Krishna Kumar Singh et al.
In briefPLAYGROUND: A method that generates design layouts step by step, adapting to the rendered canvas
SummaryThis paper proposes PlayGround, a framework for generating editable graphic design layouts. Instead of fixing the high-level plan in advance, it examines the current rendered canvas at each turn, forms the next functional group, and plans its composition, with a visual verifier providing feedback. Experiments on Crello and LICA show it outperforms existing methods, and the authors plan to release code and weights publicly.
In briefTraining Diffusion Models for Lattice Field Theory Without Data
SummaryThis paper proposes training diffusion models on their own generated samples, avoiding the need for samples from the target distribution. The method uses a target score identity and requires access to the unnormalized probability density. The authors benchmark variants on two-dimensional φ⁴ theory and find effective choices for symmetric, critical, and broken phases. The data-free approach performs comparably to training on HMC reference data.
Dennis Grinwald, Ankur Singha, Khaled Kahouli et al.
In briefDoubtLess: Training-free contrastive decoding that makes AI reasoning more concise
SummaryReasoning models often use more tokens than needed in their long chains of thought. The paper proposes DoubtLess, a training-free decoding method in which a much smaller model reads the large model's reasoning once prompted to be brief and once prompted to be verbose, and the logit difference steers the large model toward brevity. On seven benchmarks in math, graduate-level science and competitive programming, it shortens Qwen3-32B's chain of thought by 31% with unchanged accuracy. The authors also report reduced repeated checking and generalization to other model families.
WhySteers large reasoning models via small-model brief-versus-verbose logit differences, training-free; seven benchmarks show 31% shorter chains with unchanged accuracy.
In briefHybrid Reasoning Agents: LLM agents that think in latent space and act through text
SummaryThis paper introduces Hybrid Reasoning Agents (HybridAgents), which switch dynamically between latent reasoning and text-based interaction within one generation process. A lightweight HybridLink module handles transitions between the two modes, and the authors propose HybridSFT and HybridRL as post-training methods. Across 11 benchmarks covering mathematics, science, medicine, search, and agentic reasoning, the authors report a 10.4% performance gain and 35.5%–54.7% token reduction.
WhyProposes a unified agent switching between latent reasoning and text tool use, with post-training, validated across 11 benchmarks against strong baselines.
In briefAesCode: Aesthetic Code Generation Using Separate Cross-Modal Rewards
SummaryThe paper presents AesCode, a method for generating visual artifacts such as slides, posters, and dashboards as editable code. It uses an image generated from the same prompt as aesthetic guidance and represents the design as a graph, with deterministic verifiers scoring parsable properties and a VLM judge scoring the rest. Reinforcement learning optimizes these rewards to improve layout, aesthetics, and correctness. The authors report that AesCode-8B and AesCode-32B improve visual quality over same-scale Qwen3-VL baselines by 31.1 and 22.4 percentage points.
In briefVisual Tokens on Demand: A Visual Budgeting Method for Efficient Multi-Image Reasoning
SummaryMultimodal large language models become expensive when processing many images, since each image adds many visual tokens. This paper proposes mmVB, a lightweight plug-and-play module that assigns a query-conditioned token budget to each image using cross-image context. It adds only 1.6M trainable parameters, requires no retraining of the base model, and does not access attention matrices at inference. The authors report a 9x reduction in visual input tokens and 5x faster time-to-first-token, while keeping 94.4% of full-resolution accuracy across eight multi-image benchmarks.
Aimon Rahman, Joymallya Chakraborty, Vidit Bansal et al.
In briefPixelMoE: Scaling Pixel-Space Diffusion with LLM-Style Mixture-of-Experts
SummaryThis paper introduces PixelMoE, a pixel-space image generation framework that applies standard LLM-style Mixture-of-Experts to sparsely scale model capacity without extra diffusion-specific routing. It also proposes Pixel-Space Hierarchical Flow Matching, which changes spatial resolution from global structure to fine detail during denoising. At comparable active parameter counts, it outperforms dense counterparts, and on class-conditional ImageNet-1K it reports FID scores of 1.47 at 256x256 and 1.53 at 512x512 without representation alignment or perceptual losses.
