In briefBenchmark testing whether multimodal LLMs judge color accurately against CIE-computed ground truth
SummaryThis paper introduces a five-task benchmark built on the CIE colorimetric framework, covering white-balance pair judgment, skin-tone bias direction, correlated color temperature estimation, CIE Delta E 2000 pair ordering, and color naming under changing illuminants. Eight multimodal LLM endpoints were tested. The models often handle coarse binary white-balance judgments but perform weakly on numeric color temperature estimation and some fine color-difference tasks.
In briefError attribution and depth-aware safety evaluation for monocular RGB human tracking in robot navigation
SummaryThis paper studies monocular RGB human tracking used for robot stop-or-proceed safety decisions. The authors prove that a lateral-only distance rule cannot miss dangers flagged by a planar distance rule, but it causes unnecessary stops. Using ETH and UCY pedestrian videos with simulated stationary robots, they apply a two-factor Shapley attribution to separate rule and perception error contributions.
In briefEntity-anchored claim extraction for mining arguments in full-text research papers
SummaryThis paper proposes using scientific entity mentions such as tasks, datasets, and models to identify and structure claims in full-text scholarly articles. A case study on 50 sentiment analysis papers yields a human gold standard and a silver-standard dataset of 1,487 claims, generated by majority voting among large language models. Encoders fine-tuned only on the silver-standard data outperform those trained on the small human-annotated set and transfer zero-shot to NLP papers outside the training corpus.
Bhuvanesh Verma, Andy L\"uking, Mounika Marreddy et al.
In briefCombining graph structure and text meaning to predict links in text-attributed graphs
SummaryThis paper proposes Bridge-LP, a framework for link prediction on text-attributed graphs that unifies graph and language representations. It combines three components: a jumping connection of textual proximity, compressed embeddings that encode neighborhood information, and a cache embedding training strategy that reduces computation. The authors compare against 40 existing models on 9 datasets and report up to a 43.06% improvement in Mean Reciprocal Rank (MRR).
In briefCombining video generation with agent harnesses: agentic video creation, agentic world modeling, and recursive self-improvement
SummaryThis report is a review of how video generation can be combined with agent harnesses, which are executable runtimes that manage context, tools, state, feedback, and budgets. It examines three directions: agentic video generation, where harnesses coordinate scripting, shot planning, production, and editing; agentic world modeling, where harnesses mediate observation, control, and state; and recursive self-improvement in video model development. It also analyzes the agent capabilities involved and discusses open challenges such as autonomous production, lifelong memory, and evaluation.
In briefAdMuse: Skill-Guided Creative Advertising Image Generation
SummaryThis paper presents AdMuse, a multi-agent framework that distills high-quality advertisements into reusable creative skills for cross-domain advertising strategy. At generation time, it routes suitable skills, produces and selects among multiple candidate concepts, and refines the final image with vision-language feedback. Experiments across diverse product and brand positionings show that AdMuse produces visually compelling ads that communicate product positioning.
In briefAnalyzing axial attention in time series foundation models using single location regression
SummaryThis paper extends the Single Location Regression framework to two-dimensional structured data to study sparse settings where a target depends on a few tokens. The authors formulate an analytically tractable axial attention-type predictor and analyze its statistical properties and training dynamics. They prove that, under a specific asymptotic regime and with known oracle parameters, the predictor is Bayes optimal, while certain semi-linear simplifications are not. For a specific temperature scaling, projected gradient descent is shown to converge globally to the optimal parameters, and experiments confirm the importance of inverse temperature scheduling.
In briefFinding and Filling Information Gaps in Multi-Agent Systems
SummaryMulti-step web tasks often need several LLM agents to gather and combine evidence, but a fixed workflow's communication links may prevent needed evidence from reaching the agent that uses it, producing incomplete or wrong answers. The paper proposes SPLICE, which locates a possible gap for each query, adds one directed edge to deliver the relevant evidence, and re-runs the affected nodes to produce a candidate answer. A selector trained on process features then decides whether to adopt the candidate or keep the original answer. On five benchmarks, SPLICE improves web tool-use task score by 8.0% and the mean score on four reasoning and planning benchmarks by 5.7% over the static skeleton, at 1.25 to 1.44 times its cost, and transfers zero-shot to four unseen backbones.
In briefHidden Ranking Disparities: Category-Aware Fairness via Distributional Stability
SummaryThis paper studies fair ranking in recommender systems by modeling ranking policies as distributions over permutations, with user utility measured as the exposure of interacted items relative to an ideal ranking. It introduces utility parity across user groups and a category-aware variant, conditional utility parity, and shows both are equivalent to stability of utility under changes in group composition. Optimizing over mixtures of ranking policies yields a second-order cone program, for which generalization bounds are provided. A numerical study validates the theoretical findings.
Gayane Taturyan, Charlotte Laclau, Stephan Cl\'emen\c{c}on
In briefBeyond CCT: Per-shot camera color correction by recovering illuminant spectra
SummaryCameras usually correct color with pre-calibrated matrices interpolated by correlated color temperature (CCT), which reduces the light source to a single number. This paper instead recovers the illuminant's spectral power distribution using non-negative least squares, a basis dictionary, the camera white point, and a multi-channel ambient light sensor, then fits a 3x3 color correction matrix for each shot. On a synthetic benchmark of 300 illuminants and 3,100 reflectances, mean angular error falls from 1.89 to 1.20 degrees, with about a 1.7x tail reduction for LED-like off-Planckian illuminants.
In briefEventSeeColor: Reconstructing Color Images from Monochrome Event Data
SummaryEvent cameras record brightness changes under difficult lighting, but traditional event-based reconstruction yields only grayscale images. This paper presents EventSeeColor, which uses reconstructed grayscale images and aligned events to jointly guide a diffusion colorizer. In a combined evaluation on two real event datasets, it achieves the best structural similarity, learned perceptual scores, and chromatic-region color difference among the compared external methods.
In briefClosing the Log Gap: Minimax-Optimal Regret for Episodic Reinforcement Learning
SummaryThis paper studies regret minimization in episodic tabular Markov decision processes. The proposed algorithm achieves an expected regret scaling as the square root of H cubed times S times A times K for large K, matching the minimax lower bound without extra polylogarithmic factors. It is a variant of UCBVI that avoids strictly optimistic estimates and uses MOSS-style exploration bonuses, controlling error propagation in an occupancy-weighted manner. The authors also apply the approach to (ε, δ)-PAC best-policy identification, matching the minimax lower bound up to constants in some regimes.
WhyThe authors propose a UCBVI variant with occupancy-weighted error control, reporting a regret bound matching the minimax lower bound without extra log factors.
In briefFisher Contrastive Learning to reduce feature suppression, tested on leukemia diagnosis
SummaryThis paper introduces Fisher Contrastive Learning (FCL), which uses a regularized Fisher trace criterion to counter feature suppression in self-supervised contrastive learning, where dominant features overshadow other class-relevant signals. The authors prove that FCL's learned representations alleviate this effect. FCL is tested on simulated benchmarks and on pretraining with white blood cell microscopy images, then transferred to diagnosing acute promyelocytic leukemia in an external cohort. It outperforms self-supervised baselines, yielding more robust and discriminative representations.
