In briefLogic-guided graph diffusion: steering graph generation with first-order logic rules
SummaryGraph diffusion models can produce invalid graphs when data must follow strict structural constraints. This paper encodes domain knowledge as First Order Logic rules and uses classifier guidance to steer the denoising process toward satisfying them, via a fuzzy, differentiable relaxation of the rules. In experiments generating knowledge graphs, adding ontological constraints at inference markedly increased the share of generated graphs that satisfy those constraints.{
Samuel Cognolato, Luciano Serafini, Alessandro Sperduti et al.
In briefA method that boosts anomaly detection on tables using general knowledge from a pre-trained model
SummaryThis paper addresses weakly supervised anomaly detection on tabular data, where only a few labeled anomalies are available. The proposed method fuses universal knowledge from a pre-trained tabular foundation model with domain-specific features learned from the target dataset, using a prototype-conditioned modulation mechanism. Experiments on public datasets report that it outperforms competitive methods when only eight labeled anomalies are used in training.
In briefIRAH: Hierarchical Agent Tree Search for Optimizing Search Pipelines
SummaryThis paper presents IRAH, an agentic framework for tuning information retrieval pipelines used in Retrieval-Augmented Generation under limited evaluation budgets. Its core method, LATS-IR, performs UCT-guided hierarchical search that first explores pipeline structure and then refines numerical hyperparameters, using natural-language reflections to guide changes. Across four benchmarks, warm-started IRAH achieves the highest average held-out nDCG@10 of 0.4816.
Maria Ana Cardei, Aniket Singh, Amir Moghadam et al.
In briefMachine learning method for tracking how conservation laws in dynamical systems hold or break as the region of study expands
SummaryThe paper studies conservation laws in dynamical systems, noting that local invariants generically exist but few persist globally. It proposes an expanding-domain neural deflation framework that counts independent conservation laws over nested domains. The method distinguishes local invariants from persistent global structure and tracks their gradual breakdown as the domain grows.
Jimmie Adriazola, Ioannis G. Kevrekidis, Panayotis G. Kevrekidis et al.
In briefA paired dataset of camera and consumer sonar footage for monitoring fish
SummaryThis paper releases HaCKRS, a dataset pairing Garmin LiveScope consumer forward-looking sonar recordings with top-down RGB video of koi carp. It contains more than 105,000 temporally matched frame pairs, about 3.3 hours of recordings across four range and orientation configurations. The release includes frame-wise metadata, sonar geometry, synchronization data, and planar registration for three configurations. It emphasizes paired observations and calibration rather than fish annotations, aiming to support fish detection, tracking, and counting research.
Mourad Derbedrosian, Michael Hobley, Pietro Perona
In briefStochastic routing and reinforcement fusion for choosing and combining multiple models in multimodal reasoning
SummaryThe paper studies how a system can route a query to one or more models under uncertainty in multimodal settings. It proposes VL-Scout, a vision-language framework that combines Bayesian priors, policy gradient optimization, and a fusion policy that chooses between continuing and answering. Evaluated on MMMU, MMMU-Pro, OKVQA, and RealWorldQA against 11 baselines, the authors report consistent performance gains and reduced premature policy collapse.
In briefMTEdit: a multi-turn image editing dataset built with packed trajectories
SummaryThis paper introduces MTEdit, a dataset of multi-turn image editing trajectories built with a packed trajectory canvas pipeline that generates multiple editing states jointly to reduce visual-quality drift across turns. It contains 500K trajectories with up to 8 turns, averaging 5.38 turns each. The authors also propose MTEdit-Bench, a benchmark for instruction following, consistency, drift resistance, and use of history, and report that the fine-tuned Qwen-Image-MTEdit performs best among evaluated open-source models.
In briefLabel-correlation ensemble for multi-label retinal fundus disease classification, beyond multi-stage pipelines
SummaryThis paper compares three architectures for classifying multiple diseases in retinal fundus images, using a unified 39-label space built from RFMiD 1.0, RFMiD 2.0 and ODIR-5K (12,652 images). It finds that a staged cascade gives no consistent advantage, while modeling label correlations yields the largest single gain. The best configuration blends a flat model with a label-correlation model under test-time augmentation, ranking first on six of eight metrics.
Md Asif Uddin, Jannatul Ferdous, Asif Islam et al.
In briefVocalGrad: Testing Whether Audio Language Models Hear Acoustic Changes
SummaryThis paper introduces VocalGrad, a benchmark that tests whether large audio language models can perceive low-level acoustic variations, framed as a binary task on the direction of temporal change. Human annotators achieve high accuracy, while current models perform near chance. Linear probing reveals two failure modes: some task-relevant information is kept in internal representations but not used by the language model head, while other information becomes less linearly decodable across layers.
