In briefRevisiting cross-reconstruction to improve deepfake detection on unseen generators
SummaryThis paper addresses the problem that image forgery detectors often fail to generalize to unseen manipulation methods. The authors argue that the diversity of artifacts produced by different generation processes contains useful forensic cues, so their framework preserves this diversity through semantically aligned cross-generator reconstruction rather than forcing artifacts to align. They also add a masked frequency-aware reconstruction strategy that emphasizes manipulation-related residuals. Experiments on multiple benchmarks report improvements in both cross-dataset and cross-generator evaluations.
In briefSynthetic training for segmenting rare brain bleeding lesions with scarce data
SummaryThis paper addresses automated segmentation of cerebral microbleeds and cortical superficial siderosis, imaging markers of small vessel disease, which is hard due to few positive cases and voxel-level annotations. The authors propose a synthetic training framework that needs no real lesion annotations, generating image-label pairs from anatomical brain parcellations using clinical priors on lesion location and morphology. Evaluated on 10 cSS and 13 CMB cases, the proposed models outperformed classical filter baselines such as the Frangi filter.
In briefAI Assessment Sandbox Configurator: a framework for technical testing in AI regulatory sandboxes
SummaryThe EU AI Act requires Member States to set up AI Regulatory Sandboxes by August 2027 for supervised assessment involving authorities, experts and organisations. The paper derives 11 infrastructure and governance requirements and presents AI Assessment Sandbox Configurator, an open-source framework with a plug-in test catalogue, a shared data model, role-specific dashboards and audience-segmented reporting. An early pilot in a live sandbox engagement tested the harmonisation and reporting layers and contributed to an official Exit Report.
Alessio Buscemi, German Castignani, Daniele Pagani et al.
In briefFedFit: Federated fine-tuning of LLMs using compact vector banks and quantization
SummaryThis paper proposes FedFit, a method for fine-tuning large language models in federated learning while reducing communication costs. It rebuilds adapter matrices from two disjoint global vector banks, uses alternating optimization with Residual Spectral Aggregation to resolve LoRA aggregation conflicts, and adds blockwise quantization with error feedback. The authors provide theoretical convergence guarantees, and experiments on Qwen2.5 models show perplexity comparable to standard federated LoRA with compression ratios up to 100x.
In briefBalancing content and collaborative signals via graph-based Semantic IDs (GrIS)
SummaryThe paper reframes Semantic ID construction for generative recommendation as a recursive graph clustering problem rather than pure representation learning. It proposes GrIS, a unified framework in which RQ-VAE and RQ-KMeans appear as the special case of an empty graph. Two instantiations are explored: RecDMoN, using differentiable graph pooling, and RQ-GAE, which adds graph-aware representations and a graph reconstruction objective. Across multiple real-world datasets, GrIS reportedly improves over collaborative-filtering-aware state-of-the-art methods, with gains of up to 52% in Hit@10.
Aleksei Medvedev, Alejandro Ariza-Casabona, Steven Derby et al.
In briefStudying whether vision-language models stay reliable for self-driving when camera images degrade.
SummaryThis paper evaluates five vision-language models on four driving-related question-answering datasets, using single-frame, multi-view, multi-frame, and monocular inputs. It examines how visual corruption affects accuracy and the reliability of model confidence across models and input settings. The authors also test Visual Evidence Augmentation, an inference-time method, to see whether it improves reliability under degraded visuals. They report that it helps some models and datasets, but the gains are not consistent across all settings.
Manasa Mariam Mammen, Priyanka Mary Mammen, Zafer Kayatas et al.
In briefTuning when RUL predictions are made by optimizing multiple objectives
SummaryThis paper studies remaining useful life (RUL) prediction, noting that early and late prediction errors have asymmetric consequences, while standard hyperparameter tuning optimizes a single accuracy metric. The authors treat the optimization objective itself as a design variable and compare five architectures under single-objective R² maximization, single-objective NASA scoring minimization, and a joint multi-objective search. On the C-MAPSS benchmark, multi-objective optimization reduces directional imbalance between early and late predictions by about 33% at similar accuracy, while on the BackBlaze dataset the results show a substantial generalization gap.
In briefExact distinguishability of non-Markovian decision models from data
SummaryThis paper studies whether data gathered under a fixed behaviour policy can tell apart two candidate Regular Decision Processes that model non-Markovian environments. The authors prove that for observationally equivalent candidates, posterior odds equal prior odds at every sample size, and they verify these results formally in Lean 4. They also characterize this equivalence exactly and introduce PEC, an algorithm that decides it in time linear in the product automaton's size.
In briefAI Risk Observatory: Tracking AI Disclosures in Annual Reports
SummaryThis paper uses large language models to analyze 9,821 annual reports from 1,362 UK listed companies from 2020 to 2025, examining how firms disclose their response to AI. The share of reports mentioning AI risk rose from 2.8% to 41.2%, but only 4.3% of 2025 reports contain substantive AI risk disclosure. Disclosure is lower among AIM-listed firms and some critical infrastructure sectors, and harm disclosures are nearly absent.
In briefLangevin-Informed Transfer Learning: Recovering Dynamics from Black-Box Feedback
SummaryThis paper proposes LITL, a framework that recovers target Langevin dynamics from biased source samples using only black-box feedback. It learns the leading spectral structure of the target generator and the projected drift through Dirichlet representation learning, enabling kinetic reconstruction and slow-manifold gradient estimation. The authors give finite-sample guarantees in Sobolev norms and test the method on molecular simulations, generative models, and latent steering of trained neural networks.
WhyApplies Dirichlet representation learning to recover Langevin dynamics from biased samples with finite-sample Sobolev guarantees; novel and well-supported, though impact is somewhat niche.
Vladimir R. Kostic, Karim Lounici, H\'el\`ene Halconruy et al.
In briefUsing LLMs to Automatically Turn Text Rules into Logic Constraints for Neuro-Symbolic Predictors
SummaryNeuro-Symbolic predictors add logical constraints to neural networks so outputs follow domain knowledge, but writing those constraints by hand requires experts. This paper tests whether large language models can automatically convert textual knowledge into symbolic formulas. The authors introduce auto-nesy-bench, a benchmark for measuring formalization quality and its effect on prediction accuracy. They find the generated formulas often resemble expert-written ones, and syntactically valid formulas yield high-quality downstream predictions.
In briefSuperMotion: Source-Preserving Denoising for Text-Driven Human Motion Editing
SummaryThis paper presents SuperMotion, a diffusion-based framework for editing human motion from text instructions. At each reverse denoising step it reuses the aligned source motion, blended through a learned preservation gate, to keep temporal detail from being smoothed away. It reaches 33.20% full-pool R@1 on MotionFix, and ablations attribute the gain to the preservation gate.
In briefFedMIX-P: Mixing Local and Global Preconditioners for Federated Vision and Language Model Training
SummaryThis paper proposes FedMIX-P, a federated learning method that mixes shared and local preconditioners at every local training step to counter update bias caused by differing client geometries. Under smooth nonconvex objectives with stochastic gradients and partial participation, the authors establish an O(R^{-1/2}) stationarity bound and show via a two-client counterexample that fixed positive mixing can preserve a nonstationary fixed point. Experiments with SOAP, Sophia, and Muon variants on vision and language tasks report gains over local optimizers, including up to 19.47 percentage points of accuracy and lower validation loss for 60M–350M language models.
