In briefPropasafe-Hybrid: A Text Tool That Flags and Explains Propaganda Sentences
SummaryThis paper presents Propasafe-Hybrid, a sentence-level system that combines a fine-tuned transformer classifier with LLM-based technique classification to label specific propaganda strategies. It outputs highlighted sentences, technique labels, and short rationales explaining each flag. A cost-aware pre-filter sends only high-likelihood sentences to LLMs to reduce token usage. The authors argue these choices improve explainability, efficiency, and practical usability for propaganda detection in news environments.
In briefSemantic-First MixGRPO: Emotion-Controlled Image Generation via Timing Separation
SummaryThis paper studies where the stochastic sampling window sits in flow-based image generation as a way to control the emotional tone (valence-arousal) of images. The proposed semantic-first schedule uses deterministic ODE sampling for steps 0-19 to fix the content, and restricts stochastic exploration and policy updates to steps 20-23. It improves CLIPScore, cross-affect DINO consistency, and semantic drift over progressive scheduling, and cuts training cost by 33.5% versus full-path Flow-GRPO. Full-path stochastic training remains more accurate on valence-arousal targets.
In briefFG-ConvNO: A geometry-aware neural network for predicting propeller flow pressure
SummaryThis paper proposes FG-ConvNO, a neural operator that predicts propeller surface pressure from limited training data while accounting for blade geometry. It combines geometry-aware encoding, interpolation-based decoding, auxiliary global-quantity supervision, and DISCO blocks that use discrete continuous convolutions. On a procedurally generated propeller CFD dataset, the authors report the highest accuracy for both surface pressure and global thrust prediction.
In briefA Unified Framework for Every Stage of Training Large AI Models, from Data to Alignment
SummaryThis survey proposes a unified division of large language model training into stages and a six-dimensional analysis framework covering data, objectives, optimization, system constraints, evaluation, and risk governance. The authors systematically review 395 core references from 2017 to 2026 and present a map of how stages affect each other. They find that scaling-law comparability depends on data quality and mixture, and they discuss post-training methods and the structural challenges multimodal training poses to text-only approaches.
In briefBoardSafety: AI Benchmark Testing Whether Models Judge Development-Board Operations Safely
SummaryThis paper introduces BoardSafety, a benchmark for checking whether large language models can judge if an operation is safe on a specific development board. It is built from vendor datasheets and hardware guides, converted into board-specific safety rules, and used to create labeled scenarios marked safe, unsafe, or insufficient-info. Providing the safety rules generally improves accuracy, while structured-rationale prompting helps differently across models and rule-access conditions. The authors conclude that the evaluated models cannot yet reliably determine whether an operation is safe.
In briefFrom Intelligence to Autonomy: Building Self-Maintaining AI Agents
SummaryThis paper examines how AI agents can move beyond one-off tasks toward sustained autonomous operation. It defines an autonomous agent as one that maintains its goals and viability within set boundaries, and proposes a four-loop framework covering task execution, self-maintenance, metacontrol, and governed learning. Five paired simulation experiments provide preliminary evidence that these mechanisms are feasible. Broader environments and physical systems remain untested.
In briefTransferring internal signatures to detect unreliable LLM outputs across domains
SummaryLarge language models can produce hallucinations and reasoning errors, and existing detectors need substantial labeled data in each domain. This paper proposes transferring low-dimensional representational signatures across related domains, learning composition weights for domain-specific predictors via cross-fitting. With only 25 labeled questions in a new domain, the adaptive transfer improves hallucination-detection AUROC by 6.8%. For reasoning-step prediction, AUROC improves by 8.4%.
Yongtong Li, Yixuan Florence Wu, Jiawei Zhang et al.
In briefMeasuring Meaning with SCAPE-T: Semantic Complexity from Transformer Attention
SummaryThis paper defines semantic complexity as the code length needed to specify a text's meaning and measures it with SCAPE-T, an attention-based metric built on encoder-only transformers. It combines three submetrics and, with word length controlled, reaches 78.5% pairwise accuracy on children's stories versus technical Wikipedia articles, outperforming readability, syntactic/lexical, and LLM-based baselines. The authors argue it separates conceptual structure from lexical and syntactic confounds better than existing methods.
In briefSelecting training data for medical QA using perturbation-based signals
SummaryThis paper studies how to choose a compact training subset from existing medical question-answering banks. The authors propose two selectors, RSP and RSP-Drop, that rely on signals from Gaussian perturbations of the final-block representation. Using MedGemma-4B across five medical QA datasets, RSP and RSP-Drop reach dataset-averaged Macro-F1 of 47.55 and 48.28 at a 128-example budget, versus 45.11 for random selection. The results vary across datasets and budgets, and the study uses one backbone, so it does not establish clinical effectiveness.
