In briefSol-H3: Recursive Self-Improvement Speeds Up MiniMax-H3 Video Generation in Cloud and Edge
SummaryThe paper addresses the high latency and memory demands of MiniMax-H3, a 33-billion-parameter open-source video diffusion model, on cloud and edge hardware. It proposes a full-stack inference pipeline with a cross-resolution two-stage scheduler that generates low-resolution layouts first and refines details later, connected by a learned latent-to-latent mapping that avoids costly VAE decode-reencode steps. A Recursive Self-Improvement loop searches kernel fusions and memory layouts for speed and numerical agreement. The authors report up to 30x end-to-end speedup and 20% lower memory use.
In briefContext Alone Is Not Enough: Test-Time Training for Personalized Reward Modeling
SummaryThe paper finds that in-context learning based personalized reward models fail to capture the preference relations conveyed by a user's historical comparison pairs. It proposes Preference-Aligned Test-Time Training (P-TTT), which encodes these relations into user-specific fast weights for personalized reward prediction. P-TTT updates fast weights in a single forward pass without inference-time backpropagation, and the authors report it outperforms state-of-the-art methods by a large margin.
In briefMechanistic study of how language models detect internal activation changes
SummaryThe paper examines whether large language models can notice and localize perturbations to their own internal activations even when the input gives no sign of intervention. Using fixed input text, the authors inject a concept vector at one of ten token positions or apply no intervention, and ask the model to identify the position or report no change. Across three model families, they identify two small groups of attention heads: middle-layer gate heads influence whether a change is reported, while later-layer router heads help select the reported position.
In briefDoAtlas-2: A self-evolving foundation for causal biomedical discovery using population evidence
SummaryThe paper introduces DoAtlas-2, a system that organizes knowledge around causal mechanisms and evolves using external evidence from human populations. It integrates 771 research resources covering over 720,000 participants in 48 countries, plus about 4.7 million literature-derived records. The system autonomously formulates research questions from evidence gaps, prespecifies causal designs, and runs validated analyses. In the Human Phenotype Project, it formulated 4,014 candidate pathway questions and screening of the first 1,079 yielded statistical support for 756, with blood pressure identified as a convergence node linking adiposity, hepatic, and lipid phenotypes to vascular outcomes.
In briefDRIFT: A Method for Predicting Single-Cell Responses to Perturbations
SummaryThis paper proposes DRIFT, a framework for predicting how cells change after perturbations such as drug treatments. A variational encoder splits each cell into an invariant part unaffected by the perturbation and a responsive part that changes, and conditional flow matching transports only the responsive part. The authors report that across several benchmarks, the method outperforms the strongest published method for combinatorial and unseen perturbation prediction.
Mustapha Bounoua, Giulio Franzese, Pietro Michiardi
In briefE3J: An Open-Source GPU and TPU Backend for Fast Euclidean Equivariant Operations
SummaryThis paper presents e3j, a fast backend for Euclidean-equivariant operations used in geometric deep learning, with JAX bindings for GPU and TPU. It combines optimized CUDA and Pallas kernels with algorithmic improvements to reach high throughput in both forward and backward passes. The authors report up to 34% speed-up over cuEquivariance on a MACE-based water box NPT simulation, and better TPU memory bandwidth utilization than e3nn-jax. The library is released under the Apache 2.0 license.
Olivier Peltre, Armand Picard, Adrien Pichard et al.
In briefSpikeLite: Lightweight spiking neural networks for forecasting time-series data
SummaryThis paper presents SpikeLite, a time-series forecasting framework built on spiking neural networks (SNNs), which are designed to be energy-efficient. It combines a Frequency-Selective Spiking Encoder for frequency-sensitive temporal encoding and a Sparse Spiking Channel Attention module that uses a binary mask to select cross-channel interactions. Across four multivariate and eight long-term forecasting benchmarks, it reports the best aggregate performance under both evaluation protocols and the lowest reported energy consumption on the ECL dataset.
In briefDF-CBM: Region-Aware Concept Bottleneck Models for Explainable Deepfake Detection
SummaryThis paper proposes DF-CBM, a region-aware concept bottleneck model for explainable deepfake detection. It builds a vocabulary of manipulation-related concepts from textual artifact annotations, links each to facial and boundary regions, and predicts concepts from visual features using masked attention guided by parsed facial masks. The real/fake decision is made from the predicted concepts. The authors report that DF-CBM outperforms concept-based baselines and remains competitive with state-of-the-art black-box detectors.
Georgios Tsoumplekas, Vazgken Vanian, Alexandros Doumanoglou et al.
In briefBetter video-text pretraining using multi-view captions
SummaryThis paper argues that the language supervision used in video-text pretraining is of limited quality, since many datasets give each video only one sparse caption. The authors propose a framework based on a multimodal large language model that starts from 10 million videos and generates multi-view captions, combining summary and detailed descriptions. They also add a text representation with separate CLS tokens for the summary and detailed views. The reported results show improved zero-shot and fine-tuned text-to-video retrieval while using a smaller pretraining corpus than prior methods.
Fida M. Thoker, Renaud Vandeghen, Karen Sanchez et al.
In briefCan AI Automatically Extract Data for Meta-Analysis? A Study on Intercropping Research
SummaryThis paper examines whether large language models can automate extracting data from published studies for meta-analysis, focusing on crop intercropping research. The authors compare three LLM-based approaches, including direct zero-shot prompting, a staged workflow, and a multi-agent system, using six open-weight models. Direct zero-shot prompting performed best and most consistently, but no approach was fully accurate. Most combinations recovered the direction of relationships between variables but did not estimate their magnitude accurately.
Zehao Lu, Xingguo Xiong, Wopke van der Werf et al.
In briefGraphHCA: Closed-form hindsight credit assignment for long-horizon LLM agents
SummaryThis paper proposes GraphHCA, a model-free method for hindsight credit assignment that gives step-level credit to large language model agents on long-horizon tasks with sparse rewards. It reduces the hindsight ratio to a ratio of success probabilities at consecutive states and estimates a success potential through a discounted recursion over the transition graph, requiring no learned hindsight model or extra forward pass. The authors report state-of-the-art results against compared baselines on ALFWorld, WebShop, and Sokoban.
In briefContinuous Variational Synthesis: training generative DNA design models in continuous space, then quantizing them to fit chemical synthesis limits
SummaryThe paper addresses how chemical constraints on DNA synthesis force discrete parameters that hamper training generative design models. The authors train variational synthesis models in continuous space with stochastic gradient descent, then apply post-training quantization to impose hardware and wetware constraints. They demonstrate the approach on enzymes, peptides, antibody CDRH3s and regulatory DNA elements, reporting that models meet stringent reward criteria while keeping design diversity and that in vitro performance matches in silico results.
Alan N. Amin, Mattia G. Gollub, Andrei Slabodkin et al.
