In briefActionRAG: Graph Diffusion That Refines Evidence for LLM Agent Long-Term Memory
SummaryThis paper proposes ActionRAG, a graph-based memory architecture for LLM agents that re-scores retrieved triple evidence at query time using local graph structure. It is evaluated on MemoryAgentBench (MAB), where it reports an average accuracy of 55.1 across eight deterministic-scoring columns. The authors state this improves over the strongest retrieval-based memory baseline.
In briefWhether hybrid physics-neural generative models can be uniquely identified, and a meta-learning fix
SummaryThis paper provides the first theoretical analysis of identifiability for hybrid deep generative models, which combine physics-based equations with neural generative models. The authors propose meta-learning as a way to build hybrid models with provable identifiability. Experiments on synthetic and real-data benchmarks show that existing hybrid models using unconditional priors are not identifiable, while the proposed meta-learning formulations show strong identifiability.
Yubo Ye, Maryam Toloubidokhti, Sumeet Vadhavkar et al.
In briefAWARE: A Robot Dynamics Prediction and Self-Monitoring Framework for Detecting Anomalies Using Only Cameras
SummaryThis paper presents AWARE, a world-model framework that detects anomalies in robots using only external camera monitoring, estimating system dynamics and predicting states. It was evaluated on 240 real-world robotic crane trajectories, including payload collisions and motor degradation. Under noisy, camera-only state estimation, AWARE reportedly improves anomaly classification by 26% over its strongest baseline. The authors also report that decoding the inferred dynamics parameters localizes the degrading motor with 88% accuracy.
Luke Beddow, George Mavroghenis, Rodrigo Chacon Quesada et al.
In briefStrongly adaptive online learning with time-varying movement costs: a theory paper on regret bounds for sequential decisions with changing switching costs.
SummaryThis paper studies strongly adaptive regret in online convex optimization where movement costs change over time, evaluating learner performance on arbitrary sub-intervals. The authors propose an algorithm with a data-dependent regret bound, obtained via a reduction to sleeping experts with movement costs. The approach is also extended to delayed feedback and time-varying memory settings.
Andrew Jacobsen, Emmanuel Esposito, Hao Qiu et al.
In briefAI agents that mimic expert cartographers to generate hazard maps for public risk communication
SummaryThis paper presents Agentic Hazard Mapper, a multi-agent system that automates hazard-map production by emulating professional cartographic workflows and adapting design to different data, locations, and hazard contexts. A formative study with map experts informed the design, and a user study with map experts and hazard practitioners followed. Generated maps improved spatial understanding and evacuation-planning performance, reduced perceived workload and task completion time, and were generally preferred over corresponding municipal maps.
In briefFusing Two Network Architectures to Detect Coronary Artery Narrowing in X-ray Scans
SummaryThis paper proposes Fusion-YOLOv11-L, a hybrid model that combines a YOLOv11-L detector with a pretrained UNet++ through a channel attention gating module to better locate stenosed regions in coronary X-ray angiography. On the relabeled ARCADE* dataset it reaches an AP@50 of 0.585 and an F1-score of 0.596, improving over YOLOv11-L. Cross-domain testing on an Oxford AMI Study dataset showed performance drops from domain shift, and test-time augmentation raised AP@50 from 0.434 to 0.480.
In briefInternal discrepancy in contrastive learning: generalization bounds and downstream behavior
SummaryThe paper proposes internal discrepancy, a measure of the gap between the mean representation of positive samples and latent cluster centers. It proves a generalization bound for downstream supervised tasks after contrastive learning without extra assumptions. The authors also introduce a relaxed k-means loss to reduce this discrepancy and examine how contrastive learning aids imbalanced learning.
In briefCityLift: Generating Editable 3D Cities from One Image
SummaryCityLift is a framework that creates separate textured building meshes from a single urban image while keeping their spatial layout consistent. It first jointly learns building shapes and positions at a coarse level, then uses generative spatial inheritance and a flow model to refine each building's structure before placing it back into the scene. The authors also built CityLift-5K, a dataset of 4,997 synthetic scenes with masks, meshes, and placement labels. Experiments on this dataset and scene blocks from UrbanScene3D and InstanceBuilding report better building geometry and scene composition.
In briefPlayerEmbed: Learning player behavior fingerprints from raw game inputs
SummaryThis paper introduces PlayerEmbed, a framework that uses deep metric learning to create compact, distinctive player embeddings from raw video game input sequences. A hybrid CNN-LSTM encoder was trained on data from 1,000 players. A nearest-centroid classifier reached balanced accuracy of 0.950 for players seen in training and 0.879 for unseen players. Increasing the number of training players improved generalization to unseen players.
Tobias Schneider, Lorenz Sparrenberg, Priya Tomar et al.
In briefThe Butterfly Defects: How LLM Agents Cooperate in Networks
SummaryThis paper studies whether cooperation between individual large language model agents carries over to group-level cooperation. It uses the Iterated Prisoner's Dilemma, extending from pairwise play to homogeneous and heterogeneous networks of open-weight LLMs. The authors find cooperation is highly sensitive to model mix and small wording changes, so it should not be treated as a stable model-level property.
Maria Sampani, Maria Lymperaiou, Angeliki Dimitriou et al.
In briefOntology-guided knowledge graph built from climate science papers
SummaryThis paper presents an agentic, ontology-guided framework that uses specialized agents to build knowledge graphs from scientific publications. An LLM extracts triples with textual evidence, which then pass a multistage verification to conform to the ontology. Applied to 2,000 climate science publications, the framework yields a knowledge graph with 133,990 nodes and 755,308 relations, and the pipeline code and graph are publicly available.
In briefTransferring Adaptive Cruise Control Settings Across Driving Scenarios Using Flow Matching on Real Driving Data
SummaryThis paper treats tuning Adaptive Cruise Control parameters as a driver-specific problem of moving controller settings from one driving scenario to another. The authors train a flow matching generative model that predicts how a previously unseen driver will behave in a target scenario using only a few demonstrations from a source scenario. The method is compared with a linear parameter-shift baseline, and several flow-matching variants are evaluated, including ones using driving profiles extracted with a variational autoencoder.
Christian Zink, Maik Pfefferkorn, Eric Lenz et al.
In briefDo Text Differences Truly Complement Vision? Calibrated Visual Complements for Visible-Infrared Person Re-ID
SummaryThis paper studies text-enhanced visible-infrared person re-identification, where text helps compensate for weak appearance cues in infrared images. The authors argue that dense textual differences mix useful and redundant information, so direct fusion is unreliable. They propose SARC, which uses sparse representations and interference suppression to build a compact residual, then calibrates it against RGB-infrared visual differences. Experiments on Tri-SYSU-MM01, Tri-RegDB, and Tri-LLCM show consistent gains over the backbone and competitive results with text-assisted methods.
