In briefAdapting LLMs to Bashkir, a Low-Resource Language: Corpus, Datasets and Benchmark
SummaryThis paper addresses Bashkir, a Cyrillic-written agglutinative language poorly represented in modern LLMs, and releases a web corpus of 71,567 documents and labelled text classification datasets. It benchmarks eight models from 82M to 8B parameters under full fine-tuning, QLoRA, in-context learning and continual pretraining. QLoRA on 8B-scale models outperformed encoder-only fine-tuning on binary classification, and YandexGPT-5-Lite-8B performed best on multiclass classification. Continual pretraining effects varied by model and task, and all artefacts are open-licensed.
In briefConvex equilibrium potentials for Monge–Ampère equations: curvature bounds and after-the-fact evaluation
SummaryThe paper studies equilibrium neural networks, whose hidden state is the fixed point of a contraction, for solving the fully nonlinear Monge–Ampère equation. It shows that the Hessian of nonnegative softplus equilibrium networks grows at most like (1-q)^{-2}, a sharp exponent confirmed by a one-neuron example, and derives a sharp (1-q)^{-3} bound for third derivatives. Combining residual-based a posteriori estimates with interval arithmetic gives rigorous uniform bounds on the residual and solution error of a trained network. On two manufactured Dirichlet problems on the unit disk, all six trained models were certified.
In briefA Task-Centric Survey of Long-Document Visual Question Answering: Taxonomy and Future Directions
SummaryThis survey reviews long-document visual question answering by splitting it into four components: document parsing, evidence organization and retrieval, generator organization and learning, and evaluation. It gives a fine-grained taxonomy for each component and traces how research priorities have shifted from multimodal content understanding toward sufficient, high-fidelity evidence acquisition and sustained evidence management. The paper closes by outlining future directions, including evidence sufficiency, verifiable and correctable understanding, and budget-aware coordination.
In briefAI Risk Ontologies for EU AI Act Compliance: AIRO, VAIR, and Their Integration with DPV
SummaryThis paper presents ontologies and taxonomies designed to support risk management and documentation for high-risk AI systems under the EU AI Act. The authors developed the AI Risk Ontology (AIRO) to specify AI use cases and their risks, complemented by the Vocabulary of AI Risks (VAIR) for concept taxonomies. Both were informed by ISO/IEC 23894 and ISO/IEC 22989 and were integrated into the Data Privacy Vocabulary (DPV) through consensus-building, extending its scope from data to AI.
Delaram Golpayegani, Harshvardhan J. Pandit, Declan O'Sullivan et al.
In briefSafe, minimally intrusive Petri-net supervision for LLM agent tool execution
SummaryThe paper studies safe tool execution for LLM agents under prompt injection, modeling the harness between the model and its tools as a Petri-net supervisory control problem. A transition fires only when its required marks are present, so injected requests with unmet or spent preconditions are refused, and safety transfers to the running system under four assumptions. It further analyzes a replanning loop with inference costs and characterizes a least-intrusive maximally permissive supervisor. Synthesis on an autonomous release agent shows safe deployment may require a step the supervisor can refuse after rollout starts.
In briefRIDAR: A framework to detect and fix research integrity violations in autonomous research agents
SummaryAutonomous research agents can automate much of scientific work but sometimes commit research integrity violations (RIVs) that undermine validity, reproducibility, or faithfulness. The paper introduces RIVT, a taxonomy of 14 RIV types, and RIDAR-post-hoc, which detects RIVs in agent trajectories and reaches 85.2% F1 on a 60-trajectory expert-annotated benchmark. Applied to 300 trajectories, it found 3,964 RIV instances in 99.7% of them. The runtime variant, RIDAR-runtime, reduced RIVs remaining at task completion by 36.8%.
In briefUnified Swarm Control via Probabilistic Energy Modeling
SummaryThis paper proposes probabilistic energy modeling as a shared representation for robotic swarm control, composing diverse missions from reusable interaction energy terms. Decentralized policies trained in this framework are combined through lightweight policy gating without retraining. On aerial swarms, the unified policy generalized zero-shot across six mission classes, 10 to 100 agents, two hardware platforms, and indoor and GPS-denied woodland settings, with 0.96 to 1.00 task success in simulation and 1 to 10 ms per-agent inference latency.
In briefCrisiLangs: A Multilingual Dataset for Generating Crisis Warning Messages
SummaryThis paper presents CrisiLangs, which extends the English CrisiText dataset of crisis events and warning messages into Italian, Bulgarian, German, and Greek. Native-speaker experts translated warnings for 1,000 events, giving 5,000 event-message pairs across five languages. Fine-tuned multilingual models outperform a zero-shot baseline on ROUGE-L, BLEU and BERTScore in every language and configuration.
In briefMuffin: A shared memory buffer that lets AI agents work asynchronously
SummaryThis paper proposes Muffin, an asynchronous multi-agent framework in which agents exchange intermediate findings through a shared memory buffer while keeping their private reasoning local. An orchestrator routes queries by reasoning demand and sends demanding ones to the multi-agent setting. According to the authors, Muffin achieves the best accuracy in 17 of 20 benchmark settings and improves accuracy by 19.2 points over multi-agent debate while using 77.4% fewer tokens, without full-trace sharing.
Ping-Lin Kuo, Suman Navaratnarajah, Christian Frey et al.
In briefAutomating Feature Flag Cleanup with Multi-Agent LLMs: DoorDash's Industrial Experience
SummaryThis paper presents an end-to-end LLM-based system developed at DoorDash to automatically remove fully rolled-out feature flags and their gated code. The system queries live rollout data via MCP, processes Jira tickets, and creates pull requests automatically. Of 50 stale flags evaluated, 45 (90%) produced flag-removal PRs approved by engineers, with cleanups averaging $4.79 and 13.8 minutes.
In briefLearning emergent collective dynamics with graph-recurrent neural networks
SummaryThis paper introduces a graph-recurrent neural network (GRNN) that learns emergent collective dynamics directly from trajectories, without specifying microscopic equations of motion. It combines gated recurrent memory with graph message passing, so hidden states carry information beyond instantaneous observables and propagate between sites. Applied to charge-density-wave order in the semiclassical Holstein model, the method reproduces temporal and spatial correlations and dynamical scaling during domain coarsening.
In briefValuSpec: Pre-verification candidate filtering for faster tree-based speculative decoding (plug-and-play)
SummaryThis paper addresses the problem that retrieval-based tree speculative decoding produces many low-quality candidate branches that waste verification computation. It proposes ValuSpec, a plug-and-play module that filters these branches before target-model verification, with a threshold estimated from the target model's greedy trajectories on a calibration split and no parameter training. Experiments on HumanEval, GSM8K, Dolly, and SpecBench report up to 2.38x end-to-end speedup in retrieval-based tree settings and up to 2.78x when integrated into the EAGLE-3 dynamic tree framework.