In briefTesting whether people and language models generalize new verbs by linguistic form
SummaryThis paper compares how people and large language models extend a novel verb learned in active voice to passive contexts. People generalized to passives based on abstract rules of linguistic form, even when the novel verb had no plausible meaning. Language models succeeded only when a meaning similar to an existing verb could be assigned. The authors conclude that models' syntactic knowledge remains tied to specific lexical items' distribution and meaning, unlike human syntactic knowledge.
In briefConMem: Counterfactual Probing to Fix LLM Memory Misalignment
SummaryThe paper addresses memory misalignment, where long-term memories of LLM agents over-generalize, misattribute, or retain privacy-sensitive information. The authors build MemAlign, a benchmark of 1,079 memory cases from synthetic multi-session conversations, and propose ConMem, which treats retention as counterfactual causal inference. According to the paper, ConMem outperforms other methods and reduces over-retention bias without sacrificing personalization utility.
In briefDynamic Query as Budgeting: BUTR for Tiny Object Detection
SummaryThis paper analyzes how dynamic query mechanisms in DETR-based tiny object detectors work, arguing that they mainly allocate how many decoder queries are used according to object density rather than performing intricate query selection. The authors propose BUTR, which uses a lightweight, non-learned budgeter to estimate an input-dependent query budget from encoder confidences. At 800×800 resolution on an NVIDIA T4 GPU, BUTR-L reports 35.7 AP at 47.8 FPS on AI-TOD-v2 and 29.8 AP at 42.1 FPS on VisDrone.
In briefExact Unlearning After Data Curation: What a Service Must Keep to Reproduce Training After Deletions
SummaryThe paper studies training pipelines that remove similar examples before fitting a model, where deleting a raw record can readmit a previously excluded example. It shows which records can still enter training under a limit on future deletions, and builds training-statistic summaries that can be repaired exactly after repeated deletions, including finite-precision features for ridge regression. Experiments on 1,001 AskUbuntu development records found that deleting an excluded record admitted another record at 13 of 51 fixed similarity cutoffs.
In briefAn Empirical Study of Reinforcement Learning Scaling Laws: Predicting Large-Scale RL Performance from Small Experiments
SummaryThis paper empirically studies scaling laws for an actor-critic reinforcement learning algorithm in two procedurally generated environments, Starpilot and Craftax. The discovered laws predict performance, compute-optimal resource allocation, and hyperparameters for training runs up to two orders of magnitude larger in total compute. The authors find that an Elo-based score is more predictable than mean episode return, and they use the scaling laws to reach a new state of the art on Craftax, while noting a non-monotonic compute-optimal model size pattern linked to Ray interference.
WhySystematically studies RL scaling laws, proposes Elo as a more predictable metric, validates across two environments at large scale, and sets new Craftax state of the art.
In briefOnline covariance estimation for averaged Riemannian SGD on manifolds, enabling inference for parameters constrained to curved spaces.
SummaryThe paper studies online statistical inference for averaged Riemannian stochastic gradient descent when the unknown parameter lies on a manifold. The authors use vector transport to build overlapping and non-overlapping covariance estimators that combine information from successive batches while accounting for changing tangent spaces. The estimators are online, need no Riemannian Hessian and no prespecified sample size, and match the convergence rates of their Euclidean counterparts. The framework is illustrated through streaming principal component analysis on Grassmann manifolds.
In briefAre CNNs and Vision Transformers Robust for Alzheimer's MRI Classification Across Datasets?
SummaryThis paper presents a multi-dataset benchmark that compares five architectures, including ResNet-50, ReXNet-150, Swin Transformer, ViT-Base, and ViT-Large, for classifying Alzheimer's disease stages from MRI. Using standardized preprocessing and evaluation across three datasets, the authors find that ViT-Large performs best on all three, reaching up to 99.84% accuracy. Architectural differences are largest under distribution mismatch and smaller with augmented data. Lower performance mainly occurs between adjacent early-stage categories, especially Very Mild and Mild impairment.
Hafija Aktar, Molly Croes, Frank S. Menniti et al.
In briefAV-FABLE: A Benchmark for Detecting Audio-Visual Manipulation in Long Videos
SummaryThis paper introduces AV-FABLE, a benchmark of 16,650 original and manipulated video pairs totaling 998 hours, with separate audio and visual timelines for locating sparse AI edits in long-form videos. Zero-shot prompting of audio-visual large language models performs near chance, but supervised readouts of their frozen features reveal strong manipulation signals. The authors propose NFSR, which uses separate audio and visual temporal decoders on these frozen features, achieving the best audio-visual localization in-domain and on out-of-distribution settings.
In briefPerformance gaps among open-loop, certainty-equivalent, and optimal control when actions affect observations
SummaryThis paper studies a linear-quadratic system with bilinear observations, where control actions also affect what is observed. It compares the optimal policy, certainty-equivalent control, and optimal open-loop control in terms of cumulative cost. An information-strength parameter q defines distinct performance regimes. For intermediate q, the authors construct a policy that occasionally probes for information and gains an Ω(T) advantage over open-loop control, while certainty-equivalent control gains only O(1), leaving an Ω(T) gap.
