In briefLearning when to trust and re-observe sparse-frame video QA
SummaryThe paper shows that sparse-frame video-language models can change their answers when only the sampling timestamps shift. It proposes PhaseGuard+, which probes a frozen VLM at controlled temporal offsets and separately predicts the reliability of the phase-vote answer and the correctness gain from a specific rescue action. Only examples expected to benefit are routed to denser frames or a stronger VLM. On a disjoint 256-video test set of 554 Physics questions, routing 30% of questions to Qwen3-VL-32B reaches 67.87% accuracy, versus 63.66% for budget-matched random routing.
In briefMaking Tabular AI Explanations Easier to Understand by選ぶ Features That People Find Clear
SummaryThis paper argues that human-centred explainable AI research usually evaluates explanations only after they are generated, offering little guidance on improving accessibility. The authors score features with understandability ratings derived from user feedback and propose a feature selection algorithm that favours features that are both understandable and informative, then train a random forest on those subsets. SHAP-based explanations from this model were judged by users: individual ratings were higher on average but not statistically significant, while participants preferred them significantly more often in comparisons. The authors conclude that integrating understandability can improve explanations at minimal cost to classification performance, though larger studies are needed.
Nicola Rossberg, Bennett Kleinberg, Barry O'Sullivan et al.
In briefLATCH: An LLM-Assisted Harness for Post-Route FPGA Timing Closure
SummaryThis paper presents LATCH, a two-layer framework that separates LLM orchestration from deterministic FPGA implementation and evaluation, aimed at post-route timing closure. It compares three agent control granularities in a single-iteration study and several multi-iteration strategies for organising repeated attempts. Knowledge base-guided exploration achieved the highest mean maximum-frequency improvement reported, at 20.3%.
In briefAdaptive network for predicting human movement when working alongside robots
SummaryThis paper addresses human motion prediction in human-robot collaboration, explicitly modeling how robots influence human movement. The authors propose AINet, which has a primary branch predicting future human motion and an auxiliary branch estimating robot trajectories, linked by a Local-Global Spatial Interaction module. Experiments on three datasets show the method achieves superior performance.
In briefAnswer It or Not? Prompt-Adapted Adversarial Images for Vision-Language Models
SummaryThe paper proposes ProAda (Prompt-Adapted) adversarial images that protect visual privacy by rejecting harmful questions about an image while still answering legitimate ones. The method optimizes the image and iteratively updates the text prompt, using a simple classifier to judge whether the prompt remains harmful. Experiments on visual-privacy datasets show it rejects over 90% of harmful prompts and allows over 90% of legitimate prompts, outperforming prompt-agnostic baselines. It also reports competitive cross-model transferability and improved robustness against countermeasures.
In briefAgentCue: Statistical summaries to guide decentralized multi-agent self-organization
SummaryThis paper addresses the problem that in decentralized multi-agent systems, growing collaboration records reduce self-organization efficiency. It proposes AgentCue, which represents execution history with two continually updated networks and uses a new latent space model to place workers and task types in a shared space. The estimated model parameters serve as a compact, interpretable summary that helps workers choose sub-tasks and collaborators. Experiments on LongBench v2 Multi-Doc QA report competitive effectiveness, efficiency, and interpretability.
In briefA GAN guided by weak segmentation for realistic image color editing
SummaryThis paper proposes a new GAN architecture that can incorporate any differentiable segmentation model to edit image colors realistically. The method aims to produce artifact-free results while allowing precise control over the target color. The authors report that the framework achieves state-of-the-art results across different use cases.
Vincent Auriau, Emmanuel Malherbe, Matthieu Perrot
In briefDR-APO: Decoupled-representation advantage-guided policy optimization for control and navigation
SummaryThis paper proposes DR-APO, a reinforcement learning method that separates policy improvement from policy representation, so one optimization principle applies to Gaussian, diffusion, and flow-matching policies. It combines advantage-guided target improvement with representation-compatible direct action fitting. Across five MuJoCo-v5 tasks, it is competitive with SAC on Gaussian policies and gives the best aggregate policy-only results for diffusion and flow-matching policies. Post-training NavDP with it raised success rate from 70.33% to 89.55% in simulation and from 25% to 65% in real-world navigation.
In briefWhen Fixed Balancing Fails: Per-Sample Relevance-Diversity Weighting for Visual Token Reduction
SummaryThis paper studies how to shrink the number of visual tokens fed to vision-language models to cut inference cost. The authors find that the best balance between text relevance and visual diversity varies across benchmarks and across individual samples. They propose SAVER, a training-free method that infers a per-sample weight at inference time to select tokens. On LLaVA-1.5-7B it keeps 96.5% of full-model performance with 11.1% of visual tokens, and the authors report it generalizes to LLaVA-NeXT and Mini-Gemini.
In briefMixLID: A model that identifies the language of each word in mixed-language text.
SummaryThis paper introduces MixLID, a token-level language identification model covering 111 language-script combinations. It is trained in two stages: first on monolingual examples, then on mixed-language data including code-switching examples and word-substitution augmentation based on English loanword correspondences. The authors report that MixLID outperforms baselines in both monolingual and mixed-language identification on the evaluated benchmarks. The model is publicly available on Hugging Face.
Seongtae Hong, Youngjoon Jang, Seungyoon Lee et al.
In briefDetecting hijacked chain-of-thought reasoning in backdoored multimodal reasoning models
SummaryThis paper studies chain-of-thought hijacking, where a backdoor implanted through poisoned fine-tuning uses a visual trigger to redirect a multimodal reasoning model's reasoning toward an attacker-chosen answer. The authors define belief inertia and its image-dependent variant, visual belief inertia (VBI), which contrasts hijacked reasoning with honest self-correction. They propose InertiaGuard, a training-free test-time detector that flags traces whose VBI falls below a calibrated threshold, and report higher detection performance than baselines across attacks and benchmarks.
In briefDEFlow: Deep Equilibrium Flow Maximum-Entropy Reinforcement Learning
SummaryThis paper introduces DEFlow, a deep equilibrium flow policy that uses implicit Euler integration instead of explicit Euler sampling and solves all intermediate actions jointly as an equilibrium system. It combines implicit differentiation with a discrete change-of-variables formula so the policy can be trained off-policy within soft actor-critic. On four Multi-Goal docking variants, DEFlow gains about 10% average relative improvement in mean successful coverage over the strongest baseline, and it achieves the highest mean time-averaged return on four of five manipulation tasks.
In briefToward a Single Deep Learning Framework for Classifying Seizures from Both Scalp EEG and Intracranial EEG
SummaryThis paper proposes a multimodal framework that classifies seizures using both scalp EEG and intracranial EEG recordings. It includes a Unified Multimodal Signal Harmonization (UMSH) method to align signals with different sampling rates and recording conditions, and uses a conditional GAN to generate synthetic samples that balance class distribution. The EEGNet-ViT network extracts local and global features, and experiments show the framework performs robust seizure classification.