In briefCross-Rollout Bellman Closure for Long-Horizon Agent Reinforcement Learning: Step-Level Credit from Shared States
SummaryThis paper addresses group-based reinforcement learning, such as GRPO, for training LLM agents on long-horizon tasks. It proposes Cross-Rollout Bellman Closure (CRBC), which merges the rollouts of each task group into a finite empirical process and solves one linear system for its behavior-policy Bellman fixed point to propagate evidence through shared states. The resulting step-level credit is combined with group advantages without extra environment rollouts. Across ALFWorld, WebShop, and Sokoban with multiple model scales, CRBC improves final performance and learning efficiency, outperforming the strongest baseline by 5.59 points on ALFWorld with Qwen2.5-1.5B-Instruct.
In briefPropagate Then Sharpen: Refining Frozen Node Classifiers Without Features or Gradients
SummaryThis paper studies how to improve a frozen node classifier using only the graph and the class distributions it predicts, without node features, model parameters, or gradients. The authors split the Potts energy into a Dirichlet term that penalizes disagreement between neighbouring nodes and a Gini term that penalizes indecision within each node, and propose Propagate, Then Sharpen (PtS), which alternates probability propagation with mass-preserving sharpening. Across nine homophilic graphs with a frozen MLP backbone, PtS improves mean test accuracy over independently tuned APPNP by 1.71 points on clean inputs and 3.90 points under severe Gaussian feature corruption.
Preben Johnsen Bentdal, Nello Blaser, Xue-Cheng Tai
In briefRefineDrive: Learning from Self-Generated Failures for Vision-Language-Action Driving
SummaryThis paper addresses how vision-language-action driving models rely mainly on expert demonstrations and underuse their own failures. It proposes RefineDrive, a post-training framework that diagnoses collisions and drivable-area violations from simulator states, retrieves nearby safe corrections from a human trajectory bank, trains with correction supervision, and applies GRPO with a safety-layered reward. On NAVSIM v1, the 4B base policy's PDMS rises from 87.7 to 91.7, and the same checkpoint reaches 89.4 EPDMS on original NAVTEST scenes under NAVSIM v2 extended metrics.
In briefWhat Drives AI Citations of Web Pages? A Multi-Method Observational Study
SummaryThis paper analyses about 2 million citations of web pages by four commercial AI engines over six months, linked to 10,000 crawled pages from nineteen B2B SaaS workspaces. The authors test over sixty page features using nine statistical methods. They find that alignment between page content and prompts is the strongest page-level predictor, that common AEO checklist features lose their apparent effect once domain fixed effects are included, and that domain-level AI authority matters far more than page features.
In briefReCo: When to Relocate Sensor Kits Under Deployment Constraints
SummaryThis paper studies when a limited set of sensor kits should stay at their current home or move elsewhere, using non-intrusive load monitoring (NILM) as a case study. The authors propose a constraint-based relocation framework, instantiated as ReCo (Relocation by Coverage gain), which counts new operating regimes, forecasts each home's future gain, and compares the gain of staying versus moving each night. In replayed deployments on the Plegma dataset, ReCo outperformed fixed-dwell, count-based, and threshold-based schedules in every setting tested.
In briefConfidence Rising Too Early: Spotting Shortcut Reasoning in AI Models
SummaryThis paper introduces ConfLens, a framework that tracks how a large language model's confidence in its final answer changes during reasoning. The authors find that shortcut reasoning samples become highly confident in their answers at early stages. They propose DACS, a score that measures the entropy of the answer probability distribution at each step, without needing ground-truth answers or task-specific verifiers. Experiments on math and code reasoning tasks report over 4.3% F1 improvement over strong baselines for detecting shortcut reasoning.
In briefA study of how Fruchterman-Reingold graph visualization relates to agglomerative clustering
SummaryThis paper examines how the Fruchterman-Reingold graph visualization method relates to four agglomerative clustering methods using single, complete, average, and Ward linkage. The comparison uses three types of datasets in 2 and 10 dimensions, plus a PCA projection of the 10-dimensional data to two dimensions. The authors report that the relationship between the methods stayed similar across the three data types, that the agglomerative methods gave mostly similar results to each other and moderate similarity to the original data, and that the visualization resembled the original data but was less similar to the agglomerative methods.
In briefNot All Rollouts Are Worth Learning: Trajectory Valuation for Reinforcement Learning Post-Training
SummaryThis paper studies how to identify and reduce harmful trajectories during online reinforcement learning training. It proposes Dynamic Trajectory Valuation (DTV), which estimates each trajectory's usefulness at the mini-batch level using only gradient information and filters out detrimental ones. Experiments across PPO, GRPO, and DPO show DTV improves performance, data efficiency, and optimization stability.
In briefEchoes of Deeds: How an actor's past moral conduct shapes LLM choices
SummaryThe paper argues that current moral evaluations of large language models judge decisions in isolation, ignoring whether unrelated prior conduct affects later choices. The authors introduce MoralLedger, a framework that examines how moral history shapes LLM behavior under a fixed decision context, at both the output level and the internal representation level. They report that prior moral histories systematically change subsequent choices depending on their valence and intensity, and that these histories form a linearly recoverable direction in the residual stream. Intervening along this direction produces two-sided, intensity-dependent changes in decisions.
In briefDepot-closed multi-component construction for neural vehicle routing
SummaryMost neural solvers for vehicle routing build one route at a time until it is complete. This paper proposes multi-component construction, which keeps several route components at once and merges them in any order, treating each component as an implicitly depot-closed route so intermediate states are feasible. The merge cost reduction equals the Clarke-Wright saving. A policy trained only on CVRP100 reportedly outperforms representative neural solvers on CVRP100-500 and across capacities from 10 to 500 in a zero-shot evaluation.
WhyNovel simultaneous multi-component construction with CW-saving signals; extensive scaling and zero-shot experiments suggest notable influence on neural VRP solvers.
In briefTempoKV: Timing-Aware Staging of LLM KV Caches on SSD-Backed Memory
SummaryTempoKV is a timing-aware layer that decides when to reserve fast-tier memory for reusable LLM key-value caches stored on an SSD-backed flash hierarchy. It records cache hits as metadata-only claims and requests resources only when estimated retrieval time approaches the time needed to stage and protect the cache. Implemented in vLLM and LMCache, it reduces protected fast-tier byte-time per request by 63-91% versus immediate staging while keeping much of the benefit of advance staging. Compared with unmodified LMCache's Device-DAX L1 configuration, it reduces p95 TTFT by up to 48.0% and raises output throughput by up to 27.8%.
In briefBeyond Gradient Flow: Identifying and Recovering Dynamics from Distribution Snapshots
SummaryThe paper studies how to infer the dynamics of an evolving distribution from its snapshots. It shows that a single-time Fokker-Planck constraint cannot pin down the drift field, and that constraints across several time points remove polynomial gauge directions. The authors derive conditional lower and upper rates for the estimation error and propose a smooth test-function estimator with a finite-sample bound. Planted-circulation experiments are reported to confirm the predicted gauge contraction.