In briefItalian lawmaking process: a unified, structured dataset of parliamentary events
SummaryThis paper converts unstructured Italian parliamentary records into ordered, dated procedural events using an LLM-based extraction method validated against manual annotations. The dataset reconstructs the procedural histories of all 5,346 Italian laws enacted from 1985 to 2026, covering parliamentary terms IX to XIX, with 101,413 events in total. It is intended to support computational legal and legislative studies.
Matilde Contestabile, Chiara Ferrara, Giovanni Parrillo et al.
In briefLagNet: Lag-Structured Operators for Long-Term Time Series Forecasting
SummaryThis paper studies the operator that maps a past history window to a future horizon in long-term time series forecasting. The authors observe recurring diagonal patterns in sensitivity matrices that correspond to relative temporal lags, and build LagLayer, a Toeplitz-based layer, which is stacked into LagNet. Experiments on long-term forecasting benchmarks show competitive results with strong baselines, and limited-data tests indicate LagLayer is a sample-efficient alternative to linear layers.
In briefImproving Gender Inference for Chinese Names Using Hanzi-Level Information
SummaryThis paper develops gender inference models for Chinese names, using regression and classification models trained on hanzi names rated by native speakers. For Pinyin names, it proposes an aggregation pipeline and an ambiguity score that flags names whose gender cannot be resolved. Compared with Namsor, GPT, and Claude, the models show the highest agreement with human annotations. Applied to Chinese-named authors in APS journals, the models assign a gender to every name, while existing Pinyin-based methods leave 20% to 100% unclassified.
Lisette Esp\'in-Noboa, Mingrong She, Ana Maria Jaramillo et al.
In briefSynthetic face images to boost recognition of rare genetic syndromes
SummaryDeep facial phenotyping can help recognize genetic syndromes, but patient photos are scarce and unevenly spread across disorders. This study adds disease-specific synthetic facial images from the Phenotypic Disease Image Database to part of the GestaltMatcher Database and tests two models at several augmentation levels. Augmentation raised mean recognition accuracy in both setups, though gains were modest for one model and some disorders became less accurate.
In briefEscaping Parameter Space: Tight Generalization Bounds via Representation Quality
SummaryThis paper derives a tight generalization bound for deterministic neural networks using their last hidden layer representations. It evaluates representation quality with a parameter-free k-nearest-neighbor classifier, avoiding traditional parameter-space measures. The bound decomposes into three empirically measurable parts (local class geometry, decoder-network agreement, and training stability) and serves as a diagnostic tool for generalization.
WhyUses a parameter-free k-NN to certify generalization via representations, yielding non-vacuous CIFAR-100 bounds and causal evidence that class geometry drives generalization.
Niclas G\"oring, Shuofeng Zhang, Branton DeMoss et al.
In briefECG-SDI: Single-lead ECG for Continuous Sleep Depth Profiling and Sleep Health Biomarkers
SummaryThis paper presents ECG-SDI, a framework that uses a single-lead electrocardiogram to continuously track sleep depth through the night and extract digital biomarkers. A deep learning model trained on 12,392 polysomnographic recordings showed correlations of 0.828 and 0.848 with conventional sleep stages in internal and external test sets. The authors also identified a disturbed sleep subtype linked to poorer subjective sleep quality and higher all-cause mortality, and observed sleep-depth changes among participants after CPAP treatment. The findings support ECG-based sleep phenotyping and further evaluation of home CPAP response monitoring..
In briefBeyond LLMs: A Mental World Model for Simulating and Interpreting Mental States
SummaryThis paper introduces the Mental World Model (MWM) and its implementation, the Multi-Branch Mental Evolution Architecture (MMEA), which builds subjective worlds and simulates how mental states evolve under past experience. MMEA reports 96.48% accuracy on DynToM and 94.00% on RecToM, outperforming leading LLMs and specialized Theory of Mind methods. The authors present it as an interpretable framework for continuous mental modeling.
In briefRecognizing Study Lineage Is Not Enough: Over-Crediting Same-Trial Evidence in LLM Medical Checks
SummaryThis paper examines whether LLMs that correctly recognize two publications come from the same trial still double-count that trial when aggregating evidence. The authors evaluate 14 source-audited clinical families under five model conditions, using separate calls for recognition and aggregation. They report a gap between identifying shared lineage and aggregating evidence, and find that a study-grouping instruction reduces over-crediting for one model.
Mukand Krishna, Anoushka Harit, Chenghao Xiao et al.
In briefNeural Expansion: A Unified Mechanism for How Deep Networks Generalize
SummaryThis paper proposes Neural Expansion, a geometric framework that explains deep network generalization through the growing coverage of directionally selective rays in input space. It introduces Generalization Intervals to measure, for each direction, the region around training samples where predictions stay stable. Experiments across architectures and datasets show ray coverage expands during training, especially along directions with low Jacobian singular values, which the authors use to explain correctly classified, adversarial, and mislabeled or out-of-distribution inputs.
Chashi Mahiul Islam, Samuel Jacob Chacko, Mao Nishino et al.