In briefA Negotiation-Aware Hybrid LLM Framework for Intent Interpretation and Service Discovery
SummaryThis paper proposes a framework that interprets natural-language cloud intents, which may be incomplete or conflicting, and discovers matching services. The framework refines the intent and detects conflicts, structures it into a provider-independent TOSCA template, and retrieves candidate offers by deterministic matching against cloud provider catalogs. Conflicts are resolved through iterative natural-language negotiation with the user. On 70 diverse intents, it improved conflict detection and service discovery recall and F1 over single-shot and enhanced-prompting baselines, and eliminated hallucinated offers.
In briefFARE: Factor-Aware Retrieval with Error-Guided Fusion for Retail Forecasting
SummaryThis paper introduces FARE, a method for forecasting retail sales, where promotions and calendar events shift short-term demand. It retrieves historical weeks separately for each known factor of the target week and merges them into a factor reference, then uses two gates to blend this with a shape reference and correct a backbone forecast. Across the Favorita, M5, Rossmann, and Rohlik datasets, FARE reduces the weighted absolute percentage error of five backbone models by 11.46%, 4.06%, 14.63%, and 10.16% on average.
In briefTraining-free online method that breaks video into steps for step grounding
SummaryThis paper proposes Hierarchical Step Decomposition (HSD), a training-free framework for online Video Step Grounding that uses Large Multimodal Models. HSD groups streaming video segments into semantically coherent events and refines uncertain step predictions by decomposing candidate steps into finer substeps. The authors report that HSD outperforms state-of-the-art methods on HT-Step, Ego4D Goal-Step, and CrossTask, with better recognition, more accurate localization, and faster inference.
In briefThe Generative Paradox: An Anti-Prompt Approach to Human Capability
SummaryThe paper argues that generative AI in education is usually judged only by human-AI team output, overlooking whether people can still understand, challenge, and repair work when assistance changes or disappears. It introduces the "generative paradox" and proposes treating retained human capability as a design goal alongside joint performance. The Anti-Prompt paradigm reverses the usual interaction order, so AI first creates conditions for human construction when immediate generation would preempt a consequential human act.
In briefMode Dynamics: A framework linking neural network parameter changes to output behavior
SummaryThe paper introduces Mode Dynamics, a framework that links coordinated parameter motion during neural network training to changes in model outputs. It decomposes recorded parameter updates into collective modes and measures each mode's effect on outputs. Experiments show that a few leading modes can reconstruct behavior, induced effects depend on the training path, and modal complexity varies with task and data.
In briefA deep learning method for predicting demand ranges in online retail when sales data is capped and boosted
SummaryThis paper presents a deep learning framework that learns quantiles from doubly censored observations, where sales are limited by stock and raised by promotions while the target is underlying demand. It also proposes a tailored evaluation metric that correlates well with the oracle calibration measure. The authors provide theoretical guarantees and report that their estimator outperforms existing baselines. Inventory simulations and real-world applications confirm its practical value in retail operations.
In briefNeurEO: Using Neuroscience Methods to Interpret Earth Observation Embeddings
SummaryEarth embeddings compress rich Earth observation data so downstream models can handle many geospatial tasks, but their meaning is unclear. The authors adapt computational neuroscience methods, usually used to interpret neural activity, to analyse AlphaEarth embeddings. They examine embedding features alone and against land cover and elevation, finding correlational and multi-scale structure and specific directional tuning such as a "North vs South woodland edge" response. Removing the main land cover trends reveals finer-grained representational details with higher effective dimensionality.
Thijs L van der Plas, Jacob JW Bakermans, Ioannis N Athanasiadis
In briefDynaRev: A Benchmark for Testing Web Agents When User Goals Change Mid-Task
SummaryDynaRev is a benchmark of 181 tasks for evaluating web agents whose goals are revised during execution, rather than fixed in advance. Revisions are delivered at task-defined semantic milestones and include additive revisions, unconditional overrides, and runtime-conditioned overrides. Evaluation covers correctness, responsiveness, and efficiency using trajectory-based analysis, and tests of five models show agents often satisfy only part of the task requirements.
In briefVideoMemBench: How Well Multimodal Models Remember Across Video Dialogue Turns
SummaryThis paper introduces VideoMemBench, a benchmark testing whether multimodal large language models can retain and reuse information from earlier turns in a multi-turn conversation about one video. It contains 326 videos and 2,476 turns, 1,387 of which depend on earlier turns. Among over 20 evaluated models, the best, Claude Opus 5, scores only 45.7% overall.
In briefEscaping path mirages in offline goal-conditioned RL
SummaryThis paper addresses offline goal-conditioned reinforcement learning in stochastic, long-horizon settings, where agents can overcommit to trajectories that succeeded by chance. The authors propose Branch-aware Graph Planning (BGP), which turns high-variability points into graph nodes and uses goal-conditioned edges with segmented learning to avoid unreliable branches. The paper reports that BGP substantially outperforms prior state-of-the-art methods on diverse OGBench tasks.
In briefDoes system size affect uncertainty measures for machine learning interatomic potentials?
SummaryThis paper examines whether structure-level uncertainty scores from machine learning interatomic potentials can be compared across systems of different sizes. The authors characterize mean and maximum per-atom uncertainty aggregation as functions of atom count. Mean aggregation is size-independent for homogeneous systems but loses sensitivity to defects in large cells, while maximum aggregation grows with size for in-distribution systems and narrows the in-distribution/out-of-distribution gap in line with extreme-value statistics.
Chayaphol Lortaraprasert, Nontawat Charoenphakdee, Wenwen Li et al.
In briefImage-text probes versus large-scale retrieval tests for vision-language models
SummaryThis paper compares two ways of evaluating vision-language models: image-text matching probes (SVO-Probes and SugarCrepe) and large-scale retrieval. The authors judge the top-10 retrieved items for 100 queries from six models using human annotators and an LLM assessor, and vary the source of negatives to compare ranking performance. On SVO, handcrafted negatives are easier to reject than retrieved ones, while SugarCrepe shows the reverse pattern. The authors attribute SVO's low retrieval scores mainly to incomplete relevance labels, and find no association between passing a probe and retrieving a correct item first.
Ivana Be\v{n}ov\'a, Michal Gregor, Mari\'an \v{S}imko
In briefLearning a motion manifold to reduce ambiguity in unsupervised 4D medical image interpolation
SummaryThis paper addresses reconstructing intermediate anatomical states in 4D medical images from temporally sparse endpoint observations, without intermediate supervision. It proposes learning a motion manifold that restricts candidate motion fields while keeping sample-specific variation, using cross-sample spectral basis learning and sample-conditioned Cayley basis reorientation. Experiments on ACDC and 4D-Lung are reported.