In briefContinual Lasso method that preserves scores under task shift
SummaryThis paper studies continual learning for high-dimensional sparse linear regression, aiming to reduce catastrophic forgetting when past data are unavailable. It shows that recursively reusing an L1-penalized estimator can carry earlier shrinkage into later tasks and accumulate regularization bias. The authors propose SP-CL, which separates sparse estimation from state transmission by preserving the historical squared-loss score. Under squared loss, its cumulative estimator exactly recovers a full-history Lasso minimizer.
In briefCounselPlex: A Stressor-Aware Full-Duplex Counseling Model
SummaryThis paper presents CounselPlex, a full-duplex speech model for counseling that tracks the seeker's underlying stressor rather than only the latest utterance. It updates a short stressor state during seeker silence and injects it into <PAD> positions while listening, and adds RoPE Position Shift Augmentation (RPSA) to broaden positional coverage in training. LLM-judge evaluations report improved counseling quality and stressor grounding over strong baselines.
In briefCalibrating transport agent-based models in one run using online stochastic batch optimization
SummaryThis paper proposes an online calibration framework that estimates agent-based model parameters within a single simulation run, combining black-box optimization with stochastic batch optimization. Using MATSim, each simulation iteration is treated as a stochastic batch so parameters update incrementally as agents make decisions. Applied to utility function parameters for a large-scale Swiss transport scenario, the authors report that mode shares align closely with observed data within a typical simulation timeframe.
In briefSP-Guide: Spatial Priors Improve Drone Navigation Guided by Language
SummaryThe paper presents SP-Guide, a transformer-based framework that helps a UAV reach a target described in language across large outdoor scenes. It uses a Spatial Prior Cross-Modal Encoder, an Instruction Enhanced Decoder, and a Goal Convergence Filter to better align visual and language information. On the CityNav benchmark, SP-Guide achieves higher success rates than representative baselines, with the largest gains in unseen environments.
In briefNatural Language Autoencoders: Identifying Activations Without Accurate Explanations
SummaryThis paper examines Natural Language Autoencoders, which turn a model's internal activations into text and reconstruct them from that text. The authors argue that success is judged only by whether the original activation can be recovered, which tests information sufficiency rather than whether the explanation is descriptively faithful. They conclude that unverified explanations may reintroduce the interpretability problem these tools were meant to solve.
In briefConditional Neural Processes for self-supervised goal discovery in action imitation
SummaryInspired by how infants learn new tasks by imitating others, this paper evaluates whether Conditional Neural Processes can autonomously learn action representations that reflect the underlying goal. The authors show that adding an attention mechanism in the hidden space helps robust goal representations emerge early in observing an action sequence. They suggest the model is promising as a self-supervised general learner for robots.
Marco Gabriele Fedozzi, Yukie Nagai, Francesco Rea et al.
In briefWhen Does More Search Help? Empirical Study of Search Utility in Quality-Aware Machine Translation Decoding
SummaryQuality-aware decoding improves machine translation by generating several candidate translations and selecting the output with a quality-aware decoder. The authors show that extra candidates help different sentences by different amounts, and they measure search utility with total scaling gain, marginal scaling gain, and a sampling saturation point. Sentence-level signals correlate with total gain, while properties of the current candidate pool better indicate marginal gain. These signals can identify sentences that benefit more when additional search is applied selectively.
Javier Garcia Gilabert, Rico Sennrich, Andr\'e F. T. Martins
In briefPaper testing how far screening and generative design of crystals can really go
SummaryThe authors merged 55 crystal databases by structure into one database of 14.6 million crystals and 131.8 million property values, then built CGTNet to predict missing values. They found that known crystals contain many ultra-incompressible candidates lacking a recorded bulk modulus, including ReN2, measured at 428 GPa. At an equal verification budget, the generative approach's hits were mostly osmium-containing and repeated known compositions, so its novelty was overestimated.
WhyIntegrates 55 databases and completes properties, showing screening is underestimated and inverse design overestimated, backed by controlled comparisons and a released model library.
In briefRetiming robot demonstrations using a learned speed prior to remove operator-specific pacing
SummaryRobot demonstrations often vary in speed because the teleoperator, not the task, sets the pace, and policies can inherit this timing. The authors learn a single speed prior from large-scale teleoperation and handheld demonstration data, then use it to retime demonstrations before fine-tuning a vision-language-action policy. On 120 RoboLab tasks, policies fine-tuned on the learned timing succeed more often and show up to half the jerk of baselines. On a real robot, the policy stacks cups 1.72 times faster than the default policy while succeeding in 29 of 30 trials.
WhyLearns a single cross-task speed prior for retiming demonstrations, validated on 120 tasks and a real robot against baselines.
In briefEMG Tokenization: Keeping Movement Information at Lower Bitrate
SummaryThis paper studies compact EMG tokens for wireless electromyography interfaces, used to decode three-dimensional hand pose. Two causal tokenizers trained without pose labels are compared with continuous features using 320-channel data from nine participants. Tokenization left mean error nearly unchanged for one method and raised it by 0.45 mm for the learned representation, and reduced payloads 41-fold relative to raw EMG.
Raul C. S\^impetru, Amin Olamazadeh, Devon Rohlf et al.
In briefWorldCast: World-Consistent Character Video Generation with Large-Angle Camera Control
SummaryThis paper addresses the difficulty existing methods have in following large-angle camera trajectories for cinematic character video generation. It introduces CLACT, a large-scale character-centric dataset and benchmark with accurate camera geometry, and WorldCast, a framework that back-projects the initial frame into a 3D point cloud to provide geometric guidance and coverage masks. A new component, Geometry-Aligned Memory Attention (GAMA), retrieves and injects view-relevant character features to fill in unseen regions. Evaluations across diverse large-angle trajectories show accurate trajectory following while preserving character geometry, appearance, and identity.
In briefA robust offset-free control framework for keeping vision-only self-driving cars in their lane
SummaryThis paper presents a robust offset-free model predictive control framework that keeps vision-only autonomous vehicles in their lane, especially during high-speed cornering on sharp curves. Numerical simulations show the itube-CILQR solver outperforms IPOPT- and OSQP-based solvers, with a mean computation time under 5 ms. The authors also use the framework in simulated experiments to quantify the perceptual uncertainty of a multitask UNet perception model on traffic scenes not seen during training.
In briefDo Language Models Trust Their Own Explanations? A Study of Functional Consistency
SummaryLarge language models classifying text often highlight the words they consider most relevant to their predictions. This paper proposes an experimental framework that isolates, removes, or semantically inverts these highlighted tokens to check whether the model's predictions change consistently. Tests on IMDB and Steam reviews with GPT-4o-mini, Gemma3, Granite8B, and DeepSeek show that explanation reliability varies across model families and scales.
In briefThe Answer Knows Where to Look: Pre-answer zooming for multimodal LLMs under small visual-token budgets
SummaryMultimodal large language models struggle to read fine text in documents and charts when per-image visual token budgets are small. The paper finds that models know where the evidence is during answering, but this location is not readily readable from pre-answer attention. The authors propose a dedicated look token that learns, before answering, how to zoom, without box annotations or external localizers. Experiments on Qwen3-VL across five text-rich benchmarks report gains over baselines.