In briefThe Cognitive Cost of LLM Writing: Reading-Time Comparison with Human Text
SummaryThis paper compares human-written articles with their Large Language Model rewrites by estimating how long readers would take to process them, a common proxy for cognitive effort. The authors train RoBERTa-RT on human eye-tracking data to predict word-level reading times and apply it to 496 paired texts. LLM rewrites are shorter and have slightly lower total predicted reading time but higher predicted time per word.
Kamer Cekini, Elisabetta Biondi, Chiara Boldrini et al.
In briefGLOBE: A Framework for Evolving Structured Agent Memory to Optimize E-commerce Listings for Generative Search
SummaryThis paper introduces GLOBE, an agentic framework that lets language agents continually improve product listings for generative search. It stores optimization knowledge as explicit, field-scoped behaviors in an evolving handbook, with an Editor applying behaviors, a Reflector updating them from rank-based feedback, and a Moderator and Curator managing redundancy and commits. Experiments report that structured behaviors outperform monolithic prompt optimization, and the learned behaviors remain effective on unseen product categories.
In briefTool Recommendation for Multi-Step Agents: A system that helps AI agents find all the tools a task needs.
SummaryThe paper presents a deployed system that gives an agent access to about 8,200 tools through three fixed interface tools, running at the gateway between agent and tools under the user's credentials and consent. On the public ToolRet benchmark, it places every tool a query needs in the top ten 57% more often than BM25. In paired multi-step workflows, its first recommendation covered all required tools in 93% of Codex sessions, and the total bill fell in all 14 runs, by up to 40%.
In briefFedRAPP: Federated test-time prompt tuning for vision-language models with reliability-aware personalization
SummaryThis paper studies how distributed clients can collaboratively tune soft prompts for vision-language models like CLIP on unlabeled test data, without access to source training data. The authors find that standard federated aggregation struggles when client data distributions differ widely. They propose FedRAPP, which combines reliability-aware client assessment, a server-side exponential moving-average prompt memory, and similarity-guided peer aggregation. Experiments on nine CLIP benchmarks and ImageNet variants show it outperforms standard aggregation baselines under heterogeneous client distributions.
In briefBlitz: Speeding up the diffusion modules of AlphaFold3 and Boltz-2 through few-step distillation
SummaryThis paper presents Blitz, a distillation framework that cuts the roughly 200 denoising steps of diffusion-based protein structure predictors to just eight. It combines a geometric noise schedule, multistep consistency training, and inference-time physical steering with target-free candidate selection. The method is applied to AlphaFold 3 and Boltz-2 and compared with their teachers and other few-step methods under a common evaluation protocol. The authors report that AF3-Blitz keeps teacher-level structural accuracy and exceeds its teacher on several accuracy metrics.
WhyCombines geometric noise scheduling with multistep consistency distillation on AF3 and Boltz-2, compared against teachers and baselines under a unified protocol; substantial practical impact.
Vinay Saji Mathew, Soundar Kumara, Gretta Kellogg et al.
In briefVISTA: Steering LLMs Using Values Inferred from a Few Survey Answers
SummaryThe paper introduces VISTA, which infers a latent value profile from a user's full or partial survey answers and maps it through a hypernetwork to a user-specific steering intervention in a frozen LLM. The authors report that VISTA matches or beats in-context profiles on synthetic preferences, the World Values Survey (WVS), and PrimeX without adding prompt tokens. On WVS, it reportedly stops after 6% of the questionnaire while retaining 94% of the full-history gain.
In briefLogicVal: A Logic-Based Benchmark for Checking Constraints in AI-Generated Images
SummaryThis paper treats text-to-image evaluation as a constraint-satisfaction problem and defines a logical framework covering objects, attributes, relations, and the constraints generated images must meet. Four state-of-the-art models were tested across 9 constraint classes, with the best model reaching 98% accuracy on single-object constraints but only 53% on complex multi-object relations. The authors also introduce LogicVal, which combines neural visual detection with logical reasoning, achieving ROC-AUC scores of 0.88 to 0.98 on CLEVR.
Savitha Sam Abraham, Marjan Alirezaie, Hossein Rahnama
In briefEquivalence testing and how many runs to use in stochastic ML experiments
SummaryClaims that two training methods perform equally often rely on non-significant differences from only two or three runs per method. The paper proposes a protocol and power theory for certifying practical equivalence in per-run metrics such as AUROC, using a two-level seed-by-item variance model and two one-sided tests (TOST). A closed-form formula converts a margin and target power into the number of runs per method. Simulations and two case studies test the formula against real seed variation.
In briefFewer Comments, More Discussion: Multi-Perspective Collaborative Review for Automated Research
SummaryThis paper addresses limitations of automated review in AI-generated research, where single-model feedback or simple aggregation fails to capture expert roles and collective deliberation. It proposes RoundtableReview, a multi-agent framework that splits each paper into eight semantic slots and assigns six academic and industry reviewers with distinct perspectives and retrieval sources. Through independent assessment, cross-examination, roundtable discussion, weighted voting, and chair synthesis, it produces traceable feedback. Experiments on two automated research backbones report improved paper quality and better results than baselines in automatic and human evaluations, with fewer comments.