WhyIntroduces multi-time Fokker-Planck constraints resolving drift gauge ambiguity, with explicit obstructions, matching rates, and planted-circulation experiments; scope remains largely theoretical.
In briefDetecting 3D defects by modeling broken geometric relationships between neighboring parts
SummaryThis paper addresses 3D anomaly detection in point cloud data for industrial inspection. Existing methods learn the distribution of normal samples without explicitly modeling defects, which creates unclear decision boundaries. The authors propose a framework that treats defects as violations of geometric consistency among neighboring structures, learned through controlled pseudo-anomalies. Experiments on Anomaly-ShapeNet and Real3D-AD report consistent improvements over prior state-of-the-art methods in both in-domain and cross-domain settings.
In briefTIDE: Adaptive Teacher-Student Coordination of Distillation and RL for Agents
SummaryThe paper addresses training multi-turn agents with reinforcement learning, where sparse rewards limit exploration in small models and a fixed mix of teacher distillation and reward optimization may hinder progress. It proposes TIDE, which uses the trend of teacher-student disagreement to shift from on-policy distillation to RL over training and reweights teacher-guided and reward-driven updates across interaction turns. Experiments across multiple benchmarks, student scales, and ablations are reported to support its effectiveness.
In briefWhether one causal transformer can approximate sequence maps across all context lengths
SummaryThe paper asks whether a single masked transformer, with parameters independent of length, can uniformly approximate causal token-to-token maps over sequences of arbitrary length. It models tokens with α-Hölder sequences or a common modulus of continuity, and derives a generalization bound for bounded regression. Experiments on physical time series support the Hölder-regular token model at observed scales.
WhyShows one length-independent causal transformer uniformly approximates causal maps over arbitrary lengths, with a generalization bound lacking a length factor; experiments are limited to physical time
Takashi Furuya, Maarten V. de Hoop, Gabriel Peyr\'e
In briefPerceptual Quality Loss or Loss of Perceptual Quality? A study of PESQ auxiliary losses in speech enhancement
SummaryThis paper examines deep speech enhancement models trained with auxiliary PESQ loss terms to boost perceptual metrics. Objective tests show PESQ scores rise on the test set, but most other metrics stay largely unchanged. A listening experiment found that models without the PESQ loss were generally preferred across all settings. The authors also find PESQ dominates the composite metrics CSIG, CBAK and COVL, and argue for a more complete evaluation procedure.
Danilo de Oliveira, Tal Peer, Maur\'icio do V. M. da Costa et al.
In briefOPIS: A benchmark that tests whether video world models remember many specific objects over time
SummaryThis paper introduces OPIS, a benchmark for checking whether video world models keep track of several objects over time. It anchors evaluation to a fixed set of object instances from the first observation, avoiding confounds from generated histories or chosen viewpoints. The dataset has 500 cases from real-world, robotic, and game settings, with dense annotations for 12,672 rigid, articulated, and deformable instances. Across eight models, scores range from 48.65 to 56.01, and performance generally declines as the number of objects grows.
In briefEMPIRIC: Learning Residual World Models for Robot Planning Through Experiments
SummaryThe paper presents EMPIRIC, an agent that extends a physics engine with learned code for mechanisms it lacks, such as glue curing, water heating, or wind. Bayesian inference estimates the parameters and hidden states of these mechanisms from noisy observations, letting the agent predict action outcomes, choose informative experiments, and revise hypotheses. Across five simulated domains, EMPIRIC solved more tasks with fewer environment interactions than three baselines. On a physical robot, it learned wind forces and domino masses to complete a manipulation task.
WhyNovel residual world model learning missing physics mechanisms as code, validated across five simulated domains and a physical robot, though scale is limited.
In briefLEGAU: Learning Semantic Gaussian Priors for Scalable Category-level Object Pose Estimation
SummaryLEGAU is a framework that estimates an object's 6D pose and size from a single RGB-D image by jointly predicting NOCS correspondences and a canonical Semantic Gaussian Field. The Gaussian field serves as a category-conditioned structural prior, fused with visual, geometric, and category cues through a transformer module. The authors report strong results in a single-model multi-category setting, with up to 22% gains on SOPE and competitive transfer to real-world data.
In briefBA-DPO: Bias-Adjusted Direct Preference Optimization for Aligning Language Models
SummaryThis paper proposes BA-DPO, a generalization of Direct Preference Optimization that adds a bias parameter per annotator for responses carrying a declared attribute, so that annotators' systematic preferences toward attributes like names, language variety, or length are not absorbed by the model. The authors prove the objective is convex in the bias parameters and that votes identify each annotator's bias up to a shared constant. Experiments on a corpus with planted biases and on the real-annotator MultiPref dataset show BA-DPO removes most of the attribute shift caused by DPO, at no higher KL divergence and without loss in judged quality.
In briefMixed-Prior Decision Risk for Open-Set Recognition: a score that ranks uncertainty by the errors of each decision
SummaryThe paper addresses open-set recognition, where a system must either identify a probe as a known class or reject it as unknown, creating three types of error. It shows that the usual Bayesian uncertainty summary based on KL divergence does not always track decision risk. The authors propose MPRisk, which uses the same posterior but directly scores the risks of false acceptance, misidentification, and false rejection, with only four validation-tuned weights. Across nine image, audio, and text benchmarks, it achieves the best or tied-best Prediction Rejection Ratio at most operating points and beats HolUE on five benchmarks.
In briefGraph-based learning for forecasting Martian atmosphere at multiple time horizons
SummaryThis paper presents MaGMA, a framework that converts OpenMARS reanalysis data into graph-structured learning objects for multi-horizon Martian atmospheric forecasting. Local atmospheric patches serve as nodes linked by spatial, temporal, and dynamically similar relations. Evaluated on five unseen Martian years, the model reaches overall R² of about 0.73 to 0.85 in regular years and outperforms baselines for dust-column forecasting in most comparisons. In the global dust storm year, dust-column R² stays above 0.8 for the first two horizons, while broader multivariate performance declines.
In briefGUIDE-FBO: Guiding federated Bayesian optimization by sharing optimum-location distributions instead of raw data
SummaryThis paper proposes GUIDE-FBO, a federated Bayesian optimization method in which agents exchange compact distributions over the locations of their optima, inferred from local Gaussian process posteriors, rather than raw observations. The server merges and reweights these distributions before returning a subset to each agent, which then builds a Federated Interventional GP to guide local decisions. The authors evaluate the method on 12 synthetic benchmarks and three real-world optimization tasks.
In briefPersona prompts leak into unrelated user preferences in LLM simulations
SummaryThis paper tests whether changing one user attribute in a persona prompt affects only that attribute when simulating users with LLMs. Across eight black-box models, altering a target attribute also shifts responses on unmentioned attributes, which the authors call cross-attribute influence. Even when a non-target attribute is declared neutral, the target still affects choices across all five open-weight checkpoints, a gap the authors term the neutrality gap.