In briefHow Audit Rules Shape Honest Factor Explanations in LLMs
SummaryThis paper studies whether large language models truthfully report how much each input factor influenced their outputs when verification is limited by budget. The authors model this as a verification game and find that audits depending on the report create an incentive to suppress important factors. Report-independent auditing removes this incentive, making truthful reporting preferable. The framework is tested with the Counterfactual Brier Score on four NLP benchmarks, where a synthetic agent matches the theory and real LLMs respond similarly.
In briefDecision Titan: test-time training to give offline reinforcement learning agents longer memory
SummaryThis paper applies the Test-Time Training (TTT) framework, which stores memories in network parameters, to offline reinforcement learning. The authors add TTT layers to a Decision Transformer to create the Decision Titan and test it on the X-Maze environment, a sequential memory benchmark. They visualise gate values over time to study how the memory mechanism learns. The model learns dependencies up to 20 times longer than its context window and generalises to lengths 1.7 times the training data, though temporal generalisation depends on the time embeddings used.
In briefVoxelSynth3D: An interpretable 3D image-domain method for reducing metal artifacts in CT, with a synthetic test set
SummaryThis paper addresses metal artifacts that obscure bone-implant and nearby soft-tissue boundaries in postoperative musculoskeletal CT. It proposes VoxelSynth3D, a training-free method operating on reconstructed 3D CT images that combines support masking, tissue synthesis, deviation gating, and restricted edge refinement. The authors also build Synthetic CLINIC-Metal, a paired synthetic benchmark derived from metal-free CTPelvic1K volumes. On 40 held-out cases, RMSE dropped from 801.48 to 786.18 HU, improving every case, while clean-edge agreement decreased near metal.
Amritesh Banerjee, Abdul Basit, Renil Renji Joseph et al.
In briefGAW-PO: Preference Optimization with Gradient-Aligned Token Weights
SummaryThis paper introduces GAW-PO, a variant of Direct Preference Optimization that reweights each rejected token according to whether penalizing it conflicts with the preferred update direction. Tokens whose gradients align with preferred behavior receive a weaker negative penalty, while conflicting tokens keep a stronger one. Across 11 benchmarks covering mathematics, reasoning, coding, and question answering, it achieves the highest average score among the evaluated methods, 0.97 points above standard DPO and 0.65 above the strongest baseline. The authors also report that GAW-PO keeps improving as the DPO parameter β decreases, while standard DPO degrades sharply.
Andreea Dutulescu, Stefan Ruseti, Mihai Masala et al.
In briefFedCKA: Layer-wise personalization for federated 3D driving perception
SummaryThis paper addresses the problem that 3D object detectors for intelligent vehicles can degrade under domain shifts such as time of day, location, or weather. It proposes FedCKA, a federated learning method that uses Centered Kernel Alignment to measure layer-wise feature similarity between each client model and the global model, then shares only representation-consistent layers. On a unified multi-domain benchmark built on nuScenes, FedCKA improves average NDS by 7 percentage points over the strongest baseline.
In briefPost-training trade-off: RL may reduce how many tasks models can solve
SummaryThis paper studies how reinforcement learning post-training changes large language models on agentic tasks. The authors find that pre-trained models with a light inference harness often solve more tasks under repeated sampling (pass@K) than their post-trained counterparts, despite lower single-shot accuracy (pass@1). They introduce Sharpening Tax, a diagnostic metric for the loss in test-time scalability after post-training. They also propose posterior-tempered group sampling (PTGS), a Bayesian sampler that adjusts temperature per prompt, which pays a smaller tax during RL training.
WhyProposes Sharpening Tax and PTGS, validated across 42 base/post-trained cases and three agentic benchmarks, offering a notable new perspective on post-training trade-offs.
In briefMCRI: A Four-Dimensional Framework for Evaluating Agent Skills
SummaryThe paper proposes MCRI, a four-dimensional framework for analyzing agent skills, and implements it as MCRI-Eval, an LLM-based evaluation method. It evaluates 63,812 public skills from the OpenClaw skill Hub, running 58,275 skill-conditioned executions on BigCodeBench, BFCL-Fundamental, and Mind2Web. MCRI-Eval scores correlate positively with community popularity and give the highest downstream ranking agreement, while also improving top-1 skill selection on all three benchmarks.
In briefLearned End-to-End Guidance Schedules for Diffusion Models: a method that learns how strongly to guide image generation over time, so diffusion models can meet requirements with fewer sampling steps.
SummaryThe paper proposes LEEGS, which learns a time-dependent guidance scale schedule by minimizing a guidance loss over a small set of examples with stochastic gradient descent. To reduce the cost of backpropagating through guided sampling, it uses a gradient approximation that cuts training time by a factor of 4. The authors evaluate it on image inpainting, noisy inverse problems, face-ID-guided generation, and forward and inverse PDE problems. They report better results than baselines at equal budgets of 50 or 100 function evaluations, or matching constant guidance with only 10% of the steps.
In briefVTR-Bench: A Benchmark for Visual Text Rendering in Video Generation
SummaryThis paper introduces VTR-Bench, a benchmark that tests whether video generation models render text in scenes correctly, covering 300 prompts across five scenario categories such as advertisements and scientific videos. It uses an automated evaluation pipeline with human alignment to measure text fidelity and scene and motion requirements. Experiments on 11 state-of-the-art models show widespread difficulty in scene text rendering, with the best model reaching a word error rate of 0.250. The authors also propose a Keyframe-Guided Agentic Framework that refines outputs iteratively using visual feedback.
WhyIntroduces a systematic benchmark for visual text rendering in video generation, with 300 prompts, automated evaluation, and experiments on 11 models.
In briefOpenMTB-Audit: Testing AI Over-Refusal and Oncologist Views in Molecular Tumor Board Safety
SummaryThis paper introduces OpenMTB-Audit, an open-source benchmark of 500 synthetic non-small cell lung cancer cases for evaluating large language models used in molecular tumor boards. It finds that all tested model configurations widely over-refuse, failing to keep the Partially Supported label in most true cases, which yields high safety scores through label collapse. The authors then present MTB-AuditAgent, a deterministic seven-module framework that reduces over-refusal to 6.7% and reaches 91.2% accuracy.
In briefSALD: A self-referenced training method for diffusion models that uses two noise levels of one model, with no external teacher
SummaryThe paper proposes SALD, a training framework that evaluates each image-caption pair at two noise levels using the same diffusion model. The lower-noise path, evaluated without gradients, serves as a reference, while the higher-noise path provides the training gradient, adjusted by the difference between the two path errors. Its components are Advantage-Guided Diffusion, Temporal Advantage Memory, and Spectral Advantage Decomposition. Experiments across multiple architectures and datasets show consistent improvements in generation quality, and component ablations quantify each part's contribution.