In briefBeyond the Benchmark: How Qwen3-TTS Is Conditioned, Where It Works, and How It Fails
SummaryThis paper evaluates the two zero-shot voice-cloning interfaces of Qwen3-TTS, in-context learning (ICL) and X-Vector, against Chatterbox, XTTS-v2 and F5-TTS on read, meeting, emotional and air-traffic-control speech. Neither interface is uniformly better: ICL clones timbre more closely but degrades when a reference transcript is wrong, while X-Vector clones timing. Domain explains four times more variance than system choice, and an off-the-shelf spoofing countermeasure flags Qwen3-TTS clones no more often than real recordings.
WhySystematically compares two Qwen3-TTS cloning interfaces across domains and analyses failure modes with eighteen measures and listening tests, offering useful insight for TTS research.
S\'everin Baroudi, Yanis Labrak, Petr Motlicek et al.
In briefPhysRepair: Agentic Video Generation That Diagnoses and Repairs Physics Errors
SummaryThe paper proposes PhysRepair, a generate-diagnose-repair framework for video generation that fixes physical inconsistencies after the video is synthesized. It uses spatiotemporal causal constraints to locate the first invalid physical state and its affected states, then regenerates only those parts while keeping the rest. Experience-augmented in-context learning draws on earlier repairs to adapt decisions. Experiments on Physics-IQ Verified across five physics domains report improved physical plausibility over existing methods.
In briefCan AI agents reliably set priorities for healthcare scheduling decisions?
SummaryThe paper studies whether teams of LLM agents, each representing a stakeholder, can supply the trade-off weights needed for healthcare operations optimization. It proposes a framework to assess the reliability of these generated preferences and to measure how their uncertainty affects decisions. Explicit ratings varied with clinic conditions and roles, though agent deliberation could erase those differences, while pairwise choices depended largely on presentation order. Robust optimization often selected different schedules than average weights did, so the authors argue such preferences need auditing before use.
Marco Boresta, Tommaso Giovannelli, Sepehr Namirad et al.
In briefTriRouter: Robust EEG Motor Imagery Decoding via Electrode-Space Routing
SummaryThis paper addresses unreliable motor imagery brain-computer interface decoding when EEG electrode montages are sparse or corrupted. It proposes TriRouter, a framework with discriminative-preservation, mask-robustness, and latent-compensation paths, where a validation-calibrated router allocates path contributions based on observable channel availability. On the BNCI2014-001 dataset, TriRouter outperforms standard backbones and a corrupted-EEG baseline, improving accuracy by 3.68, 4.02, and 6.46 percentage points at nominal 25%, 50%, and 75% missing settings.
In briefPCDFusion: Bayesian rendering method for fusing infrared and visible images
SummaryThis paper studies infrared-visible image fusion, which combines thermal and visible observations for human perception and downstream vision tasks. The authors reformulate fusion as local Bayesian posterior rendering and propose PCDFusion, a visible-anchored framework that implements this through three rendering stages. Experiments on public benchmarks report improved fusion quality, target recovery, and downstream detection and segmentation under challenging illumination.
In briefDeep learning method for detecting beetle damage on bananas
SummaryThis study presents a lightweight YOLOv8n-based framework that locates individual bananas and classifies their visible surface as healthy or unhealthy. It uses a custom dataset of 2,473 images with 5,827 annotated banana instances collected from Bangladeshi local markets, split 80/20 for training and testing. On the test set the model reached an F1-score of 77.67% and mAP@0.50 of 87.04%, with a checkpoint size of 5.96 MB, supporting it as a compact baseline.
Md. Mehedi Hasan Naeem, Ishtiyak Ahmmad Araf, Md. Arefin Haque Mahir et al.
In briefFlexPart: Text-Guided 3D Part Segmentation at Adjustable Granularity
SummaryA 3D model can be split into parts in several valid ways, and this paper presents FlexPart, which uses text prompts to control both the segmentation granularity and which named parts are extracted. The authors build a large multi-granularity dataset with an automated vision-language model pipeline and train with randomized background dropping so the model follows semantic conditions rather than geometry alone. The paper reports higher mIoU than the strongest baseline, including a 15% gain when that baseline receives ground-truth guidance.
In briefA Language-Informed Loss for Visual Generation
SummaryThis paper asks whether language-based supervision can give a richer training signal than pixel loss for image generation. The authors introduce LILIE, an invertible image encoder trained so a vision-language model can recover captions from noised features. They use the encoder to build a language-informed loss for text-to-image generation. On the GPIC dataset, JiT models trained with this loss reduce gFD from 66.84 to 53.01 compared with a pixel-only baseline.