In briefSynthFaces: A Balanced Large-Scale Synthetic Face Dataset
SummaryThis paper proposes a pipeline for controllably generating diverse synthetic faces to train and fine-tune generative models, addressing bias and privacy issues in real-world datasets. It extends the Arc2Face diffusion model with a flow matching model operating on identity embeddings. The authors release SynthFaces, a highly balanced synthetic identity-centric dataset, and verify identity separation with a nearest-neighbor non-leakage metric. They also show its use in fine-tuning an adapter-based personalized generation method.
Sergey Karpukhin, Denis Bobkov, Kirill Morozov et al.
In briefSelfThinker: Self-Evolving Visual Reasoning Using Verified Self-Generated Thoughts
SummaryThis paper presents SelfThinker, a framework in which a unified multimodal model uses its own generated visual thoughts to improve image understanding without human-annotated helper images. The model samples candidate helper images, scores them with its frozen understanding pathway, and uses high-scoring thoughts to refine the generator and guide self-distillation. At inference, it reasons in text directly from the original image without generating intermediate images. The authors report that training on small synthetic datasets transfers to natural-image and spatial benchmarks, raising the seven-benchmark average of ThinkMorph-7B from 52.7 to 64.0.
WhyThe authors propose a self-evolving framework using self-generated visual thoughts filtered by conditional PMI, reporting an 11.3-point gain on seven benchmarks.
In briefD-MOD: Diffusion Mode Estimation for Faster MRI Reconstruction
SummaryThis paper introduces D-MOD, a fast algorithm for accelerated MRI reconstruction that estimates the posterior mode rather than returning random posterior samples. It uses a minorize-maximize procedure on a lower bound of the smoothed log-posterior, combining denoised samples from a pretrained diffusion model with measured k-space data via conjugate-gradient solves. On fastMRI brain scans (T1, T2, FLAIR), a single D-MOD reconstruction exceeds the PSNR of 10-sample MMSE baselines and stays within 0.31 dB of the strongest baselines, while being 11 to 73 times faster.
In briefCoupled Guidance: a stable way to steer flow-matching models during generation
SummaryThis paper studies guidance methods for flow-matching models, which steer the generation process. The authors show that directly adapting diffusion-based guidance to flow matching breaks the coupling distribution used in training, weakening control over sampling accuracy. They propose Coupled Guidance, a flow-matching-specific method derived from a geometric perspective that uses the coupling structure of flow-matching training. Experiments on a robotic collision avoidance task, four Adroit dexterous manipulation tasks, and image inpainting, super-resolution, and deblurring on ImageNet-128 and CelebA-HQ-256 show improved sample quality and guidance stability without extra overhead.
Mateo Cl\'emente, L\'eo Maxime Brunswic, Rui Heng Yang et al.
In briefRACE: Reward-Adaptive Co-Evolution of Policy and Scoring Rubric
SummaryThis paper proposes RACE, a framework that co-evolves a reinforcement learning policy and a rubric-based reward for open-ended tasks. It sorts rollout groups into informative, diversity-collapsed, and rubric-failed types, updating the policy only on informative groups and caching the rubric failures. It switches to rubric optimization when rewards persistently saturate and returns when reward improvement plateaus. The authors report that RACE achieves the highest average performance across eight tasks under a unified protocol, outperforming the strongest fixed-alternation baseline and improving rubric discrimination.
In briefRepConTUR: A 3D Multimodal Model Using Radiology Reports for Abdominal Tumor Segmentation
SummaryThis paper presents RepConTUR, a native 3D early-fusion multimodal model that segments abdominal tumors in CT scans while conditioning on the accompanying radiology report. The authors extend AbdomenAtlas 3.0 into a suborgan-level promptable segmentation dataset and generate report-based question-answer pairs for four jointly trained tasks: referring segmentation, report-grounded segmentation, report generation, and visual question answering. On liver, kidney and pancreas, the model reports the highest Dice coefficient, instance-wise precision and detection F1 among text-promptable baselines.
WhyThe authors condition native-3D abdominal tumor segmentation on full radiology reports, jointly training four report-derived tasks; reported metrics exceed baselines on liver, pancreas, and external c
In briefUpcycling language models into domain-specialized experts for multiple fields
SummaryThis paper proposes DS-Upcycling, a method for adapting large language models to multiple specialized domains. It adds experts and a router to a Transformer layer and introduces domain-aware supervision so the router learns to select the appropriate experts. When fine-tuning on each new domain, only the corresponding expert and router are updated while all other parameters stay frozen. Experiments report the highest average performance across adapted domains compared with baselines while retaining general performance, with routing accuracy of 90-98% using just 256 examples per domain.
In briefNeural blind deconvolution for super-resolved scanning microscopy imaging
SummaryThis paper proposes a training-free blind Deep Image Prior framework for image scanning microscopy that jointly estimates the specimen and detector-dependent point spread functions. The PSFs are parametrised through detector geometry and shared excitation and emission PSFs, and a Gaussian pre-initialisation stabilises the optimisation. On experimental tubulin data, the method recovers finer filamentous structures than multi-image deconvolution and DIP with fixed theoretical PSFs.
In briefShape-Aware Channel-Decoupled Mamba Network for Medical Image Segmentation
SummaryThis paper proposes MCDNet, a network for multi-organ segmentation in abdominal CT images. It combines adaptive local structure enhancement using deformable convolutions and channel attention with channel-decoupled long-range state-space modeling. On the Synapse abdominal CT dataset, it achieves an average Dice Similarity Coefficient of 81.61%.
In briefLearning safe robot motion policies from yes-or-no feasibility checks
SummaryThis paper addresses robot foundation models that generate actions with flow matching but do not guarantee compliance with physical limits such as joint limits and collision avoidance. It considers the case where the feasible set cannot be described analytically and is available only through black-box feasibility oracles. The authors propose a differentiable safeguard guided by these oracles, which uses binary feasibility queries to correct unsafe updates in the forward pass. Experiments on obstacle-aware manipulation report 100% constraint satisfaction while preserving task performance.
In briefUncertainty-aware coarse-to-fine segmentation of the main pancreatic duct for width measurement in 3D MRCP
SummaryThis paper presents a deep learning pipeline that automatically segments the main pancreatic duct from 3D MRCP scans and measures its maximum width. A coarse model first locates the ductal region, then a fine model segments the duct, with a five-model cross-validation ensemble used for performance and uncertainty estimation. On a holdout test set of 14 patients with 55 volumes, the best segmentation reached a DSC of 0.634, the width prediction had a mean absolute error of 0.355 mm, and dilated ducts were detected with 75.0% sensitivity and 74.4% specificity.
Zipei Zhang, Jaakko Sahlsten, Mikko Saukkoriipi et al.
In briefAgentic Process Evidence: signed, tamper-evident records of what AI coding agents did
SummaryThe paper argues that AI agents now perform many software development tasks, such as creating pull requests and refactoring code, but their reasoning and tool calls disappear when sessions end, and standard git logs cannot show what they did. The authors introduce Agentic Process Evidence (APE), a signed, tamper-evident, machine-readable attestation that links agent sessions to their outcomes. APE supports Policy-as-Code control and follows the in-toto style attestation framework, requiring human involvement only when needed.