In briefMetric conformal prediction using conditional expected local radius for non-Euclidean responses
SummaryThe paper introduces the conditional expected local radius (CELR), a measure of local conditional probability concentration, and builds a conformal prediction method for non-Euclidean responses in general metric spaces. It needs no densities or vector space structure and can handle heterogeneity and multimodality. A metric distributional random forest, fitted once, supplies all score components. The authors prove uniform consistency, asymptotic conditional validity, and finite-sample marginal coverage under exchangeability, and numerical studies show substantial computational savings over the conditional profile score method.]
In briefHarmful Responses as Jailbreak Templates: Transferring Autoregressive Attacks to Diffusion Language Models
SummaryThe paper finds that diffusion large language models can be induced to reconstruct harmful content by masking a subset of tokens in a harmful response, which serves as a jailbreak template. Building on this, the authors propose RMT-Jailbreak, a framework that transfers existing autoregressive jailbreak attacks to diffusion models through their induced harmful responses. Experiments across multiple attack methods and target models show its effectiveness and generality, revealing strong cross-paradigm transferability of jailbreak attacks.
In briefWeVisOne-VL: A model that unifies object and pixel-level visual perception
SummaryThis paper presents WeVisOne-VL, a vision-language model that unifies object-level and pixel-level perception through structured next-token prediction. It supports tasks such as object detection, referring localization, counting, OCR, segmentation, and matting, using a lightweight pixel decoder linked by a special <OBJ> token. The model is trained with a three-stage curriculum SFT followed by mixed-task RL. Across over 30 benchmarks, it shows leading performance on object-level and pixel-level tasks and robustly rejects queries about absent objects.
In briefDeep learning for random objects in metric spaces, without manifold assumptions
SummaryThe paper proposes a deep learning framework for regression with random objects in metric spaces that works directly on the response space's metric structure, avoiding manifold or Euclidean embedding assumptions. The authors establish statistical consistency and convergence rates for the estimator and report competitive simulation results against several competing approaches. Two digital health case studies with distribution-valued outcomes in Wasserstein spaces, using accelerometry and continuous glucose monitoring data, illustrate practical use. The paper also notes that embedding-based methods may be less suitable when no informative low-dimensional representation exists.
Marcos Matabuena, Xiaokai Luo, Oscar Hernan Madrid Padilla
In briefAgenticEvolve: Searching for Programs Using AI Agents That Design and Run Experiments
SummaryThis paper presents AgenticEvolve, an evolutionary program search framework in which each island is run by a persistent large language model agent that designs and runs experiments on candidate programs between evaluations. The authors compare it against a version with experiments disabled and against OpenEvolve on two heuristic design tasks, and against baselines on twelve equation discovery problems. They report better programs, reaching baseline quality with far fewer evaluations and comparable cost, and recovering the ground-truth structure of more equations.
In briefConsistent but Biased: How Chair Personas Trade Noise for Bias in Multi-Agent Clinical Decisions
SummaryThe paper tests whether a chair agent's persona prompt changes clinical decisions in a simulated seven-agent department handling 20 degenerative cervical myelopathy cases. Written as rules, an autocratic persona led to immediate surgery in 93% of runs and an avoidant one in 2%, versus 57% with no persona, mainly shifting timing rather than the surgical plan. Personas made runs agree more, but only at chance level and without a higher Fleiss κ, so the authors urge correcting agreement for chance before calling it reliability.
Yitong Bai (Columbia University), Chengyi Zhang (Swansea University), Ziyang Wang (Aston University)
In briefBridging the Data Gap: Transferring Source Knowledge to Few-Shot Diffusion Targets
SummaryThe paper introduces Meta-Diff, which learns a source-to-target relation once in a compact task space rather than retraining a generator for each target. On held-out CIFAR-100 validation tasks, correct source plus transport gives 13.9% lower mean SW than transport from the wrong source. In the final one-target comparison, it achieves 14.7% lower mean SW than target-only modelling and 32.7% lower than full fine-tuning. The authors also note that on Office-Home, transport aligns target-domain statistics without reliably preserving object identity.
In briefSpeeding up finite element model updating with hyperreduction to discover neural materials
SummaryThe paper speeds up finite element model updating, a method for learning material constitutive models from structural tests, by combining proper orthogonal decomposition with energy-conserving hyperreduction. The material law is a polyconvex neural network that is polyconvex, coercive, and non-negative by construction. Tests on two experimental datasets and one synthetic three-dimensional example show that training time drops substantially, from nine hours to 41 minutes for a two-phase specimen, without loss of accuracy.
In briefMVLoMa: Matching Local Features Across Many Views with Any Geometry Model
SummaryMVLoMa is a sparse matcher that matches one central image against many neighbouring images in a single pass, rather than one image pair at a time. It uses a frozen multi-view geometry model to add cross-view geometric context to local descriptors, then applies attention between the centre and neighbour images. Training penalizes high-confidence matches that violate epipolar geometry. On 25-view star graphs from four datasets, it achieves the best macro-average relative-pose AUC at all three thresholds and runs 2.20 times faster than sequential pairwise evaluation.
In briefStream4D: Training-free streaming 4D reconstruction that separates motion using causal KV-cache attention
SummaryThis paper presents Stream4D, a training-free streaming framework for 4D scene reconstruction that separates dynamic and static regions. It uses temporal motion cues stored in the causal attention KV cache, combined with depth priors, to produce dynamic masks, and applies soft key masking to reduce dynamic interference. The framework can be integrated into existing causal Transformer backbones such as StreamVGGT and Stream3R. According to the paper, it improves camera pose estimation and static-scene reconstruction on dynamic benchmarks and reduces ghosting over long sequences.
In briefSurveyGraph: Building Academic Surveys from a Methods-Entity Knowledge Graph
SummarySurveyGraph is a framework that generates academic surveys using a knowledge graph of methods and their relations, such as follow-on work and competition. The graph guides paper discovery, retrieval, outlining, and text generation. In the reported tests, graph-aware retrieval nearly doubled recall compared with flat lexical baselines, and 80.6% of audited generated sentences had full or partial citation support.
In briefHow well fast, fixed-choice AI models handle clinical decisions, and where they fall short
SummaryThe paper evaluates four "System 1" models, two proprietary and two open-weight, zero-shot on 86 public clinical tasks across 14 domains and 20 languages, covering 179,111 cases. Scores were high for text and medical-knowledge tasks but lower for imaging, especially volumetric findings. Triage nurses outperformed the best model on accuracy, and accuracy dropped sharply for some non-English languages.
Xiufeng Yang, Konstantinos Vrettos, Zeineb Ben Chaaben et al.