In briefTesting whether routing entropy signals when a Transformer's prediction may be wrong
SummaryThis paper checks whether per-example routing entropy in Attention-Residual Transformers (Swin-Tiny and DeiT-Small trained on CIFAR-10/100) tells us about prediction correctness beyond the model's own confidence. No test survived multiple-comparison correction, and apparent effects did not replicate across training seeds. The authors conclude the results show control-dependent gains and incomplete estimator recovery, not that conditional routing information is absent.
In briefLetting attention heads communicate across edges
SummaryThe paper studies matrix-valued transport maps in Sheaf Neural Networks, connecting them to multi-head attention through quiver representations. It shows that standard multi-head attention only lets each source head contribute to the matching receiver head, and proposes Topological Attention (Top-A), which learns edge-dependent routes between different heads. Experiments on relational reasoning, heterogeneous graph learning, and algorithmic reasoning find cross-head transport helps most when tasks need interaction-dependent transformations, while heterophily alone gives no systematic benefit.
Riccardo Ali, Alessio Borgi, Mario Severino et al.
In briefNo Model Needed: Text Entropy Rate Filtering to Curb Iterative Fine-Tuning Collapse
SummaryThis paper studies model collapse, where repeated fine-tuning on synthetic data narrows output diversity. The authors propose a model-free training-data filter based on the non-parametric Kontoyiannis entropy rate estimator h_k, computed from raw text via match-length statistics. In a six-generation QLoRA experiment on Llama-3.1-8B, h_k filtering improved text-diversity metrics, while a logprob-based baseline gave no significant benefit.
In briefQ-SPT: Learnable query-based compression for low-frame-rate speech tokenization
SummaryThis paper proposes Q-SPT, a low-frame-rate dual-stream speech tokenizer for speech language models. It uses separate learnable query-based compressors for semantic and acoustic streams, with an autoregressive text loss supervising the semantic compressor to keep linguistic content. Experiments report the best reconstruction among compared codecs at the same frame rate, and the best speech recognition accuracy and text-to-speech perceptual quality in downstream speech language models.
In briefHuman audit of web agent evaluation on WebArena-Lite: outcomes and trajectories
SummaryThis paper manually audits all 165 WebArena-Lite tasks, comparing GPT 5.5 and an untrained Qwen3.5 9B model under six evaluation conditions. Human review corrects success cases the automatic evaluator missed and identifies the first consequential error in each trajectory. The authors also study MASM, a mechanism that maintains explicit execution state, and Guide Text, which gives task-relevant procedural guidance. The findings suggest final scores alone give an incomplete picture of web agent behavior.
Chengguang Gan, Zimeng He, Yoshihiro Tsujii et al.
In briefIs the Persona Still Speaking? Identity Reversion in Persistent AI Agents
SummaryThe paper studies a persistent personal AI agent that, after repeated automated heartbeat checks, stopped responding as its assigned persona and identified itself as the underlying harness. Controlled experiments show that persona continuity depends jointly on system-level anchoring and conversational context, and that restoring the anchor reversibly restores the persona. The authors distinguish represented identity, which can remain in conversation history, from enacted identity, the identity bound to "I."
In briefPredicting listener backchannels in multi-party meetings: diagnosis, benchmark, and performance limits
SummaryThis paper extends backchannel prediction, which has mostly been studied in two-person conversation, to multi-party meetings. The authors build a benchmark from the AMI corpus with person-disjoint held-out splits. A dyadic model applied zero-shot to meeting audio performs at chance (AUROC 0.499), but retraining improves it to 0.751. Listener conditioning helps only for listeners seen in training, and the authors attribute this partly to speaker identity being entangled with useful backchannel cues.
In briefFiVOS: A fish segmentation method using interactive video segmentation with filters
SummaryThis paper proposes FiVOS, an improved interactive video object segmentation method for tracking fish in aquaculture videos, and builds two fish-specific datasets. Its mask block filter detects and corrects erroneous propagated mask blocks early, and serialized noise filters remove further mask noise to improve robustness. According to the authors, experiments show FiVOS reaches state-of-the-art performance on fish video segmentation tasks.
In briefALFRED: A requirement-driven open-source mobile robot arm for repeated plant monitoring
SummaryThe paper presents ALFRED, an open-source mobile manipulator built from commercial parts to repeatedly observe the same plants outdoors. It was refined through four builds against six requirements: durability, modularity, repairability, sensing reach, endurance and reproducibility. Analysis of the last three builds shows the usable share of reachable arm poses rising from 34.0% to 66.1%, and only the final build keeps the frame-mounted LiDAR's view clear in both directions. Over a year it completed 528 monthly forest traversals without missing a scheduled collection.
Ciar\'an Miceal Johnson, Christopher Quail, Garry Ellard et al.
In briefZero Flux: A flow-based way to compare high-dimensional discrete distributions
SummaryThis paper introduces the Zero Flux criterion, a discrepancy measure for comparing two high-dimensional discrete distributions. It extends flow-matching-based comparison, previously used for continuous distributions, to the discrete domain using local probability fluxes. Under independent coupling, all local fluxes vanish at the midpoint if and only if the two distributions are identical. The discrepancy can be estimated from samples, with finite sample error bounds, and experiments on synthetic and real categorical data show reliable recovery of sparse dependence signals and stable tracking of distribution shifts.
In briefWhen a Second AI Reviewer Helps: Testing Cross-Model Checks of AI Output
SummaryThe paper asks when a second review by a different large language model helps catch errors. In a controlled experiment with 30 artifacts, 150 planted errors, and 900 review sessions, a top-tier cross-model reviewer performed not significantly differently from a same-model reviewer in a fresh session. The two reviewers found partly different errors, but the study could not separate model difference from reviewer capability.
In briefLESS: A fast evolutionary method for finding neural network designs in minutes
SummaryLESS (Lightweight Evolutionary Supernet Search) is a low-cost neural architecture search method that combines a brief fair hard-path warm-up with discrete search using CMA-ES. On NAS-Bench-201 it reaches 93.189±0.467% CIFAR-10 test accuracy in 409.1 seconds, within 0.04 points of FairNAS at about 1/24 of its search time. The frozen configuration transfers without tuning to CIFAR-100, ImageNet16-120, and the larger DARTS space, where searches take about 43.5 minutes on one GPU.
In briefNextMe-800: Predicting Personal Behavior from Months of First-Person Video
SummaryThe paper introduces NextMe-800, an approximately 800-hour first-person dataset recorded by one volunteer over 126 days, with images, gaze, and audio captioned at five abstraction levels. The authors frame personal action anticipation as open-vocabulary K-step sequence prediction. They also build NextAct, a 1,500-point benchmark combining NextMe-800 with the multi-person EgoLife dataset, and evaluate models using an embedding-based soft edit distance.
In briefTight bounds on transition time for separable logistic regression at the edge of stability
SummaryThis paper studies logistic regression on linearly separable data trained by gradient descent with a large constant step size. In such settings the loss can first oscillate in an Edge of Stability phase before decreasing steadily. Prior work gave a step-size-independent bound in dimension two and conjectured the same in higher dimensions. The authors disprove this conjecture, showing the worst-case transition time scales as the logarithm of the step size raised to a power depending on sample size and dimension.
WhyDisproves a conjecture by proving a tight step-size-dependent transition time bound with matching hard instances.