In briefBRACHIO: Multi-centre Imaging Models for Head and Neck Cancer (HECKTOR 2026 Report)
SummaryThis report describes team nucli's submission to the HECKTOR 2026 challenge on head and neck cancer imaging. The tasks cover tumour volume segmentation, TN staging, and recurrence-free survival prediction using multicentric PET/CT scans. The team used self-supervised pre-trained encoders and combined segmentation, staging, and prognosis models. Reported results include a mean Dice of 0.5409 for segmentation, balanced accuracies of 0.3962 (T-stage) and 0.6432 (N-stage), and a test Harrell's C-index of 0.6577 for survival.
Roxanne De Maeyer, Ine Dirks, Yannick Brackenier et al.
In briefLocalizing Visual Concepts in Autoregressive Image Models (LoCo)
SummaryThis paper introduces LoCo, a model-agnostic framework for locating where and when specific visual concepts emerge within next-scale visual autoregressive models, across layer, scale, and position. The authors also present LoCoBench, a dataset spanning 10 categories, and probe models including Infinity, HunyuanImage-3.0, and InfinityStar. They apply the localization method to concept erasure, model personalization, and adversarial concept injection.
WhyThe authors propose a model-agnostic concept localization framework for next-scale autoregressive models, probing layer, scale, and position with three applications.
In briefReinforcement learning when key deployment variables are unobserved, covering policy learning and off-policy evaluation
SummaryThis paper studies reinforcement learning in settings where variables that drive decisions cannot be collected at deployment time, a situation called runtime confounding. Such constraints can arise in healthcare, e-commerce, and robotics due to legal, ethical, or logistical concerns. The authors contribute policy learning algorithms for this setting, semi-parametric efficient off-policy estimation procedures, and diagnostics for understanding the cost of runtime confounding.
Raphael Kim, Michele Santacatterina, Ramin Zabih et al.
In briefOne Vector or Many? Steering Language Models With Clustered Optimal Transport Under Quantization
SummaryThis paper studies methods that edit a language model's hidden states during inference to reduce toxic outputs. It compares the common difference-of-class-means vector approach with clustered optimal-transport maps on Gemma-2-2B and Granite-3.1-2B, including under INT8 and NF4 quantization. A per-token variant reduces toxicity more than the mean-difference baseline in most settings, though results vary across toxicity classifiers.
In briefRethinking Pre-pretraining: Limited Benefits of Synthetic Data as a Separate Training Stage
SummaryThis paper reports that, in compute-matched experiments, synthetic pre-pretraining (PPT) never outperformed pure natural-language training. It tests four synthetic curricula across three training regimes and finds that PPT gains from prior work depend heavily on weight initialization. The authors conclude that synthetic pre-pretraining offers limited benefits as a distinct stage and suggest rethinking how such data is used.
Alexander Stern, Tim Rockt\"aschel, Acyr Locatelli et al.
In briefEPC-RawDC: 3D diffusion model using scanner physics and raw projection data to convert CBCT scans into CT-like images
SummaryThis paper proposes EPC-RawDC, a conditional 3D latent diffusion framework for synthesizing CT images from cone beam CT (CBCT) scans. It adds two constraints derived from the CBCT scanner's forward model: a geometry equivariant projection constraint and raw data consistency using noisy projection measurements. On two clinical cohorts, it reached an SSIM of 0.97, improved PSNR over DDPM by 2.66 dB (SynthRAD2025) and 2.10 dB (JUST), and reduced mean MAE by 20.3% and 19.0%, respectively.
In briefUPPAAi: LLM-assisted modelling, verification, analysis and synthesis for UPPAAL
SummaryUPPAAL is a tool for modelling and analysing real-time and cyber-physical systems, but building models, formalising properties and interpreting counterexamples still require substantial expertise. This paper presents UPPAAi, an extension of UPPAAL in which users construct, explain, debug and specify models through dialogue with a large language model, which proposes and explains while UPPAAL verifies. The authors report seven case studies covering timed automata, stochastic hybrid automata and strategy learning. They found that the LLM was helpful not only in generating solutions but also in proactively suggesting relevant UPPAAL features the user had not asked about.
Peter Dolog, Kim G. Larsen, Marius Miku\v{c}ionis et al.
In briefHybrid AI patterns: thirteen ways to combine language models with rule engines
SummaryThis paper catalogues thirteen design patterns for coupling a language model with an executable rule language, using the format of object-oriented design pattern catalogues. Each pattern is described by its intent, applicability, participants, consequences and implementation. The authors provide a decision guide that leads from a problem to a suitable pattern, based on their measurements, including cases where a model alone performed as well as or better than the hybrid, and they report what each design cost.
In briefAI Cosmos: An agentic infrastructure that supports learning in realistic scenarios
SummaryThis paper presents AI Cosmos, an agentic infrastructure for situated learning built on scenario-based participation and cumulative practice assets. Learners and role-specialized agents work together within learning scenarios, and an asset-memory loop stores practice processes and outcomes for reuse. Two preregistered randomized controlled studies, using the same underlying LLM across conditions, examined legal reasoning and programming skills. AI Cosmos produced greater learning gains than a conventional LLM in legal reasoning and higher mastery of unseen programming problems than an LLM-as-Teacher baseline.
In briefA single measure, the value of information, for cheap memory retrieval and credit assignment in reinforcement learning
SummaryThe paper addresses two memory problems in reinforcement learning: costly attention over the whole memory store and crediting the acquisition of useful memories across long sequences. The authors propose estimating the value of information from archived rollouts using a learned world model. This yields a provable lower bound on the optimal value of the acquiring action and defines the optimal retrieval set. A pilot retrieval system is presented that solves a memory task with reduced context.
Daniel Strauss, Nikolaos Sotirakis, Zhenshan Bing et al.
In briefA Mamba-based U-Net guided by Segment Anything Model context for medical image segmentation
SummaryThis paper proposes SAMBA-UNet, a framework that combines context knowledge from the Segment Anything Model with Mamba's hierarchical feature learning to improve both global representation and local detail in medical image segmentation. It introduces an attention-based Hierarchical Context Enhancement Module and uses LoRA and Adapter fine-tuning to adapt SAM to medical imaging. Experiments on Synapse, ISIC17, and ISIC18 datasets report state-of-the-art performance.
In briefHEMO: A Self-Evolving Framework for Multi-Objective Combinatorial Optimization Operator Design
SummaryThis paper introduces HEMO, a method that uses large language models to automatically design search operators for multi-objective combinatorial optimization. Unlike prior approaches that evolve operators while keeping the discovery process fixed, HEMO treats the design harness itself as a self-evolving target, updating modules such as experience retrieval, operator improvement, parent selection, and search orchestration. A persistent memory preserves candidate outcomes and feedback across updates. Experiments on eight optimization tasks with two evolutionary backbones show it outperforms conventional evolutionary, neural, and LLM-based baselines and transfers effectively to related tasks.