In briefWhen LLMs Forget What They Believe: State Integrity Collapse in Long Dialogues
SummaryThis paper examines why large language models can remember an instruction yet stop following it in later turns of a long conversation, a failure it calls State Integrity Collapse. The authors built StateTrack-Bench, a controlled set of dialogues testing numerical, format, persona, factual, and logical constraints across seven model snapshots. Strict joint retention ranged from 62.4% to 100.0%, and most failures were behavioral drift rather than forgetting. The authors conclude that single-turn instruction following does not reliably indicate performance over long dialogues.
In briefIdentifying Causal Effects from Stationary Data in Linear SDEs with Feedback and Hidden Variables
SummaryThis paper studies whether causal effects among variables in linear stochastic differential equation models can be identified from observational data drawn from the stationary distribution. Previous results mainly covered fully observed systems without directed 2-cycles. The authors give new graphical conditions under which causal effects are generically identifiable even when reciprocal interactions (directed 2-cycles) are present. They also develop an instrumental-variable framework to identify a target direct causal effect under partial observability with latent variables.
In briefCompressing per-layer embeddings via shared cross-layer codes
SummaryThis paper studies extreme compression of Per-Layer Embeddings (PLE), which make up 35.2% of Gemma 4 E4B's parameters. The proposed OPTIC method stores one quantized code per token and decodes it at each layer with a layer-specific affine decoder, fitted to the PLE output path. At about 100× compression, OPTIC retains 97.8% of the uncompressed model's aggregate performance across seven benchmarks, compared with 66.3% for product quantization, the strongest baseline.
WhyProposes output-path-fitted cross-layer shared codes for PLE; retains 97.8% performance at about 100x compression, well above conventional quantization baselines.
In briefA generalized Bayesian learning rule using non-KL geometries
SummaryThe paper extends the Bayesian learning rule, which normally relies on KL divergence, to mixed-space Bregman or Fenchel-Young discrepancies that still allow additive updates. Using Tsallis negentropy with quadratic statistics, it derives bounded-support, heavy-tailed q-Gaussian posteriors with a Newton-like implementation. Experiments up to CIFAR-100 show controllable support and tail behavior with competitive performance.
In briefSubgoal-aware memory graph for GUI agents that reuses past experience.
SummaryThis paper proposes a subgoal-aware memory graph to help GUI agents reuse past experience in long-horizon mobile tasks. It distils past trajectories into structured, reusable evidence at the subgoal level, where each subgoal is a sequence of consecutive actions sharing one intent. These subgoal experiences are stored as graph nodes that can be merged, extended, and retrieved based on the current observation. On SPA-Bench, the method improves success rate by 11.11% over the baseline.
In briefEGGPlacer: Energy-Guided Diffusion for Chip Placement with Thermal Awareness
SummaryThis paper presents EGGPlacer, an energy-guided diffusion framework for chip macro placement. It injects analytical placer feedback (standard cell distribution and full netlist wirelength proxy) and differentiable thermal feedback as composable energies during reverse diffusion, without retraining the generative model. On public benchmarks, it achieves competitive mixed-size HPWL and obtains the lowest average routed wirelength and peak temperature among the evaluated methods.
In briefSuperintelligent software engineering: governance that scales with AI capability
SummaryThe paper argues that capability alone does not improve software engineering and defines superintelligent software engineering as building software with agents exceeding human engineers, with governance scaling alongside them. A simple model shows that beyond fixed review capacity, escaping changes grow linearly with throughput, and governance can lower this slope while capability alone cannot. The authors also present OSTEON, a discipline with governed knowledge repositories, but report no measured effect of it.
In briefCola2: Co-Adaptive Hierarchical Language Model Compressing Text into Continuous Semantics
SummaryThis paper proposes Cola2, a hierarchical language model that splits text into semantic spans, compresses them into continuous latent representations, and generates text with a discrete block-causal masked diffusion decoder. It jointly addresses compression rate, the balance between reconstruction and generation, and co-adaptation of continuous and discrete components. The authors report experiments on about 210 billion training tokens validating core designs, and an unconverged checkpoint trained on 346.03 billion source tokens averaging 60.10 across eight tasks, surpassing Llama-3-8B on HellaSwag and StoryCloze despite a pretraining budget reported as over 43 times larger.
In briefA survey of data-driven vision-language models for remote sensing.
SummaryThis survey systematically reviews the current state of remote sensing datasets built for vision-language models (VLMs). It argues that the field has shifted from model-centric to data-centric innovation and offers a practical guide covering data construction, fine-tuning, and evaluation. The authors encourage building more diverse, higher-quality datasets toward a comprehensive remote sensing vision-language corpus.
In briefHow memory helps language agents reuse past experience in future decisions
SummaryThe paper develops a decision-theoretic framework for the value of memory in language agents, tracing how useful distinctions from past experience persist, get used, and can be recovered or revised. It expresses value in units of avoided task loss and includes a simulation and minimal counterexamples. Diagnostics on LoCoMo and other cases examine failures of preservation, use, and measurement. It proposes testable principles for allocating memory, preserving recovery options, and revising evidence dependencies.
In briefPiHand: Latent hand-object interaction priors for physics-based robot and VR hand control
SummaryPiHand is a framework that synthesizes physically grounded dexterous hand-object interactions for VR/AR, gaming, and robotics. It trains expert policies to track motion-capture demonstrations, distills them into a goal-conditioned VAE prior, and solves manipulation tasks by bounded residual exploration in the latent space. The authors report that, using a single reward, the policies generalize to unseen objects and trajectories on GRAB and DexYCB, and the same prior transfers to object catching and bimanual handover. The paper states that PiHand outperforms baselines on all tasks and that processed data and model checkpoints will be released.
In briefRecallGS: Recalling Past Views to Add Geometry to Online Semantic 3D Scene Mapping
SummaryOnline semantic 3D Gaussian Splatting enables open-vocabulary scene understanding from streaming observations, but existing methods rely on an extra learned 3D network to recover geometry missing from image-based semantic features. RecallGS is a lightweight, training-free method that treats past observations as a non-parametric geometric-semantic prior, retrieving relevant views with CLIP and weighting geometrically supported views via alpha-compositing. On ScanNet and ScanNet++, it achieves strong semantic segmentation performance at 2.48 FPS without an additional learned 3D branch.
In briefCognitively constrained partners for zero-shot coordination in real-time control
SummaryThe paper proposes training teaming agents against simulated partners whose diversity comes from procedural constraints based on human reaction delays and perceptual limits. In a modified Lunar Lander environment, preliminary results suggest these partners behave more like real people than unconstrained agents, and agents trained against them match strong cross-play baselines. The authors also describe a planned human-participant study to test whether cross-play rankings predict real human-agent teaming.