In briefGenerating graphs with exactly matching topological persistence
SummaryPersistent homology summarizes how a graph's topology changes, using persistence diagrams (PDs). Existing topology-aware graph generators use soft regularization, so their PDs only approximately match a target. This paper proposes two methods that generate graphs realizing a target PD exactly: an iterative local method based on vertex degrees, and a global constraint-satisfaction formulation solvable by constraint programming solvers. Experiments on four standard graph generation benchmarks confirm that both methods reproduce the target PD exactly.
WhyFormalizes exact persistence-diagram-matching graph generation, with two provably sound methods validated on four standard benchmarks.
In briefTwo Clocks: Why Larger Rollout Batches Learn Slower but Fail on Schedule in RLVR
SummaryThis paper studies how batch size and learning rate interact in reinforcement learning with verifiable rewards (RLVR) for language models. It proposes a two-clock framework that separates the samples needed for progress from those consumed before training collapse. Controlled GRPO experiments on Qwen2.5-Math-1.5B across a 16x batch range support both predictions, and the authors derive a simple multi-step schedule that starts below the critical batch and switches to a larger one before failure.
In briefMFG: Multi-turn Feedback Guided Molecular Optimization
SummaryThis paper introduces MFG, a reinforcement learning framework that trains a large language model to optimize molecules over multiple turns using feedback from earlier turns. It uses a feedback vector covering property improvements and similarity to the source molecule, along with a reward function aligned to that feedback. On ten multi-objective molecular optimization tasks, GeLLM³O-MFG reached an average similarity-constrained success rate of 81.50%, versus 70.81% for the base GeLLM³O model.
In briefWorldRecon: generative multi-view mesh reconstruction of large 3D scenes
SummaryThis paper presents WorldRecon, a framework that reconstructs complete 3D meshes of large-scale scenes from unposed RGB images. It uses a coarse global point cloud and camera poses as a scaffold, generating the scene block by block with neighbour-aware inpainting to keep overlapping blocks coherent. The authors report reduced Chamfer distance and improved F-score over the strongest baselines on synthetic and real-world indoor scenes.
In briefCan AI Agents Build Earth Observation Models Themselves?
SummaryThis paper proposes that Earth observation (EO) models can be built by a coding agent instead of human experts. Humans only supply a harness, which defines what the agent may use, and a verifier, which defines success, and the agent then builds and recursively improves the model. On six EO benchmarks covering optical, SAR and multispectral time-series imagery, agent-built models outperform the prior state of the art on every one, and on land-cover mapping they use only 11% of the training labels.
WhyProposes agent-built EO models via harness and verifier, claiming state-of-the-art on six benchmarks; evidence is only summarized in the abstract.
In briefRepProbe: Can Video Models Count Repeated Actions? A Diagnostic Benchmark
SummaryThe paper introduces RepProbe, a benchmark of 42K real and synthetic video clips, each containing exactly 1 to 20 complete repetitions of an action. The authors find that evaluated video language models' counting accuracy collapses at high counts. Linear probes trained on final-layer representations still decode substantial count information, suggesting models encode repetition structure but do not reliably report it in their outputs.
In briefLIBERO-RIG: a benchmark testing vision-language-action robot models against hardware and setup changes
SummaryThe paper introduces LIBERO-RIG, a benchmark that changes only a robot's hardware and setup, such as camera mounting, calibration, control rate, and gripper pads, while keeping LIBERO tasks fixed. It covers eleven perturbation axes with five severity levels each, giving 210 conditions, and evaluates nine vision-language-action models. Pooled success drops from 0.932 to 0.767 at middle severity and 0.509 at the highest severity. The authors also report that training on visual perturbations does not address faults that leave camera images unchanged.