Suchir Agarwal, Keshigeyan Chandrasegaran, Ziyu Chen et al.
In briefIn Human Memories: A Study of AI, Creativity and Ethics in Composition
SummaryThis paper examines how artificial intelligence is used in contemporary composition, through two works: Alexander Schubert's Convergence (2020/21) and Francesco Vitucci's In Human Memories (2024). It traces how composers have historically adopted new technologies to expand expressive possibilities. The authors argue that AI now acts as a participant that poses questions to the composer, and they outline an ethical path for treating technology as a creative tool.
Francesco Vitucci, Giuseppe Silvi, Francesco Abbrescia et al.
In briefJAXtari: A Fast, Modifiable Atari Environment for Reinforcement Learning
SummaryThis paper introduces JAXtari, an open-source, GPU-based reimplementation of Atari 2600 environments. Using massive GPU parallelization, PPO agents can complete 100M environment steps on a single A100 GPU in under an hour. It also offers object-centric observations and a flexible system for modifying environments, supporting neuro-symbolic research and testing for shortcut behavior.
In briefScalable machine learning potential with electrical response for long-range material simulation
SummaryThis paper presents DPA4C-LR, a machine-learning interatomic potential that separates long-range electrostatic forces and Born effective charges into a global electrostatics calculation plus local derivatives of learned latent charges, preserving spatial decomposability. The framework scales to 5.2 million atoms on four GPUs at about twice the cost of its short-range backbone. The authors apply it to infrared spectra of water, polarization hysteresis in a relaxor ferroelectric, ionic conductivity in an amorphous solid electrolyte, and interfacial thermal transport across Si/diamond interfaces.
WhyAnalytically separates long-range electrostatics from local derivatives, scaling equivariant MLIPs to 5.2M atoms and validating on water spectra and relaxor hysteresis.
In briefDiversity Without Labels: Task-Agnostic Coreset Selection for Perception and Action
SummaryThe paper introduces CoreDNA, an unsupervised, task-agnostic method for coreset selection that pools neuron activations from vision foundation models into per-image and per-dataset distributions. It selects high-quality samples that concentrate in the tails of the last layer's activation statistics. Across image classification, semantic segmentation, depth estimation, and imitation learning for driving and mobile manipulation, CoreDNA matches or outperforms state-of-the-art methods.
Julia Hindel, Benjamin Ramtoula, Kyuhwan Yeon et al.
In briefNo-Regret Language Generation: A Generator Whose Mistakes Grow Sublinearly
SummaryThis paper introduces no-regret generation, requiring a generator to make a sublinear number of mistakes across all demonstration sequences and compatible target languages, with an agnostic variant for incompatible demonstrations. It defines tree-based mistake-forcing dimensions that exactly characterize finite-horizon minimax complexity and gives necessary and sufficient conditions for no-regret generation. For language classes with finite Littlestone dimension d, it proves realizable mistake bounds of O(d log(1+T/d)), which are optimal up to universal constants, and agnostic regret bounds of O(√(dT log(1+T/d))).
WhyProvides exact characterizations of no-regret generation and optimal mistake bounds via Littlestone dimension, a strong theoretical contribution with mainly theoretical impact.
In briefDeep LLM layers reveal passage-level context in silent-reading EEG
SummaryThe authors trained EEG decoders on every layer of Llama-3.1-8B, using 19-channel EEG recorded during silent reading. Word-level decodability peaks in early layers and declines toward the last layer, while the context-tracking component rises with depth. Even a single 0.5-second EEG window tracks the text passage. Controls indicate this context tracking mostly follows the text rather than session-specific signatures or target geometry.
In briefPresenceBench: An Exploratory Test of Whether Language Models Adapt to User State
SummaryThis paper introduces PresenceBench, a tool for measuring whether an AI's intelligence layer adapts appropriately to a user's functional state, such as readiness, flow, and stability, while preserving task quality and safety. An exploratory study compared four models with and without a synthetic state signal, plus a mismatched-state control, using human preference judgments. Supplying a state signal gave preference estimates above 0.5 for all four models, but specificity to the correct state was not established.
Pe\~na Arroyo Gallego, Mauro Ruiz Moreno, Alejandro Acien et al.
In briefRethinking multi-task active learning: gradient conflict and orthogonal decomposition
SummaryExisting multi-task active learning methods usually combine objectives such as uncertainty, diversity, and gradient conflict linearly, which the authors argue cannot capture Pareto-optimal trade-offs. The paper proposes not scoring conflict during acquisition and instead resolving it with gradient surgery during training, splitting acquisition into informativeness filtering followed by coverage maximization, and introducing Cross-Task Mutual Information computed from MC-dropout task entropies. The authors report consistent improvements over recent methods on COCO, Cityscapes, and NYUv2.