In briefTHEIA: A Dataset and Benchmark for AI Analysis of Analog Chip Layouts
SummaryThis paper introduces THEIA, a dataset of thousands of analog integrated circuit layout images paired with question-answer conversations, plus a benchmark. The authors fine-tune a vision-language model to analyze GDSII layout files so designers can query physical layouts interactively. Across five tasks using thousands of analog designs, the fine-tuned model outperforms general-purpose VLMs by up to 73%.
Giuseppe Chiari, Michele Piccoli, Federico Viola et al.
In briefEvaluating recommendation rankings offline when examination depends on position, using an examination-relevance split
SummaryThis paper studies off-policy evaluation, estimating how a recommendation ranking policy would perform using logged click data. The authors decompose clicks into examination and relevance and propose two estimators, LE-IIPS and ED-DR. Experiments show ED-DR achieves lower mean squared error than existing methods with large sample sizes, especially when examination depends on ranking. The authors also note limitations under small samples or cascade user behavior.
In briefJRDB-AVR: A Benchmark for Active Visual Reasoning by Robots in Real Environments
SummaryThe paper builds JRDB-AVR, a benchmark from the real-world JRDB robotics dataset, where an agent with limited view requests observations by timestamp and viewing angle. It scores both the final answer and the visual evidence supporting it. The authors also present JRDB-AVR-Agent, a reference method using an observation-grounded graph-based world model. Experiments show a substantial gap between answer accuracy and evidence accuracy in current baselines.
In briefPEAR: Progressive evidence-based automated search improvement for industrial systems
SummaryThis paper proposes PEAR, a framework for iteratively improving industrial search systems through automated experimentation. It replaces a simple keep-if-better rule with evidence-driven research states that revise hypotheses, and uses a four-level confidence-gated evaluation ladder so only candidates with statistically significant positive effects advance to costly online tests. In two A/B experiments on a real-world industrial search system, PEAR-optimized strategies increased Main Order/DAU by 2.7336% and 3.2957% relative to their baselines.
In briefFast learning-rate transfer in shallow linear networks when training length grows with width
SummaryThe paper studies whether learning rates tuned on narrow networks transfer to wider ones when the training horizon also grows with width. It analyzes a shallow linear network with one trainable hidden matrix trained by full-batch gradient descent, under extra spectral assumptions. The authors prove fast learning-rate transfer when T=o(√n) and characterize transfer rates, and they derive limiting distributions for the optimal learning rate and optimized loss.
In briefVEX-Bench: Benchmark Measuring How Hard LLM-Generated Misinformation Is to Verify
SummaryThis paper introduces VEX-Bench, a benchmark that evaluates how complex LLM-generated misinformation is to verify, as perceived during screening. It scores content across dimensions such as checkability, harm potential, source credibility signals, imposter legitimacy, and expected verification effort, and combines them into a VEX score. The authors evaluate 7 frontier LLMs and 7 generation methods on 5,880 articles, finding that no single method dominates all dimensions and that high-VEX content can be generated far more cheaply than it can be verified by agent-based methods.
In briefWebPageBench: Event-Log Verified Benchmark for Web Agents with Controlled UI Variants
SummaryThis paper presents WebPageBench, an open framework for evaluating web agents whose tasks are verified by matching typed events in each interface's own log, without a judge model or page scraping. It uses six instrumented mock sites and supports controlled UI variants that change one control's implementation while keeping the prompt and success conditions fixed. The release includes 152 tasks, a runner covering several browser and GUI-agent configurations, and a public leaderboard. On the 152-task leaderboard, the gap between agents' declared completion and log-confirmed completion reaches 41 points.
Anton Emelyanov, Maria Tikhonova, Zaven Martirosian et al.
In briefAutoDataBench: Can AI Agents Write Data for Self-Improving Models?
SummaryThe paper introduces AutoDataBench, a benchmark testing whether agents can write new agentic tasks that meet data-pipeline acceptance standards for validity, novelty, difficulty and behavioural coverage. Across three benchmarks, no evaluated agent scored above 20 out of 100 at the default 45-minute budget. Giving the strongest agent four times as long improved its score substantially, while the cost per usable task stayed nearly unchanged.
In briefProxy2World: Generating Videos from Lightweight Scene Proxies Without Paired Training Data
SummaryProxy2World is a controllable world model that generates video from lightweight scene proxies, which set layout and motion while leaving appearance open. It learns only from ordinary posed RGBD videos, without authored proxy-video pairs, using cross-modal flow matching for depth-conditioned RGB and joint RGBD generation. The authors also introduce ProxyBench for evaluation. Experiments report a better balance of structural adherence and visual quality than camera-controlled and geometry-conditioned baselines.
In briefDetecting adversarial attacks hidden in super-resolution models using spectral features
SummaryThis paper proposes a spectral method to detect adversarial attacks embedded in super-resolution model weights. It uses radially-averaged power spectral density as a feature to train an XGBoost detector. The detector is benchmarked against magnitude- and phase-based Fourier spectrum detectors across several training and cross-architecture scenarios. The authors report that their method outperforms the comparisons in most scenarios and that high-frequency features are most informative for detecting AdvSR attacks.
Emma J. Reid, Haley Duba-Sullivan, and Tony G. Allen
In briefOptimal transport masking for resolving spatial indistinguishability in spatiotemporal forecasting
SummaryThis paper tackles spatial indistinguishability, where sensor nodes with similar past patterns follow different futures, which hurts forecasting accuracy. It proposes STOT, a self-supervised framework that uses optimal transport-guided structured masking to emphasize ambiguous positions during pre-training. Experiments on six real-world datasets show competitive performance against state-of-the-art baselines, along with improved interpretability via transport-plan visualizations.
In briefLanguage-Forcing: Aligning Language with Whole-Body Robot World Models for Humanoid Tasks
SummaryThe paper proposes LANGUAGE-FORCING, a training method that uses language to make a whole-body World Action Model focus on task-relevant visual changes. A frozen vision-language model describes task progress and likely next state changes, and these text embeddings are aligned with action-conditioned future visual features during training only. The method reaches a 79.17% average success rate on HumanoidArena and raises real-world success on five tasks from 53.33% to 71.33%.
In briefTesting Whether Vision SAE Concepts Are Truly Interpretable
SummaryThis paper examines whether vision sparse autoencoders (SAEs), which aim to split model features into single human-understandable concepts, are actually interpretable. The authors adapt the Autointerpretability Score (AIS) to vision tasks and validate it in a user study. They find that common SAE interpretability metrics correlate neither with each other nor with AIS, so no single reference-free metric is sufficient to verify vision SAE interpretability.
Teodor Chiaburu, Franz Motzkus, Frank Hau{\ss}er et al.
In briefPath-Conditioned Attention Mobility: Relocating Attention Heads Across Depth in SegFormer
SummaryThis paper studies whether attention heads in the hierarchical vision transformer SegFormer-B5, used for human segmentation, can be moved to nearby depths, finding this works only when their computational context is preserved. The authors consolidate 120 retained Stage-3 heads from 40 execution depths into 20 execution hubs, with mIoU matching a control within noise. A FOLDED_VEC deployment method reduces end-to-end NVIDIA L4 latency at batch sizes 1, 4, and 8.