WhyA new hardware-perturbation benchmark for VLA models, with 210 conditions, 500 episodes each, and statistical tests showing substantial degradation useful to the subfield.
In briefWho Improves the Robot? Toward Physical Recursive Self-Improvement
SummaryThis paper studies Physical Recursive Self-Improvement (Physical RSI), where a robot's physical experience leads to lasting updates of both its capabilities and its improvement procedure, so later improvement works better across tasks. The authors propose an L0–L5 framework separating capability-level from meta-level updates, and reserve L5 for independently verified recursive gain. They conclude that the surveyed literature does not yet demonstrate Physical RSI as defined, and they identify priorities in reliable feedback, causal attribution, safety preservation, and benchmarking.
In briefFaster Offline Q-Learning Guarantees Using KL Regularization
SummaryThis paper studies offline Q-learning for finite-horizon Markov decision processes with KL regularization. It proposes KLQ, a variant of classical Q-learning, which achieves a sample complexity of Õ(1/ε) without explicit pessimism, though with exponential dependence on the regularization parameter and horizon. The authors then introduce KLPQ, which adds pessimism and keeps the same rate while removing that exponential dependence. The paper claims these are the first fast-rate guarantees for offline Q-learning in tabular finite-horizon KL-regularized MDPs.
In briefLarger online optimal transport minibatches to stabilize flow matching training
SummaryFlow models need stable training, and how source samples are paired with data affects that stability. The authors propose a scalable online entropic optimal transport approximation that allows much larger source-target couplings while keeping the benefits of optimal transport. They prove that, compared with independent random pairing, it strictly reduces full stochastic-gradient variance and integration error, with gains growing as the coupling minibatch grows. Experiments on single-cell gene interpolation flows and image flows support these predictions.
In briefPlan-Bench: Can Current Vision-Language Models Reason and Plan Beyond Perception in Embodied Settings?
SummaryThis paper introduces Plan-Bench, a benchmark that tests vision-language models' high-level reasoning and multi-step planning for embodied tasks, separating planning from action execution. It spans three complementary dimensions, 16 subclasses, and 699 curated samples. Evaluating 28 models, the strongest reaches only 75.0% accuracy, and larger models generally perform better.
In briefRoboSimStudio: A toolkit for quickly building robot task simulations from real-world observations.
SummaryThis paper presents RoboSimStudio (RSS), an interactive toolkit that makes solver, contact model, and physical parameter choices explicit and editable. It lets both humans and coding agents create and refine simulations based on observed physical behavior. The toolkit is built on the open Newton physics simulation system. Experiments cover human-guided real-to-sim using visual evaluation, force feedback, and agentic Real2Sim2Real.
In briefGame2Real-Bench: Do Toy Games Reveal and Transfer to Real-World Visual Reasoning?
SummaryThis paper introduces Game2Real-Bench, a benchmark that tests whether vision-language models' reasoning on procedurally generated toy games matches their performance on real-world visual tasks. It covers fine-grained perception, spatial reasoning, constraint reasoning, and sequential planning. Across 25 models, game and real-world results correlate strongly (0.935), and one-shot game demonstrations improve constraint reasoning for several models while reducing fine-grained perception.
In briefOmniCompass: A Comprehensive Benchmark for Testing AI Understanding of Video, Audio, and Locations
SummaryThis paper introduces OmniCompass, a benchmark for evaluating how well multimodal AI models understand videos. It has a vision-language track and an audio-visual track, each covering temporal grounding, spatial grounding, and question answering. The benchmark includes 820 videos for the vision-language track and 240 for the audio-visual track. Tests on 50 models show open-source models lag behind proprietary ones, especially in audio-visual understanding, and that precise grounding remains hard with inconsistent predictions across paired target descriptions.