In briefDINOv2 for PET: A Pre-Training Benchmark for Medical PET Imaging Tasks
SummaryThis paper introduces PET-DINOv2-2D, the first DINOv2 model pre-trained on 2D PET images. It is benchmarked against the original DINOv2, trained on natural images, and OmniRad, trained on CT, MRI and ultrasound, using identical lightweight heads. On normal/abnormal classification, lesion segmentation and recurrence-free survival prognosis, it matches or outperforms both, with the largest advantage when few labels are available.
Roxanne De Maeyer, Yannick Brackenier, Vicky De Ridder et al.
In briefFIDE: Fiber-informed differentiable design of experiments for reference-relative dynamics
SummaryThe paper proposes FIDE, a method for choosing measurement designs while keeping an existing dynamical model as the baseline. It completes each candidate design's aggregate constraints by information projection onto a common frozen reference, then ranks adequate designs by the minimum dynamical revision needed to realize the completed paths. Experiments show this preference changes sensor selection, can help or harm downstream predictions depending on baseline reliability, and remains computationally usable for a 32-dimensional configuration law.
In briefGreenscapes: A Method for Optimized Plant Placement in City Scenes
SummaryThis paper presents a method for placing vegetation, from shrubs to large trees, in 3D urban scenes. It uses a vision-language model to generate a top-down vegetation layout from individual lots, which is converted into a 3D scene. An optimization loop using CMA-ES then adjusts plant scales and assets, guided by surrogate models of shade and acoustic exposure and view-dependent perceptual measures.
Patryk Bartkowiak, Marcin Marlewski, Pawe{\l} W\'ojcik et al.
In briefTreating Human-AI Interactions as Structured Data: A Formal Model and Runtime for Explainable AI Collaboration
SummaryThe paper argues that exchanges between people and AI assistants are often treated as transient and should instead be stored as durable, structured data. It introduces HAIDL, a formal language that represents each interaction with its context, request, prompt, response, feedback, provenance, and model metadata. The authors implement HAIDL-X, a JupyterLab middleware that captures and stores these interactions, and add a learning layer that derives reproducible measures of user behavior from them.
In briefEfficient Learning of Mechanism Designs: Reusable Constraint Modules for Limited-Data Prediction
SummaryThe paper introduces Assur-KB, a representation of planar mechanisms built from irreducible modules, boundary interfaces, and solve dependencies while keeping source geometry. Under the same multilayer-perceptron setup, Assur-derived features achieve a geometric-mean 1.8× sample-efficiency advantage over flattened adjacency and position inputs across 50 to 10,000 mechanisms. The authors also report invariance under tested joint relabelings and higher valid-generation rates under dyadic composition.
Jo\~{a}o Pedro Gandarela, Thiago Rios, Stefan Menzel et al.
In briefHow Much a Vision-Language-Action Policy Should Trust Its Reasoning
SummaryThis paper studies reasoning vision-language-action policies, which generate a reasoning trace before choosing an action. It proposes CoC-CFG, which applies classifier-free guidance between the full context and the context without the reasoning span, using a per-step weight to strengthen or weaken the reasoning's influence. The weight choice is framed as a Markov decision process with action values measured by rollouts, optimized by Q-adaptive guidance. Results are reported on the Alpasim driving simulator, the Waymo end-to-end driving benchmark, and a hidden test set with Alpamayo 1.5.
WhyThe authors treat reasoning guidance weights as an MDP with measured action values, and report Q-adaptive guidance beating fixed weights and baselines on Alpasim and Waymo.
In briefImproving Atrial Fibrillation Detection Through Hyperparameter Tuning
SummaryThis study examines how hyperparameter optimization can improve detecting atrial fibrillation with machine learning and deep learning models. It compares Decision Tree, Random Forest, XGBoost, and several neural network architectures, tuned using four HPO methods. Experiments on PhysioNet electrocardiogram datasets used inter-patient and intra-patient splits. The Hyperband-optimized XGBoost performed best, reaching an F1 score of 96.95%.
In briefAn uncertainty-aware physics-informed surrogate model for p-n junction diodes
SummaryThe paper builds an uncertainty-aware physics-informed surrogate for a one-dimensional p-n junction diode, conditioned on device length, doping, and bias. From the device's internal state predicted by the surrogate, it defines quantities of interest such as terminal current, depletion width, peak field, and stored charge. For the symmetric one-dimensional junction, the surrogate achieves a median L1 error of 0.11% on electrostatic potential and 1.5% and 0.61% on electron and hole densities. The authors also present an uncertainty analysis that determines whether surrogate results fall within a user-specified tolerance.
In briefAccumulating Pruning Constraints: New Exact Polynomial-Space Max 2-SAT Algorithms
SummaryThe paper presents deterministic polynomial-space algorithms that solve Max 2-SAT exactly in time roughly 2^((1-γ)n) on several classes of instances. These include formulas with many biased variables, semirandom 2-CNF formulas, and Max-Cut on regular expander graphs and stochastic block model graphs. The method extends DPLL branching by accumulating pruning constraints along each branch, which eliminates nearly all assignments.
WhyThe authors extend DPLL with accumulating pruning constraints, reporting polynomial-space exact algorithms with 2^((1-γ)n) time for dense Max 2-SAT classes including semirandom formulas.
In briefBlind video denoising for vocal tract real-time MRI using a self-supervised ensemble
SummaryThis paper addresses noise in real-time MRI of the vocal tract, where clean reference images cannot be acquired, by proposing a self-supervised denoising framework. It combines spatiotemporal super-resolution learning with pseudo-supervision from a variance-stabilized non-local means filter, and fuses two models to balance noise suppression and preservation of high-frequency articulatory detail. On the USC-75 dataset, the ensemble raises signal-to-noise ratio by 11.9 dB on unseen subjects while retaining more high-frequency content.
Julian Andres Hernandez-Potes, Tom\'as Arias-Vergara, Daiqi Liu et al.
In briefEvidential masked pretraining that aligns confidence with observability for reliable VRU motion prediction
SummaryThe paper introduces eCoMAE, an evidential masked pretraining method for multivariate time series that ties predictive evidence to what is observed. It combines block-wise time-step masking, a Normal-Inverse-Gamma decoder, and three observability-aware mechanisms. On the IMPTC benchmark for VRU intention prediction under leave-one-environment-out weather shift, it raises OOD accuracy from 0.7971 to 0.8068 and lowers OOD AURC from 0.0718 to 0.0648. It also achieves the best fine-tuning MSE on pedestrian trajectory forecasting across IMPTC, Waymo, nuScenes, and inD.
In briefMTG-AD: Dataset and Evaluation for Inline Audio Descriptions of Videos
SummaryThis paper studies how well multimodal large language models generate inline audio descriptions, which must fit into a video's existing silences to help blind and low-vision viewers. The authors introduce MTG-AD, a dataset of 36 hours of videos with over 8,000 descriptions co-created by blind and sighted professionals, and MTG-AD-Eval, a reference-free framework that renders scripts to audio to check their fit with silences. Results show models describe visual events accurately and greatly reduce speech overlap, but still struggle with narrative pacing and flow. Automated metrics correlate with human ratings, and users slightly prefer human-written descriptions while finding model-generated ones very helpful for comprehension.