In briefRethinking Probability-Based Reinforcement Learning via Posterior Concentration
SummaryThis paper studies how probability-based rewards for verifier-free reinforcement learning become unreliable on long reasoning traces. The authors identify the Posterior Concentration Phenomenon, where the probability of a reference answer collapses to a low-variance interval as traces grow longer, making rewards nearly indistinguishable. They propose RLCPR, which combines uncertainty-aware data sampling and concentration-aware regularization, and report gains of up to 4.0% over a state-of-the-art baseline on six of seven benchmarks along with higher token efficiency.
In briefStreaming algorithms for robust max-min diversity selection
SummaryThis paper studies max-min diversification, which selects k points from a metric space to maximize their minimum pairwise distance, a goal that is sensitive to noisy points. It identifies three weaknesses in an existing robust streaming algorithm: offline coreset construction needing memory linear in n, possibly returning fewer than k points, and only probabilistic outlier exclusion guarantees. The authors propose a deterministic coreset-based algorithm that, under a natural inlier-outlier separation assumption, returns exactly k inliers with a (2+ε)-approximation. For wide parameter ranges, its memory does not depend on n.
Andrea Pietracaprina, Geppino Pucci, Stefano Zanon
In briefCLEAN: Consistent Incremental Cognitive Diagnosis as Concepts Expand
SummaryThis paper addresses cognitive diagnosis in intelligent education, where new items introduce unseen concepts and existing incremental models forget past diagnoses. The authors propose CLEAN, which freezes historical diagnostic functions, uses orthogonal column masking to block gradient interference, and adds expandable branches for new concepts, structurally guaranteeing invariance of past diagnoses. Experiments on three large-scale educational datasets report zero representation drift, with old-item metrics identical to static anchors and new-item performance competitive with or better than strong continual-learning baselines.
In briefReusing Outdated Model Representations After Deployment
SummaryThis paper addresses deployed deep neural networks whose task requirements change, making part of their output space obsolete. It proposes Deep Repurposing (DR), a post-hoc framework that estimates the latent geometry of obsolete and retained regions, removes components supporting obsolete behaviour, and reallocates retained-compatible evidence through an analytic repair map without gradient updates. The authors report that across classification benchmarks DR matches or exceeds unlearning and editing baselines in retained accuracy, eliminates obsolete predictions, and adapts up to 60 times faster.
In briefUncertainty-guided human-in-the-loop correction for more accurate glioma segmentation with less manual effort
SummaryThis paper proposes a human-in-the-loop framework that uses test-time augmentation uncertainty and topological filtering to flag high-risk structural errors in automated brain glioma segmentation. On an out-of-distribution clinical cohort (N = 362) under a simulated human oracle, Whole Tumor Dice rose from 0.891 to 0.914 and HD95 fell from 5.82 mm to 4.76 mm. Tumor Core HD95 dropped from 17.96 mm to 14.83 mm. The median interactive workload was 11.3% of the target volume. The authors describe this as a simulated upper bound without real-world cognitive friction.
In briefA multi-agent AI system that interprets personal health checkup results and gives guidance.
SummaryThis paper presents a multi-agent large language model system that identifies several intents in a health checkup query, assigns each to a task-specific agent, runs them in parallel, and combines their outputs. On 120 Korean compound queries using synthetic records, the multi-agent version raised the weighted LLM-judge score from 1.695 to 1.797 (p = 0.027), with gains in usefulness and consistency. Medical safety showed no significant difference, critical failure rates were similar, and latency and cost increased by 1.31 and 2.02 times.
In briefIdentifying Code-Switched Speech Languages as Multi-Label Set Prediction
SummaryThis paper addresses language identification for code-switched speech, where people mix languages, which can slip past monolingual filters used to build large speech corpora. The authors propose a set generator that directly outputs the set of languages in an utterance and compare it with atomic-pair and score-based baselines. The oracle Top-k baseline performs best, while thresholding fails to separate code-switched from monolingual speech. The proposed model predicts the correct number of languages on unseen pairs but has lower exact set accuracy than oracle Top-k.
Shunsuke Mitsumori, Matthew Wiesner, Shigeo Morishima et al.
In briefibUMAP: A Coherent, Scalable Field-Based Evaluation Method for UMAP Optimization
SummaryUMAP uses stochastic negative sampling for scalable layout optimization, but the sampling order can make embeddings unstable across reruns. This paper proposes ibUMAP, which computes attraction and repulsion synchronously from a shared embedding snapshot and evaluates a degree-weighted repulsive field efficiently with an interpolation-based FFT scheme. Reported results show median speedups over umap-learn on CPU and over cuML on GPU, greater run-to-run stability, and measurable trade-offs in local-global fidelity.
In briefDRelay: A Method That Repairs Early Draft Errors in Parallel Speculative Decoding for LLMs
SummaryThis paper addresses parallel speculative decoding for large language models, where a single early selection error can stop later correct predictions from being used and limit the accepted prefix length. The authors propose DRelay, which uses global information from the whole draft block to selectively repair candidate choices before target-model verification. Across eight benchmarks, DRelay improves average acceptance length and end-to-end decoding over DFlash, Domino, and DSpark, with reported SGLang speedups of 14.7%-16.8%, 8.7%-9.3%, and 8.1%-9.3% respectively.
In briefHow processing settings affect high-accuracy drone mapping
SummaryThe study tested 768 processing setups on ten drone datasets collected over 1.5 years in a 220-hectare area. The best setup reached a 16 mm 3D error, while the weakest reached 303 mm. Ground control point count, extra camera calibration corrections, and the PPK GNSS method for camera positions had the largest effects.
Pawe{\l} \'Cwi\k{a}ka{\l}a, Edyta Puniach, El\.zbieta Pastucha et al.
In briefA Deterministic, Auditable AI Security Risk Assessment Framework Using MITRE ATLAS-Aligned Rules and Formal Verification
SummaryThis paper presents an evidence-driven framework for assessing AI system security as a deterministic, auditable decision function. Engineering artefacts are normalised into a standard control taxonomy and mapped via a pinned MITRE ATLAS snapshot to technique-level feasibility and impact outcomes. The evaluator's boundedness, totality, semantic consistency, and monotonicity are formally verified. Tested on five public open source AI projects, the framework shows feasibility decreasing as observable controls are strengthened, while worst-case residual feasibility remains when core controls are absent.
Yixuan Huang (1), Basel Halak (1), Boojoong Kang (1) ((1) University of Southampton et al.
In briefDistilling tabular foundation models into fast, dataset-specific predictors
SummaryTabular foundation models make accurate predictions via in-context learning, but they must repeatedly process labeled data, which makes inference costly. This paper studies how to distill their predictive ability into lightweight student models, examining how to build teacher supervision and whether broader query coverage helps. The proposed recipe uses the full labeled training set as teacher context and trains students only on teacher predictions for observed and synthetic queries. On TabArena the students outperform tuned-and-ensembled supervised counterparts by 57-98 Elo points, and on TALENT they improve most datasets while achieving 3.0-21.6 times faster median inference than their teachers.
WhyDistills tabular foundation models into lightweight students with a new recipe, outperforming supervised baselines on TabArena and TALENT across hundreds of datasets.