In briefWorldOracle: A benchmark testing whether action-driven world models match real robot outcomes
SummaryWorldOracle is a benchmark for evaluating action-conditioned world models (ACWMs) against paired real-world executions. Given the same initial observation and low-level action sequence, it checks whether a predicted rollout reproduces real physical consequences and task-state transitions. Models are assessed on physical state correctness, interaction spatial alignment, key interaction area fidelity, and multi-view action coherence. Experiments on seven representative ACWMs reveal distinct strengths and failure modes.
In briefNeural method jointly selecting chemical reaction models and estimating their rate parameters
SummaryThis paper presents a neural inference architecture for Bayesian inference in Chemical Reaction Networks, aiming to distinguish overlapping reaction pathways and estimate rate constants at the same time. It uses shared 1D-CNNs and a permutation-invariant Set Attention mechanism to learn low-dimensional summary statistics from variable-length, multi-experiment time series. These statistics feed a model selection network and model-specific conditional Invertible Neural Networks that estimate parameter posteriors. The approach is illustrated on a real protodeboronation example involving competing catalyst deactivation pathways.
J. Arturo Esquivel, Jorge A. Campos-Gonzalez-Angulo, Sergio Pablo-Garc\'ia et al.
In briefRead the Scene First: Choosing LLM Sampling Temperature from a Scenario Representation
SummaryThis paper proposes Scenario-Actor, a framework that selects the sampling temperature for large language models. It extracts scenario-relevant structures from Transformer hidden states and determines temperature from a scenario embedding, a geometric representation of the communicative setting. The authors report that on Llama-3.1-8B-Instruct it achieves strong goal completion in interactive social scenarios, significant gains on open-ended instruction following, and competitive multi-turn conversation performance against several baselines.
In briefDuplexBridge: Coordinating Understanding and Generation for Unified Image Editing
SummaryThe paper presents DuplexBridge, which lets a unified model use its own understanding ability to enhance prompts for its own image generation, without an external understanding module. It combines a source-grounded bridge with separate local supervision for understanding and generation, plus reciprocal trust gates for shared feedback. The authors introduce CE-Bench, a 1,200-task benchmark, and report 21.75 to 25.75 percentage point gains in complete compliance over direct editors.
In briefBack-Translation Still Matters: Using Monolingual Data to Improve LLM Translation for Low-Resource Languages
SummaryThis paper studies how to use moderate amounts of monolingual data to improve LLM-based machine translation for low-resource languages. The authors apply back-translation, a pre-LLM technique, where the fine-tuned model translates monolingual sentences to create synthetic parallel pairs for training in the reverse direction. With 1K parallel pairs on Odia and Swahili, it outperforms continued pretraining in domain. Across five low-resource languages and two 3B backbones, it improves translation into the low-resource language by 16 to 173% in XCOMET-XXL over parallel-only fine-tuning, and human evaluation on Odia and Sanskrit agrees.
In briefSelf-evolving agent harness: improving agent software scaffolding while keeping model weights fixed
SummaryThis paper studies how an agent's software scaffolding, including prompts, memory, workflows, and executable components, can be revised based on feedback while model weights stay fixed, with accepted changes persisting across runs. The authors frame the search as evolutionary optimization and analyze its reliability using continual PAC-learning, deriving an expected generalization bound, a retention bound, and a high-probability admission rule based on fresh validation data. The work also discusses how persistent revisions affect knowledge sharing, organizational workflows, and scientific reliability, physical safety, and accountability.
In briefA statistical study of the grokking phenomenon in a tractable setting
SummaryThis paper studies grokking in a simple, analytically tractable setting to separate the effects of initialization, signal strength, weight decay, sample size, and input dimension. It identifies conditions under which grokking provably occurs and derives asymptotic expressions for the lengths of generalization plateaus. The theoretical results are supported by extensive experiments.