In briefEnvironmental conditions needed for artificial life agents to use social information via evolved plastic neural networks
SummaryThis paper uses artificial life models based on Evolved Plastic Artificial Neural Networks (EPANNs) to study when social information use might evolve in foraging environments, measured by generations needed to reach a fitness criterion. The experiments used NEAT, an algorithm that evolves both network topology and weights. A simple topology based on nearest food location let agents meet the criterion quickly, so social information was unnecessary. The authors conclude that social information use would require greater environmental complexity to evolve.
In briefTermJudge: A Document-Level Metric That Judges Rather Than Counts Terminology Errors in Machine Translation
SummaryExisting metrics for terminology in machine translation penalise any deviation from a fixed reference, mixing real errors with valid variation that human translators produce. TermJudge assigns each term occurrence an interpretable verdict: glossary-conforming uses are settled deterministically, while deviations are judged by an LLM in two steps, first labelling terminology errors and then separating valid document-level variation from inconsistency. It ranks first in system- and segment-level meta-evaluation against expert annotations and MQM scores, and is released as open-source code.
Nicolas Dahan (ISIR, ALMAnaCH), Fran{\cc}ois Yvon (MLIA et al.
In briefAutomated feature engineering, AutoML and decision-focused learning for energy consumption forecasting
SummaryThis thesis addresses the reliance of energy consumption forecasting models on expert-designed features. It introduces AutoEnergy, a domain-tailored automated feature engineering algorithm that creates interpretable features from timestamps and lagged consumption and integrates with AutoML. Across eighteen real-world energy datasets, it reduces forecasting error by 19.52% to 84.72% relative to baseline methods while running 1.31 to 4.41 times faster. Combined with decision-focused learning for a battery energy storage problem, it reduces operating costs by 22.9% to 56.5% on a real UK property dataset.
In briefPrice stability in the EU: a systemic analysis using random matrix theory
SummaryThe paper treats monthly inflation indices across Eurozone countries as a single system of prices and applies Random Matrix Theory to study the statistical and asymptotic properties of its eigenvalues. It finds the system cannot be assumed randomly generated and is noise-dominated, with small persistent country-level variations. Two countries are identified as distorting inflation dynamics, alongside three distinct regional country groups.
In briefShort-term memory convolutions for on-device keyword spotting: lower compute and power for voice wake-word detection
SummaryThis paper applies the Short-Term Memory Convolutions (STMC) framework to a modular CNN so it can run online with LSTM-like inference for keyword spotting on resource-constrained edge devices such as wearables. The authors report up to 82% MCPS reduction versus equivalently frequent standard CNN execution and up to 46% versus vanilla STMC. The best configuration reaches 93.8% accuracy on the 11-class Google Speech Commands task, and 97.1% with zero-padded data.
Pawe{\l} Warlewski, Artur Czeczko, Artur Szumaczuk et al.
In briefLearning to Act When Cameras Cut Out: Vision-Language-Action Models Under Visual Interruptions
SummaryThis paper studies how vision-language-action robot models perform when a camera stops delivering frames during a task. The authors introduce MAIL-Bench, which interrupts different cameras at several stages of successful reference trajectories to measure policy robustness. They also propose MINT, which trains policies to work with missing views and, at inference, fills gaps using optical-flow extrapolation or an action-conditioned world model before withdrawing unreliable predictions. Experiments on π0.5 and GR00T N1.5 report significantly higher task success under camera loss, and real-robot results on AgiBot G2 are also presented.
In briefFrom One-Shot Generation to Incremental Music Composition
SummaryThis paper explores using a general-purpose instruction LLM as a reusable editing operator on a persistent symbolic score in ABC notation, rather than as a one-shot music generator. The authors fine-tune Llama 3.1 8B Instruct with LoRA on about 496,000 operation-aware dialogue records from Irish traditional music, covering chord addition, inpainting, and transposition. Checker admission rises from 29.37% to 99.37%, but the paper does not establish superior musical quality or human-AI co-creativity.
In briefComposable Decoding on the Probability Simplex: Theory and Implementation
SummaryThis paper frames language model decoding as an optimisation over next-token distributions on the probability simplex, trading off model score against regularisation under support constraints. It introduces CompoSimplex, a library for building and evaluating compositional decoders. The authors report that combining decoding objectives yields trade-offs between single-sample quality, multi-sample quality, and diversity that individual objectives do not achieve.
Xiaotong Ji, Ahmed Khaled Khamis, Rasul Tutunov et al.
In briefMaPP: A Unified Framework for Data-Efficient Reinforcement Learning with Verifiable Rewards
SummaryThis paper addresses a problem in GRPO, a reinforcement learning method for training large language models with verifiable rewards: randomness in sampled group outcomes adds composition noise that limits gradient accuracy and weakens prompt selection. The authors propose MaPP, which uses a shared Beta posterior and closed-form Beta-Binomial marginalization to denoise per-response advantages and derive an uncertainty-aware prompt selection score without extra rollout cost. Experiments on mathematics, planning, and visual geometry across five model backbones show consistent gains over GRPO and strong selection baselines, with up to +2.45 average accuracy improvement under the same rollout budget.
In briefRight Lesson, Right Step: Learning Localized Control Updates for Self-Evolving Agents
SummaryThe paper introduces EvoCUE, a framework that represents an agent's workflow as an editable state-machine controller and learns localized instruction or skill edits from completed execution traces. Each candidate edit is evaluated at the node where it would act by resuming parent and edited controllers from the same checkpoint and comparing final outcomes, and accepted edits are stored with applicability rules. Experiments on AppWorld and PAST-Bench office workflows report improved task success and execution quality.
In briefNürnberg NLP's ChildSafeAds 2026 system: structurally different voter ensembles under four data-access levels
SummaryThis paper describes the Nürnberg NLP team's system for the ChildSafeAds 2026 shared task, which asks how well a system can monitor commercial content in child-facing YouTube videos at different data-access levels. The system uses per-subtask ensembles of nine voters arranged in three branches that differ in backbone, adaptation method, and class scope. It achieves the best scores on two of the three subtasks and ranks third on the task mean, and the authors also compare four access levels.
In briefORPG: Combining Several Reward Goals by Adjusting Gradients Per Objective
SummaryThis paper proposes ORPG, which builds a separate clipped policy objective for each reward and merges the resulting gradients into one policy update. Compatible gradients are coordinated by a cosine-dependent interpolation, while conflicting gradients are projected according to task priorities. The method is evaluated on helpfulness-safety alignment and on correctness-cost optimization for mathematical reasoning, where it reports gains over the strongest external baseline.