In briefMWOP: A Method That Prunes Attention and Feed-Forward Computation Separately by Modality to Speed Up Multimodal Models
SummaryMultimodal large language models are costly to run when processing long sequences of images and text. This paper proposes MWOP, which prunes the visual-to-visual, text-to-visual, and text-to-text attention paths separately within each layer, and selects feed-forward channels separately for visual and textual inputs. Pruning uses a first-order Taylor criterion followed by LoRA-based recovery training. On LLaVA-OneVision-7B, MWOP alone gives a 1.6x prefill speedup while retaining 99.7% average performance across 12 benchmarks, and combined with token compression methods it raises their speedups to 2.9x and 2.7x.
WhyPruning attention paths by modality pairs and separate visual and textual FFN channels refines prior compression; validated on 12 benchmarks with clear prefill speedups.
In briefLearning ab initio phase-field models: a framework that derives mesoscale equations from molecular dynamics using neural networks
SummaryThe paper presents a framework for learning ab initio phase-field models, deriving a mesoscopic equation from molecular dynamics via a Mori-Zwanzig projection onto species-density fields. The nonlocal free energy and mobility are parametrized by neural networks and trained on short trajectories from machine-learning interatomic potentials of ab initio accuracy. The method is demonstrated on an iron-boron melt and on hydrogen-helium mixtures under planetary conditions.
WhyDerives a mesoscopic phase-field equation from molecular dynamics via Mori-Zwanzig projection, learning nonlocal free energy with neural networks; validated on two systems, though comparisons are limi
In briefGeneralization as Stability: Multi-Axis Evaluation of LLMs
SummaryThis paper argues that generalization in large language models should be judged by whether outputs stay consistent and semantically stable when the same input is rephrased, rather than by aggregate accuracy. The authors introduce the Stability-Aware Generalization Objective (SAGO), which measures how model behavior changes across variations along dimensions such as generation consistency, internal activations, confidence, and response mirroring. They report that many common models show statistically significant, consistent generalization instability, that different behavioral axes capture independent failure modes, and that cross-dataset variation can reverse model rankings.
Nagham Omar, Mahmoud Jabarin, Maya Rozenshtein et al.
In briefSHAMS: A Benchmark for Levantine Arabic Pronunciation Using Aligned Audio
SummaryThis paper introduces SHAMS, a benchmark of 1,300 utterances from open audio corpora, balanced across five Levantine Arabic varieties. Each utterance has four aligned layers: audio, unvocalized orthography, diacritized text, and phonetic transcription. The benchmark supports tasks such as diacritization, grapheme-to-phoneme conversion, speech recognition, and audio-to-phoneme conversion. The authors evaluate open and proprietary models on these tasks and release the benchmark publicly.
In briefLeast-time Gradient Flow: Studying an Optimization Method That Prescribes How Fast Risk Decreases
SummaryThe paper proposes a gradient flow dynamics that forces the risk to decrease at a prescribed rate, regardless of the shape of the loss landscape. To make the minimum-time problem well posed, the authors add a regularization term. They prove that the optimal solution exists, is unique, and is a linearly scaled cycloid, and they describe its behavior near zero risk.
Alessandro Betti, Marco Gori, Stefano Melacci et al.
In briefWhy train-validation gaps emerge: update-pressure dynamics in pretrained models
SummaryThis paper proposes that during continued fitting of pretrained models, update demand shifts from broadly reusable features toward narrower features that transfer poorly to held-out data, which widens the gap between training and validation performance. The authors test this account with a controlled simulation, NLP models (RoBERTa, DeBERTa, Qwen) on six datasets, and ResNet-18 on three vision datasets. They report that training-probe dispersion readouts are positively associated with the accuracy gap across the runs studied.
In briefSpikeMoE: Brain-Inspired Competitive Routing for Flexible Spiking Mixture-of-Experts
SummaryThe paper proposes a spike-based k-WTA router, inspired by competition-inhibition in the hippocampal CA1 region, that selects Top-K experts from discrete spike counts. It builds SpikeMoE, which combines neuron-scale spiking dynamics with model-scale expert selection, and adds a two-stage module for missing-modality inputs. Experiments on vision, language, and multimodal benchmarks report state-of-the-art results among SNN baselines and performance matching or exceeding ANN counterparts.
In briefBeyond Memory: Explicit Belief States for Long-Horizon LLM Agents
SummaryThis paper introduces PoS, an inference-time framework that builds and continually maintains explicit belief states for LLM agents, combining an estimate of the current world state with unresolved task requirements. PoS checks belief consistency and monitors task progress to detect Belief Trapping, where the agent keeps acting without meaningful progress, and then applies tailored recovery. Experiments on four benchmarks show PoS achieves the highest overall performance with all three LLM backbones.
In briefOptimal Transport and Reinforcement Learning: A Survey
SummaryThis survey reviews how optimal transport (OT) is used inside reinforcement learning objectives and algorithms. OT compares distributions by the cost of moving probability mass from one to another, which can work better than common divergences when distributions overlap weakly. For each method, it identifies the role of OT, the distributions compared, the OT formulation, and how temporal structure is handled. It also discusses cost design, computational challenges, and open problems.
Yujie Zhu, Charles A. Hepburn, Matthew Thorpe et al.
In briefEstimating when a pretrained object detector's bounding boxes are unreliable
SummaryThis paper proposes a lightweight post-hoc evidential meta-model that learns when predicted object localisations are uncertain while the base detector stays frozen. It uses saliency-guided image modifications to build an increasingly challenging curriculum, with targets combining localisation error, modification level, and prediction instability. The authors state that no detector changes are needed and original outputs are preserved. Across adversarial attacks, TP-FP AUROC improves by up to 22% over the strongest comparator in some cases, with in-distribution performance maintained.
In briefCoFlow: smoother flow matching by pushing generation paths apart during training
SummaryThis paper addresses trajectory crossing in Flow Matching, which makes the target velocity field have large local Lipschitz constants that are hard for neural networks to fit and cause integration errors in few-step inference. The authors propose CoFlow, which applies a contrastive learning idea by adding a repulsive drift term derived from a stochastic differential equation, pushing positive trajectories away from negative ones. Experiments on ImageNet 256x256 report that CoFlow significantly reduces FID compared to standard Flow Matching in few-step inference (e.g., 20 steps) with no added training overhead.
In briefContrastive Attention Reduces Spectral Bias in Spiking Transformers
SummaryThe paper finds that spiking neurons and spiking self-attention both act as low-pass filters, dissipating high-frequency components, which contributes to the performance gap with artificial neural networks. It proposes Spiking Contrastive Attention (SCA), which enhances high-frequency information through global contrastive aggregation and local differential refinement. Experiments report consistent gains across image classification, semantic segmentation, and event-based tracking, with lower complexity than original spiking self-attention.
In briefReusing past samples in PPO: when and how it helps, explained
SummaryThis paper studies whether reusing past samples improves Proximal Policy Optimization (PPO), a widely used reinforcement learning method. The authors build two variants, wPPO-U and wPPO-BH, within a multiple importance weighting framework that reuse only samples from a window of recent iterations. They derive policy improvement lower bounds for both and empirically analyze when and how this reuse improves sample efficiency or final performance on continuous control tasks.