In briefDepth Memory: A Unified Framework and Linear Depth Memory for Cross-Layer Information Retrieval
SummaryThe paper treats mixing representations across network depth as storing and retrieving information across layers. It introduces Depth Memory, a unified framework describing the PreLN transformer, mHC, and AttnRes through streams carrying representations, and proposes Linear Depth Memory (LDM), which solves cross-depth retrieval as a recursive least squares regression. A hardware-aware implementation is presented to reduce peak activation memory and memory traffic. According to the paper, LDM achieves the lowest perplexity among the baselines at 400M and 1.5B parameters, with competitive zero-shot accuracy.
David Gonz\'alez-Mart\'inez, Philipp Nazari, T. Konstantin Rusch et al.
In briefTuning Team Diversity Online in Multi-Agent RL via Extremum-Seeking Control
SummaryThis paper proposes Adaptive Diversity Control (ADiCo), a two-timescale framework for cooperative multi-agent reinforcement learning that adjusts behavioral diversity during training. A fast loop regulates diversity toward a target, while a slower Extremum Seeking Controller updates that target using only mean episodic returns. Evaluated on six configurations across two cooperative tasks, ADiCo maintains task-appropriate diversity and matches or exceeds the mean performance of grid-searched DiCo and Unconstrained DiCo.
In briefSkillScaffold: A Training Scaffold for Agentic RL with Selective Skill Injection
SummaryThis paper proposes SkillScaffold, a training framework for agentic reinforcement learning in large language models. It reuses outcome-verified trajectories to generate candidate skills and estimates their utility through outcome-aware likelihood comparisons, avoiding costly rollouts. Based on this utility, the method decides whether and which skill to inject, and runs skill-free when no candidate is estimated to help. Experiments on ALFWorld, WebShop, ScienceWorld, and τ²-Bench show gains over baselines, and the code is publicly available.
WhyEstimates skill utility via likelihood comparisons instead of rollouts, enabling selective injection; evaluated on four benchmarks against strong baselines with ablations.
In briefRegistration as a Representation Problem: Whole-Body PET/CT Alignment
SummaryThe paper proposes establishing correspondence between images using complementary anatomical representations instead of raw intensities. It combines dense per-voxel features from the 3D foundation model anatomix with coarse labels from off-the-shelf segmentation models, driving a GPU-accelerated FireANTs solver in a multistage framework. The authors also introduce lesion-specific volume preservation schemes, and the system ranked first on the Learn2Reg 2026 PSMAReg leaderboard without task-specific training.
In briefλ0: Scaling Open-World Egocentric Human Data for Humanoid Whole-Body Loco-Manipulation
SummaryThe paper introduces HumanVerse, a 500-hour dataset of open-world human loco-manipulation with egocentric video and synchronized body and hand motion. A human-to-humanoid pipeline aligns camera viewpoints and arm appearance and refines motion references and controller inputs offline. The resulting foundation model, λ0, reports 90.0% success on SIMPLE and 62.5% across real-world tasks, with zero-shot transfer to Unitree G1.
Lambda-humanoid Team: Chongyang Xu, Jin Chen, Jinhui Ye et al.
In briefSEAR: A Framework and Benchmark for Self-Improving Language-Model Robot Agents
SummarySEAR (Self-Evolving Agentic Robotics) is a framework and benchmark for testing whether language-model agents genuinely learn to improve robots, or merely make more attempts. The agent runs in three modes: writing control code as the policy, acting directly in the control loop, or fine-tuning a vision-language-action model. Across seven frontier models, 330 simulated tasks, five physical tasks, and four years of agent runtime, the study finds that learning depends on time budget and task difficulty, that control code transfers across tasks and simulators, and that different models evolve in different ways. The authors release the benchmark and codebase.