In briefActionUNet: Making Vision-Language-Action Robot Models More Robust with Efficient Multi-scale Fine-tuning
SummaryVision-Language-Action (VLA) models map language and vision to robot actions but struggle to align coarse semantics with fine-grained timing, limiting robustness in cluttered scenes. The paper proposes ActionUNet, a fine-tuning framework that adds a lightweight temporal U-Net to fuse hierarchical priors, followed by a conditional SIREN decoder with second-order smoothness constraints to reduce motion jitter. Experiments on RoboTwin 2.0 and LIBERO-Plus, plus real-world tests, report absolute success-rate gains of 9.8%, 6.1%, and 11.4% for π0.5, and generalization to the OpenVLA-OFT backbone.
WhyCombines a temporal U-Net with a conditional SIREN decoder for VLA fine-tuning, gaining 6-11 points on pi0.5 across RoboTwin, LIBERO-Plus, and real robots, and generalizing to OpenVLA-OFT.
In briefWhat Makes World Action Models Generalize? A Study of Test-Time Future Modeling
SummaryThis paper studies whether world action models must generate future video frames during inference. Latent models that skip future generation match explicit models on in-distribution tasks but generalize worse across environmental perturbation, data efficiency, and task generalization. The gap comes mostly from the first denoising step, so preparing the future matters more than generating it. The authors propose Simple-WAM, a single forward pass over fully noised video tokens, which performs well on generalization with efficiency comparable to latent models.
In briefOne Sensor, Whole Body: 3D Body Pose from One Consumer Earbud
SummaryThis paper examines how much 3D body pose can be recovered from a single head-mounted inertial sensor in a consumer earbud (AirPods), and whether adding consumer insole IMUs helps. The authors built a synchronized multimodal capture pipeline and a 35-take single-subject benchmark. They find the head IMU recovers lower-body pose well, while adding foot IMUs does not significantly improve results and can degrade them, with sensor orientation quality identified as the limiting factor.
In briefSPIDER: Pruning Visual Tokens and Skipping Sub-Layers to Make Multimodal LLMs More Efficient
SummaryThis paper addresses efficiency problems in multimodal large language models, which come from both redundant visual data and redundant computation. It proposes SPIDER, a training-free framework that combines multi-layer semantic visual token pruning with adaptive sub-layer skipping. According to the paper, on LLaVA-NeXT-7B, SPIDER reduces FLOPs by 79% while keeping 96% of the baseline performance.
In briefInspector: A Conversational, Lightweight Tool That Uses LLMs and CNNs to Analyze Analog Circuit Layouts
SummaryThis paper presents Inspector, a framework that combines fine-tuned large language models and convolutional neural networks to analyze GDSII files of analog circuits. It enables designers to interact with the tool through a conversational interface. Experiments on thousands of analog designs across four realistic tasks show it outperforms general-purpose large vision-language models by up to 81%.
Abril Cano Castro, Giuseppe Chiari, Michele Piccoli et al.
In briefTeach to Learn: Hint Annealing for Self-Improving LLM Reasoning
SummaryThis paper addresses GRPO training, where difficult queries yield only incorrect rollouts and provide no learning signal. It proposes HATCH, a single-policy framework in which the model generates its own hints, re-solves failed queries with them, and uses online weighting to anneal the influence of hinted trajectories, with gradient projection resolving conflicts with hint-generation updates. On mathematical reasoning benchmarks, the authors report gains over prior state-of-the-art methods on Llama-3.2-1B-Instruct and Qwen3-1.7B, and a further gain on Qwen3-8B.
In briefAPEX-Voice: Can Voice Agents Complete Professional Workflows Through Full-Duplex Interaction?
SummaryThis paper introduces APEX-Voice, a benchmark of 120 interactive professional workflows across ten work types, such as form completion, negotiation, and interviewing. Five frontier real-time voice agents were tested in a stateful environment. None exceeded 25% Pass@1, and the best Reliable@3 was 10.8%. The authors find that stateful coordination is the main failure point across systems.
In brief"Just MLPs": Lightweight Networks for Efficient Visual State Reconstruction in Multimodal Models
SummaryThis paper addresses the high computational cost of visual tokens in multimodal large language models. The authors find that after blocking visual-to-text attention, restoring only a few low-rank directions recovers most lost accuracy, and lightweight MLPs can closely approximate layer-wise visual states. Building on this, they propose δ-Vision, which replaces repeated Transformer processing of visual tokens with low-rank adapters while keeping all visual tokens available for text retrieval. On image and video benchmarks, it reports higher accuracy than visual token pruning baselines at comparable or lower computation.
In briefAction-Space Shaping for LLM Agents: Measuring and Mitigating Tool-Schema Bias
SummaryThis paper examines how LLM agents behave when the same executable action is described by different but functionally equivalent tool schemas, a phenomenon the authors call schema bias. They introduce a framework that rewrites tool schemas using nine operators while keeping tasks, actions, and reachable states fixed, and evaluate eleven LLMs on up to 32 schema variants. Success rates vary from complete failure to 97% depending only on the schema, and training repairs a variant only when that variant appears in the training data.
In brief"Grab a Coffee": Future-Aware Guidance for Discrete Diffusion (the COFFEE framework)
SummaryThis paper introduces COFFEE, a framework that guides discrete diffusion models toward sequence-level objectives during generation. It separates sequence dependence from the objective, using a compiled finite-state model to pass global preferences to unresolved token positions without retraining the diffusion model. The authors evaluate it on symbolic, language, and biological benchmarks, reporting strong control with task-dependent trade-offs between quality and diversity.
Hua (Edward) Xu, Dongxin Li, Gwen Yidou-Weng et al.
In briefDiagnosing and mitigating emergent misalignment in LLMs via parameter-space geometry
SummaryThis paper studies emergent misalignment (EM), where safety-aligned LLMs fail on unrelated domains after narrow fine-tuning. The authors track training trajectories and find that directional Hessian curvature concentrates on semantic pivot tokens. They propose a Geometric Mitigation Framework that orthogonally projects the harmful gradient subspace out of parameter updates. On Qwen2.5-14B-IT, the method suppresses free-generation EM by up to 80.0%.
In briefRoboFL: federated assembly of expert adapters for world-action robot models
SummaryThe paper proposes RoboFL, a federated learning framework for training vision-language-action and world-action robot models across institutions without sharing raw interaction data. It installs locally trained LoRA adapters as expert branches of a server-side mixture-of-experts, using routing distillation and path-consensus aggregation to keep expert selection stable. Experiments on RoboTwin 2.0, RLBench, and a real Franka arm are reported, with the method outperforming a centralized PEFT baseline on the Franka arm and reducing per-round client communication relative to MoE-based federated baselines.
In briefWhy measuring whether LLM answers state the same fact is key to judging their uncertainty
SummaryThis paper examines how semantic uncertainty methods for large language models sample several answers and measure their disagreement. The authors argue that off-the-shelf comparison tools, such as NLI models or generic sentence encoders, are the main bottleneck because they do not accurately judge factual equivalence. They train a single encoder contrastively on synthetic data to isolate target facts, which improved performance in 120 of 126 evaluation settings.