In briefAutomatic labeling of pedestrians and cyclists in low light using thermal-RGB stereo cameras
SummaryThis paper presents an automated pipeline for labeling vulnerable road users in low-light conditions, using paired thermal (LWIR) and RGB images. Labels come from the thermal images, since people stand out from the background due to body temperature. The method uses a LWIR-RGB stereo setup instead of a complex alignment rig, which lowers hardware cost and eases reproduction. Models trained on the generated labels outperformed off-the-shelf baselines and models fine-tuned on a hand-labeled subset on a manually labeled test set, for two LWIR detectors.
Dimitrios Bouzoulas, Pyry J\"appinen, Pejman Habibiroudkenar et al.
In briefGRAFT-EEG: Physics-Based Transfer of Brain-Signal Decoding Across Datasets
SummaryThis paper addresses cross-dataset EEG decoding, where changes in electrode montage, reference, and active sensors alter the measurement setup itself. The method transfers motor-imagery electrode masks through a template motor leadfield and removes the common-reference gauge, using a classifier fitted only on source data with no target signals, labels, or statistics. On five leave-one-dataset-out targets (253 participants), it reached 60.865% balanced accuracy versus 52.684% for a strict-clean BIOT baseline. Claims are limited to five binary motor-imagery domains and the tested operators.
WhyModels montage and reference changes as observation-operator shifts with physical field transport, validated by leave-one-dataset-out tests and multiple controls, though limited to binary motor imager
In briefThetreStageAI: a system that drafts structured Bengali theatre production plans using retrieved references
SummaryThis paper presents ThetreStageAI, a retrieval-augmented framework that generates structured production plans for Bengali theatre. It keeps separate retrieval views for scenes, actor blocking, and lighting, embeds Bengali text with BGE-M3, and passes retrieved references with a brief to Gemini, which outputs schema-constrained JSON checked for structure and meaning. The authors also built a 500-record synthetic dataset expanded to 1,500 retrieval documents and compared single-combined and multi-view retrieval under matched settings. Source code, dataset, and evaluation materials are publicly available.
Md. Mehedi Hasan Naeem, Mst. Kamrunnahar Ruma, Shakila Sultana et al.
In briefGEMS: Identifying Molecular Structures by Learned Search Over Isomer Graphs
SummaryThis paper addresses identifying small-molecule structures from molecular formulas and tandem mass spectra, a key problem in metabolomics. It introduces GEMS, which refines initial candidate structures through learned 2-switch bond edits that keep the molecular formula and atom valences intact. GEMS uses a pretrained spectrum-to-fingerprint model and a neural policy to score possible edits. The authors report Hit@1 and Hit@10 accuracy exceeding published baselines on NPLIB1 and MassSpecGym, with much shorter runtime than their FRIGID comparison.
Utku Umur Acikalin, Yingheng Wang, Goncalo J. Gouveia et al.
In briefLUMINR: Faster Medical Image Registration with Local Intensity-Conditioned Meta-Learned Neural Representations
SummaryThis paper proposes LUMINR, a meta-learning framework that speeds up deformable medical image registration. It uses an implicit neural representation that maps a coordinate, plus local intensity and gradient features, to a deformation vector, so conditioning varies across the domain. On 3D brain MRI registration from the OASIS dataset, fine-tuning from the learned initialization matches the accuracy of training from scratch in about a tenth of the gradient steps and wall-clock time, with smooth deformation fields and negligible folding.
In briefSign Code: Hidden Binary Structure Within Deep Networks
SummaryThe paper defines the "sign code" as the set of positive or negative signs of all neuron outputs in a trained network. For overparameterized vision classifiers, a nearest-neighbor classifier using only this binary code achieves accuracy comparable to the full model. Replacing all magnitudes with random noise barely changes residual network performance, whereas flipping 20% of signs reduces accuracy to chance. The authors suggest discrete sign codes may be a more natural unit of analysis than continuous weight space.
Niclas G\"oring, Shuofeng Zhang, Roi Holtzman et al.
In briefBeyond Action Prediction: Learning Search-Free Planning in Latent Space
SummaryThis paper proposes HIJP (Hindsight Inverse JEPA Planning), a framework that replaces costly online search over action sequences in latent world models with a goal-conditioned inverse model that predicts actions directly from current and goal observations. During training, future observations along offline trajectories are relabeled as goals, and predicted actions are kept consistent with the latent predictor's next-state representations. The abstract reports 86.8% overall success across four visual control tasks and a 1.9 to 16.5 times planning speedup over search-based planning.
In briefTrajectory-triggered sparse collaborative decoding: calling a larger model only at key moments to improve accuracy.
SummaryThe paper studies sparse collaborative decoding, where a draft model generates most tokens and a target model is called only for short corrective segments. The authors find that local uncertainty signals such as entropy and log-likelihood predict failures near chance, while aggregate trajectory-shape features of the draft's confidence path reach an AUC of 0.86. Using online learned triggers, the method improves over the best open-loop schedule by 5.3 to 20.0 accuracy points on five GSM8K draft-target pairs, with smaller gains on MATH-500 and StrategyQA that depend on model competence.