In briefTrust Before Autonomy: Rethinking Agentic Remote Sensing for Crisis Response
SummaryThe paper argues that autonomy for agentic Earth-observation systems should be granted in proportion to demonstrated reliability, epistemic honesty, and verifiability. An evidential meta-analysis of twelve peer-reviewed agentic remote sensing benchmarks finds that most lack repeated runs, unanswerable tasks, or calibration measures, and that task success often overstates tool-call correctness. The authors propose three EO-specific trust properties, an autonomy ladder for crisis tasks, and a readiness protocol that replays archived emergency activations, concluding that current evidence supports only assistive roles.
Georgios Angelidis, Aggelos Psiris, Spyridon Evangelatos et al.
In briefSituHOI: A Large Dataset for Scalable Capture of Full-Body Human-Object Interaction in Real Scenes
SummaryThe paper introduces SituHOI, an RGB-D dataset of 18,429 annotated clips from 16 indoor rooms, totaling about 86 hours, recorded with a single moving camera. It includes static scene scans, object geometry, language descriptions, and estimated human and object trajectories, all aligned in a shared frame. The authors also propose FlexHOI-Bench, a benchmark for predicting future human and object motion, with SharedHOI as a reference model that outperforms four adapted baselines.
In briefWhere LLM Ensemble Gains Come From: Splitting Candidate Value, Diversity, and Aggregation
SummaryThis paper examines where gains from ensembling large language models come from, comparing direct generation, single-candidate, duplicate-candidate, and multi-candidate conditions. It shows that overall candidate gain splits into an additional-candidate contrast and a candidate-context contrast, and that this split depends on the chosen control. The authors report paired results on MultiHopRAG and FRAMES and argue that evaluations should state the reference control.
In briefPack Light, CarryOn: Object-Oriented Context for Long-Horizon Agents
SummaryThis paper introduces CarryOn, an object-oriented context runtime that organizes long-running agent tasks around reusable task objects instead of interaction histories. It maintains continuation carrier records across context folds, revising them against bound evidence and environment checks. On OSWorld 2.0 and BrowseComp-Plus, the authors report gains of 5.8% and 7.0% with total token usage reduced by 51.9% and 22.9%, respectively.
In briefBLADE: Auditing How Information Is Used to Attribute Failures in LLM Multi-Agent Systems
SummaryLLM multi-agent systems combine specialized agents, tools, and workflows, but collaboration makes failures harder to diagnose. This paper introduces BLADE, a training-free framework that audits how agents use received information, not just where it came from, to distinguish error introduction, propagation, and repair. Tested on 1,228 failed runs across five Who&When and Who&When Pro subsets, BLADE achieved the best or tied-best step accuracy among eight strategies without reference answers, outperforming All-at-once by 2.2 to 11.1 percentage points.
In briefGAD: Geometry Attention Distillation for More Consistent Video Generation
SummaryVideo diffusion models produce visually appealing short clips but often fail to keep scene geometry and object structure stable over time. The paper introduces Geometry Attention Distillation (GAD), which transfers cross-frame correspondences from a frozen geometry model into a video model's attention, so it focuses more consistently on the same surfaces or moving objects. Experiments on three video backbones show better long-range scene consistency and dynamic temporal consistency. On Wan2.2 5B, GAD improves cross-frame matching accuracy by at least 30% over the Wan baseline on two long-range benchmarks, without adding inference modules.
In briefJeva: Fast Decision-Making Models Built from First Principles
SummaryThis paper introduces Jeva, a decision-native architecture for fast, direct "system one" decisions. The authors argue that existing methods that predict answers from serialized text do not respect the structure of decision problems. Jeva is designed to be null-aware, permutation-invariant, scalable to many options, and cacheable across requests. The paper reports strong typed-decision performance, near-perfect null detection, no decision changes under tested permutations, and substantially lower latency through reuse.
In briefFeTS++ Challenge: a competition for training tumor imaging models across institutions without sharing data
SummaryThis paper presents the FeTS++ challenge, a federated learning competition in medical imaging where models are trained locally at each institution and data never leaves its site. It uses multi-institutional multi-parametric MRI data from the BraTS 2025 Lighthouse challenge, covering gliomas and meningiomas. Tasks are segmentation and classification, and submissions are ranked by model utility and an efficiency-aware score.