WhyAnalyzing the localization gap between answer tokens and pre-answer attention yields a novel look-token zooming method, validated across three scales and five benchmarks, with practical relevance for
In briefSemGeo-Gen: Unsupervised Generation of Approximate Cross-Instance Correspondences
SummarySemGeo-Gen is an unsupervised method that automatically generates approximate semantic-geometric correspondences between different 3D object instances, avoiding costly manual annotation. It lifts multi-view DINOv2 features to 3D, aligns instances with continuous piecewise-affine registration, and prunes matches using semantic consistency. The generated correspondences reach 93% PCK@0.10 and 97% PCK@0.15 against KeypointNet annotations, and using them for synthetic pretraining improves state-of-the-art models on SPair-71k by up to 5 PCK@0.10 points after fine-tuning.
In briefLLMs Tackle High-Energy Theory Physics, but Gaps Remain: Quantum Field Theory
SummaryThe paper introduces QFT-Bench, a benchmark of 75 graduate- to research-level quantum field theory problems requiring extensive derivations, each with a closed-form reference answer. Across 39 proprietary and open-source models, the strongest exceed 95% accuracy, though even strong models make calculation, convention, and instruction-following errors, while weaker models often fail on physical concepts. The authors also find nontrivial reasoning unfaithfulness, where models reach correct answers without their reasoning logically supporting them.
In briefWiMoT: Replay-free continual Wi-Fi sensing by transporting feature statistics
SummaryWiMoT stores per-class feature means and covariances from earlier domains instead of raw channel state information samples, so no old data needs to be kept. When the feature representation changes, an affine map applied to a frozen basis updates these stored statistics accordingly. Reported average final accuracies are 94.02%, 89.61%, and 95.16% on 3DO, WiMANS, and Widar3.0, with at most 0.58 points of forgetting. Storing covariances as rank-32 factors reduces memory to 6.4%.
In briefDT-JEPA: Predicting Future Representations to Improve Offline Decision-Making
SummaryThis paper proposes DT-JEPA, an offline pretraining and online fine-tuning framework for reinforcement learning built on Decision Transformer. It learns from reward-free offline trajectories and uses a probabilistic predictor to model the distribution of future representations in latent space. Experiments on D4RL benchmarks show it adapts efficiently across offline RL tasks and outperforms DT-based and other offline RL methods.
In briefSolving Drift-Diffusion Equations for Semiconductor Devices with Physics-Informed Neural Networks
SummaryThis paper applies Physics Informed Neural Networks (PINNs) to solve the drift-diffusion equations used to model semiconductor devices. It tests the method on abrupt PN junctions in one and two dimensions under various bias conditions. The authors report errors comparable to traditional simulation software and also apply the solver to inverse problems, retrieving doping profiles from desired electric fields or potentials under non-zero bias.
Luke Ragan, Matteo G. C. Alasio, Enrico Bellotti et al.
In briefWL-SeqVAE: Generating Graphs as Sequences Using Variational Autoencoders and Weisfeiler-Leman Colors
SummaryThis paper introduces WL-SeqVAE, a sequence-based variational autoencoder for generating graphs. It uses node colors from the Weisfeiler-Leman algorithm as discrete tokens describing local structure, turning each graph into a color sequence that is encoded into a Gaussian latent space and decoded autoregressively. Evaluated on molecular and general graph datasets, it is competitive with established baselines on several local structural and molecular quality metrics.
Iakovos Evdaimon, Giannis Nikolentzos, Achilleas Tsimichodimos et al.
In briefComputational imaging through turbid, bubbly and flowing water close to a telescope
SummaryResearchers photographed displayed images and physical objects through a 25 mm water layer placed 45 mm in front of a telescope, under seven combinations of flow, bubbles and turbidity, with clear-water recordings as reference. Turbidity mainly removed contrast, reducing a Siemens star's contrast by 4.4 to 4.8 times. Averaging eight frames improved structural similarity (SSIM) by 0.032 over a single frame. The paper proposes averaging frames, classifying raw frames directly, and restoring images with networks trained on measured and synthetic pairs.
Alexander Ingold, Sabina D. Menon, Aadhi Umamageswaran et al.
In briefInferSlice: A Telco WAN Reference Architecture for SLO-Driven LLM Inference
SummaryThis paper presents InferSlice, a reference architecture for running large language model inference under service-level objectives on telco infrastructure, combining network slicing, GPU partitioning, elastic serving and observability. It was implemented on the JOINER UK national wide-area network and evaluated across service tiers, elastic scale-out, client-edge bandwidth limits, and parallelism strategies over the WAN. The authors report that pipeline parallelism tolerates near and mid-range cross-site hops, whereas tensor parallelism and prefill/decode disaggregation fail beyond a single subnet hop.
Juan Parra-Ullauri, Anthony Chiu, Hari Madhukumar et al.
In briefResidual Expert Routing for Text-Only Domain Adaptation of Factorized Transducer ASR
SummaryThis paper addresses adapting a factorized transducer speech recognizer to a target domain using only target-domain text, which improves target accuracy but hurts the source domain. The authors propose conservative residual routing, treating the adapted predictor as a residual correction to the base model, with a 16k-parameter gate deciding per prefix how much to admit. On a 40-hour conversational corpus, routing retains nearly all the text-only gain on two held-out domains while keeping source-domain WER essentially unchanged.
Shashi Kumar, Esa\'u Villatoro-Tello, Thibault Ba\~neras-Roux et al.
In briefCan AI Agents Help Robotic Engineers? Lessons from Industry
SummaryThis case study examines how a small and medium-sized company that programs industrial robots used AI agents for multi-step programming tasks. The agent did not raise engineers' productivity, except for a task requiring highly repetitive code, but results improved over time as engineers learned to use it and the agent was adapted to the domain. The authors conclude that AI agents are not plug-and-play, so companies must train engineers, plan for a learning curve, and match agent configurations to the domain.
In briefEditJev: One-Pass Routing for Choosing Image Editing Models
SummaryThis paper studies routing for image editing, which dynamically selects the best editor from a pool of candidate models for each image and instruction. The authors propose EditJev, which predicts per-model preference scores in a single forward pass and combines them with model speed scores to balance quality and latency. On ImgEdit and GEdit-En, EditJev raises aggregate quality over the strongest fixed editor (4.540 to 4.592, and 8.194 to 8.255) while reducing estimated end-to-end latency by 60.8% and 31.7%.
In briefWhy VICReg's standard-deviation hinge is safer than a variance hinge
SummaryVICReg prevents representation collapse with a hinge on each embedding dimension's standard deviation, and its authors argued a variance hinge would fail. This paper tests that claim on ResNet-18 with CIFAR-10 and CIFAR-100, changing only the hinge and sweeping its weight over five orders of magnitude. At the default weight both reach similar linear-probe accuracy, but the standard-deviation hinge's collapse threshold is 35 to 130 times lower, so it keeps training at weights where the variance hinge collapses.