Alessandro Montenegro, Riccardo Venturelli, Marco Mussi et al.
In briefAF-Muon: An AdamW-free Muon optimizer for models with tied vocabulary embeddings
SummaryThis paper proposes AF-Muon, an extension of the Muon optimizer that does not rely on AdamW. It keeps Muon's matrix update for hidden weight matrices, uses a support-aware finite-cap linear minimization oracle for tied vocabulary tables, and uses RMS-normalized updates for one-dimensional auxiliary parameters. Across nine tied-token settings, AF-Muon improves mean validation loss and perplexity over Hybrid Muon and a SCION-style Sign endpoint, while saving around 20% optimizer-state memory and adding about 1% step-time overhead.
Arash Lagzian, Paniz Halvachi, Junming Zhang et al.
In briefLLM-assisted discovery of typed semantic links between ontologies
SummaryThis paper proposes an end-to-end framework that automatically discovers intra-domain and inter-domain relationships between ontologies, combining domain-adapted DistilBERT embeddings, clustering-based pre-filtering, and GPT-4o relationship generation. Applied to ReproduceMeON, a network of 33 ontologies, it reduces about 800k concept pairs to 95k candidates. Expert validation of 429 generated relationships gives 80.19% precision and an F1 of 0.890, well above the best similarity-based baseline (F1 0.581).
In briefAiSearch: Interactive Multimodal Image and Video Search Using Vision Language Models
SummaryThis paper presents AiSearch, a multimodal retrieval framework that uses the zero-shot abilities of Vision Language Models (VLMs) to search images and videos with natural language. Users can refine results interactively through feedback to better match their intent. The framework also allows visual benchmarking across multiple VLMs so users can choose the most suitable model for their task.
Ali Koksal, Mei Chee Leong, Vicky Sintunata et al.
In briefLearning sound localization from camera movement in real videos
SummaryThis paper proposes a method for learning binaural sound localization using egomotion, the camera's own movement, as a supervisory signal. As the camera moves during a video, the direction to a sound source changes, and the audio model is trained to predict sound directions consistent with camera motion estimated by traditional multi-view geometry methods. This weak but plentiful supervision is combined with conventional binaural cues. The authors also introduce a real-world audio-visual dataset with egomotion, and show that the model learns from real data and performs well on sound localization tasks.
In briefRobust Evidential Learning via Latent Consistency
SummaryThe paper proposes CLEAR, a lightweight, task-agnostic post-hoc method that improves the robustness of Evidential Deep Learning without retraining or changing base predictions. It uses held-out calibration data to characterise the group-conditioned geometry of the latent space, then measures latent conflict from perturbation views at inference to reduce evidential strength for unsupported inputs. On ImageNet→CUB, CLEAR improves OOD and adversarial AUROC by 8.29 and 5.01 points and runs 17.4 times faster than competing post-hoc methods, while preserving performance on classification, regression, and object detection.
In briefPRISM: A Framework for Measuring and Refining Multimodal Analogies Using Category Theory
SummaryThe paper introduces PRISM, a framework that represents analogies as relational mappings grounded in category theory and measures their alignment with a pullback score. On the AnaloBench benchmark, selecting the correct analogy using only this score achieves 82.5% accuracy. An iterative refinement loop uses the score as feedback to revise generated images, and human participants preferred the refined output in 57.65% of pairwise comparisons. The authors also note that refinement can favour visually crowded compositions over genuinely deeper relational correspondences.
Mirella Zeisler, Ojas Shirekar, Mircea Lic\v{a} et al.
In briefGacha Decoding: Getting Diverse Outputs by Following Instructions
SummaryThis paper introduces Gacha Decoding, an inference-time method for producing more varied language model outputs, with diversity that grows as the model's instruction-following improves. Tested on open-ended tasks such as chat, creative writing, image-generation planning, and protein design, it outperforms prior diversity methods at equal quality. The approach treats diversity as an instruction-following problem by combining the model with randomness from an external random number tool.
WhyNovel idea of treating diversity as instruction following with external randomness, showing strong sample-efficiency gains across several open-ended domains.
In briefSupraTITO: A transferable generative model that simulates how peptides self-assemble over long timescales
SummaryThe paper introduces SupraTITO, a generative molecular dynamics framework for supramolecular systems, demonstrated on peptide self-assembly. It learns transfer operators conditioned on peptide sequence, molecular topology, and periodic geometry, so configurations can be propagated over much longer intervals than a standard MD step. On a dipeptide benchmark it generalizes to held-out sequences and keeps molecules intact over long rollouts. Compared with direct ensemble prediction trained on the same data, it more accurately reproduces assembly structures and their temporal evolution.
In briefGeneration Provenance Before Behavior Attribution: Auditing Synthetic Speech Datasets
SummaryThe paper proposes a generation-provenance framework that links each synthetic speech item's source specification, generated content, waveform, target, fact requirements, quality signals, review history, and manifest identity. It audits this approach in a private Japanese care-handoff pipeline covering 113 assets, about 1.552 hours of synthetic speech. The authors argue provenance is necessary but not sufficient for behavior attribution, which still requires frozen training runs and intervention or influence evidence.
In briefMinimax-optimal regret for causal logistic bandits under counterfactual fairness
SummaryThis paper studies causal logistic bandits where a learner must satisfy counterfactual fairness constraints but observes only factual rewards. It shows that some coverage condition is necessary, and under a weaker full-rank condition it defines an information scale V⋆ with matching lower and upper regret bounds. The bounds agree in their leading dependence on T, V⋆ and min{log K,d}, up to logarithmic factors.
In briefShared Interaction Schema for Predictive and Generative Personalization
SummaryThe paper proposes an action-on-item schema that pairs a mapped interaction role with a content embedding, so one set of update parameters can process user histories from movies, news, and dialogue. On PENS, the same history encoder supports both next-news prediction and personalized headline generation. In a PENS-to-MovieLens transfer experiment, a frozen source-trained core outperforms an identically structured random core by 15.23 MRR points after fitting the target consumer.
Parthiv Chatterjee, Kashish Kanjaria, Vashisth Purani et al.
In briefLearning commute-time-preserving world models for planning
SummaryThe paper introduces Commute-Time-Preserving World Models (CTWMs), which let agents plan in latent space by reaching goals with representations that mirror environment commute times. It argues that existing self-supervised methods that encourage isotropic representations distort the eigenvalue-dependent scaling needed for an accurate graph Laplacian embedding. CTWM combines a latent displacement predictor with a log-determinant regularizer to prevent collapse, and is shown to recover the correct scaling under reversible deterministic dynamics. In simulations on complex continuous goal-reaching benchmarks, it matches or outperforms the task-agnostic baseline LeWM while using half the parameters.
Michael Hauri, Peter Buttaroni, Fabian A. Mikulasch et al.
In briefFrom Redundancy to Minimality: Reducing Learned Piecewise-Linear Dynamics Guided by Fixed Points
SummaryThis paper studies learning nonlinear dynamical systems from time series using almost-linear recurrent neural networks (AL-RNNs), where only some units use ReLU nonlinearities. The authors propose a fixed-point-guided hierarchical reduction that starts from a network with more ReLU units, progressively linearizes some of them, and retrains each reduced model from its parent. On the 3-scroll Chua system, this learn-reduce-retrain approach raises the seed-level success rate of obtaining minimal realizations from about 20% with direct training to about 71% at the same final nonlinear capacity.