In briefSafe Greenhouse Climate Control with Constrained RL and Kolmogorov-Arnold Networks
SummaryThis paper treats greenhouse climate regulation as a constrained decision problem and applies the Lagrangian safe reinforcement learning method RCPO-PPO, which separates economic optimization from cumulative safety constraints so penalties adjust without manual tuning. Kolmogorov-Arnold Networks replace standard neural networks for the policy and value functions, and sinusoidal time features capture daily cycles. In simulations of a winter lettuce greenhouse driven by 40 days of real weather, the method reduced cumulative climate violations by 18.65% and increased lettuce economic profit by 2.91% compared with vanilla penalty-based PPO.
In briefPhase Conditioning for Cyclical Medical Signal Restoration
SummaryMany physiological signals, such as cardiac and brain recordings, have statistics that change periodically with their underlying cycle phase. The paper proposes encoding this cyclic structure through a shift-covariant wavelet representation and dense per-sample phase conditioning inferred from corrupted input, applied to diffusion-based restoration. It also introduces a training-free cyclostationarity index that predicts when phase conditioning helps, and antithetic coupling of reverse trajectories that achieves comparable performance with fivefold fewer network evaluations.
Samuel Ruiperez-Campillo, Michele Copetti, Jorge da Silva Goncalves et al.
In briefALICE: In-Context, Zero-Shot Mutual Information Estimation
SummaryThis paper presents ALICE, a foundation model for estimating mutual information from samples without retraining for each distribution. Trained only on synthetic distributions, it estimates the velocity field of an unseen distribution from its samples, and mutual information is then computed via a fixed identity. The authors validate it on a standard benchmark and on biology, genetics, and neuroscience data, reporting accuracy competitive with per-distribution neural estimators and support for varying data dimensions and sample sizes.
In briefProofLoom: Proof-obligation-driven autoformalization of research-level stochastic optimization
SummaryThis paper presents ProofLoom, a fully automated LLM-agent system that builds Lean models and supporting theory for research-level stochastic optimization from a published algorithm, target theorem, and source proof. Open proof obligations guide the creation of definitions, lemmas, and proof plans, while an independent Judge rejects unsupported assumptions and weakened conclusions. On fifteen tasks it received mean human ratings of 6.3/7 and 6.4/7, above the best baseline's 4.9/7 and 5.0/7. Across 33 developments it produced 490,693 lines of sorry-free Lean code and identified 28 errors in published sources.
WhyProposes proof-obligation-driven Lean theory construction with cross-task accumulation; outperforms baselines on 15 tasks and finds errors in published sources, with fairly solid evidence.
In brief"AX Is the New AEO": Paper Finding AI Agents Favor Businesses Whose Own Sites They Can Read
SummaryThe paper argues that agent experience (AX), meaning whether an AI search agent can fetch and read a business's own website, matters more than merely being mentioned when the agent drills down to answer buyer questions. The authors ran 37,927 agent journeys across 1,056 real businesses. Agent-ready businesses had answers built from their own pages 78% of the time, versus 56% for others, and were clearly recommended 1.9 times more often.
In briefVD-DeepStack: Linking Visual Comparison and Language Reasoning for Few-Shot Anomaly Detection
SummaryThis paper proposes VD-DeepStack, a method for few-shot visual anomaly detection that makes language reasoning explicitly depend on dense visual differences between a query image and normal reference images. It combines DINO features with the visual hierarchy of a large vision-language model to strengthen fine-grained representations. Experiments on 4 industrial and 2 medical anomaly benchmarks report substantial gains over baselines that rely on textual comparative reasoning.
In briefProactive Dialogue Policy Optimization via Cognitive-State Transition
SummaryThis paper proposes Cog-Sim, a user simulator that tracks a user's cognitive and affective states and generates feedback through constrained state transitions across turns. It also introduces CSTPO, which represents each action as a high-level strategy label plus an utterance and estimates advantages at both levels using sparse complete-branch sampling. Across three tasks, Cog-Sim shows monotonic dose-response relationships and is preferred over prompt-based simulators for naturalness. CSTPO raises Qwen3-14B's performance to a level comparable with GPT-5.5-based planning methods.
In briefAdjoint Guidance Flow: Lightweight Critic Guidance for Robot Action Policies
SummaryFlow-based Vision-Language-Action (VLA) policies are usually trained by behavior cloning, so they do not explicitly optimize long-term task return. The paper proposes Adjoint Guidance Flow (AGF), which distills trajectory-aware critic guidance into a lightweight guidance network while keeping the pretrained VLA policy intact. At inference, each step needs only one guidance-network forward pass. Across LIBERO, RoboCasa, and LIBERO-Pro, AGF consistently improves pretrained VLAs, and compared with QGF it runs 3.6 times faster per guidance step with 7.0 times fewer parameters.
In briefTask-conditioned routing that picks input resolution per query to make multimodal LLMs faster
SummaryMultimodal large language models slow down sharply on high-resolution images because they produce many visual tokens. This paper proposes Task-Conditioned Resolution Routing (TCRR), a lightweight router that uses the text query to choose the minimal sufficient visual compression level. The authors built a 500k-sample dataset across 12 task categories for training. On Qwen3-VL-8B, TCRR reportedly cuts visual FLOPs by 40.9% and latency by 53.7% while keeping competitive performance, without changing the backbone model.
In briefXMatch: Improving Time Series Forecasting with Tree-Structured Matching of External Variables
SummaryThis paper proposes XMatch, a forecasting model that matches future external-variable patterns with similar historical patterns and uses the associated internal time series patterns to improve predictions. It organizes these historical correspondences into a tree structure (ProtoTree) and adaptively selects how many external variables to use for matching. Experiments on 12 real-world datasets show that XMatch outperforms state-of-the-art baselines.
In briefHow neural language models learn contextual distributions of dependency structures, a statistical view of compositionality
SummaryThis paper examines how neural language models acquire structural meaning encoded by grammar, independent of lexical semantics. It proposes that learned dependency structures become new distributional units for further statistical learning. Using a synthetic grammar and BERT-style masked language models, the authors show models learn contextual distributions that token statistics alone cannot reveal, with dependency relations learned before their contextual features.
In briefTaoTex: Improving Texture Detail Accuracy in Native 3D Material Generation
SummaryThis paper presents TaoTex, a diffusion-based model that generates native 3D materials and aims to reproduce fine textures. The authors build a dataset of high-frequency textured 3D assets to address gaps in public data, and add a multi-level feature fusion module and a latent-to-pixel loss transition. Learnable viewpoint embeddings extend the method to multi-view inputs. Experiments reported in the paper show better texture detail preservation than existing approaches in both single- and multi-view settings.
In briefDecomposing How Language Models Memorize Training Data
SummaryThis paper examines how language models develop memorization of training sequences over the course of training. Using the Pythia model family, the authors decompose the loss trajectories of memorized sequences across training and model parameters, studying both duplicated sequences (recitation) and rare ones (recollection). They find that both forms rely on sequence-level gradient alignment, with recitation suffering forgetting due to misalignment with other training influences, and that lower layers are most involved. The decomposition also predicts memorization better than a cross-entropy baseline, and intervening on a small set of influential parameters can ablate memorization in the final model.