In briefInterpretable Singing Emotion Recognition Using Acoustic and Vocal Technique Features
SummaryThis paper proposes an interpretable framework for recognizing emotion in singing, using a compact set of singing-specific features covering prosody, spectral features, phonetics and vocal technique. The authors use SHAP to explain model decisions and GRAD-CAM to qualitatively assess learned vocal-technique representations. They report that these features reduce cross-singer variability and that vibrato and pharyngeal techniques appear as salient cues of emotional delivery.
Fotini Kiriakopoulou, Vassilis Lyberatos, Spyridon Kantarelis et al.
In briefA new acquisition function for linear-surrogate Bayesian optimization on the unit cube
SummaryThe paper examines the complexity of globally solving acquisition subproblems for Bayesian Optimization over the unit cube, presenting both positive and negative results. Based on this analysis, it proposes a new acquisition function that can be globally optimized in significantly less time than the standard multi-start quasi-Newton approach, while maintaining or improving performance. The method is competitive with Gaussian-Process-based BO in accuracy per evaluation and outperforms it in accuracy per walltime on black-box functions of low to moderate computational cost.
In briefLGMT: A Tool for Testing Whether LLMs Reason Consistently
SummaryLanguage models can answer a logic question correctly yet contradict themselves when the same problem is restated equivalently, which static benchmarks miss. This paper presents LGMT, an open-source tool that uses first-order-logic-based metamorphic relations to generate equivalent follow-up test cases and checks whether entailment predictions agree, without ground-truth labels. Tested on six widely used LLMs with a 3,091-group benchmark, it found inconsistency rates from 20.84% to 29.44%.
In briefCan Experience Transfer Across Protein Pockets? Reusing Molecular Optimization Experience for Agentic Drug Design
SummaryThe paper studies LLM agents that optimize ligands in protein pockets by step-by-step editing and docking, and asks whether probing experience stored in a reusable base transfers to unseen pockets. Analyzing one full run's log on CrossDocked2020 test-set pockets, the authors find that descriptors such as edited site, action, and round separate improving from worsening edits, but this separation is largely explained by molecular size change rather than the target itself. They conclude the experience base is more useful for prioritizing the agent's own docking proposals than for directly prescribing what to build.
In briefEarth embeddings: decomposable land surface records
SummaryThis paper analyzes two widely used annual Earth embedding products, learned vectors that compress a year of satellite observations for each ten-meter pixel. The authors show the vectors contain individual ecological features such as fire, forest clearing and crop type, validated against independent reference data across six continents. The two products differ in what they detect: one finds small, quickly recovering fires, while the other separates crops more strongly.
In briefDriveOSS: A Dataset and Benchmark for Spotting Unknown Objects and Labeling Known Ones in 3D Driving Scenes
SummaryDriveOSS is a real-world multimodal dataset and benchmark designed for 3D anomaly segmentation and open-set semantic segmentation in autonomous driving. It contains 12,661 synchronized, annotated frames from six surround-view RGB cameras, a 128-beam LiDAR, and a GPS/IMU unit, drawn from 52 driving scenes in varied weather and lighting. The data is split into in-distribution scenes with only known classes for training and out-of-distribution scenes with unknown objects for validation and testing. Experiments show that unseen objects remain a major challenge for current 3D perception systems.
Zhiran Yan, Savankumar Bhanderi, Sourabh Lolge et al.
In briefInterpreting Supervised Pairwise Graph Representations: Local Evidence and Confounding
SummaryThis paper examines whether supervised pairwise graph representations, which turn pairwise predictions into sample-specific graphs, truly rely on higher-order topology. In a Transformer error-detection task, these graphs beat attention-based adjacency, but their advantage over model confidence vanishes after controlling for sequence length, a nearly fully recoverable confound. On HCP Emotion data, topology-aware augmentations and a tuned GCN add little beyond a signed node-summary baseline, and the authors propose a Fisher post-hoc corrector with distribution-agnostic guarantees. They conclude that higher-order topology should be credited only after local supervised evidence and nuisance variables are excluded.
Ilya Brekin, Ariela Moldavskaya, Elina Sidorova et al.
In briefHair and skin analysis from close-up photos to guide laser treatment settings
SummaryThis paper presents a system that segments hair and measures skin features in macroscopic dermatological photos, aiming to replace subjective or uncalibrated manual settings for laser treatments. It combines a custom YOLOv9 instance segmentation model with classical image processing, and uses spectrophotometer-based color calibration to estimate phototype, hair density, color, and thickness. On a custom dataset of short, shaved hairs, it reports a mAP@50 of about 0.80.
Alex Navarro Navarro, Marc Arriaga Hidalgo, Carlos Moreno Rubio et al.
In briefAmpe-DB: A Dataset for Analyzing Human Activity and Interaction in the Ghanaian Game Ampe
SummaryThe paper introduces Ampe-DB, a multimodal dataset based on Ampe, a traditional Ghanaian game involving jumping, leg movements and two-person interaction. It contains 637 clips from 15 participants with RGB video, pose data and movement labels. Evaluations under fixed, pair-held-out and participant-held-out protocols show that strong fixed-benchmark scores can hide poor generalization across people and pairings, with pose-based models proving more robust than RGB models.
In briefLLaCO: Continuous diffusion language model with competition-aware semantic scheduling
SummaryThe paper introduces LLaCO, a continuous diffusion framework for text generation that addresses temporal compression caused by embedding norm decay. It proposes Competition Threshold Theory to describe pairwise token competition and Semantic Retention Scheduling to reallocate denoising time. An 8B instantiation reportedly outperforms previous continuous diffusion baselines and rivals discrete diffusion models of comparable scale.
In briefTrajectory Matching: Discrete Flow Matching with Structured Actions
SummaryThis paper proposes trajectory matching, a method that conditions discrete flow models on a full sequence of actions linking start and end states. It supports complex actions such as attaching molecular fragments, inserting nodes with edges, or adding hierarchy children, which cannot be split into independent component edits. The authors prove the objective recovers the marginal flow and subsumes factorized discrete flow matching. On tree and molecular graph generation, it raises MOSES validity from 92.8% to 99.2% and transfers without retraining to substructure extension and linking.
Niklas Kemper, Marten Lienen, David L\"udke et al.
In briefAnalyzing turbulence in both space and time, with neural networks
SummaryThis paper uses time-resolved direct numerical simulations of turbulence, exceeding 28 TB per run, to examine how numerical choices shape the resolved (3+1)-dimensional physics of turbulent flows. It then studies how neural networks represent this recorded turbulent motion. The authors find that spatial and temporal assessments can disagree for both numerical data and learned representations. They urge the machine learning community to treat (3+1)-dimensional turbulence as a four-dimensional phenomenon.