In briefLLM-Driven Multi-Agent Control for Smart Manufacturing: A Study on Skill-Based Factory Automation
SummaryThe paper proposes pairing each factory module with an LLM-based agent and an MCP tool server that exposes the module's skills via OPC UA, with agents coordinating over MQTT using real-time factory state. Three agent architectures were compared across nine production challenges in a simulated six-module hexagonal factory. The monolithic and peer-to-peer designs reached the highest mean solve rate of 93%, while the orchestrator alone resolved a silent conveyor-belt fault in all ten runs.
Kay K\"ohle, Darko Anicic, Thomas A. Runkler et al.
In briefDegree-corrected joint matrix factorization for finding communities in multilayer networks
SummaryThis paper proposes a method for detecting communities in multilayer networks, where the same nodes interact across several layers. It approximates each layer's graph with a nonnegative symmetric matrix trifactorization, with constraints that make communities disjoint and shared across layers, while letting each layer keep its own connection patterns and node degrees. The authors also develop an algorithm to solve the problem efficiently. Experiments on synthetic graphs from the multilayer degree-corrected stochastic block model show reliable detection across diverse settings, whereas existing methods are often limited by restrictive assumptions.
Alexandra Dache, Manon Rustin, Arnaud Vandaele et al.
In briefFold'EM: Inferring atomic structures directly from cryo-EM particle images
SummaryThis paper introduces Fold'EM, an inference-time framework that combines priors from protein generative models directly with cryo-EM single-particle images. It determines atomic models from a small number of particles, skipping intermediate density map reconstruction and later model building. The authors report accurate structures on synthetic and experimental datasets, both with known particle orientations and when orientations are inferred jointly. On heterogeneous data, it also separates distinct conformational states from mixed particle populations.
WhyDirectly combines protein generative priors with cryo-EM particle images, bypassing density reconstruction; validated on synthetic and experimental datasets, with heterogeneous-state results.
Advaith Maddipatla, M\"art-Erik M\"aeots, Marco Pegoraro et al.
In briefPort-Hamiltonian neural networks for systems with several stable equilibria
SummaryStable port-Hamiltonian neural networks guarantee asymptotic stability by design, but their single-minimum Hamiltonian limits them to systems with one attractor. This paper parametrises the Hamiltonian as a product of Bregman divergences from one input-convex network, allowing systems such as double-well energy landscapes to be modelled. The authors prove local Lyapunov stability and report that on three systems the method recovers energy surface features and converges 1.8 to 8.5 times faster.
Simon Heilig, Jens P\"uttschneider, Mohammad Itani et al.
In briefDiscrete Wasserstein Flows: a discrete drifting method for one-step generative modeling
SummaryThis paper proposes a framework for one-step generative modeling on finite state spaces, extending drifting from continuous to discrete domains. It uses discrete Wasserstein geometry to define a target-relative KL gradient flow over transitions of a reversible Markov kernel, realized at the particle level via Markov jumps. The transport updates are amortized into a latent-conditioned generator, so iterative dynamics are used only in training while inference stays one-step. In a controlled setting with exactly computable distributions, the authors verify KL dissipation and numerical scaling, and show a finite-capacity neural generator can track the exact transport targets.
In briefDoes AI-generated science writing follow human argument patterns? A CARS-based comparison of paper introductions
SummaryThis paper uses Swales' CARS model to analyze the rhetorical structure of research article introductions. It compares original introductions from published linguistics papers with versions generated by large language models for the same papers. Human introductions vary more in which moves they use and their order, while generated ones are more uniform. Providing the models with the CARS definitions makes their structure more rigid.
Abdelrahman Sadallah, Narjes Sheikh Asadi, Lonneke van der Plas
In briefMMVistaReason: An open-data post-training recipe for multimodal reasoning
SummaryThis paper introduces MMVistaReason (MVR), an open-data post-training method for multimodal reasoning that covers analytical and real-world reasoning, builds SFT and RL datasets (MVR-SFT-528K and MVR-RL-63K), and merges separately trained RL experts through multi-teacher on-policy distillation. Across 15 multimodal benchmarks, the 4B model averages 72.8 and the 9B model averages 74.4, while using about 70% fewer samples than MMFineReason. The authors report that smaller and larger models respond differently to supervision difficulty and mixed-domain training.
In briefARCCS: A System That Automatically Checks Documents Against Regulations
SummaryThis paper presents ARCCS, an automated legal NLP system that checks whether a target document meets regulatory obligations. It breaks regulatory text into traceable atomic requirements and evaluates documents using retrieved evidence, confidence scores, and human-readable justifications. In a GDPR policy-document evaluation, LLM-based judges found its decisions and justifications consistent in up to 96.67% of cases. On an EU public-procurement benchmark with over 1,200 rule checks, it reached 98.8% accuracy in violation detection.
Giorgos Filandrianos, Jos\'e Menezes, Chrysoula Zerva et al.
In briefReachability-guided reinforcement learning for multi-impulse interplanetary transfer trajectories
SummaryThis paper proposes Reachability Analysis-Informed Reinforcement Learning (RARL) for deterministic multi-impulse interplanetary transfers, where the policy selects intermediate waypoints inside an ellipsoidal set derived from first-order reachability maps, and Lambert reconstruction yields the connecting maneuvers. On a two-body Earth-Mars benchmark, three independent training runs achieved a mean maneuver cost of 10.23 km/s, 1.72% above a validated local sequential convex programming reference. Multi-state policies completed all 10,000 held-out Monte Carlo departures without impulse-cap violations, compared with a 6.49% mean feasibility rate for single-state policies.
In briefRobust scheduling without knowing job lengths, guided by a classifier
SummaryThis paper studies single-machine scheduling that minimizes total job completion time when each job's processing time is unknown until it finishes. It uses the confusion matrix of a job classification model to represent uncertainty as sets of permutations within predicted classes. The authors propose an optimal non-adaptive strategy that is oblivious to three robust criteria. They also show that adaptive and randomized algorithms can outperform it when the matrix has certain structural properties.
In briefCLASP: One hypernetwork generates continual low-rank adapters for spatially placed concepts
SummaryThis paper addresses continual personalization of text-to-image diffusion models, where new concepts must be learned without forgetting earlier ones. The authors propose a rehearsal-free method that uses a single fixed-size hypernetwork to generate concept-specific adaptations for a frozen diffusion model. The framework also lets users specify where a personalized concept should appear, without adding per-concept components. Experiments reportedly show strong retention of learned concepts and reliable spatial grounding, scaling to long sequences of tasks.
Wojciech Gromski, Patryk Krukowski, Jan Miksa et al.
In briefA shared ontology for cross-sector crisis management, including a public health module
SummaryThis paper presents the European Crisis Management Ontology (ECMO), a modular OWL-based framework for disaster risk reduction and response across sectors. Its core module defines concepts such as hazard, event, exposure, impact, and response measure, and a public health module aligns with SNOMED CT and ICD-11. The authors used ECMO to build a pipeline that turns unstructured epidemiological news from the JRC's Epidemic Intelligence system into a knowledge graph. Initial results suggest it supports consistent knowledge representation and integration.