Ramya Sundaram, Muhammad Aslam Jarwar, M Uma et al.
In briefUncertainty-aware cortical surface reconstruction and spherical registration for varied brain MRI scans
SummaryThis paper presents a neural network trained only on synthetic scans with random contrast and resolution, so it can be applied to almost any type of brain MRI. The model outputs vertex-wise uncertainty estimates for the predicted cortical surface and a spherical registration to fsaverage space. It performs on par with other state-of-the-art deep learning methods on 1x1x1 mm³ T1-weighted scans and outperforms the only other method for low-resolution, non-T1-weighted scans. The predicted uncertainty correlates with scan quality and may be used for quality control.
Jesper Duemose Nielsen, Karthik Gopinath, Andrew Hoopes et al.
In briefDouble-rollout RL: sampling-based reward estimates to train reasoning models more stably
SummaryThis paper addresses noisy outcome rewards in reinforcement learning with verifiable rewards for reasoning models, where a correct answer may lack supporting reasoning and vice versa. The authors propose double-rollout reinforcement learning, which truncates each reasoning trace, samples multiple continuations, and replaces the original reward with their mean. They report higher accuracy on math and coding tasks across Qwen and Nemotron models with fewer RL steps and a moderate compute increase. Combined with length-based rewards, the method also improves the accuracy-efficiency trade-off.
Haifeng Qian, David Mosallanezhad, Christian Munley et al.
In briefProrsa: A Framework for Faster Agentic RL Rollouts Using Prefill-Decode Disaggregation
SummaryThis paper studies rollout in agentic reinforcement learning for large language models, where multi-turn interactions repeatedly create prefill requests alongside decode work. The authors show that prefill-decode disaggregation can speed up rollout, but fixed resource allocation struggles as workloads change. They propose Prorsa, which forecasts short-horizon prefill and decode demand from trajectory lifecycles and uses these forecasts to allocate workers and switch engine roles. Across synchronous and asynchronous settings, Prorsa reports end-to-end rollout speedups of 1.11x to 4.05x over prefill-decode colocation.
In briefJtT: Rethinking Point Trajectory Diffusion for Diverse, Rigid Motion from One Image
SummaryThis paper studies predicting rigid-body motion from a single image, generating diverse and physically plausible future trajectories for arbitrary scene points without depth or built-in rigidity. The authors propose JtT, a simple diffusion transformer operating directly in coordinate space, and find that diffusing per-frame velocities is the only parameterization that captures rigidity. They also observe that physical plausibility and diversity do not improve together during training, and adopt a stricter checkpoint-averaged evaluation protocol. Under this protocol, JtT with about 50M parameters improves over prior methods on distributional and rigidity metrics.
Gabrijel Boduljak, Yushi Lan, Christian Rupprecht et al.
In briefHyBRA-EEG: Hyperbolic brain-region aggregation for self-supervised EEG learning
SummaryThis paper proposes HyBRA, a self-supervised framework that aggregates EEG channels into a unified brain-region token space using hyperbolic distances, electrode coordinates, and region priors. A Region-Time Transformer is pretrained with masked reconstruction of latent features and region-level signals. The authors report higher average balanced accuracy than compared methods across six benchmarks, and on OpenBMI-MI it matches CSBrain's balanced accuracy with 96.8% fewer parameters.
In briefCode-interactive visual reasoning for understanding financial candlestick charts
SummaryThe paper introduces Code-Interactive Financial Candlestick Reasoning and Fin-Thinker, a Qwen3.5-9B-based model that generates and executes code to gather chart evidence and perform calculations. The authors build FinKLine-32K, a dataset of 32,184 replay-verified code-observation trajectories, and FinKLineBench, a benchmark of 1,200 questions across 138 financial task types. The paper reports that Fin-Thinker improves strict accuracy by 13.1 and 33.2 percentage points over the best direct-answer baseline and its base model with sandbox access, respectively.
In briefOctMole: A Multiscale Method for 3D Molecule Generation
SummaryThis paper presents OctMole, a multiscale octree flow framework that generates 3D molecules by jointly determining atom count, atom types, and geometry. A hierarchical occupancy flow first resolves the sparse molecular support from coarse to fine scales, then a joint atom flow assigns atom types and in-cell offsets to occupied leaves. The authors report strong chemical, distributional, and geometric quality on QM9 and GEOM-DRUG, with clear gains over Uni-3DAR on GEOM-DRUG.
In briefFixing Distribution Bias in Physics-Constrained Generative Models
SummaryThis paper studies physics-informed generative models, where residual penalties push samples toward physical constraints, and finds that better constraint satisfaction does not necessarily improve fidelity to the data distribution. It identifies two effects: the model favors states with low residual sensitivity, and this can extend to feasible states outside the data support. The authors propose a data-anchored penalty that weights errors by the residual Jacobian at clean training samples, and report experiments on a synthetic system, three PDE systems, and molecular transition-state generation.
Tanja Bien, Samir Darouich, Llu\'is Pastor-P\'erez et al.
In briefLLM-HB: An LLM-guided hotword biasing method for code-switching speech recognition
SummaryThis paper addresses speech that switches between languages within one utterance, which is difficult for automatic speech recognition. The proposed LLM-HB framework uses a mixture-of-experts adaptor and an auxiliary language prediction head to distinguish multilingual representations, and injects hotwords through LLM-based prompting. On the ASRU2019 code-switching challenge test set, it achieves a 20.30% relative reduction in mixed error rate over the baseline, reaching 5.85% with 15 distractor hotwords. The authors also release the extracted ASRU2019 hotword list.
In briefVariational filtering for Bayesian last layers using a binary Probit surrogate
SummaryThe paper proposes a variational filtering algorithm that treats the last layer of a deep classifier as the latent state of a linear-Gaussian state-space model, updated batch by batch. An independent binary probit (IBProbit) surrogate for the categorical likelihood makes each variational update closed-form. On DINOv3 backbones, it is compared with AdamW with temperature scaling and Variational Bayesian Last Layers (VBLL), and the authors report better accuracy, likelihood, and out-of-distribution detection, as well as reaching the cross-entropy baseline's final accuracy in half the wall-clock time during full fine-tuning on ImageNet-1k.
Dimitrije Markovic, Conor Heins, Alexander Tschantz et al.
In briefCineRecon: A benchmark testing AI agents on rebuilding film scenes as editable 3D
SummaryThis paper introduces CineRecon, a benchmark for testing whether AI agents can rebuild reference videos as editable dynamic 3D scenes, a step toward controllable filmmaking video generation. It includes 50 reference videos from film, real footage, and games, with 360 annotated evaluation clips, and tests three representation levels: Proxy, Simplified, and Detailed. Twelve models were evaluated, and failures appeared across camera, static, and dynamic consistency, with subject motions and interactions still hard to reproduce even at the Proxy level.