Maximilian Plattner, Fabian Paischer, Johannes Brandstetter
In briefNeural networks that universally approximate symmetric polyconvex energies and envelopes
SummaryThis paper introduces symmetric polyconvex neural networks (SPCNNs) for modeling small-strain elastic energies and approximating symmetric polyconvex envelopes. The authors show that ReLU-based input-convex networks with the required monotonicity restrictions fail to universally approximate the lifted representers, while the new architecture combines convexity with monotonicity. They prove a universal approximation theorem for symmetric polyconvex energies without requiring isotropy, and derive reduced architectures for isotropic cases. Numerical examples cover stress-based regression and approximation of a double-well potential's envelope.
In briefSpeculative Forking: Running Parallel Setup Attempts to Speed Up Repository Environment Setup
SummaryThis paper presents SpecBuild, an LLM-based agent system that automatically sets up the runtime environment needed to run software repositories. It explores several candidate setup branches in parallel, keeps the first committed environment, and also prepares repair branches for predicted command failures. Tested on 334 tasks from 40 repositories across nine languages, it improved success rates by 6.4 to 12.0 points over the strongest baseline and cut average setup time by 22.4% to 41.6% across three backbone models.
In briefChoosing Between Predicting and Generating Scenarios for Decisions Under Uncertainty
SummaryThis paper compares two ways to make decisions under uncertainty using observed covariates: predict-then-optimize (PtO), which learns expected objectives, and generate-then-optimize (GtO), which learns uncertainty and optimizes over generated scenarios. The authors analyze how the learning target affects statistical accuracy and optimization tractability, presenting regret bounds and a tradeoff between smoothness and convexity preservation. Synthetic experiments and a real case study show the preferable approach depends on data availability and problem structure, and the paper offers practical guidance rather than a universal winner.
In briefBatched NN 4D-Var: a faster-learning hybrid model for weather forecasting
SummaryThis paper studies online learning of state-dependent model-error corrections in the ECMWF Integrated Forecasting System, using a neural network embedded in the forecast model and trained through 4D-Var data assimilation. The authors propose a batched variant that processes assimilation windows in random order and averages parameter updates across windows. The batched approach improves out-of-sample forecast skill, while sequential training improves only the training period, and larger batches speed convergence.
Alban Farchi, Marcin Chrust, Massimo Bonavita et al.
In briefTransferring task-specific behavior between Transformers of different sizes without retraining
SummaryThe paper introduces Ariadne, a training-free method that transfers fine-tuned behavior between Transformers with different depths and widths. Instead of moving parameters, it characterizes fine-tuning by the activation changes it induces at matched blocks, aligns these effects with local Procrustes maps, and synthesizes a native target update via closed-form ridge regression. The authors report improved target-model performance across expansion, reduction, and language-model transfer, with substantially lower time and memory cost than previous cross-architecture rebasin approaches.
Filippo Rinaldi, Aniello Panariello, Giacomo Salici et al.
In briefPHYRD: A Living Benchmark for Reproducing and Extending Theoretical Physics Research
SummaryThis paper introduces PHYRD, a continuously updated benchmark for theoretical physics reasoning, containing 63 Reproduce problems and 43 Discover problems. Reproduce problems test derivation of published results, while Discover problems test consequences of modified theoretical conditions. The tasks are built with FormulaForest and include reference solutions and subquestion-level rubrics. Across five models, whole-problem accuracy ranges from 49.21% to 68.25% on Reproduce and from 18.60% to 74.42% on closed-book Discover.
In briefRegional fine-tuning of open camera-trap models for UK wildlife on modest resources
SummaryThe authors fine-tuned the open-weight DeepFaune classifier on 28 UK wildlife classes and paired it with the unmodified MegaDetector v6 for camera-trap image classification. On 20,212 hand-verified MammalWeb images, DeepFaune-UK v1.2 reached 88.8% strict image-level accuracy, above two baselines, and the v1.3 version built on a newer DINOv3-based classifier reached 90.6%. The authors note that the largest like-for-like gain came from a newer base model trained by others, so this approach depends on large actors continuing to release open weights.
In briefUsing pseudo-labels as rewards rather than targets for semi-supervised speech recognition
SummaryThe paper addresses adapting speech recognition to domains with only untranscribed audio, where a teacher model creates pseudo-labels to fine-tune a student. It proposes using these pseudo-labels as reward references for Group Relative Policy Optimization (GRPO) instead of as supervised targets, making training more robust to label noise. In the banking domain, GRPO achieves 29.3% to 56.2% relative WER improvements over supervised fine-tuning on Whisper-generated labels, with the advantage holding across several model sizes, teachers, domains and languages.
Shashi Kumar, Yanis Labrak, Sergio Burdisso et al.
In briefReusing Existing Models Through Feature Distillation for Efficient Multi-Sensor Geospatial AI
SummaryThis paper explores feature distillation to reuse existing representations and reduce the cost of pretraining geospatial foundation models. The authors distill frozen CROMA and TerraMind teacher models into smaller Vision Transformers through patch-level feature matching. They also introduce TEDI, which combines TerraMind for multispectral imagery and DINOv3 for high-resolution RGB inside one shared backbone without requiring co-registered data. The reported results show TEDI-Large reaching 62.51 mIoU on the PANGAEA-9 comparison.
In briefShared Is Not Enough: Predictor Width and Factor Readability in Masked Latent Prediction
SummaryThis paper uses a simulator where all image factors are known to study which information masked latent prediction methods such as I-JEPA retain. Small encoders trained with I-JEPA, SimCLR and BYOL on the same view pairs were probed for a factor both views share. With a predictor half the encoder's width, I-JEPA's median score was about 40% of SimCLR's. Widening the predictor to the encoder's width reversed this ordering at every training seed.
In briefCompressing 360-degree images with an equal-area sphere-based neural codec
SummaryThis paper proposes an overfitted image codec for omnidirectional (360-degree) images, aimed at real-time decoding in virtual reality. It replaces the equirectangular projection with a multi-scale HEALPix representation that samples the sphere with equal areas, reducing redundancy near the poles. The method adds a cross-scale conditional entropy model and a multi-ring spherical convolution. On a public dataset, it reduces BD-rate by 16.3%, 26.5%, and 19.5% over VVC in V-PSNR, V-MS-SSIM, and V-LPIPS, and decodes 2.6 to 3.6 times faster than the evaluated learned autoencoder codecs.
In briefHow LLM agents using external tools hold up when MCP tools fail, and why bigger models don't fix it
SummaryThe paper introduces MCP-Injector, a middleware that inserts realistic faults (permanent, transient, intermittent) into the Model Context Protocol tool-call path, with fault rates calibrated from 57,240 issue and pull-request threads across 1,774 open-source MCP repositories. In experiments with two model sizes, three agent frameworks and three fault conditions, permanent-heavy faults cut task-completion scores from 8.08 to 3.23 out of 10. The 120B model showed no better fault tolerance than the 20B model. The authors argue that capability benchmarks alone can hide operational brittleness and recommend protocol-aware fault injection as an added evaluation dimension.