In briefDISC: A learned way to combine time-series dataset similarity metrics for predicting transfer performance
SummaryThe paper observes that many time-series dataset similarity metrics capture different relationships and align unevenly with downstream forecasting errors. The authors propose Domain-Integrated Similarity Cost (DISC), which uses calibration transfer observations to select, weight, and combine these metrics. On the reported split, DISC achieves higher alignment with downstream errors than single-metric references. The results suggest that dataset similarity is task-dependent and that learned metric integration is a stronger proxy for transfer behavior.<br>
Yen-Ku Liu, Hongjie Chen, Franck Dernoncourt et al.
In briefRetrieval-augmented textual memory for answering questions about long first-person videos
SummaryThis paper studies how retrieval affects question answering over long egocentric videos stored as temporally grounded text captions. It systematically compares retrieval units, query formulations, context budgets, and reranking strategies on QAEgo4D-Closed while keeping the generated memory fixed. The best configuration, Parent, reaches 56.4% accuracy, matching full-memory and strong retrieval baselines while exposing only a small part of the memory. The authors also find that larger retrieval budgets can improve temporal recall yet lower accuracy by adding plausible distractors and uncertain evidence.
Raffaele Cali, Giuseppe Lando, Rosario Forte et al.
In briefReCoG: A Relational Graph That Helps AI Tell Similar Custom Concepts Apart Without Retraining
SummaryThis paper addresses training-free personalization, where a frozen vision-language model recognizes user-defined concepts from a few examples. Existing retrieval methods treat concepts in isolation, so retrieved evidence can fail to distinguish confusable categories. The authors propose ReCoG, which links concepts with directed edges encoding discriminative cues and resolves ambiguity through pairwise comparisons at inference. ReCoG outperforms prior methods on the new FINGER benchmark for confusable concepts and performs best on captioning and VQA benchmarks.
In briefBeyond Accuracy: Separating Task Decomposition from Typed-State Quality Checks in Multi-Agent LLM Pipelines
SummaryThis paper examines where gains in multi-agent LLM pipelines for requirements engineering come from, comparing a single-call baseline, a two-call variant reusing specialized prompts without validating intermediate output, and a pipeline with a Pydantic-validated typed state. The F1 improvement is associated mainly with the two-call, specialized-prompt variant (up to +0.075), while the typed state adds a small difference significant in only one of six comparisons. The findings are associative and based on three 7–8B models and one dataset.
Maria Gabriela Alves Zuppardo, Vinicius Cardoso Garcia
In briefCoWriter: Guiding Human-LLM Co-Authoring With Proven Writing Methods
SummaryThis paper introduces CoWriter, a proactive agent that applies expert co-authoring methods from cognitive science to guide the full writing process while encouraging active user participation. In an exploratory user study against a standard LLM baseline, authors using CoWriter better expressed their perspectives, felt greater document ownership, and had 2.2 times more successful agent engagements. External readers also rated its texts higher for human-touch qualities such as storytelling and personal experience.
In briefRIPPLE: a training method that helps coding agents recover from mistakes
SummaryLong-horizon coding agents can spot a failure and write a reasonable critique, yet still keep following the same wrong path. RIPPLE is a training framework that teaches a single policy to recover by itself, using a teacher that sees only the prefix of the trajectory, followed by reinforcement learning with stage-specific credit for reflection, diagnosis, and corrective action. Across four repository-level software engineering benchmarks, it raises the base policy's arithmetic mean score by 4.8 points and needs no auxiliary critic at deployment.</summary>
In briefPretraining Robot Behavior Priors on Non-Robotic Synthetic Data
SummaryThis paper examines whether procedurally generated synthetic data without any robot embodiment can provide useful structural priors before a model sees real robot behavior. The authors build three pipelines modeling mathematical abstractions of robotic control and pretrain a Transformer on robot action prediction. The model trained only on this non-robotic data achieves non-trivial zero-shot prediction on unseen physical robot tasks via in-context learning, and control-specific priors outperform generic causal priors.
Mateusz Odrowaz-Sypniewski, Michael Amir, Jasmine Bayrooti et al.
In briefPointerBench: A Diagnostic Benchmark for Fine-Grained GUI Grounding
SummaryThis paper introduces PointerBench, a suite of 1,500 screenshot-instruction pairs that tests whether multimodal models can locate small interface elements, character-level text in long documents, and mixed controls across 100 professional applications. The authors evaluated general-purpose multimodal models and 39 specialist GUI grounding configurations. Text grounding proved the hardest setting, with the three highlighted models reaching 69.6%, 53.4%, and 25.8% success there, well below their overall averages. The benchmark and evaluation code are publicly released on GitHub.
Valentino Sacco, Saurabh Pandey, Alexander Florian Jercher et al.
In briefMemory-Aware Reward Modeling: Learning Human Preferences for LLM Memory Use
SummaryThis paper examines whether memory-conditioned responses from language model assistants match what users prefer. Researchers collected about 1,000 pairwise judgments from 87 participants and found that memory-based responses were not consistently preferred, with some users strongly favoring responses without memory. The authors also introduce a memory-aware reward model trained on user judgments, which outperforms existing baselines in preference prediction.
In briefNotation3: Defining the Meaning of Graph Terms in RDF
SummaryNotation3 extends RDF to allow rules over triples and statements about graphs. Its original informal semantics left the meaning of graphs used as terms, called graph terms, undefined. This paper examines how graph terms are used and what formal requirements they must meet, then gives an abstract syntax and defines semantics in two steps: basic entailment and N3-entailment, the latter extensible as new predicates are added.
In briefDeep ReLU networks may lack the best approximation property for continuous functions
SummaryThe paper constructs a continuous function on [-1,1]² for which the infimum of the L² approximation error over ReLU networks with two hidden layers and architecture (2,2,1,1) is not attained. This shows that deep ReLU networks generally do not have the best approximation property, even for continuous targets. The result contrasts with shallow networks, where ReLU guarantees global minima in typical regression tasks. The proof relies on the fact that breaklines of second-hidden-layer neurons can bend, which breaks the symmetry arguments used for one-hidden-layer networks.
WhyConstructs a continuous counterexample showing deep ReLU networks can lack best approximation, overturning the shallow-network intuition with a rigorous proof.
In briefA systematic comparison of search space complexity and efficiency in end-to-end speech recognition models
SummaryThis paper compares search space complexity and efficiency for CTC and neural transducer speech recognition models under different pruning strategies and language model settings. The models usually work with a small search space but show large temporary peaks in hypothesis count during uncertain periods. Score-based pruning adapts the search to model uncertainty and significantly lowers the Real-Time Factor, while the authors also highlight the limits of small fixed-size beam pruning and offer practical guidelines.