Subhashini Venugopalan, Jenny Tan, Taylor Roper et al.
In briefWhen to spend privacy: optimal allocation of a fixed privacy budget during private learning
SummaryThis paper studies how to distribute a fixed differential privacy budget across iterations of Differentially Private Stochastic Gradient Descent. The Gaussian noise-variance schedule is formulated as open-loop control under a composed zero-Concentrated Differential Privacy constraint. The authors introduce Privacy-Budget Adjoint Mirror Descent, which optimizes the schedule, and report improved tracking and predictive performance over several allocation rules on the tested problems.
In briefUniRA: a unified method that aligns representations and auto-encodes them to speed up diffusion image generation.
SummaryDiffusion transformers produce high-quality images but train slowly and cost a lot to run. This paper proposes UniRA, which combines Representation Alignment (REPA) and Representation Auto-Encoders (RAE) into one objective, learning a tokenizer whose output serves as both the diffusion target and the alignment anchor. The tokenizer, diffusion backbone, and external representation model are optimized together. The authors report that on ImageNet at 256 and 512 resolution, UniRA matches or beats SD3, REPA, and RAE in FID while using up to 5 times fewer training steps.
In briefElasticFormer: A transformer that adaptively allocates computation to detect objects in gigapixel wide-field images.
SummaryThis paper addresses object detection in High-Resolution Wide (HRW) gigapixel images, where foreground objects are extremely sparse and their numbers vary widely. It proposes ElasticFormer, a sparse backbone whose ElasticSelector module scores windows by variance and predicts foreground proportion to select target-containing windows. A new loss and a three-phase training strategy let the selector work without bounding box annotations. On the PANDA benchmark, ElasticFormer reduces backbone FLOPs by 80% and significantly improves AP50 over fixed-ratio sparse methods.
In briefCan LLMs Ignore What They Know? Measuring Knowledge Leakage in Persona Simulation
SummaryThis paper examines whether language models simulating a persona with no prior knowledge of an item can actually avoid using knowledge that the persona would not have. The authors pair real items with fabricated, surface-matched twins and run a rating experiment across three knowledge-handling conditions and five API endpoints. They find a persistent rating gap favouring real items even when the persona is explicitly stated to have never heard of them, and an unfavourable-item contrast that reverses sign across domains.
In briefBiHashFormer: A Hash-Based Dual-Branch Transformer for Efficient Object Detection in Gigapixel Wide Images
SummaryThis paper addresses object detection in High-Resolution Wide shots, where foreground objects are sparse, images are gigapixel-scale, and foreground and background are intermixed. The proposed BiHashFormer uses an RoI Selector to pick object-containing windows, a compression branch that uses HashAttention to discard half of the key-value pairs, and a compensation branch that uses HashMiner for low-cost hash searches to recover missed objects. On the PANDA gigapixel benchmark, it reduces backbone FLOPs by 37.5% and raises AP50 to 81.0%.
In briefMLLM-GS: Fast novel view synthesis in extreme environments guided by multimodal LLM priors
SummaryThis paper addresses novel view synthesis in degraded scenes, reconstructing 3D scenes from multi-view images affected by various degradations and rendering clear images from unseen viewpoints. The proposed MLLM-GS framework uses a frozen multimodal large language model's visual priors to enhance degraded views, with a lightweight correction network reducing pixel shifts and color distortions. It also optimizes 3DGS rasterization, Gaussian memory layout, and parameter updates to lower reconstruction time and memory consumption.
In briefH²R: Reinforcement learning with hierarchical hint rewards for math reasoning
SummaryThis paper proposes H²R, a framework that uses the sub-question structure of math problems as intermediate signals during reinforcement learning with verifiable rewards. The authors also introduce HiMath, a benchmark of hierarchically structured math problems, to validate the approach. Using hierarchical hints, Qwen2.5-Math-7B's accuracy on HiMath rises from 30.4% to 39.1%, and reaches 45.6% with Oracle Hierarchical Hints.
In briefLow-bit quantization of recurrent state in GDN linear attention, with gradient-aware bit allocation and a pure-write sliding window
SummaryThis paper studies how to quantize the fixed-size recurrent state of linear attention to low bit widths without hurting long-reasoning performance. It proposes Global Gradient-aware Allocation (GGA), which assigns head bit widths under a model-wide budget using quantization residuals and gradient sensitivity to future prediction loss. It also introduces a Pure-Write Sliding Window (PW) that keeps recent writes in FP8 and merges them into the low-bit state after they leave the window. On Qwen3.5-4B/9B across seven evaluations, the combined method improves task averages over STEPQuant@4 and shortens generated outputs.
In briefLearning calibrated confidence so multi-agent systems can judge each other's answers
SummaryThis paper studies multi-agent systems that answer complex web questions, such as multi-hop question answering, by splitting work among specialized agents. It proposes MASCAL, in which each agent learns via reinforcement learning to output a calibrated correctness probability and passes it to the downstream agent at handoff. Across three collaboration architectures, MASCAL improves task accuracy and calibration for both agents and generalizes to deeper chains and different model scales.
In briefSpectral Invariant Fusion and Refinement (SIFR) for Domain Generalized Semantic Segmentation
SummaryThis paper addresses domain generalized semantic segmentation, which requires accurate pixel-level prediction on unseen domains without any target-domain data during training. The authors propose SIFR, a frequency-aware dual-stream framework that handles domain shift during both feature fusion and decoding. Experiments on synthetic-to-real, real-to-real, and adverse-weather benchmarks show consistent gains over previous state-of-the-art methods.
In briefPhysAg3D: Linking 3D generative and agentic models to build detailed, physically interactive 3D assets
SummaryPhysAg3D is a framework that generates interactive 3D assets from a single image, combining a 3D generator with a vision-language agent. The agent infers semantic parts, possible interactions, and physical priors, while the 3D generator produces a textured whole-object mesh. The agent then uses geometric and physical tools to add joints, deformable regions, and surface bindings, revising them based on rendered results within a fixed budget. The framework supports both rigid articulation and compliant deformation.
In briefComparing prompting and token probabilities for testing language models' focus sensitivity
SummaryThis paper compares two ways of assessing language models' linguistic competence: explicit prompting and token-probability methods. It tests semantic focus-sensitivity, such as contrasts between "Lucy only SWAM yesterday" and "Lucy only swam YESTERDAY", across models of various sizes. In some larger models, prompting outperformed token probabilities by a wide margin. The authors attribute this to models rarely producing out-of-the-blue focus marking, and show that token probabilities degrade more than prompting when standard focus markers are replaced with random ones.
In briefTokenizer Tokens Tell Training Tales: Detecting Dataset Membership via Tokenizer Vocabularies
SummaryThis paper examines whether a BPE tokenizer's vocabulary reveals if a candidate dataset was part of its training corpus. It argues that rare strings become tokens only when corpus frequency is high enough, so membership signals should appear in the candidate dataset's rare tokens. The authors propose RTAM-Fusion, which calibrates rare-token mass using shadow tokenizers, and report better AUC and balanced accuracy than Vocabulary Overlap on a tokenizer membership benchmark with fewer shadow tokenizers.