Amit Ranjan Trivedi, Sruthikesh Surineni, Kyong Ho Lee et al.
In briefSIT-Eval: A Situation-Based Turn-by-Turn Framework for Testing Dialogue Agents
SummaryThis paper introduces SIT-Eval, a framework for evaluating conversational agents by defining desired behavior as situational rules and uniformly sampling situations at each dialogue turn. An LLM judge, validated against human-verified synthetic dialogues, scores replies using the applicable rules. Benchmarking eight LLMs on 1,292 turns shows that models with similar aggregate scores differ widely in how well they follow individual rules. All models performed worst on ambiguity clarification and Socratic questioning.</summary>
Ravil Mussabayev, Saveliy Zavertyaev, Dmitriy Parpulov et al.
In briefHow many objects, and how certain? Class competition as a training-free confidence signal for density counting
SummaryMulti-class density estimators can count objects in crowded scenes, but their outputs do not show how reliably density is assigned to each class. The paper proposes a training-free confidence proxy based on local competition between the strongest class channels, requiring only predicted density maps. Tested on DSACA predictions for the Hicks et al. dataset, attenuating contested density at inference reduced overall MAE from 0.920 to 0.846, suggesting class competition carries information about count reliability.
Villanelle O'Reilly, Georgios Leontidis, Marc Hanheide
In briefNot every mention is a claim: rethinking object hallucination detection in vision-language models
SummaryObject hallucination detectors usually treat every object named in a caption as a claim that the object is in the image, but some mentions, such as "a bench for people to sit on", name objects without asserting them. The authors build NAME-Bench, with 25,497 object mentions from captions that five vision-language models generated for COCO images, labelled as concrete, hallucinated, or non-assertive. All evaluated detectors flag non-assertive mentions as hallucinations, and adding them as training negatives largely fixes this on LLaVA-1.5 but needs five to ten times more data on four newer models.
In briefADA 2025: A Multimodal Dataset for Detecting Water-Deficit Stress in Sugar Beets
SummaryThis paper publicly releases ADA 2025, a dataset of over 500 GB of young sugar beet plants grown under healthy and water-deficit conditions. It combines thermal, hyperspectral, RGB, and depth imagery with environmental data and metadata. Relative water content and leaf thickness serve as ground truth. Exemplary results show that differences between stressed and healthy plants can be detected in the data.
In briefWhen Magnitude Suffices: Phase-Free Microwave Signal Reconstruction on a Chest Phantom
SummaryThe paper asks whether a magnitude-only spectrum analyzer with a tracking generator can stand in for a vector network analyzer when reconstructing time-domain-gated microwave S-parameters. The authors train a fully connected network on one calibrated configuration and test on a torso phantom with eleven dielectric states, each held out in turn. For S11, reconstruction from spectrum analyzer input matches that from vector network analyzer input, and for S21 it preserves the ordering of states. Explicitly predicting phase performs much worse.
Maciej Cicho\'n, Patrycja Go\l{}a\'s, Maksymilian Bartosik et al.
In briefTeaching Models to Tell Scene Structures Apart Makes Self-Supervised Depth Learning Scalable
SummaryThis paper addresses the problem that self-supervised monocular depth estimation often degrades when trained on diverse scenes, because models learn dataset-specific structural biases instead of transferable geometry. The authors propose MaSS, which uses a structural representation pathway decoupled from contextual features and guided by a scene-structure discrimination objective. Trained on four datasets, MaSS substantially outperforms prior self-supervised methods and is the first to consistently benefit from mixed-domain training, with strong zero-shot results on eight additional datasets.
WhyProposes a decoupled structural pathway with a scene-structure discrimination objective, validated on four datasets plus eight zero-shot sets, first to benefit consistently from mixed-domain training.
In briefCognitive Differentiation for Self-Improving LLM Agents
SummaryThe paper notes that existing skill-evolution methods for LLM agents organize knowledge around whole skills and overlook functional differences within them. Inspired by cognitive science, it proposes CDSE, which splits a unified skill into four coordinated components (Meta, Experience, Tool, Script) and updates them from successful and failed trajectories. CDSE achieves the highest scores in all 36 tested settings across six benchmarks, three agent models, and two execution harnesses, improving over the Base agent by 7.6 percentage points.
In briefDual-channel retrieval with RL-guided time windows for long-term agent memory
SummaryThe paper proposes TIDER, a time-aware retrieval framework for long-horizon LLM agents, in which a reinforcement-learning-trained locator predicts temporal windows where the evidence lies. These windows define a temporal retrieval channel whose candidates are merged with semantic results and reranked by proximity to the windows. The authors report that a lightweight 1.7B locator reaches 61.52% accuracy on LifeBench (+8.82%) and 78.96% on LoCoMo (+3.66%).