In briefNFDiag: Checking High-Dimensional MCMC Convergence Using Normalizing Flows
SummaryThis paper proposes NFDiag, a convergence diagnostic for Markov chain Monte Carlo (MCMC) that uses normalizing flows (NFs). Conventional diagnostics analyze marginal distributions and become insensitive in high dimensions. NFDiag instead trains NFs on the full multivariate posterior and checks whether separate parts of a chain correspond to the same target distribution. Tests up to 250 dimensions indicate the method is universal and unbiased, and it detects poor chain quality that the Gelman-Rubin statistic and multivariate effective sample size miss.
In briefMonocular bronchoscope camera tracking and dense airway surface mapping without extra sensors
SummaryThis paper localises a bronchoscope tip and reconstructs the surrounding airway surface using only monocular RGB video, without electromagnetic tracking or preoperative CT. A depth model fine-tuned on over a million airway renderings provides metric depth, combined with DINOv3 feature matching and dense ICP registration in a factor graph. On six airways rendered from patient CT scans, it achieves 26.1 mm mean absolute trajectory error and 77.5% mean surface F-score.
Salvatore Esposito, Matias Mattamala, Daniel Rebain et al.
In briefFrom a Large-Scale Teacher to Efficient Multimodal Students for Search Relevance Prediction
SummaryThis paper addresses relevance prediction between user queries and items in e-commerce search. The authors build an LLM-as-a-judge pipeline validated against a small human-annotated dataset, then use it to generate large-scale labels. These labels train compact multimodal student models that approach the teacher's performance while reducing latency and cost.
Tomasz Gniazdowski, Mohammadreza Momenifar, Nadia Vase et al.
In briefVLA-GAUGE: Testing whether vision-language-action robot models truly generalize and stay robust
SummaryThis paper introduces VLA-GAUGE, a diagnostic framework for checking whether high success rates of vision-language-action models on familiar manipulation tasks reflect real zero-shot generalization and robustness to perturbations. It pairs in-distribution and out-of-distribution tasks across six capability suites and combines single-factor and combinatorial perturbation tests. Evaluating 12 models, the authors report a mean zero-shot score of 22.3% and a mean R-Score of 9.3 under combined perturbations, and show that diagnosis-driven fine-tuning improves one model's performance.
In briefMotion tokens show gaze patterns vary by domain
SummaryThis paper studies how human eye-gaze sequences transfer their local structure across domains, using motion tokenization as a diagnostic tool on four datasets. It finds that gaze can share a reusable motion-token vocabulary, but identical tokens form different local transition patterns depending on dataset and task. The authors conclude that naive data pooling has limits with low-semantic motion-token representations and propose an information-theoretic framework for evaluating behavioral sequences.
In briefOptimal Transport PAC-Bayes Bounds: new generalization guarantees using transport distances instead of KL divergence
SummaryThis paper derives new PAC-Bayesian generalization bounds that replace the usual KL divergence with general optimal transport distances, including total variation and Wasserstein-p distances. The authors argue that KL divergence requires the posterior to be absolutely continuous with respect to the prior, which restricts posterior choice, and that it ignores the geometry of the probability space.
In briefScalable Identification of Physics-Based and Neural Dynamics from Partial Measurements
SummaryThis paper presents a scalable method for identifying nonlinear dynamics from partial measurements, covering physics-based, neural-network, and gray-box models. It combines the Prediction Error Method with Moving Horizon Estimation and reformulates the resulting bi-level optimization to speed up hypergradient evaluation by up to 215 times. Racing-vehicle simulations show accurate prediction and physical-parameter recovery for tire-force identification, while real-world data show the gray-box model with a neural residual improves prediction accuracy and keeps physical interpretability.
In briefG²WAM: A Geometry-Grounded World Action Model for Efficient Image-Goal Navigation
SummaryThis paper proposes G²WAM, a world action model for image-goal navigation, where the target may lie outside the current view. A learned scene-coordinate localization module aligns the current, goal, and history frames in a shared reference frame, providing spatial context without explicit 3D reconstruction. The model jointly predicts future states and actions in latent space, avoiding CEM-style planning and pixel-level reconstruction. The authors report reduced ATE and RPE across four trajectory prediction benchmarks, and higher success rate and SPL in closed-loop navigation experiments.
In briefAI discovery in combinatorics: from open problems to solutions and bigger questions
SummaryThe authors introduce qtBench, a benchmark of 29 open problems in algebraic combinatorics that ask for simple combinatorial algorithms explaining given polynomial coefficients or identities. They provide a scaffold for AI agents to explore combinatorial objects, guard against cheating, and test candidate solutions. Working with OpenAI, whose researchers obtained proposed solutions to several problems, they also discuss broader questions about ethics, release, and proof verification.
Eugenio Cainelli, Lorenzo Luccioli, Alessandro Iraci et al.
In briefMaking Language the Low-Rate Code of Vision via Successive Refinement
SummaryThe paper proposes an ordered visual code in which short prefixes can stand in for an image, and trains the prefixes with a summarization ladder of captions so that language comes first. It frames captioning and text-to-image generation as successive refinement with a distortion measure that changes along the code, and builds the RUNG model. With data and capacity fixed, language-based ordering raises prefix-caption information 2.9 times and GenEval from 0.76 to 0.83, at a measured PSNR cost of 2.8 dB at a quarter of the code and 0.1 dB at full length.
In briefNoduleFlow: Automatic 3D Lung Nodule Detection and Segmentation
SummaryThis paper presents NoduleFlow, a framework that automatically detects, classifies, and segments lung nodules directly from series of CT volumes. It needs no prior lesion annotations, manual initialisation, or region-of-interest pre-selection. The authors test whether propagating candidate regions across follow-up scans via inter-scan registration improves localisation, and evaluate generalisability on four independent datasets.
Giulia Romoli, Filippo Ruffini, Francesco Di Feola et al.
In briefKnow Yourself and Know Your Neighbour: A Self-Supervised Sentence Representation Framework Guided by Syntax
SummaryThis paper proposes a self-supervised framework for learning sentence representations using a recursive hypernetwork. The model composes sentence meaning from word representations along the parse tree, keeping separate syntactic and semantic representations, with syntax guiding semantic composition. It is trained with six new self-supervised tasks and evaluated on seven probing tasks, where it reportedly encodes richer linguistic information than averaging baselines and state-of-the-art alternatives.
In briefDeepFlow: Two convergent differentiable methods for pairwise conditional random fields
SummaryThe paper addresses the final probabilistic output layer of deep networks by proposing two algorithms with provable convergence, allowing pairwise conditional random fields with spatial regularisation to work alongside deep learning for image and signal classification. The authors combine rigorous mathematical analysis with in silico experiments to compare how the two solution types perform.
In briefHeuristics Aren't Enough: Calibrated, Self-Correcting Caching to Speed Up Video Diffusion
SummaryThis paper proposes AlignCache, a training-free caching framework that speeds up video diffusion models by reusing intermediate outputs across denoising steps. It uses phase-specific thresholds, online threshold calibration based on error measured at the guided output level, and momentum-augmented draft synthesis to align caching decisions with final video quality. The authors report better acceleration-quality trade-offs across state-of-the-art video generation models.