In briefIndoNutriQA: An Indonesian Clinical Nutrition Question Answering Benchmark
SummaryThis paper introduces IndoNutriQA, an expert-curated Indonesian clinical question answering benchmark of 911 question-answer pairs built from real clinical-nutrition reports. Eight standardised question types following the Nutrition Care Process are answered with free-form clinical prose, and an expert validation study finds 6-10% of reference answers contestable. The authors report baselines using fine-tuned models and eight LLMs, and compare an LLM-as-judge rubric against three expert raters. The dataset is released to support Indonesian clinical NLP research.
Tirana Noor Fatyanosa, Ummu Ditya Erliana, Mahendra Data et al.
In briefTumour localisation on intra-operative CT using weak signals from pre-operative MR
SummaryTumours are hard to see on intra-operative CT, which complicates liver cancer surgical navigation. The method trains a classifier on pre-operative MR images to detect tumour presence, then applies it to windows of intra-operative CT to weakly guide a reinforcement learning agent toward tumour locations. Among direct transfer, few-shot adaptation and synthetic transfer, synthetic transfer (translating CT into synthetic MR first) performed best on an external hepatocellular carcinoma cohort, and it needs no paired MR-CT training images.
In briefQuaNAS: Reusing Past Architecture Searches with Uncertainty Awareness for Efficient Neural Architecture Search
SummaryOn a new dataset, neural architecture search often allows only a few dozen short training runs, so prior knowledge matters a great deal. QuaNAS treats earlier searches in the same search space as priors of uncertain reliability and uses a Gaussian process to re-estimate how far each prior is trusted after every evaluation. On 12 new datasets, it reduced anytime regret on the short-training objective by 2.7 to 3.9 times compared with five from-scratch baselines, and its final models matched or exceeded reference models on nine datasets.
Brama Seta Putra Satria, Dwi Sunaryono, Misbakhul Munir Irfan Subakti et al.
In briefCombining a tree model with neural experts to reduce multi-horizon solar power forecast errors
SummaryThe paper studies whether a weighted combination of a strong tree model (LightGBM) and several neural forecasters can lower solar power forecast error over 15 to 240 minute horizons. Combination weights are nonnegative and sum to one, so unhelpful candidates can receive zero weight. On 56,078 windows from two public PVDAQ systems, the authors report that a historical shared-weight policy reduced scaled MAE relative to the tree. The authors note that these results do not establish independent replication, weather-specific causation, or universal gains from eight experts.
In briefReward-Weighted Preference Optimization: a reinforcement learning training method for problems where the solver must also choose which points to visit.
SummaryThe paper addresses selective combinatorial optimization problems, such as the Orienteering Problem, Prize-Collecting TSP, and Knapsack Problem, where solutions differ widely in objective value. It proposes Reward-Weighted Preference Optimization (RWPO), which weights pairwise preference losses by the normalized reward gap between candidate solutions from the same instance. Experiments report consistent gains over unweighted preference optimization and policy-gradient baselines, and on the Orienteering Problem the method exceeds the mean reward of the Compass metaheuristic at scales from 100 to 2,000 nodes.
Haohui Zhang, Xinyu Hu, Wouter van Heeswijk et al.
In briefMuting Tokens: Token Sparsification for Efficient Audio Transformers
SummaryThis paper comprehensively evaluates token sparsification methods, including pruning and merging, for audio transformers across multiple datasets and model sizes, in both post-training and fine-tuning settings. It also proposes a new approach that combines attention scores with spectral cues from token embeddings. The experiments report that frequency-aware methods based on spectral flatness and low-frequency energy consistently outperform attention-only methods. Spectrogram patch intensity-based sparsification is reported to be especially effective for longer audio inputs.
In briefLightweight Frame Compensation for Real-Time Polyp Segmentation Under Frame Skipping
SummaryThis paper addresses accuracy loss in real-time polyp segmentation during colonoscopy when frames are skipped to increase speed. It proposes LFC, a 77K-parameter plug-and-play module placed before a frozen segmentation backbone, which fuses three consecutive frames using spatio-temporal attention. LFC is trained only on pseudo-labels from a differently seeded model, and on three polyp video benchmarks it recovers much of the accuracy lost to frame skipping while keeping a 2.05X to 2.55X throughput gain.
In briefCross-dataset benchmark of five retinal vessel segmentation methods on performance, structure and reproducibility
SummaryThis paper compares five retinal vessel segmentation pipelines on 244 images from six public datasets, assessing pixel agreement, cross-dataset stability, thin-vessel retention and structural metrics. No single pipeline was best in every setting: FSG-Net led on the overall metrics, while SA-UNetv2 had the highest sensitivity and thin-vessel sensitivity. The authors conclude that pipeline choice should be task-conditioned and validated on the intended target domain rather than based on one universal ranking.
In briefGLOW-FM: Fast One-Step Generative Enhancement for Dark Images
SummaryGLOW-FM is a framework that enhances low-light images in a single inference step by performing flow matching in the latent space of a frozen Tiny-VAE. It uses Terminal Velocity Matching (TVM), which the authors state is being applied to image restoration and low-light enhancement for the first time. Experiments on real-world benchmarks report competitive image quality and state-of-the-art cross-dataset generalization, with the small variant requiring 34 GFLOPs.
In briefQICOM: open software for reproducible image quality checks, blinded review, and radiomics comparison
SummaryQICOM is open-source software for quantitative image comparison, paired radiomics analysis, and blinded assessment. It stores processing settings, image identities, measurements, and response mappings so that a comparison can be reopened and recalculated. The paper tests reference agreement, responses to controlled perturbations, and recovery of saved analyses using synthetic experiments and public reference cases.
In briefFinding patient groups where a risk model is inaccurate, using a localized AUC
SummaryRisk prediction models are usually evaluated with AUC, which hides differences in performance across patient groups. This paper introduces Localized AUC (LAUC), a weighted version of AUC, and a method called LOCATOR that uses it to identify cohorts for whom the model is inaccurate. The method is tested on real-world data and in simulations. The authors also provide a statistical foundation for using LAUC for this purpose.
Wenye Ren, Anna Gy\H{o}rffy-Kerekes, Tancredi Dorsi et al.
In briefTrion-1: A Diffusion Language Model Unifying Language, Protein Sequence, and Structure
SummaryThis paper introduces Trion-1, a native unified multimodal diffusion language model for proteins that represents sequence, structure, and natural language in one discrete token space. Instead of strict left-to-right generation, it uses bidirectional or semi-bidirectional token interactions to model global residue communication and long-range contacts. The authors report broadly competitive results across six capabilities, including folding and sequence understanding.
In briefSegmenting Actually Moving Regions Under Camera Motion, Not Whole Moving Objects
SummaryThe paper introduces MoRSe, which jointly analyzes two frames to predict a mask of moving regions, separating actual scene motion from apparent motion caused by the camera. It uses a single RGB encoder, so no optical flow, point tracks, or separate mask generator is needed at inference. The authors release MoRSe-5K, a benchmark of 42 sequences covering static, dynamic-static, and dynamic settings.