Florian Eichin, Philipp Mondorf, Andrei Mircea et al.
In briefJazzSAMBA: A multi-take jazz band dataset for live music models
SummaryThis paper introduces JazzSAMBA, a multitrack dataset of 76 jazz standards performed by eight musicians on drums, bass, piano, trumpet, and saxophone. It includes both overdubbed (asynchronous) and live ensemble (synchronous) recordings, musician-selected preferred and alternate takes, and timed annotations for bars, chords, sections, and soloists. The authors demonstrate the dataset with a jazz combo source-separation baseline and a chart-conditioned accompaniment ablation. The dataset, code, and samples are available from the project demo page.
In briefPDEU-Bench: A Benchmark for the Personalized Planning Lifecycle of Tool-Calling LLM Agents
SummaryThis paper introduces PDEU-Bench, a benchmark that tests whether tool-using AI agents can define, execute, and update personalized plans while respecting user preferences. It contains 214 long-horizon tasks across 12 everyday domains and 94 tools. Evaluating 15 open-source and closed-source LLMs, the authors find models can apply preferences in single tool calls but struggle with coherent plan construction and updating. Existing personalization and memory-augmentation methods do not reliably keep preferences across the full planning lifecycle.
In briefTraining language models to sample from the distributions they describe, using per-rollout MMD rewards
SummaryLanguage models are often asked to sample from a specified distribution, such as simulating survey respondents or generating synthetic data. The paper observes that instruction-tuned models can state a target distribution correctly yet still fail to sample from it, and that prompting and decoding changes only partly fix this. The authors propose the witness advantage, a per-rollout advantage derived from maximum mean discrepancy, and use it as the reward in GRPO. On unseen target distributions, the method substantially reduces total variation distance to the target while largely preserving general model capabilities.
In briefVoxParity: Testing Whether Voice Agents Act on How a Caller Sounds
SummaryThis paper introduces VoxParity, a benchmark testing whether voice agents change their actions based on audio, not just transcript words. Across 183 scenarios from 14 sectors, the transcript is held fixed while the audio varies, such as a mayday call or a child's voice placing a bet, and the correct tool call is checked. Only 11 of the 23 systems that can also run on transcripts pass the words-only null test, and errors mostly lean toward following the words.
In briefAudit the Scaffold, Not the Checkpoint: Recursive Self-Improvement in Coding Agents
SummaryThe paper proposes a stationarity dichotomy: if an agent's reachable set of edits stays fixed, iterative self-modification yields strictly diminishing returns, and escaping requires that set to expand. Rewriting scaffolding such as tools and verifiers expands reach without changing weights. The authors also show that best-of-k orchestration is capped by the best worker, and that same-family workers have overlapping failures. Measurements on SWE-bench and 401 production sessions show improvement and churn saturating over time.
In briefPredicting drug-target interactions using chemical and protein language models with cross-attention
SummaryThis paper presents a sequence-only model for predicting drug-target interactions, using ChemBERTa on drug SMILES strings and ESM-2 on protein sequences. It applies convolutional layers and bidirectional cross-attention to model interactions between the two, followed by attention-based pooling for binary prediction. The model achieves the best results on BIOSNAP, matches the best AUROC on Davis, and is competitive on BindingDB, with 25.2 million trainable parameters. Ablation and cold-start experiments suggest the approach generalizes to unseen proteins and drugs.
Khadidja Henni, Hamza Abdelali, Abdelkrim Aries et al.
In briefRISE: Stress-testing modern text-to-image models by evolving reusable attack strategies
SummarySuccessful policy violations on production text-to-image systems are rare, and earlier human-written seed prompts are often blocked. The paper argues that existing automated red-teaming methods fail because their success judges are unreliable and their exploration is weak, and proposes RISE, which evolves reusable strategies for generating prompts. On DALL-E 3, Nano Banana 2 and GPT-Image-2, RISE reaches up to 13% human-verified attack success, while prior methods reporting around 30% fall to near zero under the same calibrated evaluation.
Dmitrii Kharlapenko, Sergei Bratchikov, Konstantin Korolev et al.
In briefNeural network framework for simulating avalanche flow depth and speed
SummaryThis study develops a physics-informed neural network (PINN) for the Savage-Hutter depth-averaged granular flow model, verified first against 1D analytical solutions and then validated against 2D laboratory experiments. The coupled 1D formulation learned flow height and velocity without prescribed data, with mean RMSEs of 0.043 and 0.079 in non-dimensional units. In 2D, physics-only training converged to a trivial zero solution, so 10 sparse points from final deposit profiles were added. Across four experimental cases, global height RMSE ranged from 2.7 to 6.7 mm and mean wetted-area IoU from 69% to 81%..
Pradyumn Singh Sikarwar, Vishal Sharma, Gaurav Bhutani
In briefMuon Resolves the Edge of Stability in LLM Pretraining
SummaryThis paper studies Muon, an optimizer increasingly used for language-model pretraining, and finds that the classical edge-of-stability picture for gradient descent does not capture its large-step behavior. For stochastic no-momentum Muon, the authors derive a coherence-corrected conditional loss-neutral boundary, and controlled experiments show loss balance and temporal alignment respond differently to learning rate and batch size. In 130M and 1B LLM runs, update directions stay far from coherent reversal even as training continues to improve.
In briefEMMA: Event-Driven Memory System for Long Video Question Answering
SummaryThis paper addresses the difficulty multimodal large language models have with hours-long videos that exceed their context windows. EMMA segments videos into semantic events using visual drift and spoken transitions rather than fixed windows, then builds a three-level memory of raw video, clips, and events. A multi-agent retrieval process searches this hierarchy from coarse events down to precise clips. The authors report substantial gains on three benchmarks, and code is publicly available on GitHub.
In briefDon't Throw Away the Tail: Reusing Discarded Robot Actions to Speed Up Policies
SummaryRobot policies predict a chunk of future actions but execute only a prefix, discarding the rest before replanning. This paper proposes Action Upcycling, a training-free method that reuses those discarded actions by extending the execution horizon while the action velocity stays smooth. In simulated and real-world manipulation experiments across several Vision-Language-Action models and a World Action Model, it reduced policy calls by 1.2 to 1.7 times without loss in success rate.
In briefSincDPNet: A compact, interpretable model for recognizing bathroom sounds to assist daily living
SummaryThis paper introduces a seven-class dataset of 21,387 annotated bathroom sound clips recorded in five environments. It proposes SincDPNet, a small raw-waveform classifier with learnable sinc filters whose passbands can be read directly in hertz. Recording sessions were separated across training, validation, and test sets to reduce room-specific leakage. The best model reaches 80.2% accuracy and 0.760 macro-F1 on the held-out environment with 14,040 parameters, while a compact version uses 2,848 parameters. Errors are attributed to spectral overlap among water-related events and similar transient structure in some door and walker/crutch sounds.