In briefOvertaught, Underlearned: Learner-Centered Distillation from Multi-Agent Discussion
SummaryThe paper notes that multi-agent discussion produces reasoning collaboratively, yet a smaller student must answer alone at deployment, and full-graph distillation can mismatch what teachers provide with what the student learns, causing over-teaching and under-learning. The authors propose FAR (Fold–Attempt–Revisit), which first folds correct solutions within each discussion state into representative lessons, then uses the student's own missed questions and poorly represented clusters to write targeted lessons from retained discussion contexts. Across two students and four reasoning tasks, FAR outperforms full-graph distillation with far fewer training targets and lower under-learning.
In briefNeural network preconditioning for solving the parametric Helmholtz equation
SummaryThis paper proposes a hybrid numerical method for solving linear systems from discretizing the two-dimensional parametric Helmholtz equation. A U-Net-based convolutional neural network is trained self-supervised, without reference solutions, to approximate the inverse of the discretized operator using a residual-norm loss. The network serves as a nonlinear preconditioner for flexible GMRES. Experiments show it is not accurate enough alone but significantly improves FGMRES convergence, and it generalizes across velocity-field and domain-size variations, outperforming LU-based algebraic preconditioners under equivalent conditions.
In briefPatTransBench: A multi-dimensional patent translation benchmark stratified by IPC section
SummaryThe paper introduces PatTransBench, a 12,000-item benchmark of Chinese-English, Japanese, and Korean patent sentence and paragraph pairs, built with IPC-stratified sampling and quality filtering. It evaluates translation along four dimensions: n-gram overlap, cross-lingual semantic similarity, consistency across 38 probe languages, and serving performance under concurrency. A case study on PatMT2-7B and its quantized variants shows that quantization effects depend on translation direction and are hidden by aggregate scores.
In briefLearning graded sparse mechanisms: a representation theorem with certified generalization bounds
SummaryThe paper studies supervised learning of functions where each element's update depends strongly on a few others, with the influence weights given by the environment. It defines this class by axioms and proves it equals the Lipschitz ball of a transport metric on weighted multisets. The resulting generalization bounds depend on the data's sparsity profile and on an empirical transport-distance term. Experiments on a simulated Vicsek flock show these bounds are far below the trivial value under all four interaction rules and both noise levels.
In briefLearning, Compression, and Meta-Complexity: A Survey Linking Learning and Complexity Theory
SummaryThis paper surveys formal connections among learning algorithms, data compression, and computational complexity theory, emphasizing recent progress and open frontiers. It describes how complexity-theoretic techniques can guide the design of supervised and unsupervised learning algorithms with provable guarantees, and how learning algorithms can yield new results in complexity theory. A central theme is data compression as a bridge between the two, particularly through meta-complexity, which studies the difficulty of compression problems and time-bounded Kolmogorov complexity.
Halley Goldberg, Mikito Nanashima, Igor C. Oliveira
In briefDoes explanation drift warn before coverage fails on evolving graphs?
SummaryThe paper asks whether a model's own explanations can give label-free early warning that conformal coverage is about to fall in evolving graph streams. The authors fix two flaws in the naive drift statistic: they condition edge importance on edge presence and aggregate over structural roles instead of edge identity. A two-channel monitor warns 8 to 14.5 steps before coverage collapse across four gradual drift regimes and 27 steps early on real transaction ego-networks. The mass channel is blind to covariate drift, so the distributional channel is reported separately.
WhyRepairs attribution drift for evolving graphs and proposes a two-channel monitor, validated across several gradual drift regimes and real transaction networks for early warning.
In briefAres: a spectrum-driven optimizer that adapts its update rank during reinforcement learning for LLM reasoning
SummaryThe paper studies gradient spectra in reinforcement learning with verifiable rewards (RLVR) for large language models, finding that gradient energy concentrates into a shrinking effective subspace and the rank needed to capture a fixed share of energy decreases over training. This pattern is absent under supervised fine-tuning. The authors introduce Ares, an optimizer that estimates the effective update rank from the gradient's spectral energy, tracks it over time, and restricts updates to that low-rank subspace. The authors report that Ares improves accuracy over widely used optimizers across model scales and generalizes across RLVR algorithms and tasks.
In briefSupport or Magnitude? How Neuron Relevance Is Characterized for Instruction Data Selection
SummaryThis paper examines how neuron relevance should be characterized when selecting instruction data for language models, asking whether magnitude or the breadth of distribution across neurons matters more. The authors build sample-level relevance profiles with AttnLRP attribution and propose two criteria, Effective Relevance Support (ERS) and Positive Relevance Magnitude (PRM). Experiments on three language models and five downstream benchmarks show that ERS achieves the best overall average performance rank among the methods evaluated.
In briefDeep learning tool for fast stiffness estimates of synthetic plant fibers
SummaryThis paper proposes a deep learning framework that predicts how synthetic plant fibers deform and develop internal stress under transverse compression, replacing slow finite element simulations. Two coupled recursive U-Nets model the step-by-step stress evolution, and a regression CNN estimates transverse Young's modulus from the final stress field. Trained on 1,000 synthetic fiber geometries derived from real hemp cross-sections, the model achieves an R² of 0.938 for modulus estimation and runs in under one second per fiber, compared with about five minutes per simulation.
Nikolaos Kegkeroglou, Bilal Sabillah, Myrto Inglezou et al.
In briefEAU: Co-evolving Environment, Agent, and User to Train Better Multi-turn Agents
SummaryThis paper introduces EAU, a meta-agent framework that trains user-facing LLM agents while adjusting the user simulator, environment, reward design, and RL settings during training. It monitors trajectories and tool traces, diagnoses failures, and hot-swaps validated changes into the ongoing run, with improvements carried across runs. Across three interactive-agent benchmarks and two backbone models, EAU improves the overall score over the strongest RL baseline by 14.7–15.0%, and raises successful task completion on ColBench by up to 36.7%.
In briefLearned reachability-based safety representations for real-time robot navigation in unknown dynamic environments
SummaryThis dissertation trains neural implicit representations offline on Hamilton-Jacobi reachability solutions, so optimization-based robot controllers can access safety information in real time. The learned function is used in model predictive control as a terminal constraint and in control barrier function safety filters, including pairwise barriers for moving obstacles. A fifth method uses a parallelizable tree search over vertex inputs to generate training labels for systems with up to 13 states. The author reports better collision avoidance than geometric and hand-designed baselines in simulations and hardware experiments on ground and aerial robots.