In briefTokenRepel: Generating Varied Image Sets Without Lowering Quality
SummaryThis paper presents an inference-time method that increases the visual variety of image sets produced by text-to-image models from the same prompt. A token repulsion mechanism pushes feature tokens apart across samples, while a token density preservation mechanism keeps updates within high-density regions learned by the model to protect image quality. The approach is tested on three pretrained text-to-image models and reports substantial gains in diversity with comparable image fidelity.
In briefCovFedAvg: Local Low-Loss Curves for Federated Model Selection
SummaryThis paper proposes CovFedAvg, a federated learning method for clients that each observe different parts of a prediction task. Clients keep local models, fit a shared endpoint from aggregated gradients, and select a candidate model using distributed training losses without centralized calibration data. On the participating gas-sensor batch at width 16, macro balanced accuracy rises from 88.88% for FedAvg to 89.84%. The authors state that the results do not establish an advantage of the curve-based approach over equally costly direct optimization.
In briefData-driven random-feature approximation of Koopman generators for nonlinear continuous-time systems
SummaryThis paper proposes a data-driven random-feature method that builds a compact, effective function representation for Galerkin approximations such as gEDMD, targeting Koopman generators of continuous-time dynamical systems. Numerical experiments show the approach is accurate for forecasting and system analysis and faster than iterative optimization. The authors prove convergence of the approximated vector field and resulting trajectories, give a general convergence criterion for the gEDMD semigroup, and present a counterexample showing gEDMD does not generally preserve exponential boundedness.
Ana \v{C}ukarska, Maximilian Gollwitzer, Viola Muning Li et al.
In briefTACFormer: A Transformer model that generates target-specific PROTAC molecules from scratch
SummaryThis paper introduces TACFormer, a transformer-based generative model for de novo design of PROTACs, molecules that direct a target protein to degradation. The model conditions a GPT-style molecular decoder on pretrained ESM-2 representations of protein sequences, so the target shapes generation of the full molecule rather than only the linker. The authors report that the generated PROTACs are highly novel and chemically diverse, with chemical space shifting consistently depending on the target protein.
In briefsDMD: Sinkhorn-based distillation that speeds up diffusion image and video generation
SummaryThis paper proposes sDMD, a method for distilling multi-step diffusion models into generators that sample in only 2 to 4 steps. It links Distribution Matching Distillation to Wasserstein gradient flow and adds an optimal transport term, estimated with Sinkhorn divergence, to guide outlier samples and improve mode coverage. A discriminator adaptively activates this correction. The authors report that on FLUX.1-dev, Qwen-Image-20B and Wan2.1-14B, sDMD outperforms distillation baselines on GenEval, DPG, WISE and VBench and improves sample diversity over DMD2.
In briefKARMA: A training-free method that evolves knowledge modules for industrial advertising agents
SummaryKARMA is a framework that evolves six typed knowledge modules for advertising agents without changing model weights or executable harness code. A trace verifier maps observed errors to candidate modules, and a critic compares successful and failed trajectories to guide an evolver that makes atomic knowledge edits. Edits are kept only when their measured gain reaches a threshold, and the paper reports that average hold-out scores rose from 30.19 to 39.15 across 18 domain-model settings.
In briefWhich regions survive when labels are downsampled? A method that keeps coarse segmentation targets exact
SummaryThe paper presents MatchDown, a method that builds coarse training labels for segmentation networks by matching regions to the grid cells they occupy at minimum cost, so as many regions as possible are kept. On COCO, Cityscapes and BBBC038, MatchDown-P loses exactly the minimum number of regions, while common downsampling methods lose several times more. On nucleus datasets, a model trained on MatchDown targets achieves higher instance accuracy than one trained on strided targets.
In briefTuring LLM post-training: calibrated distillation and efficient mixture-of-experts execution
SummaryThis paper studies post-training for dynamically routed mixture-of-experts models, aiming to combine separately trained specialist models efficiently. It proposes calibrated multi-teacher on-policy distillation together with an optimized execution path. On Turing-20B-A2B, the model reaches a four-domain integration score of 0.9988 and outperforms Qwen3-8B and a shared SFT baseline on four math and code benchmarks. Consolidation step time falls from 398.7 to 139.3 seconds, an observed 2.86x speedup over serial execution.
In briefIncorporating Machine-Others: An Enactive Look at Human-LLM Interaction
SummaryThe paper argues that understanding how people are changed by interacting with large language models requires new interactional concepts, since standard theories of technological impact and social cognition do not fit. Drawing on enactive theory, especially the theory of Linguistic Bodies, the authors describe two ways people incorporate machine-others: as tools that enable skills, and through adopting LLMs' ways of languaging. They introduce the notion of "susceptive sense-making" to describe the extra semiotic labour arising from the low friction of these services.
Elena C. Cuffari, Marta P\'erez-Verdugo, Xabier E. Barandiaran
In briefParametric Gaussian Noise Mixer (PGNM): a unified model for camera noise and film grain synthesis
SummaryThe paper proposes an explicit statistical model fitted directly from paired clean and noisy images, rather than physical image-formation models or neural generators. A normalized convolution correlates a Gaussian field across pixels and color channels, and a lookup maps each Gaussian value to a residual based on clean intensity, with the two components fitted separately. On SIDD camera noise it achieves the lowest average distribution and correlation errors among the evaluated methods, and it remains competitive on simulated film grain. A denoiser trained on its noise comes closest to one trained on real noise, and a 1080p frame renders in 10 ms on GPU.
In briefM-PFT: An Angular-Margin Method for Calibrating Prompt-Tuned Vision-Language Models
SummaryPrompt-tuned vision-language models often become underconfident on base classes and overconfident on novel classes. M-PFT enlarges the angular margin by minimizing the maximum cosine similarity between tunable class embeddings and by pulling each tunable embedding toward its zero-shot counterpart. Experiments across 9 prompt-tuning methods, several backbones, and 15 datasets show lower average calibration error on novel classes, with competitive base-class performance and comparable accuracy.
Shihab Aaqil Ahamed, Mohamed Ansaff, Nang Khin Myat Lay Phyu et al.
In briefOpen-set multi-view seahorse re-identification from underwater photo sequences
SummaryThis paper presents new methods for species classification, sex determination, and open-set re-identification of individual seahorses from image sequences. The models were trained and evaluated on a dataset of 35,914 photographs of 474 seahorses observed in the wild. The authors report that combining information from multiple photographs achieves state-of-the-art performance.
Tristan Fauvel, Sylvie Louisy, Patrick Louisy et al.
In briefProbabilistic meshless PDE solving on manifolds, with convergence analysis
SummaryThis paper extends probabilistic meshless methods, which solve PDEs by conditioning a Gaussian process prior on collocation nodes, from Euclidean domains to compact manifolds. It studies convergence for known manifolds and for implicit manifolds given only as point clouds, proposing a generalized finite difference scheme with a free stencil radius. The authors derive error bounds for the posterior mean and confirm the theory on the circle, sphere and flat torus. On real cortical thickness data, the method matches established discretizations while conditioning on fewer than a tenth of the observations.