In briefTesting whether biomedical reranking helps LLMs answer patient questions from long clinical notes
SummaryThis paper evaluates whether biomedical reranking improves finding relevant evidence for patient-specific questions over longitudinal clinical notes. It uses a locally deployed retrieval-augmented generation pipeline combining dense and lexical retrieval, rank fusion, and MedCPT cross-encoder reranking, tested on 1,000 question-answer pairs from 200 bariatric surgery patients. Reranking raised Hit@10 from 46.6% to 60.6%, while answer correctness rose only from 44.8% to 48.6%, so retrieval gains did not translate proportionally into better answers.
Maryam Shahbaz Ali, Laura B. Strachan, Caitlin Sherman et al.
In briefTRACE: A Method for Multi-Turn Chatbot Safety Training
SummaryThe paper observes that safety-aligned language models often refuse harmful requests in a single turn but comply when the same goal is split across several turns. It gives conditions under which suppressing risk in single-turn contexts bounds risk in multi-turn trajectories, and proposes TRACE, a token-level training objective based on this principle. Across five open-weight models and seven multi-turn attacks, TRACE achieves the lowest attack success rate in all 35 model-attack pairs, with utility on MMLU and HellaSwag dropping by at most 1.23 points.
In briefClifford sheaf neural networks: an equivariant model for geometric graphs
SummaryThis paper introduces the Clifford Sheaf Neural Network (CSNN), which places a Clifford algebra on each stalk of a cellular sheaf and transports multivector features along edges for equivariant learning on geometric graphs. The authors replace algebra homomorphism and versor conjugation with a K-term sandwich restriction map, yielding a sheaf Laplacian that is positive semidefinite and still mixes grades. They characterize this map family by which grades it couples, how much of the endomorphism space it reaches, and its conditioning, with the number of terms controlling expressivity.
In briefProtoFlow: Forecasting Multivariate Time Series with Prototype-Guided Flow Matching
SummaryThis paper introduces ProtoFlow, a framework for forecasting multivariate time series. It maps sequences into a discrete latent space with vector quantization, builds a structured prior from the learned codebook, and uses a DiT-based rectified flow to transport that prior to future latent representations conditioned on past observations. The authors argue this avoids the mismatch of autoregressive token generation and speeds up training. They report consistently better forecasting performance and efficient inference on benchmark datasets.
In briefFeature-selective model collapse in diffusion models: full replacement versus fixed-budget training
SummaryThis paper studies model collapse in diffusion models trained repeatedly on synthetic data. It compares full replacement of real data with synthetic data against a fixed-budget protocol that keeps all past data but trains each generation on a fixed-size sample. Experiments on a 2D spiral dataset and on MNIST, Fashion-MNIST, and CIFAR-10 show that the fixed-budget protocol degrades data only partially, preserving some features.
In briefPrediction-powered Neural Architecture Search: combining scarce labels with cheap zero-cost proxies to rank candidate networks
SummaryThis paper addresses the trade-off in neural architecture search between costly performance labels and cheap but noisy zero-cost proxies. The proposed PPNAS method fuses a small set of architectures with observed performance and a large set with proxy information. It builds extra pairwise ranking supervision from the proxies' ordinal information and uses prediction-powered inference to correct systematic gaps between proxy-based and true rankings. The authors report state-of-the-art results for predictor-based NAS under limited evaluation budgets.
In briefWhat Wins a Vote? Formatting, Length and Vocabulary in the French Compar:IA LLM Arena
SummaryThe paper analyzes 137,293 decisive French-language votes from the July 2026 Compar:IA release to see how answer presentation relates to winning. Bold text and moving-average type-token ratio (MATTR, a vocabulary-variety measure) show the most stable associations across specifications, while formatting, length and lists often occur together and are hard to separate. The authors find that adjusted rankings do not better match external benchmarks, so they recommend publishing raw and adjusted rankings side by side.
Simonas Zilinskas, Maayeesha Farzana, Christophe Benavent
In briefEP-Flow: Predicting Disordered Crystal Structures Without Site-Level Labels
SummaryThis paper introduces EP-Flow, a generative method that predicts disordered crystal structures from chemical formulas alone, without requiring site-level disorder annotations. It represents ordered crystals, solid solutions, vacancies, and interstitial occupancy with a unified occupancy distribution matrix and generates occupancies, fractional coordinates, and lattice parameters jointly. On formula-conditioned benchmarks derived from COD and MPDS, it outperforms adapted ordered-crystal generators.
In briefARROW: a model that reconstructs and tracks 3D points from arbitrary images
SummaryARROW is a feed-forward model that performs 3D reconstruction and 3D point tracking together from any set of images. Its core is an order-invariant querying method that links queries to observations across arbitrary inputs. The authors report that training on more diverse input sets improves performance and enables generalization to tasks such as multi-view tracking. It sets a new state of the art in 3D tracking on WorldTrack and TAPVid-3D, and code and weights are publicly available.
WhyOrder-invariant querying unifies reconstruction and tracking, reaching new state of the art on WorldTrack and TAPVid-3D benchmarks.
Ilya Fradlin, Christian Schmidt, Jens Piekenbrinck et al.
In briefDAYJOB: A Benchmark for Testing AI Agents on Long Professional Tasks
SummaryThis paper introduces DAYJOB, a benchmark of 130 tasks created by professionals in healthcare and finance. Tasks are estimated to take about 13.6 hours (healthcare) and 16.6 hours (finance) on average, and each is judged against expert-written binary criteria. The strongest of 30 model configurations passes about 24.7% of healthcare and 23.9% of finance attempts, while the median configuration passes far fewer. The authors also report that agents sometimes accept premises contradicted by the record and carry wrong inputs through their analyses.
WhyIntroduces a new long-horizon professional benchmark with 130 expert tasks, systematically evaluating 30 model configurations plus case studies.
Stephanie Finley, Liudas Panavas, Thomas Mikkelson et al.
In briefFederated LLM Learning over Mobile Networks: Transport Challenges and Fixes in the RAN
SummaryThe paper argues that federated LLM fine-tuning creates recurring, deadline-sensitive traffic at the network edge, but conventional transport networks treat asynchronous device updates as unrelated flows. It proposes that future RANs act as learning-aware traffic shapers, using in-network aggregation at the gNB to reduce transfers. The aggregated traffic could then trigger temporary optical circuit provisioning during transfer windows.
Emilio Paolini, Andrea Pinto, Flavio Esposito et al.
In briefSTAGE: A Training-Free Method for Erasing Concepts from Text-to-3D Generators
SummaryThis paper introduces STAGE, a training-free, closed-form framework for erasing specific concepts, such as shapes, materials, or objects, from native text-to-3D generative models. Each edit is confined to a low-dimensional subspace and applied to the structural stage, the appearance stage, or both, depending on the concept type. On TRELLIS, across 15 shape, material, and object concepts, STAGE achieves a composite score of 66.7, compared with 53.2 for the strongest adapted baseline.