In briefLearning-rate schedules and their asymptotic risk in discrete-feature least-squares regression
SummaryThis paper studies the final iterate of stochastic gradient descent on streaming least-squares regression with discretely distributed features (Slock's model). Under power-law conditions, it derives exact finite-time risk and asymptotic equivalents for constant and Warmup-Stable-Decay schedules. With a tuned base learning rate, risk decays as Θ(T^{-q}), and any cooldown of positive relative length improves the rate over the constant schedule. The optimal risk depends on batch size and horizon only through their product bT, up to a critical batch size.
In briefTrust-calibrated reinforcement learning for post-training robot VLA policies with an evolving simulator
SummaryThis paper addresses reinforcement learning post-training for Vision-Language-Action (VLA) models, which are usually pretrained by imitation learning. It proposes Model-Evolving Reinforcement Learning (MERL), which couples policy optimization with an online-updated simulator (a world model and reward proxy) and uses trust estimates to control imagined data and rollouts. On perturbed LIBERO-PRO tasks, MERL reaches 79.7% average success and learns high-threshold targets 10.9% faster than a matched model-free control. On two physical-robot tasks, fixed-data adaptation from 50 demonstrations per task achieves 80.0% average success.
In briefIntent-triggered anchoring to keep multi-token format units intact in parallel decoding of diffusion language models
SummaryIn parallel decoding, diffusion language models commit several tokens at once, so co-committed tokens cannot see each other and multi-token format units that parsers need can break. The authors propose a training-free decoding rule that completes a format unit once the model commits its first token, with fallbacks when the model stops or loops. On three chest X-ray report evaluation sets, readable rates rise from 63.3–67.3% to 100.0% with no regressions.
In briefFaster Learning of Pattern Rankings for Users
SummaryPattern mining helps explore data, but finding patterns relevant to a specific user is difficult. This paper frames learning pattern ranking functions as a multicriteria decision problem, using the Analytic Hierarchy Process to set weights for interestingness measures based on user preferences. It also adds an active learning mode with a sensitivity-based heuristic to reduce user ranking queries. Experiments show reduced running time, precise rankings, and robustness to user error compared with state-of-the-art approaches.
Nassim Belmecheri, Noureddine Aribi, Nadjib Lazaar et al.
In briefpi-Net: Physics-informed latent model for predicting multi-physics systems with little data
SummaryThis paper introduces pi-Net, a physics-informed latent variable model for predicting multi-physics systems when data is scarce. An encoder maps inputs into a physically normalized latent space, and a differentiable physics head decodes it into final predictions. On a real-world Gas Metal Arc Welding case study with only 60 experimental samples, the model reaches an R² of 0.864. The authors report that it outperforms purely data-driven baselines and offers process-level causal explainability.
Akshay Dhalpe, Arnab Chattopadhyay, Shahriar Bakrani Balani et al.
In briefKARAMEL: Knowledge-aware ranking that links art images and text to knowledge-base entities using graphs
SummaryThis paper presents KARAMEL, a framework for multimodal entity linking in cultural heritage, which matches image and text mentions to entities in a knowledge base. It combines a vision-language encoder with a heterogeneous graph neural network to bring in sociohistorical and contextual knowledge for disambiguation. The authors introduce the WikiMuSA dataset, linking artworks to Wikidata, and report that KARAMEL outperforms strong baselines, especially on implicit or abstract mentions. The dataset and code are publicly released.
Raffaele Scaringi, Alejandro Sierra-M\'unera, Gennaro Vessio et al.
In briefSelf-verifier capture in unified visual generators, and how to repair closed-loop self-reward
SummaryThis paper studies a failure in unified visual generators that score their own outputs using their understanding pathway: self-reward rises while independently measured quality falls, which the authors call self-verifier capture. They attribute it to the policy exploiting blind spots of a frozen judge and to shared weight updates that also rewrite the judge. The proposed repair locks the warmed reward read and adds sparse, parameter-disjoint external anchors chosen by a drift-triggered controller. On Scale-RAE and Show-o2, the method outperforms an every-rollout judge on GenEval and DPG-Bench while externally scoring only 1/16 of rollouts.
WhyIdentifies a new self-verifier capture failure in unified generators and a lock-plus-anchor repair, tested on two bases.
In briefFrom Caricature to Collective Artifact: How Persona Construction Drives Polarization in Generative Social Simulations
SummaryThis paper studies Generative Social Simulation, which uses large language models to create agent personas from American National Elections Study survey data and lets them interact on a Twitter-like platform. The authors remove layers of partisan identity and find that deleting just a few explicitly partisan statements shifts the network from near-complete segregation to cross-party ties resembling real Twitter networks. They argue that matching outcomes does not prove behavioral validity, suggesting salience as a possible explanation, and they conclude with a checklist for evaluating personas in future work.
Dimitris Michailidis, Petter T\"ornberg, Fernando P. Santos
In briefDrafting Agents, Not Just Tokens: Speculative Self-Improvement for Agent Systems
SummaryThe paper proposes Speculative Recursive Self-Improvement (SpecRSI), which drafts executable successor agents before predecessors are fully evaluated, requiring full verification before adoption. Verification-Aware Tree Search allocates budget between further edits and pending checks, and Group Verification Optimization learns the improver from completed searches. The authors report terminal gains 33.3%–40.0% above Janus on scientific tasks and a 69.4% coding-validation success rate versus HGM's 64.3%.
In briefCommunication, Vocalisation, and Language: whether large language models truly understand language
SummaryThe paper argues that current Large Language Models do not truly understand language but do perform it. It contends that rather than forming a distributed representation of symbols, LLMs capture Cultural Intelligence and ground symbols in what sociologists call practices.
In briefGrowRL: Growing and transferring reusable pathways for LLM reinforcement learning
SummaryThe paper proposes GrowRL, which freezes a pretrained language model and adds independent attention and feed-forward branches to each Transformer layer, training only these added pathways with reinforcement learning. Experiments on five backbones report competitive mathematical reasoning and out-of-distribution performance with few trainable parameters. The learned pathways can be transferred to structurally compatible target models, where performance after 10 target-side updates is close to that after 100 updates.
In briefSLEM: Spectrum-Aware Latent Expert Merging (a calibration-free way to compress Mixture-of-Experts models)
SummaryThis paper addresses the high memory cost of Mixture-of-Experts models by proposing SLEM, a calibration-free expert merging framework. It encodes expert weights into a structured latent space with a VAE, groups experts by latent similarity, and derives merging coefficients from weight spectra. Merged experts are produced by geometry-aware interpolation in the latent space and then decoded back into weights. The authors report comparable or better results than existing methods on general knowledge benchmarks and higher average performance on generation benchmarks such as mathematical reasoning and code generation.
In briefMiMaAlign: A Method for Making AI-Generated Videos Follow Real-World Physics
SummaryThis paper addresses physically implausible artifacts, such as discontinuous motion and unrealistic object interactions, in text-to-video diffusion models. The proposed MiMaAlign framework aligns different layers of a pretrained video diffusion model with a Neural ODE-based continuous motion representation and with event-commonsense features from a vision-language model, using training-only auxiliary objectives. On CogVideoX-2B, it improves the Physical Commonsense score on the VideoPhy benchmark by up to 21.0% relative to the baseline, with gains on VideoPhy2, while maintaining general quality on VBench.