In briefBenchmark testing whether multimodal AI models can locate eye disease signs in retinal OCT scans
SummaryThis paper introduces RetinaGroundBench, the first benchmark for checking whether multimodal large language models can visually localize biomarkers in retinal OCT images. It contains 815 B-scans annotated by ophthalmologists with 4,537 bounding boxes across 13 biomarkers and 8 diseases, and proposes the AggIoU metric for evaluating overlapping predicted boxes. Testing 27 open-source models, the authors find grounding remains largely unsolved, with the best general model reaching 23.7% macro F1. Their fine-tuned RetGrounder model raises grounding F1 to 33.9%.
Ronald Fecso, Jos\'e Morano, Botond Fazekas et al.
In briefPredicting music valence and arousal with text and multimodal in-context learning
SummaryArousal can be reliably predicted from music audio, but emotional valence is harder to model because it is subjective and context-dependent. The authors use in-context learning with large language models, feeding demonstrations from an auxiliary dataset to refine the predictions of a small task-specific base model. They test a text variant using lyrics and generated audio captions, and a multimodal variant that adds raw audio, using the entire MER60 dataset as a held-out test set. Text-based ICL raises valence R² from 0.110 to 0.427 with 80 shots, and the multimodal variant reaches 0.346 with only 10 shots.
In briefTesting how LLMs' value answers shift with country and language cues
SummaryThis paper examines how large language models answer value questions when two cultural identity cues are given at once: a country persona in the system prompt and a second identity in the user turn, either by naming a country or through the query language. The authors test World Values Survey multiple-choice questions and open-ended sociopolitical generation across 1.6 million responses. They conclude that country and language cues should be evaluated as distinct signals rather than interchangeable proxies for culture.
In briefTask- and timestep-aware mixed-precision quantization for diffusion language models
SummaryThis paper studies how quantization affects diffusion language models, whose iterative denoising makes sensitivity vary by task and timestep. The authors show these sensitivity patterns can be predicted without ground-truth labels or task-specific metrics. They propose T²Quant, a mixed-precision framework that allocates activation precision across denoising steps using a label-free sensitivity signal under a fixed bit-width budget. Experiments on three models and six benchmarks report accuracy gains of up to 6.7 percentage points over existing quantization methods.
In briefInside-JEPA: From Prediction to Relational Grounding in Latent World Models
SummaryThis paper introduces Inside-JEPA, which adds relational channels for ordered entity pairs inside the latent state of a joint-embedding predictive architecture. The authors find that latent prediction alone does not ground these channels, while an inverse-dynamics loss that recovers actions from channel changes improves relational AUROC across several datasets and settings. The model also ranks correct actions for one-step relational goals better than prediction-only latents and a frequency prior.
In briefMaking weaker faithfulness usable: recovering causal links that structure learners discard
SummaryThe paper addresses structure learners discarding variables that predict an outcome only jointly, not individually. It compares two methods: selective protection in an adjacency learner and explicit retention of alternative directed graph explanations. In a noisy XOR experiment, combining both changes raised complete parent-pair recovery from 0/20 to 20/20, while adding false edges. The authors also report that protection can add only error and that computations for larger variable counts remained unfinished.
In briefWhen More Metadata Misleads: Testing How LLMs Ground Museum Descriptions Field by Field
SummaryThis paper studies how the composition of metadata affects the artefact descriptions that language models generate for museum collections. It uses a two-phase framework: first analysing natural metadata variation in a large collection, then running controlled probes that keep object identity fixed while varying metadata quantity, field type, and image access. The authors find that metadata quantity alone does not explain generation behaviour, and that field type determines whether models ground, omit, or infer information.
In briefRaSFNet: A Ray-Aligned, Scan-Aware Network for Segmenting LiDAR Point Clouds
SummaryThis paper presents RaSFNet, a network for semantic segmentation of LiDAR point clouds used in autonomous driving. It uses a ray-aligned encoding of point distributions, a scan-topology-aware Mamba block for long-range context, and range-constrained fusion between point and frustum features. The authors report mIoU scores of 70.2% on SemanticKITTI and 77.7% on nuScenes validation sets.
In briefParticle physics foundation models for track finding and particle identification
SummaryThis paper evaluates whether foundation models can serve as a tool for analyzing particle physics data. The authors test a foundation-model-based reconstruction of charged particles on several downstream tasks, using spacepoints from a Time Projection Chamber. They report that this approach significantly outperforms traditional methods. They also show it can be adapted to tasks not directly related to its original training objective.
In briefPrompting effects in AI judging depend on the judge model
SummaryThis paper studies whether prompting strategies that work for large language models acting as evaluators hold up across judge models and languages. The authors tested 9 prompting strategies with 4 open-weight judges, 27 languages, 4 tasks and 13 datasets, producing over 1 million scores checked against human judgments. They find the choice of judge model matters more than the prompting strategy, strategy effects can reverse across judges, and stability drops in low-resource languages where aggregate reports hide these effects.
Asal Mehrabi, Mohammad Aliannejadi, Clemencia Siro
In briefMemory before Goal: How pre-history memory helps humanoid robots move and manipulate objects
SummaryThis paper addresses humanoid robots performing loco-manipulation in dynamic scenes, where object cues can be hidden or leave the view during movement. The authors propose WholeBodyMem, an object-guided perceptual memory that keeps the history of object events before the robot acts on its goal. They also introduce HuMeGo, a humanoid memory benchmark with simulated and real Unitree G1 tasks. The method outperforms memoryless and uniform-memory baselines in both simulation and real-robot experiments.
In briefHow network structure shapes consensus and vulnerability in LLM populations
SummaryThis paper studies how groups of large language models reach agreement depending on who interacts with whom. The authors run the naming game with policies from four LLMs on fully connected, random regular, and heterogeneous networks with 1,000 to 100,000 agents. Sparse networks slow consensus by up to twelvefold and sometimes change which convention wins. The study also finds that network structure governs vulnerability to adversarial agents, and for most models networked populations are easier to tip than well-mixed ones.
Roberto Passaro, Ariel Flint, Romualdo Pastor-Satorras et al.
In briefCalibrating computer models with controlled prediction uncertainty
SummaryThis paper proposes a two-layer framework for calibrating computer simulators that directly controls predictive error. A conditional flow-matching model learns an amortized posterior over calibration parameters, and conformal risk control selects prediction bands that achieve a prescribed time-averaged miscoverage. In a synthetic benchmark the bands are about 20% the width of KOH-MCMC bands with miscoverage near the 10% target. In calibrating an age-structured SEIR model to the 2010 Malawi measles outbreak, the bands are again sharper and near-nominal, and the uncovered weeks coincide with the asymmetric epidemic decline.