Antonio Ken Iannillo, Joshua Segun Owotogbe, Roberto Natella et al.
In briefHGWM: A hierarchical graph-guided world model for robots finding objects without prior training
SummaryThis paper proposes HGWM, a method that lets a robot search for a target object in unfamiliar spaces without task-specific training, using an LLM-generated room prior and persistent spatial-semantic memory to choose search directions. On the HM3D and MP3D simulation benchmarks, the authors report higher success and path-efficiency figures than the WMNav baseline, with modest gains. These are author-reported aggregate results, and evaluation is limited to simulation.
In briefExact global minimax risk when a nuisance parameter is unidentified
SummaryThe paper studies squared-error estimation where a nuisance parameter is unidentified on a submodel and affects a target that vanishes there. It gives sufficient conditions under which the largest critical-path minimax risk equals the exact asymptotic minimax risk over the full parameter space. An estimator attaining the leading constant uniformly is constructed. The conditions are verified in two models, yielding exact constants for each.
In briefBeyond Channel Data: Radio Foundation Models Enriched with Configuration Information
SummaryThis paper adds channel configuration (CFG) information to radio foundation model pre-training, producing CFG-enriched embeddings. Under label scarcity, these embeddings act as effective inductive priors for downstream tasks. The authors compare them with raw channel tokens and several baselines on LoS/NLoS classification, timing advance estimation, and beam selection, across three data regimes and in-distribution and out-of-distribution configuration settings. The proposed embeddings consistently outperform or are competitive with the baselines in both seen and unseen configuration scenarios.
Shrihari Vasudevan, H. G. Ranjani, Raaghul Ranganathan et al.
In briefDetecting Active Speakers Using Context and Registered Voices
SummaryThis paper presents ROAM-ASD-C and ROAM-ASD-CR, extensions of ROAM-ASD for detecting active speakers. ROAM-ASD-C reasons jointly over all visible faces, so other faces can supply evidence when the target is ambiguous or partly missing. ROAM-ASD-CR adds speaker representations from enrollment speech, allowing detection even when the speaker is not visible. The authors report state-of-the-art results on five benchmarks and up to 13.3 mAP improvement over ROAM-ASD under 75% target visual removal.
In briefStarRTT: A real-time framework that predicts Starlink satellite round-trip delay without identification gaps
SummaryStarlink-style low Earth orbit satellites switch serving satellites often, causing large round-trip time (RTT) swings. The paper proposes StarRTT, an end-to-end framework that identifies the serving satellite every second and predicts RTT, removing the identification blind zone, with an LSTM-based model called RT-LSTM. In strict real-time back-testing, RT-LSTM achieves a mean absolute error of 1.405 ms, reducing error by 4.3%, 11.6%, and 22.1% versus LSTM, StarNet, and ASTM respectively.
In briefIGG: A Benchmark Testing Whether Agents Can Reveal Hidden GUI Targets Before Clicking
SummaryThis paper introduces Interactive GUI Grounding (IGG), a benchmark where an agent must first make a target visible through interaction, such as scrolling or hovering, and then click it. It contains 601 tasks across 16 GUI environments and seven visibility-constraint sub-types. Humans reach 94.6% accuracy, GPT-5.5 reaches 79.4%, and the strongest evaluated open-source VLM reaches 31.6%. The authors find the bottleneck is executing the reveal interaction rather than recognizing the exposed target, and report that this capability can be strengthened through learning.
In briefInference-time prior adjustment for classifying imbalanced, drifting data streams
SummaryThis paper introduces PAI-KNN, a k-nearest neighbor classifier for data streams that face concept drift and class imbalance. Instead of modifying its memory, it keeps a sliding-window buffer and adjusts its decision rule at inference time using class priors estimated from recent observations. Tests on many synthetic streams show performance competitive with state-of-the-art methods, with particularly strong results under dynamic imbalance and concept drift.
Tobias Callies (1), Eirini Ntoutsi (1), Arthur Zimek (2) ((1) University of the Bundeswehr Munich et al.
In briefVERA-Hand: A Verified Multidimensional Benchmark for 2D Hand Pose Estimation in Dynamic Gesture Videos
SummaryThis paper presents VERA-Hand, a benchmark for 2D hand pose estimation in third-person gesture videos, using manually verified frame-level annotations for 10,690 hands. It evaluates estimators across detection, localization, identification, and temporal stability, and introduces a joint Pose-F1 metric requiring correct detection and accurate localization. Tests of six estimators show no single method dominates, and the benchmark reveals failures that localization-only evaluations miss.
In briefPrincipled Nonlinear Concept Erasure and Steering: A Method That Removes Sensitive Attributes From Representations While Keeping Task Content
SummaryLearned representations can encode sensitive attributes such as gender or race, which may be unwanted. The paper proposes a conditional autoencoder framework for nonlinear concept erasure, aiming to remove only the sensitive attribute and keep everything else. The authors state that, under their assumptions, perfect reconstruction with population-level independence suffices for perfect nonlinear guardedness, with preservation guarantees under imperfect reconstruction. Experiments report a better utility-erasure trade-off across datasets, attributes, and language models without target-task labels. The framework also enables concept conversion (representation steering) by changing the target concept at inference time.
WhyProposes a conditional autoencoder for nonlinear concept erasure with formal guarantees and steering, validated across datasets and language models; impact is mainly within representation erasure.
In briefAnalysing Austrian Parliamentary Speeches by Predicting Party Affiliations
SummaryThe paper uses a fine-tuned BERT model to estimate, for each speech by Austrian National Council members, a probability distribution over parties, aggregated by speaker and party. Parties that govern together become more similar in speech characteristics, while opposition parties move away from governing parties. Members who switch parties usually change their speech characteristics only after the switch. The authors also release a dataset of party affiliation predictions for speeches from 1996 to 2018.
In briefFaster Thompson sampling and information-directed sampling for simulator-based experimental design
SummaryThe paper proposes amortizing Thompson sampling before data collection, so an approximation of the sampling distribution is available as soon as observations arrive. The method also enables efficient estimation of expected information gain about the optimal action, which supports information-directed sampling alongside immediate regret. The approach is demonstrated on two simulator-based design problems in synthetic biology where likelihoods are intractable.
In briefPunjabi Geet Toxicity Dataset (PGT-FG): Fine-Grained Labels for Toxic Content in Punjabi Songs
SummaryThis paper presents PGT-FG, a segment-level dataset for detecting toxicity in Punjabi songs collected from public YouTube videos. It contains 2,396 segments from 392 songs, annotated by four native speakers with binary toxicity, eight fine-grained categories, severity, sentiment, and uncertainty. Of these segments, 29.42% are labelled toxic, with violence being the most common category. The dataset is publicly available on Zenodo, while original lyrics and copyrighted audio or video are not redistributed.