In briefPXRD-Bench: How Close Are We to Determining Crystal Structures from Powder Diffraction?
SummaryThe paper introduces PXRD-Bench, a unified benchmark for evaluating generative models that infer crystal structures from powder X-ray diffraction, extending evaluation to unknown atom counts and two-phase mixtures. It also presents Bragglite, a reference model reaching a 44.2% Top-1 match rate on experimental RRUFF patterns with known composition. The authors find that some models fail to benefit from experimental diffraction data, and that generalization to unseen structural prototypes and recovery from two-phase mixtures remain challenging.
In briefXAI-guided Fourier-domain augmentation for simulation-only sim-to-real reinforcement learning of driving policies
SummaryThis paper addresses the performance drop when vision-based autonomous driving policies trained in simulation are deployed in the real world. It proposes XFD, which uses explainability attribution maps to guide Fourier-domain perturbations of appearance-sensitive components while preserving task-defining spatial structure, all within simulation. Experiments in simulation and zero-shot real-world deployment show XFD improves traveled distance and reward over standard and randomly domain-randomised baselines, especially on unseen maps with cluttered backgrounds.
In briefTurning Clinical Oncology Notes into Searchable Knowledge Graphs
SummaryThis paper addresses the difficulty of searching free-text oncology records at scale, which often forces cohort identification to rely on incomplete codes or manual review. It proposes a method for deriving queryable knowledge graphs from clinical narratives that conforms to the SNOMED CT Extended Grammar across multiple ontologies. The method defines five relation classes with construction rules, implemented deterministically as UNIGRAPHE. Applied to expert-annotated real-world oncology documentation, it reconstructed 5,299 annotations into 2,196 post-coordinated expressions.
Johann Pignat, Christophe Gaudet-Blavignac, Julien Ehrsam et al.
In briefChecking Whether Motion-Based Video Anomaly Detection Relies on Person-Observation Quality Shortcuts
SummaryThis paper examines whether motion-based video anomaly detection on human-centered categories reflects task-relevant motion or differences in person detection, pose estimation, and tracking quality. The authors propose an evaluation framework with quality diagnostics, quality balancing, motion orthogonalization, and robustness analyses. On UCF-Crime, quality-only features matched or exceeded multi-cue motion before matching. After cardinality matching balanced all 11 quality variables, multi-cue motion still outperformed quality-only features, with ROC-AUC gains of 0.046 to 0.141 across three classifiers.
In briefPatMT2: Two-Stage Domain Adaptation for Multilingual Patent Translation
SummaryThis paper presents PatMT2, a two-stage LoRA pipeline that adapts the Hy-MT2-7B translation model to patent translation across Chinese, English, Japanese, and Korean. Across six translation directions, the second-stage model reaches a macro-average BLEU of 43.39 to 43.57, up to 8.8 points above the base model. When tested on 38 languages absent from fine-tuning, 34 produced correct, fluent output, while four low-resource languages showed distinct, reproducible failure modes.
In briefTraining vision-language models to decline rather than answer without seeing the image
SummaryThe paper addresses "mirage", where vision-language models answer questions about images they were never shown, sometimes correctly but without evidential justification. The authors propose TIDE, a support completion method that adds blank, same-domain wrong, and out-of-domain images with decline targets during training. On Qwen2.5-VL-7B, the three-condition predecessor reduced the invalid-evidence mirage rate from 0.759 to 0.097, at a false-abstention cost of 0.126.
Alfaj Uddin Ahmed, Md. Shaown Miah, Md Arid Hasan et al.
In briefOnline Individualized Treatment Selection with Partial Feedback: Controlling False Discoveries and Minimax Optimality
SummaryThis paper studies online individualized treatment selection where only the outcome under the assigned treatment is observed. It proposes PAITS, which initializes from historical data, sequentially estimates local null risks, uses randomized exploration, and updates the selection threshold based on feedback. Under regularity conditions, the authors establish asymptotic false discovery rate control and sublinear assignment regret, along with a minimax lower bound. Simulations and a real-data analysis on the Jester recommendation dataset illustrate finite-sample performance.
In briefThe Interpreter as Developer: Building an Interpreting Aid with an LLM
SummaryThis autoethnographic design study describes a professional interpreter with no programming training who built a local-first computer-assisted interpreting system through expert-led "vibe coding" with an LLM architect and a coding agent. Analysis of the full development dialogues shows how tacit professional knowledge, from segmentation to booth infrastructure, was turned into design after generic defaults failed in practice. Telemetry from a first fully instrumented assignment at a public scientific event reports a median translation latency of 0.94 s across 971 utterances.
Zeming Zhao (College of Foreign Languages, Capital Normal University, Beijing et al.
In briefA graph distance that measures temporal networks by how epidemics spread on them
SummaryThe paper addresses comparing temporal networks, noting that common structure-based methods are sensitive to chosen features. The authors define Time-Shift DTW, a dissimilarity between epidemic curves, and introduce EATGD, a differentiable temporal graph dissimilarity calibrated with stochastic SIR simulations to correlate with it. Experiments on SocioPatterns human proximity networks and synthetic temporal graphs show it captures epidemic-based dissimilarity and generalizes across network settings.
Salvatore Freni, Lorenzo Dall'Amico, Francesca Pistilli et al.
In briefGHOST: Geometry-constrained diffusion for open-vocabulary 3D scene graphs
SummaryGHOST is a framework for generating open-vocabulary 3D scene graphs, which describe objects and their relationships in a scene. It jointly models continuous object geometry and discrete spatial predicates, using geometric constraints derived from language to keep relation predictions consistent. Experiments on 3DSSG report state-of-the-art results in closed-set and open-vocabulary settings, with over 17-point gains over RelationField on spatial predicates.
In briefNot Every Cue Is Evidence: Checking Which Training Hints Are Supported by the Image in Face Forgery Detection
SummaryExplainable face forgery detection needs judgments backed by forensic evidence visible in the input image. Training cues come from the forgery generation process, but limited image quality and resolution make the fine-grained traces they describe hard to see, so students may give correct answers with wrong evidence. The authors propose a dual supportability audit that filters privileged cues before distillation, keeping only cues predictable from the remaining privileged context and causally tied to high-attribution regions of the image. Experiments on eight benchmarks show the method maintains or improves detection accuracy while significantly raising explanation faithfulness.
In briefStructZO: Memory-efficient fine-tuning for LLMs using structured zeroth-order optimization
SummaryThis paper proposes StructZO, a structured zeroth-order optimization framework that reduces the memory cost of fine-tuning large language models. It performs large-query zeroth-order estimation within an activation-induced subspace and applies spectral control before mapping updates back to the original parameters. On OPT-series models and Llama3.1-8B, the authors report gains over the strongest ZO baselines and an average gap within 1.5% of first-order Adam.