In briefHiSTEVE 1.0: Human-Interface VLA Agents for Open-World Games
SummaryThis paper presents HiSTEVE, a hierarchical vision-language-action agent for playing open-world games like Minecraft. A vision-language model planner selects subtasks, while a single VLA policy executes them using only visual input and native keyboard-and-mouse actions. The authors report 85.0% on MineStudio Bench and a 37.4% average success rate on OpenHA, and say the agent can complete long-horizon tasks such as crafting an iron pickaxe from scratch.
In briefWearable sensor systems for recognizing construction worker activities in the field
SummaryThis paper studies wearable inertial measurement unit systems that recognize construction workers' activities, using the VTT-ConIoT benchmark. It compares classifiers, sensor placements and window sizes under a leave-one-subject-out protocol, and tests personalised calibration and data augmentation. Per-person accuracy varies widely, and multi-sensor fusion, longer windows and personalised calibration are reported to help, though all data were collected in a laboratory.
In briefBenchmark for AI agents writing video-based research reports
SummaryThis paper introduces video-centered deep research report generation, where an agent takes a video and a user request, forms research questions, retrieves external multimodal evidence, and produces an interactive report linked to relevant video moments. The authors release OmniDR-Bench, with 50 Chinese and English tasks and 6,414 evaluation criteria, along with the Omni-ART rubric method. Among 18 agent configurations tested, the best scored 63.3 out of 100.
In briefHuman-robot interaction challenges in aged care homes: This chapter examines how robots and ambient intelligence can support older adults in residential aged care and private homes.
SummaryThe chapter examines how robotic systems and ambient intelligence can support older adults in residential aged care homes and private homes, which it treats as challenging settings for human-robot interaction. It reviews systems that assist with daily living, safety monitoring, and social engagement, including hybrid setups where robots work alongside fixed and wearable sensors. It also covers interaction design for users with low digital literacy, privacy measures such as on-device inference, and real-world evaluation metrics like false alarms per day.
David Ahmedt-Aristizabal, Mohammad Ali Armin, Akila Pemasiri et al.
In briefHow wording changes model rankings when models use tools
SummaryLanguage models are often ranked on question-answering benchmarks, but rewording a question can flip which model scores higher. The authors propose a matched protocol that separates answer readout, reasoning, the decision to call a tool, the tool's answer, and the report made without a call. Testing six open models across five domains and five wordings, they find tools raise accuracy where a program can answer, but widen wording sensitivity in many pairwise comparisons, following call frequency. They recommend reporting call frequency and comparison stability under several wordings alongside accuracy.
In briefPriMa-Causa: Causal Reasoning for Prescriptive Maintenance
SummaryThis paper proposes PriMa-Causa, which adds causal reasoning to prescriptive maintenance in manufacturing so that maintenance actions are ranked by their estimated interventional effect rather than only predicting failures. The system combines a causal reasoning layer, a parametric simulator that supplies interventional ground truth, and a graph-conditioned tabular causal foundation model pre-trained on a synthetic production-line prior. According to the authors, it achieves lower prescription error than associational baselines, with the largest gains under weak overlap and sparse context, and a partially disclosed causal graph reduces the error further.
In briefImproved Fisher vectors that encode second-order multi-scale CNN features for scene image retrieval
SummaryThe paper addresses the difficulty of turning CNN responses, which lack geometric invariance, into a compact and discriminative vector. It proposes a codebook-free method that encodes second-order information of multi-scale CNN features within the Fisher Kernel framework. Experiments on the INRIA Holidays and UKBench benchmarks show significantly higher retrieval accuracy than state-of-the-art approaches, with only a slight increase in vector dimension.
Alexy Bhowmick, Sarat Saharia, Shyamanta M. Hazarika
In briefRankAlign: Aligning Vision and Language Representations via Rank Transformation
SummaryThis paper proposes RankAlign, a rank-based normalization method for matching vision and language models at the class level without paired cross-modal supervision. Each within-modality similarity matrix is converted into an ordinal relational representation before permutation matching, making the result invariant to strictly increasing rescalings of similarity values. Experiments with diverse frozen encoders and datasets show more accurate and stable class-level correspondence recovery, including under label corruption and domain shift.
In briefGM-GReFEL: Geometry-guided state-space learning for robust video facial expression recognition
SummaryThis paper presents GM-GReFEL, a model for recognizing facial expressions in unconstrained videos. Its backbone is a grouped-scan state-space network whose cost grows linearly with clip length, and a reliability module gives more weight to facial geometric anchors when the classifier is uncertain. In five-fold cross-validation on DFEW, MAFW and FERV39k, the authors report competitive recognition scores, with some results ahead of compared methods.
In briefPos4CHD: A Positive-Free Mask-Guided Framework for Synthesizing Fetal Heart Defect Images
SummaryThis paper addresses the scarcity of positive cases for fetal congenital heart defect diagnosis by proposing Pos4CHD, which synthesizes pseudo-positive ultrasound images using only normal samples. It builds on FE-Gen, a mask-guided generator for fetal heart ultrasound, and first alters semantic masks to simulate structural anomalies before rendering consistent images. Experiments report improved visual fidelity and anatomical consistency, along with gains in zero-shot classification, lesion detection, and segmentation.
In briefMixing Matters: Position Bias of Evidence Across Sequence Mixers in Long-Context QA
SummaryThis paper compares how Transformer, state-space, and hybrid Mamba models use answer-bearing evidence depending on its position in a long prompt, by moving one gold document among nine distractors. The evaluated Transformers showed an advantage for evidence at the beginning that the state-space models lacked, while both families favored evidence near the end. The authors describe these results as exploratory and do not isolate the sequence mixer as the sole cause.
In briefIs High Local Accuracy Enough? A Federated Learning Method for Global and Personalized Recognition
SummaryThe paper proposes an evaluation protocol for personalized federated learning that adds global, rare-class, and unseen-class tests beyond local accuracy, and measures performance of the worst-performing clients. It introduces FedGRC, combining consensus-guided representation learning, multilayer metric synthesis, and support-adaptive shrinkage personalization. Experiments on CIFAR-10, CIFAR-100, and the FedRS dataset against 21 methods report gains in global recognition and personalization floors.
In briefPost-Audit Mirage: Limits of Offline Tests for Agent Safety
SummaryThe paper argues that offline tests may fail to show whether an agent update will stay safe after deployment, calling this the post-audit mirage. The authors build matched safe and harmful lifecycles that present identical verifier-visible evidence while an isolated scorer records opposite deployment outcomes. They introduce IdentifiedRangeMonitor, which uses fresh group-labelled harm data, and find it can recover the missing distinction in favorable settings but may stay undecided for rare groups, small effects, or missing observations.