In briefMaking Fast Image Generators Useful for Training Vision Models
SummaryDistilled text-to-image models generate images quickly, but classifiers trained on their images perform worse than those trained on images from slower base models. The authors add a low-rank update to released distilled generators so their outputs better suit training data while keeping fast generation. Across three generator families, the adapted images train better classifiers, with ResNet-50 top-1 accuracy on ImageNet-1k rising substantially.
Patryk Bartkowiak, Dominik Michels, S\"oren Pirk et al.
In briefRedesigning DINO's Data Augmentation for Glaucoma Classification from Eye Images
SummaryThis paper examines how DINO, a self-supervised learning method, performs on retinal fundus images when using its default image augmentations. The authors find these augmentations yield unreliable representations that degrade glaucoma classification. They redesign the augmentation strategy to emphasize diagnostically relevant features, then train a linear classifier on frozen representations. The method reportedly generalizes well across seven benchmark datasets.
In briefQuantum-classical link prediction on dark web hyperlink graphs
SummaryThe authors analyze HTML content and hyperlinks from an offline Tor crawl of 383 onion hosts and 2,067 directed cross-host links. They compare classical graph and feature models with a four-qubit variational quantum classifier for link prediction, using degree-controlled evaluation. Combined graph and content features show some predictive signal, but under source-conditioned degree matching all methods perform close to chance.
Vojt\v{e}ch Nov\'ak, Radovan Re\v{c}ka, Ond\v{r}ej Knebl et al.
In briefStudy of how well anatomical landmarks in laryngeal endoscopy videos can be tracked under NBI versus WLI imaging
SummaryThis paper compares how reliably anatomical landmarks can be visually tracked in laryngeal endoscopy videos recorded with Narrow Band Imaging (NBI) versus conventional White Light Imaging (WLI). Using 61 videos from 39 patients, it evaluates the vocal folds and mucosa with a prominent vascular pattern with a forward-backward error framework across classical and learning-based trackers. NBI consistently reduced tracking error in both regions, with statistical significance for the vocal folds, while failure rates remained low.
Jonas Below, Piermarco Gerli, Anna J\"otten et al.
In briefBAARC: Uncertainty-driven adaptive partial spectral measurement for full reconstruction and material classification
SummaryThis paper addresses spectroscopic instruments with limited wavenumber coverage, which often need precise scanning or repeated measurements. It proposes BAARC, a jointly trained framework that measures small spectral patches step by step, reconstructing full Raman spectra and classifying materials at each step. Model uncertainty selects the next measurement position and determines when to stop. Experiments reportedly achieve high-accuracy identification using only a fraction of spectral channels.
Adithya Ashok Chalain Valapil, Aishwarya Venkataramanan, Joachim Denzler
In briefDRIVE: A Sample-Weighting Method for Fine-Tuning That Balances Forgetting and Adaptation
SummaryThis paper addresses catastrophic forgetting during fine-tuning when the original pre-training data is unavailable. It proposes DRIVE, a framework that assigns per-sample weights through a unified differentiable objective, splitting each sample's contribution into anti-forgetting and adaptation parts. A covariance-based diversity term and a sample-characteristic prior further shape these weights. The authors report that DRIVE offers better stability-plasticity guarantees, faster convergence, and stronger overall performance and robustness than existing weighting schemes.
In briefDistortion-aware spatiotemporal learning for real-time wide-angle video correction
SummaryWide-angle lenses widen the field of view but stretch people near image edges due to nonlinear distortion. The paper proposes FastVC, a full-body video rectification framework that uses a distortion-aware thin-plate spline module and spatiotemporal learning to correct frames while reducing inter-frame shaking. The authors also release the first benchmark dataset for wide-angle full-body video correction. They report state-of-the-art rectification fidelity and visual stability at 57.0 FPS.
In briefRandom Jump Forests: a nonparametric method for multi-state models
SummaryThis paper introduces random jump forests, a nonparametric tool for estimating covariate-dependent cumulative transition rates and occupation probabilities in multi-state models. A single tree partitions the covariate space for all transitions at once, using Nelson–Aalen estimators at terminal nodes, and occupation probabilities are obtained via the product integral without the Markov assumption. Uniform consistency and a functional central limit theorem are established under conditions such as finitely many covariate values, and simulations and four biomedical datasets are used for illustration.
In briefMusicTwins: Training Audio-Language Models to Listen Like a Specific Person
SummaryThis paper introduces MusicTwins, a framework that combines persona summaries, personal memory, and post-training so audio-language models can generate responses specific to one person. Persona representations are built from human writings covering background, views, personality, writing style, and a musical profile. The authors post-train Audio Flamingo Next using supervised fine-tuning and reinforcement learning. Evaluation on ten held-out personas shows notable gains in persona fidelity, including personality, writing style, and responses to music.
In briefK-CLUE: A Segmentation-to-Clustering Framework for Ki67 Cell Typing
SummaryKi67 is a marker of tumor proliferation in breast cancer, and quantifying it from immunohistochemistry images usually requires classifiers trained for each dataset. This paper presents K-CLUE, which segments individual nuclei, encodes them with frozen pretrained encoders, reduces dimensions with PCA, and separates Ki67-positive from negative nuclei using k-means, without annotations or task-specific training. The authors find clustering is highly sensitive to the encoder choice, and whitening the principal components removes this dependence. K-CLUE outperforms DeepLIIF and IHCScoreGAN on the MIST and BCData datasets.
Ana Beatriz Vieira, Sara P. Oliveira, Diana Montezuma et al.
In briefAccordion: A Language Model with Elastic Depth via Fixed-Prefix Layer Looping
SummaryThis paper introduces Accordion, a mixture-of-experts model that supports five inference budgets by looping a fixed prefix of its layers. Pretraining samples one budget per micro-batch, and later supervised fine-tuning and reinforcement learning merge three domain experts into one model. After 3.6T tokens of pretraining, increasing inference depth raised the five-category macro-average across 37 tasks by 3.01 points, with the middle budget capturing 89% of that gain.
In briefL0-norm CycleGAN for unpaired brain CT-to-MR image synthesis
SummaryThis paper proposes replacing the L1-norm cycle-consistency loss in CycleGAN with a smooth, bounded approximation of the L0-norm for unpaired brain CT-to-MR synthesis. On 102 training images and 9 test images, it reduces mean squared error from 0.03464 to 0.02746 (a 20.7% reduction) compared with an otherwise identical L1 baseline. It also raises mean PSNR from 14.77 to 15.80 dB and mean SSIM from 0.5407 to 0.5689, suggesting modifying the cycle-consistency penalty can improve reconstruction accuracy on this dataset.
In briefA multi-endpoint drug-protein activity dataset that distinguishes protein constructs
SummaryThis paper presents Curate-DTI, a dataset built from the 1 July 2026 BindingDB export, with IC50, Ki, Kd and EC50 processed separately. It contains about 2.92 million measurements, 1.29 million standardized ligands and 9,668 sequence-defined protein entities, distinguishing experimental constructs rather than merging by UniProt accession alone. The resource supports endpoint-specific classification and regression, and preprocessing code and validation summaries accompany the release.