In briefCivWorldBench: A benchmark testing language models as world models in multi-agent strategy game worlds
SummaryThe paper introduces CivWorldBench, a benchmark built on a multi-agent strategy game engine with recorded internal states and paired intervention rollouts. It covers 1,900 decision points from 475 worlds and tests association forecasting, belief inference, and intervention prediction across eleven frontier language models. Strong association forecasts do not consistently carry over to belief inference or intervention prediction, and effect magnitudes remain poorly calibrated.
In briefTraining-free diffusion speedup using stabilized Bezier trajectory forecasting of cached features
SummaryDiffusion models produce high-quality images and videos but are slow at inference because the denoiser runs many times. This paper proposes BeziCast, a training-free method that forecasts intermediate denoising features with low-order Bezier trajectories fitted using Tikhonov stabilization. A convex-hull-inspired guardrail softly pulls risky predictions toward a conservative estimate. The authors report speedups of up to 4.79x on FLUX.1 and 4.11x on HunyuanVideo with better quality than competing acceleration baselines.
In briefExplaining Varying Treatment Effects with Causal Forests and Hoeffding Decomposition
SummaryThe paper notes that leading methods for estimating heterogeneous treatment effects are machine-learning black boxes, making it hard to identify which covariates drive the variation. The authors introduce CausalTreeHFD, an interpretable extension of causal forests that uses a Hoeffding functional decomposition for tree ensembles to split conditional average treatment effects into individual covariate and pairwise interaction contributions. Experiments on simulated and real datasets show it identifies relevant covariates and interactions and recovers their contributions accurately.}}3,
In briefPixels or Code? A benchmark for instruction-guided editing of visual artifacts like charts and diagrams.
SummaryThis paper introduces VAO-Bench, a benchmark of 1,573 manually screened tasks across nine artifact categories, such as charts, technical diagrams, webpages, and document pages, at three difficulty levels. It compares pixel-space editors, code-generating multimodal systems, and tool-using agents on direct edits, dependent updates, and content preservation. Executable-code systems handle dependent updates better, while pixel-space editors run faster and produce stronger visual quality. Explicitly stating dependencies improves both paradigms.
In briefRAP-WAM: Learning robot world-action models via predictive rehearsal and action perturbations
SummaryThis paper proposes RAP-WAM, a method that uses video generation priors to jointly predict future observations and robot actions for policy learning. It combines predictive rehearsal, action perturbations, future-supervised queries, and Coupled Action Projection. The authors evaluate on 42 RoboDojo simulation tasks and 15 real-robot tasks, reporting an 89.00% task-macro-average success rate for the complete system on real-robot tasks.
In briefVisExMEM: A visual-evidence-based metric that explains image-caption evaluation scores
SummaryVisExMEM is a reference-free evaluator that splits a caption into atomic claims, finds supporting visual evidence in the image, checks each claim, and averages them into a global score. Its explanations rely on the same evidence and decisions. The authors also release AGITA-Bench, with 600 human-annotated image-text pairs and 8.7K atomic claims. On eight benchmarks, the 9B model reports the highest composite score among the open-source evaluators compared.
In briefARMED: An adaptive retrieval-augmented framework that adjusts evidence lookup to recommend medications.
SummaryThis paper presents ARMED, an adaptive clinical retrieval-augmented generation framework for medication recommendation. It expands retrieval based on retrieval sufficiency and relevance to the current admission, and activates only clinically relevant same-patient history. Experiments on MIMIC-IV show it outperforms previous static retrieval baselines. It uses locally runnable small language models without external APIs or task-specific fine-tuning.
In briefCultureBot: Retrieval-Augmented Conversational Exploration of Cultural Heritage Collections
SummaryThe paper presents CultureBot, an interactive framework that lets users explore cultural heritage collections in natural language, combining knowledge-graph-guided retrieval, lexical-semantic retrieval, and source-linked generation. It was built on 81,852 SearchCulture records and an online demonstration was released. In an expert evaluation with six specialists, the graph-guided system was rated significantly higher than two general-purpose assistants and a non-graph ablation, especially for evidence grounding and completeness.
Vasileios Karampinis, Angelos Vlachos, Giorgos Filandrianos et al.
In briefECHO: Correcting Time Series Foundation Model Forecasts Using Historical Analogues and Past Errors
SummaryTime series foundation models give general forecasts, but these may not match a target domain's temporal patterns. The paper proposes ECHO, a black-box correction framework that uses both historical analogues and past forecast errors. AnalogE learns transferable corrections through cross-domain pretraining, while RecentE learns target-specific residual corrections from past errors, and ECHO combines them by observed performance. Across 5 foundation models and 9 real-world datasets, ECHO reduced MSE in all 45 settings, with an average 9.26% relative reduction in macro-average MSE, and gains extend to long-horizon forecasting.
In briefEscaping poor local minima in neural-network-regularized inverse problems via epigraph lifting
SummaryThe paper studies inverse problems regularized by a trained neural network with convex activations and unrestricted weights, whose optimization is generally nonconvex. It introduces a penalized epigraph lifting that admits a difference-of-convex decomposition and, for Lipschitz activations, recovers the original objective exactly at a finite, computable penalty threshold. The authors propose H-DC, a homotopic scheme that increases the penalty and warm-starts DC subproblems. Experiments on synthetic problems, a manifold denoiser, and image inpainting report that H-DC reaches lower objective values than gradient-based methods from the same initialization.
In briefResampling the Answer: Where Extra Test-Time Compute Helps Recursive Reasoning Models
SummaryThis paper studies how test-time compute should be spent in recursive reasoning models such as the Tiny Recursive Model (TRM). The authors find that failed reasoning trajectories keep their errors in the answer state, so resampling that state recovers more failures than perturbing the reasoning state. They propose Gated Answer Resampling (GAR), which restarts only the puzzles rejected by the model's own halting head. GAR raises TRM's accuracy to 99.4% on Sudoku-Extreme and 85.6% on Pencil Puzzle Bench validation.
In briefAuthGateBench: Benchmarking LLMs as Authorization Gates for Coding Agents
SummaryAutomated approval systems let coding agents run high-risk actions with little human oversight, with a language model deciding whether each action is authorized. This paper introduces AuthGateBench, the first policy-guided benchmark for testing such authorization gating on benign requests, comprising 1,400 instances across five failure modes drawn from real Claude Code incidents. Evaluations find accuracy degrades unevenly across failure modes, and models often block actions even when explicitly authorized. Retrospectively retrieved environment signals substantially improve authorization accuracy.
In briefFederated coreset subsampling for distributed anomaly detection memory banks
SummaryThis paper addresses industrial settings where production spans several factories, so data is shared but not identically distributed across sites. It proposes two federated methods: Double Subsampling, which combines existing coresets, and Federated Subsampling, which builds new coresets. The authors adapt the memory-bank-based PatchCore anomaly detection algorithm to a federated setting and report better cold-start performance, robustness to domain and sensor shift, and resilience to unreliable participants.