Wenyi Tang, Carla Pitarch-Abaigar, Siddhesh Thakur et al.
In briefMIST: Predicting Spatial Gene Expression from Tissue Slides Using Multi-Scale Context
SummaryThis paper introduces MIST, a Transformer-based model that predicts spatial gene expression from standard H&E histology images. It adds two context pathways to each layer: a local pathway over spatial nearest-neighbor spots and a slide-level summary pathway, predicting all spots in one pass. Evaluated on six human organ cohorts from HEST-1k, MIST reports higher mean Pearson correlation than the strongest baseline across gene panel sizes, along with gains in leave-one-organ-out transfer and domain recovery.
Koushik Howlader, Tirtho Roy, Md Tauhidul Islam et al.
In briefRubrics From the Task, Verdicts From Evidence: Tool-Augmented Pointwise Rubric Reward Models
SummaryThe paper proposes ReVeR, a pointwise rubric reward model that derives hard rules and graded principles from the instruction alone and judges each response independently with tool-augmented evaluation. It is trained via label-blind distillation, producing the large-scale ReVeR-Traces dataset. Experiments report higher average preference accuracy than pairwise rubric reward models, most clearly on instruction following, along with gains in test-time scaling and downstream policy optimization.
In briefLayerCanvas: Generating Editable Infographics for AI Learning Assistants
SummaryThis paper presents LayerCanvas, a method for automatically generating infographics in AI tutoring assistants. It keeps SVG for structure and text, while using diffusion models to produce the visual appearance of non-text elements. Elements are split into z-ordered, non-overlapping layer sets that are generated in parallel and deterministically recomposed. The authors report better text fidelity than state-of-the-art raster generators and the highest overall aesthetic quality, along with non-destructive editing, localization, asset reuse, and animation.
Keshav Gupta, Jeremy Owen, Christopher Lloyd et al.
In briefLooking Complete Is Not Being Grounded: Checking LLM Disaster Sitreps Against Human Reports
SummaryThe paper derives a 15-slot information schema from professional humanitarian situation reports and evaluates 121 human and 261 LLM-generated reports across five disasters. On Cyclone Idai, the best machine output scores slightly above the human ceiling, but the difference is not statistically significant. The authors also find that several automatic metrics are uninformative or unreliable for this task.
In briefBeyond the Verdict: Agentic Code Judging with Verifiable Feedback
SummaryThis paper proposes a task where a model judges and diagnoses competitive programming solutions using only the problem statement and source code, producing checkable counterexamples as evidence instead of consuming official submissions. The authors build USACO-JUDGE from USACO 2026 contests and train ARBITER, a 35B mixture-of-experts verifier, with reinforcement learning. Training improves counterexample success on incorrect programs and binary verdict accuracy, and ARBITER outperforms baselines in code selection and refinement, including an IOI 2026 case study.
In briefPreciseCoder: A Debugging Training Method Beyond Unit-Test Rewards
SummaryThe paper finds that rewarding code models only for passing unit tests teaches them to make more edits and to hallucinate bugs in already-correct code. The authors propose PreciseCoder, which adds a reward for semantically verified, localized edits to the unit-test reward. On 4B, 9B, and 27B models, it cuts average edited lines from 7-15 to 3 while keeping pass rates comparable to unit-test-only training.
In briefLoop geometry emerging inside gauge-equivariant graph neural networks
SummaryThis paper studies the statistical ensemble of closed lattice walks produced by recursive message passing in gauge-equivariant graph neural networks, independent of training. Analytical and numerical results show that contour length grows much faster than spatial spread, with deeper networks repeatedly traversing compact regions. Loop expansions of fermionic observables give a physical interpretation of this hierarchy.