Federico Pavesi, No\'emie Jaquier, Antonio Candelieri
In briefAttention, Dynamic Wave and Hybrid Sequence Mixers in Language Models: a study of replacing attention with a recurrent layer
SummaryThe paper proposes a Dynamic Wave recurrent layer using input-dependent exponential damping and cosine phase modulation as a sub-quadratic alternative to attention. It compares three roughly 10.94M-parameter models, a pure Transformer, a pure Wave model, and a hybrid, on TinyStories. The hybrid achieves the best perplexity (5.95), while the efficiency results depend heavily on the attention implementation used as the baseline.
In briefKV cache quantization for diffusion language models
SummaryThis paper studies how to quantize the key-value cache used in diffusion language models during fast inference. Reusing cached states across denoising steps can introduce errors, because bidirectional token representations change as masked tokens are revealed. The authors adapt KIVI and KVQuant and introduce TRACE, which reconstructs feasible reveal trajectories for calibration, and dLLM-KVQ, which keeps selected per-layer K-cache entries in high precision while applying standard low-bit value quantization. On LLaDA and Dream across GSM8K, HumanEval, MATH, and MBPP, dLLM-KVQ achieves the highest average score among quantized methods in the reported 4-bit and 2-bit settings.
In briefDWISE: A scheduler for efficient serving of diffusion LLMs based on denoising workload
SummaryThis paper presents DWISE, a scheduling engine for serving diffusion large language models, which generate text by denoising many tokens in parallel. The authors argue that scheduling policies designed for autoregressive models, such as first-come-first-served and prefill priority, are ill-suited to these models. DWISE schedules requests by predicted remaining work, using a prompt-based predictor refined by hidden states from the first denoising round. Implemented in SGLang, it reduces mean end-to-end latency by 5.4 times on average and up to 14.9 times while preserving task accuracy.
In briefLiDAR-SHIELD: A Dockerized Framework for Testing LiDAR 3D Perception Models Against Adversarial Attacks
SummaryThis paper presents LiDAR-SHIELD, a unified, fully Dockerized framework for assessing how robust LiDAR-based 3D perception models are to adversarial inputs. Instead of installing each target model into the toolkit, it exposes any model through three HTTP endpoints (/info, /forward, /gradient), so connecting a new model or dataset needs only a single loader or adapter. The framework supports both 3D semantic segmentation and 3D object detection, and integrates attacks including PGD, Square Attack, point perturbation, point removal, point injection, and LiDAttack. The codebase is publicly available on GitHub.
Alexandros Gkillas, Ioulia Kapsali, Christos Anagnostopoulos et al.
In briefSymbolic Regression for Predicting Sow Farrowing Time in Precision Livestock Farming
SummaryThis study uses acceleration sensor data from sows around farrowing to detect nest-building behaviour and then predict the remaining time until farrowing. Symbolic regression generates interpretable mathematical formulas for this prediction. It achieved a mean absolute error of 9.28 hours, outperforming closed-box models such as support vector regression and neural networks in both performance and interpretability.
Elisabeth Mayrhuber, Maciej Oczak, Michael Iwersen et al.
In briefSHPRR: A Pure Exploration Algorithm Under Constrained Regret
SummaryThis paper studies best arm identification under both a fixed sampling horizon and a regret budget measured relative to a trusted baseline. The authors theoretically show that the regret cost of exploration decreases across the phases of Sequential Halving, a widely used pure-exploration algorithm. Based on this, they propose SHPRR, which front-loads the regret budget to allow more exploration when many arms remain. Experiments on synthetic and real-world instances show improved best-arm identification over uniform regret allocation and existing baselines.
Georgina Nouli, Flore Sentenac, Maximilian Schiffer
In briefAligning Medical Vision-Language Models at the Capability Boundary with Entropy-Adaptive Rubric Grounding
SummaryThe paper introduces MedCBR, a two-stage method for aligning medical vision-language models so their diagnostic reasoning is linked to visual evidence and clinical knowledge. The first stage builds counterfactual preference pairs at the model's capability boundary and optimizes them with difficulty-weighted preference learning. The second stage distills clinical rubrics from boundary cases and applies them based on prediction entropy within a reinforcement learning framework. The authors report an average accuracy of 61.2% across six medical benchmarks, with the two stages combined yielding larger gains than either alone.
In briefDeep reinforcement learning for selecting RIS configurations in MIMO systems without channel information
SummaryThis paper proposes a deep reinforcement learning framework for blind passive beamforming in RIS-assisted millimeter-wave MIMO communications. A Deep Q-Learning agent learns an adaptive policy to choose near-optimal RIS configurations while sounding only a few codewords and requiring no channel state information. The approach is reported to outperform Q-Learning and conventional baselines in convergence, performance, and overhead reduction.
Ouma\"ima Bounhar, Thomas Ch\^ene, Ghaya Rekaya-Ben Othman et al.
In briefAegisFlow: Training-Free Non-Myopic Path-Safe Guided Flow Matching
SummaryThis paper proposes AegisFlow, a training-free guidance framework that keeps pretrained diffusion and flow matching models within constraints across their entire generative path, such as preserving identity in image editing or avoiding obstacles in robotics. It treats guided generation as a global safety-critical optimal control problem and integrates a time-varying barrier function into the global cost functional. The authors derive a tractable approximation of the Hamilton-Jacobi-Bellman equation to compute this control. Experiments on text-guided image manipulation and robotic planning are reported to show a better balance between target alignment and constraint satisfaction than existing methods.
In briefPossibility-Based Inference: A Deterministic Framework for Reliable Uncertainty Quantification in Decision-Making
SummaryThis paper addresses decision-making under uncertainty in machine learning, especially for safety-critical systems where reliable evaluation of model behavior matters. Drawing on fuzzy set theory, it proposes a possibility-based contrastive grade of prediction as an alternative to latent posterior-based methods for parameter estimation and model selection. Instead of representing uncertainty probabilistically as in Bayesian inference, the approach uses a possibility-based representation that can capture uncertain and conflicting information more flexibly. The framework aims to support data mining and to reduce overly confident but implausible predictions from specialized deep learning models.
In briefEvoTeam: Multi-agent collaboration through evolving teams
SummaryThis paper proposes EvoTeam, a framework for LLM-based multi-agent systems that focuses on collaboration and evolves team organization, interaction rules, and individual agent skills through execution experience. It maintains a library of reusable teams, selecting or creating teams and guiding their collaboration with team-specific interaction rules. Experiments on four benchmarks demonstrate the framework's effectiveness.
In briefGROME: A Framework for Grounding Factorized Memory in Video World Models
SummaryThis paper introduces GROME, a factorized memory framework that helps video world models stay consistent over long time horizons. It separates past information into static scene geometry and dynamic entity states, retrieves spatially aligned history, and injects this memory into a frozen backbone without additional training. Experiments on long-horizon revisit scenarios report better scene persistence and dynamic entity fidelity while preserving visual quality and controllability.