In briefWhat Remains Without Test Labels? Re-evaluating a Graph-Level Anomaly Detection Benchmark
SummaryThis paper audits the UB-GOLD graph-level anomaly detection benchmark and later work that reuses its results, finding that test labels were often used to select checkpoints and hyperparameters. The authors re-run 21 detectors under selection rules with and without test labels and show that choosing checkpoints by test AUROC inflates reported performance. They recommend treating results selected without test labels as the main comparisons and releasing per-run outputs for verification.
In briefNew methods for finding boundaries in high-dimensional point clouds
SummaryThis paper introduces two methods for detecting boundaries in high-dimensional point cloud data with uneven density, high dimensionality, and noise. The first fits a separating hyperplane on locally reduced neighborhoods and works best when intrinsic dimension can be reliably estimated; the second measures directional uniformity of local neighborhoods and performs best overall on synthetic data. Both produce pointwise boundary scores and estimated normal vectors. The methods are validated on synthetic data, spatial transcriptomics data for tumor boundary detection, and LLM activation geometry.
Abigail Hickok, Valerio Galanti, Osama Al Dalahmah et al.
In briefSemPurify: Semantics-Aware Data Cleaning to Defend Against Hidden Backdoors in Black-box Models
SummaryThis paper addresses backdoor attacks, where hidden triggers in training data cause a model to behave maliciously on specific inputs. The authors propose SemPurify, a data purification framework that uses a text-to-image model guided by multi-level semantics, including low-level embeddings and high-level image tags, to restore images after purification. Experiments on classification models, CLIP-style encoders, and generative models show that it removes backdoors while preserving semantic information useful for downstream tasks.
In briefGliomaINK: Tumor- and Anatomy-Conditioned Flow Matching for Synthesizing Missing 3D Brain MRI Sequences in Glioma
SummaryThis paper introduces GliomaINK, a method that uses anatomy-guided conditional flow matching to synthesize missing 3D brain MRI sequences for glioma patients when scans are incomplete. It was trained on 1,083 patients from the BraTS-AG dataset and independently tested on an in-house clinical cohort. The authors report the highest PSNR across all six MRI translation directions, along with results on tumor segmentation and IDH1 mutation-status prediction.
Moinak Bhattacharya, Annie Singh, Mohammad Chaudhry et al.
In briefImproving retrieval-augmented medical visual question answering with gated multimodal query representations
SummaryMedical vision-language models answer medical visual questions but can sometimes give factually inconsistent responses. This paper proposes a multimodal retrieval-augmented generation framework that uses a Gated Multimodal Unit to learn a joint representation of medical images and clinical questions for retrieving external clinical evidence, while keeping the base model frozen. On benchmark datasets, the approach consistently outperforms a reference medical vision-language model and remains competitive with state-of-the-art methods.
Daniela Jimenez-Pano, Hugo Jair Escalante, Delia Iraz\'u Hern\'andez-Far\'ias
In briefOccupational Gender Bias Across LLM Training Stages: How Post-Training Changes Gendered Narratives
SummaryThis paper examines how occupational gender bias changes across the training stages of OLMo 3, a fully open language model family, by having each checkpoint write narratives about synthetic personas. When gender is stated, post-training increases semantic diversity more for male-coded than female-coded profiles. When the model must infer gender, its assumptions skew female relative to US labor-force demographics, and each post-training stage raises the odds of a female label. The authors find the supervised fine-tuning data has only a 7.7% female-token surplus, which is too small to explain these downstream effects.
WhyTracks occupational gender bias across OLMo 3's post-training stages with corpus provenance analysis, offering moderately novel, well-supported findings useful to the subfield.
Manthan Gadhia, Alexander Hoyle, Karolina Sta\'nczak
In briefMARS: A Framework That Jointly Chooses Memory and Model for Long-Running AI Assistants
SummaryThe paper proposes MARS, a framework that uses session state to coordinate how an LLM assistant builds its memory and which model answers each turn. A Query and Topic Manager estimates decision-relevant state that guides a memory policy and an online router. On RealMem, MARS keeps 98.2% of full-context GPT-4o's QA performance while using 14.5% of its generation API cost, including controller calls. On MultiChallenge it trails the strongest static system overall but comes closer on a mechanism-focused subset at lower cost.
In briefCLEAR: Cooperative latent flow for accurate region-preserving visual counterfactual explanations
SummaryThis paper introduces CLEAR, a latent-space Flow Matching method for visual counterfactual explanations that changes a classifier's prediction while keeping images realistic and unrelated content intact. It separates classifier guidance into cooperative and conflicting parts using the flow's velocity, keeping the cooperative part and attenuating the conflicting part. CLEAR also identifies editable regions and directly preserves factual content outside them to reduce VAE-induced changes. The authors report that across five bidirectional tasks on two datasets, it matches or exceeds strong image-space baselines with a 2.5x speedup over the strongest baseline.