In briefContext-aware Bayesian fusion for multimodal object detection
SummaryThis paper addresses pedestrian and road-user detection for autonomous vehicles, where poor weather and low light degrade detector and sensor performance. The authors propose a context-aware multimodal late fusion method that adds a lightweight gating network to a probabilistic calibration framework, adjusting pre-trained detector outputs based on inferred environmental context without retraining. Evaluated on a modified KITTI dataset with camera and lidar data augmented for varying fog levels, the method achieves about a 14% relative improvement over non-adaptive late fusion baselines and a consistent average precision of 56% across fog levels.
Arunachalam Venkatachalam, Mhafuzul Islam, Marwan Yusuf et al.
In briefTransferring convergence rates from SDEs to stochastic gradient descent
SummaryThe paper introduces a rate-transfer method that converts finite-time Lyapunov estimates for a related stochastic differential equation into finite-time convergence bounds for stochastic gradient algorithms. Under generic moment and smoothness conditions, the convergence analysis reduces to verifying a Lyapunov generator inequality for the SDE. The method recovers established rates for classical stochastic gradient methods and can be applied to derive convergence rates for new optimization algorithms.
In briefToken-level importance sampling for off-policy RL with verifiable rewards
SummaryThis paper studies off-policy training for reinforcement learning with verifiable rewards, where data may come from asynchronous generation or reused responses. It proposes Proximal KL, which uses token-wise importance weighting and replaces PPO clipping with bounded future KL credit, and supports the method with convergence and finite-sample analyses. Experiments on Qwen3-1.7B and Qwen3-4B report higher micro-averaged accuracy than GRPO across eight math benchmarks, with up to 2.58% gain in matched off-policy comparisons.
WhyReplaces PPO clipping with future KL credit and proves a unique stationary policy; multiple matched comparisons and ablations support the claims, offering practical value for RLVR.
In briefDetecting Concealed Handguns Using WiFi Signals
SummaryCamera systems cannot see hidden weapons, and metal detectors cause congestion in crowds. This paper uses WiFi Channel State Information to detect concealed handguns and introduces the WiFire dataset of aligned WiFi signals and RGB images. It also proposes WiFireNet, a deep learning model trained with vision-based supervision. The results suggest WiFi-based detection is feasible as a complement to cameras and metal detectors, and the dataset, model, and code are released.
In briefSP²ec: Adaptive Self-Speculative Decoding Speeds Up Vision-Language Models
SummaryThis paper addresses slow inference in vision-language models by proposing SP²ec, a training-free bandit algorithm. It adapts the draft configuration during generation, exploiting the single-peak throughput pattern seen in existing self-speculative methods. Experiments on Qwen2.5-VL-7B-Instruct and LLaVA-NeXT-Mistral-7B show throughput gains of up to 11% over the original methods.
In briefLearning When to Transfer: RL for Carbon-Footprint Modeling in New Manufacturing Domains
SummaryThis paper proposes a reinforcement-learning method that selectively transfers data from existing manufacturing sources to a new target domain, avoiding negative transfer. A validation gate deploys transferred predictions only when they beat supervised alternatives. Tested on 729 material-extrusion additive manufacturing experiments across five materials, it reaches a mean test RMSE of 5.2 g CO2e per part.
In briefSegment Where You Look: Using Language Model Attention to Guide Image Segmentation
SummaryThe paper notes that vision-language models compress a target's location into one [SEG] token, losing spatial detail. The authors find that this token's attention over image patches carries a better location signal, and they propose AttnSeg, which uses this attention as the decoder's spatial prompt with direct supervision. Applied to Sa2VA, AttnSeg improves gIoU on ReasonSeg and GS-Eval, with the largest gains on object parts, multiple objects, and single objects.
WhyReplaces the [SEG] hidden state with attention maps as segmentation prompts, showing clear gains across benchmarks via a two-stage recipe.
Eslam Abdelrahman, Hamza Omar Alkaf, Yuheng Li et al.