In briefWildDepth-v2: A Synchronized RGB-LiDAR-Radar Dataset for Measuring Wildlife Distance and Motion
SummaryThis paper extends the WildDepth dataset with imaging radar recordings, including 37 RGB-LiDAR-4D radar sequences of 19,716 frames. It provides a frozen, checksummed hidden test set for RGB-only distance estimation and RGB plus sparse-LiDAR depth completion. The authors also lift 2D animal tracks into metric 3D trajectories using LiDAR, radar, or monocular depth. Object-level radar calibration improves monocular depth over scene-level LiDAR calibration without LiDAR at inference, though animal-level LiDAR remains most accurate.
Muhammad Aamir, Sangyun Shin, Naoya Muramatsu et al.
In briefSetCommit: Coordinated Token Commitments for Faster, More Consistent Diffusion Language Model Decoding
SummaryDiffusion language models can generate several token positions in parallel, but when decoding takes only a few steps, individually reliable tokens may be jointly inconsistent. SetCommit is a training-free method that selects a set of tokens to commit by balancing each token's reliability against local joint risk, and it swaps vulnerable earlier commitments when new context supports better alternatives. Tested with LLaDA-8B and Dream-7B on GSM8K, MATH-500, MBPP, and HumanEval, it reports accuracy or pass@1 gains of up to 7.4 points at 32 model evaluations.
In briefCombining constraint reasoning and LLMs to explain optimisation-based decisions
SummaryThis paper addresses optimisation-based decision support systems and proposes a method that combines constraint reasoning with Large Language Models (LLMs) to produce contrastive explanations for recommendations derived from Mixed-Integer Program (MIP) solutions. Constraint reasoning first extracts the information needed for an explanation, which is passed to the LLM as a graph of reasons capturing relationships among constraints. Results on real-world scheduling domains indicate that this structure leads to explanations with more relevant reasons and fewer errors than explanations generated without it.
Roger X. Lera-Leri, Filippo Bistaffa, Athina Georgara et al.
In briefProductWebBench: A Benchmark for Website Continuity Beyond Task Completion
SummaryThis paper introduces ProductWebBench, a benchmark that tests frontend coding agents on living repositories, scoring both whether the requested change is built and whether the rest of the website keeps working. It includes 320 Change tasks on existing code and 80 long-horizon Build tasks, evaluated over 8,737 runs from 13 models. The authors report that 28.2% of Change runs passing all requested-work checks still fail a continuity check, and 10.8% break browser states that had previously passed.
WhyIntroduces website continuity as a new evaluation axis; 8,737 runs show task completion and continuity diverge, with substantial relevance to frontend coding agent evaluation.
In briefThrough a PRISM, Clearly: Step-Sequence Reasoning for Procedural Video Mistake Recognition
SummaryThis paper introduces PRISM, a method for recognizing mistakes in multi-step procedural videos. It describes video segments in language and maps them to procedural states using a lightweight hidden Markov model trained on correct executions, then identifies step-content and order errors through reference alignment and semantic checks. Evaluated on three benchmarks, semantic role labeling improved Any-F1 in all six matched comparisons, though gains varied across datasets.
In briefA dataset and tool mapping grounds of appeal to Court of Cassation reasons
SummaryThis paper releases a dataset of 121,536 pairs of grounds of appeal and court reasons, drawn from 86,464 rulings of five civil chambers of the French Court of Cassation from 2016 to 2025. A deterministic pipeline pairs each ground with its reasons and assigns it to a family in the established French taxonomy. The authors use this to describe which grounds succeed, which legal texts they cite, and how these patterns change over the decade. The dataset, code and an interactive consultation tool are made available.
Avrile Floro, Tamara Dhorasoo, Nils Holzenberger et al.
In briefExplaining Next-Event Predictions with Neuro-Symbolic Rules
SummaryThis paper addresses next-event prediction in event sequences, noting that existing methods often sacrifice interpretability for accuracy. The authors propose RULES, a neuro-symbolic model that learns sequential rules directly from data, and combine it with a residual network to form RUNES. On synthetic data, RULES accurately recovers ground truth rules, and on real data it discovers compact, interpretable rule sets. RUNES achieves competitive accuracy while transparently showing active rules and how residuals adjust predictions.
Tim Nico Bauerschmidt, Isabel Valera, Jilles Vreeken
In briefCan AI Co-Design Structures? An AI Agent for Early-Stage Structural Design
SummaryThis paper presents a proof-of-concept workflow in which an AI agent, connected to a COMPAS-based structural modelling API through an MCP server, turns natural language prompts into graph-based structural models made of explicit building elements. The resulting models can be inspected, edited, serialised, and linked to downstream AEC workflows. The authors demonstrate the approach with prompt-to-structure examples and editable design sequences resembling iterative structural design. The results are preliminary but suggest that combining language, agentic workflows, and structural data models is a possible first step toward AI-assisted structural co-design.
Ricardo Maia Avelino, Mennatallah El-Assady, Tom Van Mele et al.
In briefWhere Long Image Captions Fail: Position-Aware Diagnosis and PAGER Reinforcement Learning
SummaryThis paper studies where errors appear in long, detailed image captions generated by autoregressive vision-language models. The authors decompose captions into atomic visual claims and check each position for supported, hallucinated, and generic content. They find that middle sections tend to contain unsupported specifics, while final sections drift into vague language. They propose PAGER, a GRPO training objective with position-specific rewards, which improves position-wise grounding and pairwise preference on detailed-caption benchmarks.
In briefDRAFT: Reconstructing 3D Room Layouts from fMRI and Camera Trajectories
SummaryThis paper introduces DRAFT, a method that reconstructs the 3D layout of unseen rooms using only camera trajectories and fMRI signals from participants watching indoor walkthroughs, without any visual frames. It extracts a compact layout representation from the early visual cortex and fuses it with free space derived from camera motion. On ten held-out scenes with 20 participants, DRAFT achieves a layout correlation of 0.480, outperforming the camera path alone (0.446) and seven fMRI decoders (at most 0.429).
In briefLearning What to Ask for Tomorrow: Active Preference Acquisition for Web Agents
SummaryThis paper studies how web agents can actively ask users for preferences that will be useful across many future tasks. The authors propose Active Preference Acquisition (APA), which selects questions by estimating their cumulative utility gain over future tasks minus the cost of asking. Evaluated on the controlled ActiveWebPref benchmark, APA acquires far more valuable preferences than a current-task-only baseline and achieves higher net utility under matched asking costs.
In briefWho Drives Consensus? Influence Dynamics in Multi-Agent LLM Systems
SummaryThis paper studies which model most shapes the final consensus when several language models work together in one system. Across seven model families, opinions of other models were absorbed into those of the larger, stronger model, even though the task needs no pretraining knowledge or correct answer. The authors model agents as weighted particles on a simplex, arguing that larger models have greater weight and stronger in-context learning, which makes them the influencer.
Gouki Minegishi, Jingyuan Feng, Hiroki Furuta et al.