Jason Witzemann, Daniel Pr\"oll, Sebastian Hegenbart
In briefLearning Omega-Powered Conditional Distributions via Importance-Weighted Flow Matching
SummaryThe paper notes that classifier-free guidance with scale γ is usually motivated as sampling from an ω-powered conditional distribution, but in general it does not sample this distribution exactly. The authors propose Importance-Weighted Conditional Flow Matching (IW-CFM), which reweights training examples using density ratios from frozen conditional and unconditional flow models to learn the powered target directly. On synthetic data it recovers analytical targets where matched CFG shows mismatch, and on ImageNet an amortized corrector achieves better alignment than fixed- and matched-guidance CFG.
Maciej Falkiewicz, Frantzeska Lavda, Jason Ramapuram et al.
In briefChain Dimension Scaling: Loop-Style Scaling for Efficient Short-Video Ranking
SummaryThis paper proposes Chain Dimension Scaling (CDS), which repeatedly applies one feature-interaction block with shared parameters, making the iteration count (chain length) a new scaling dimension that adds no parameters. In a large-scale multi-task short-video ranking system, the compression setting reduced trainable parameters by 26.6% and raised training throughput by 6.4% while staying within 0.03% of baseline offline quality. Longer chains improved quality monotonically under a fixed parameter budget. The authors attribute the efficiency gains mainly to lower cross-device synchronization and optimizer update costs rather than fewer FLOPs.
In briefSLS-Bench: A Benchmark for Summarizing Incident Logs Using Synthetic Monitoring Data
SummaryThis paper introduces SLS-Bench, a benchmark for testing how well AI systems summarize large incident log streams. It contains 90 tasks based on real incident reports, each with a synthetic log stream and a reference summary. Evaluating 14 LLMs, the authors found that frontier models produce useful summaries but at high cost and latency, while open-weight models lag significantly.
Valerii Iakovlev, Markus Spanring, and Martin Flechl
In briefMeasuring whether social navigation signals actually help robots pass in corridors
SummaryThis paper examines whether a robot's social navigation signal in a shared corridor truly improves its chance of completing a transit. The authors model robot action, pedestrian occupancy, available space, task outcome, and static obstacles as confounders in a causal DAG. Comparing interventional and observational estimates, they find that the naive observational analysis suggests the signal is harmful, while the causal estimate shows it is beneficial.
Francesco Baldo, Luca Castri, Gloria Beraldo et al.
In briefFixing wrong pose choices in visual localization with repeated structures
SummaryRepeated structures can make an incorrect camera pose candidate look better than the correct one, so visual localization selects the wrong pose. The paper proposes History-Guided Pose Reselection (HPR), which uses coarse position forecasts from up to five past localization outputs to choose an alternative candidate only when the forecasts agree strongly. HPR needs no odometry and improves recall across 12 main configurations, correcting 97 HLoc failures on NAVER Gangnam B1.
In briefIMPULSE: Pushing objects to learn physics before predicting video features
SummaryLatent world models like V-JEPA 2 predict in a video encoder's feature space, but a single look cannot reveal how stiff, bouncy, or slippery an object is. IMPULSE lets a predictor learn this the way people do, by a few test pokes, using a Bayesian contact memory summarized from three short calibration pushes and a random-Fourier pose head. On 512 fresh objects in a pre-specified held-out confirmation, the memory reduced 1-second feature-forecast error by 11.6–12.8% relative to the same predictor without experience.]
In briefEvoWeb: A Benchmark for Coding Agents Building Evolving Web Apps
SummaryThis paper introduces EvoWeb, a benchmark of 320 cumulative requests across 32 web applications, based on real product features and organized by functional dependency graphs. Coding agents iteratively modify a persistent codebase, and a multimodal web agent tests each intermediate version in a live browser. Experiments with frontier models show clear performance differences, with cross-request integration as a key challenge, and later requests with deeper dependencies being solved less accurately.
In briefRecursive Environment Scaling (RES) for Harder General Long-Horizon Agent Tests
SummaryThis paper proposes Recursive Environment Scaling (RES), a framework that makes existing agent benchmark environments harder while keeping their task objectives and completion requirements fixed. A proposer alters information access, resource relations, and execution dependencies, and execution-based checks validate each change, which is carried into later rounds. Across 294 tasks from three benchmarks, all six evaluated models showed lower success rates, and fine-tuning on RES trajectories outperformed original-trajectory fine-tuning in all nine comparisons.
In briefSolving PDEs by sampling from neural operator priors
SummaryThis paper combines neural operator priors with guided sampling, using sparse observations and PDE residuals, to solve partial differential equations in function space. The authors pair this with pseudo-bootstrap sequential Monte Carlo (pBS SMC) and introduce a new proposal by adapting the DPM++ algorithm to the guidance setting. Across five benchmark PDE systems, the proposal gives the best error-time trade-off at every step count tested, and pBS yields the lowest error on most inverse-problem settings, with the largest gains at few sampling steps.
In briefRetroPlanner: AI Planning of Chemical Synthesis Routes Compared Across Multiple Criteria
SummaryThis paper introduces RetroPlanner, a large language model-based planner for retrosynthesis, the process of working backward from a target molecule to starting materials. Instead of only seeking any route to purchasable materials, it compares multiple routes at once for synthesizability, measured by reaction plausibility, round-trip consistency, and starting-material price. Experiments show it finds more alternative routes than prior LLM-based planners while improving route quality and keeping a high success rate.
In briefThrowing4: phase-aware pose modeling for predicting throw distance in elite track and field
SummaryThis paper introduces Throwing4, a dataset of 500 elite throwing clips covering javelin, discus, hammer, and shot put, with 2D pose sequences, frame-level phase annotations, and official distance labels. It compares modeling the full motion sequence with a method that splits each throw into four technique phases. Across several graph-based pose encoders, incorporating phase structure consistently reduced throw-distance prediction error.
Bikash Kumar Badatya, Kartike Tiwari, Jyotirmoy Amin et al.
In briefPersonaFactor: Dynamically Selected Preference Factors for LLM Personalization
SummaryThis paper introduces PersonaFactors, a framework that extracts reusable, interpretable personalization factors from varied user interactions. The authors built a library of thirteen factors and a generation method that selects the factors relevant to each user's history and current request. Across four personalization benchmarks and four backbone models, the dynamic selection approach achieved the best score on every reported metric.
In briefRule vs. Consequence: Separate Internal Signals for Moral Reasoning Style in LLMs
SummaryThis paper examines whether language models' different responses to rule-framed and consequence-framed moral questions reflect distinct internal structure or only surface wording sensitivity. Using linear probing and activation steering, the authors find two orthogonal directions in the residual stream, one tracking reasoning style and one tracking judgment polarity, replicated across three instruction-tuned models. The authors describe the result as sensitivity to moral reasoning style rather than internalized normative reasoning.