In briefState Estimation as In-context Tabular Prediction
SummaryThis paper reframes power system state estimation as an in-context tabular prediction problem, using historical measurement vectors and their state estimates as context to predict the state for a new measurement. It employs tabular foundation models pretrained on synthetic data with frozen weights, so power system data serves only as context. Tested on 14- to 118-bus transmission and distribution networks, the method reports lower errors than weighted least squares under low observability and bad data.
In briefEfficient zero-shot video anomaly detection using adaptive signal-aware fusion
SummaryThis paper presents SAAF, a framework for detecting anomalies in videos without training data while keeping computational cost low. It selects frames using two signals, sudden frame-to-frame changes and drift from a rolling temporal context, and uses a cascaded design that runs costly video-level analysis only when needed. Across four benchmarks, it reports the best AUC-ROC among training-free zero-shot methods, along with substantially higher throughput than the prior best zero-shot method.
In briefPCFBench: How Far Vision-Language Models Can Go in Physics-Aware Photonic Design
SummaryThis paper introduces PCFBench, a benchmark for testing whether vision-language models can perceive, reason about, and design photonic crystal fibers. It includes 26 tasks built from over 220K FDTD simulation samples across 18 fiber families. The models can recognize fiber geometry to a moderate degree but often fail to turn that visual understanding into physics-based design actions, and errors from early perception stages carry into later design steps. The authors conclude that larger models or simple agent chaining are not enough, and that explicit physical grounding is needed.
In briefWhen AI Agents Fail: Is the Model or the Harness to Blame?
SummaryThis paper argues that LLM agents combine a model with an execution harness, so harnesses should be judged by how they interact with models. The authors propose five measures, including Valid Completion Rate, Model-Harness Fit, and Token Load, and evaluate four general-purpose harnesses with four core models across six benchmarks. They report that harness rankings vary by model and benchmark, and that task performance and valid completion can favor different pairings.
In briefCASE: A Frozen-Model Method That Reduces Confusion Between Similar Emotions
SummaryMultimodal large language models often confuse semantically similar emotions when recognizing emotions from multimodal cues. The paper proposes CASE, a framework that keeps a sparse, self-evolving library of pair-specific emotion skills outside the frozen model and retrieves the relevant skill at inference to support boundary-aware reasoning. Experiments on CA-MER, DFEW, and EMER show effectiveness and transferability, with Qwen3-VL-8B's average accuracy on CA-MER rising from 46.80% to 59.13%.
WhyFirst agent-skill framework for emotion recognition, using reusable pair-specific skills validated across datasets and frozen MLLM backbones.
In briefPathJev: A decision-focused vision-language model for histopathology
SummaryThis paper introduces PathJev-P, a vision-language model for histopathology that takes a question and candidate options as input and returns a normalised probability distribution over those candidates in one forward pass. It is built on MedGemma-1.5-4B with LoRA adapters and a decision head, trained on 40,000 patch states from three public collections. PathJev-P achieves higher balanced accuracy (0.541) than a generative model, the CONCH model and a general medical decision model. Its calibration advantage depends on how the generative model's probabilities are read out, and it remains over-confident on unseen organs.
In briefCauProphet: An Agent That Builds Causal Understanding Like Humans
SummaryThis paper introduces CauProphet, a large language model agent that forms and updates causal knowledge rather than only retrieving past records. It combines stored causal knowledge with current evidence to build a structured Causal Belief for decision-making, and revises both as new observations arrive. The authors report a 25.4% improvement over a baseline agent in financial research and an increase in prediction accuracy from 21.3% to 28.6% in dynamic real-world settings.
In briefBeyond Post-Hoc Rewards: Privileged Reasoning Exploration for Multimodal Embeddings (PILOT)
SummaryThis paper presents PILOT, a framework that treats reasoning as a latent variable, implementing a prior mode conditioned only on the input and a posterior mode that also uses training-time relational evidence within one shared multimodal LLM. It derives an importance-weighted variational objective and co-evolves both modes under a shared inference-compatible reward, selectively transferring posterior trajectories to the prior. On the 78-task MMEB-V2 benchmark, PILOT outperforms the primary baseline in all 12 task groups at 2B and 7B scales, raising overall scores by 3.2 and 3.3 points, with the largest gains in visual-document retrieval.
Tianyu Yang, Yijun Tian, Lars Benedikt Kaesberg et al.
In briefFrom Perception to Pixels: Human-Based Distinct Colors for Better Dominant Color Extraction
SummaryThis paper proposes a dominant color extraction framework based on human perception, built on a Human Color Dictionary (HCD) of 64 perceptually distinct colors derived from a perceptual experiment with 110 participants. Integrating this dictionary into RGB histogram and K-means methods improved alignment with human perception on a dataset of 4,562 images, measured by Earth Mover's Distance and best-match distance. The study found no meaningful gender differences in perceived unique colors, and that highly saturated colors are most distinguishable while dark colors are less reliably differentiated.
Uldana Shyndali, Nurdaulet Toleugaliyev, Kuat Zharykbayev et al.
In briefOnline continual learning for edge-cloud IoT systems that adapts models to changing data while limiting forgetting.
SummaryThis paper addresses deployed IoT and edge/cloud machine learning models that must adapt online to non-stationary data streams without access to the full original training data. It proposes an edge-cloud methodology that combines Continual Learning with Out-of-Distribution detection and Concept Drift detection, using ODIN and ECDD to select informative samples for annotation and trigger rehearsal-based model updates. Experiments on automotive, healthcare, and agricultural datasets indicate the approach is applicable and can support adaptation while mitigating catastrophic forgetting.
Charalampos Davalas, Dimitrios Michail, Christos Diou et al.
In briefALLMTS: An LLM-Based System That Runs Clinical Triage Guidelines for Phone Calls
SummaryThe paper presents ALLMTS, a system that decides from a brief intake note whether a caller needs an ambulance, emergency department, appointment, or home care. Instead of asking a large language model to triage directly, it makes the model select and evaluate questions from a commercial telehealth guideline that nurses use. On 566 expert-validated synthetic scenarios it reached 0.95 accuracy versus 0.47 for a raw-LLM baseline. On a 50-scenario open subset it reached 0.98, close to a five-nurse panel's 0.944.
Micha{\l} Remigiusz Janiszewski, Artur Stopa, {\L}ukasz Lenkiewicz et al.
In briefOPTARI: Open-source software for making optoacoustic imaging usable in clinical studies
SummaryThis paper presents OPTARI, an open-source graphical application that helps researchers analyze combined optoacoustic and ultrasound clinical imaging data without programming skills. It supports vendor and open HDF5 formats and offers features such as region-of-interest annotation, reconstruction, chromophore unmixing, and deep-learning tissue segmentation. The authors report that automated muscle segmentation was comparable to expert labeling and that batch processing cut analysis time to under one second per scan. The tool is already being used in ongoing clinical studies.
Moritz Schillinger, Maximilian Bader, Adrian Buehler et al.