Leo Maynard-Zhang, Sunmook Choi, Yahya Sattar et al.
In briefBenchmarking fast decision-making AI models on time-series tasks, using Jev as a case study
SummaryThe paper evaluates structured System 1 models, which directly output answers and probabilities, on time-series tasks using Jev 1.13.0 as a case study. Tasks include time-series classification, event type prediction, oil temperature direction, and stock movement prediction, with continuous targets converted to discrete decisions. The models perform better on some explicit temporal patterns and event type prediction, but near chance on many numerical sequence classification tasks and stock movement prediction.
In briefData Optimization for LLMs: A Survey, a review of techniques for improving the data used to train and run large language models
SummaryThis survey reviews data optimization techniques for large language models, which the authors argue have received less attention than other aspects of LLMs despite data's key role. It examines data properties along four dimensions: quality, difficulty, diversity, and redundancy. The existing techniques are grouped into data selection, augmentation, reweighting, and other methods, and the interconnections among these groups are analyzed. The paper concludes by identifying challenges and future directions compared with conventional deep learning.
In briefCOPE: A tool that predicts which Mixture-of-Experts offloading policy runs fastest on a given model and hardware setup, before implementation.
SummaryThe paper presents COPE, a tool that selects the best mixture-of-experts (MoE) offloading policy for a target model and platform before runtime integration. It splits the design space into two independent axes, expert residency and MoE execution strategy, and estimates candidate performance by combining profiled operator schedules on the target platform with offline routing traces. Compared with the evaluated baselines, COPE improves fastest-policy selection accuracy by 24%-100%, reduces performance-prediction error by 1.7x to 6.8x, and cuts policy-specific implementation effort by up to 74%.
In briefA federated method for generating faithful mixed-type tabular synthetic data (FITS)
SummaryThis paper studies generating synthetic tabular data with mixed numerical and discrete variables in federated settings, where preserving distributional fidelity is a key challenge. The authors propose FITS, which uses a two-step federated algorithm based on local linear regression to estimate conditional CDFs, a probability chain rule synthesis procedure, and a continualization method that maps discrete variables into a continuous space. Experiments show FITS-generated data improves downstream task performance for all clients, with benefits depending on how the synthetic data is used.
In briefBidirectional cross-frequency network for enhancing ultra-high-definition low-light images
SummaryExisting methods for enhancing ultra-high-definition low-light images often model only the spatial or frequency domain, leading to inconsistent illumination and residual noise. This paper proposes BiFreqNet, which adds a dual-domain feature interaction module to multi-scale features of a pretrained variational autoencoder so low- and high-frequency features reinforce each other. Experiments on four general and UHD low-light enhancement benchmarks show competitive or superior performance compared with state-of-the-art methods.
In briefRethinking Microscopy Generation: Co-Designed Diffusion for Interpretable Single-Cell Synthesis
SummaryThis paper addresses the lack of control over individual cell appearance in microscopy image generation. It generates images at the single-cell level and designs input controls and the model's latent representation together, organizing both around biologically meaningful semantic axes such as size, elongation, brightness, and texture. The authors introduce BioVAE, an encoder-decoder whose latent space is aligned with these measurements. Experiments on LIVECell show that semantic-axis conditioning improves generative fidelity over high-dimensional conditioning, a blind study with biologists confirms realism, and per-axis traversals yield stable, interpretable phenotypic changes.
Shumpei Takezaki, Duway Nicolas Lesmes-Leon, Ryoma Bise et al.
In briefFast Instinct, Slow Reason: Self-Configuring Satellite Video Tracking
SummaryThis paper addresses multi-object tracking in satellite video, where offline-tuned trackers cannot adapt onboard within tens of seconds. It proposes a dual-speed framework: AUSPEX, an offline-compiled method that selects a tracker configuration per sequence using a label-free ranker, and CORVUS, an onboard runtime that cascades reflex, skill, and language-agent layers so verified skills commit within 2.5 seconds. The authors report IDF1 and MOTA gains on SatMTB and SatVideoDT and operation on Jetson Xavier NX.
In briefHow to Build a Modern Diffusion Language Model: A Survey of Discrete Diffusion for Text
SummaryThis survey reviews key ideas in discrete diffusion and explains how modern diffusion large language models are built. Unlike autoregressive models that generate one token at a time, diffusion models generate tokens in parallel and iteratively refine the whole output, which the paper says enables fast generation, built-in error correction, and controllable generation. The article covers foundations, algorithmic and architectural building blocks, applications in biology and language, and implications for future LLM scaling laws.
Volodymyr Kuleshov, Marianne Arriola, Yair Schiff et al.
In briefCost of an unknown stopping time for clipped SGD under heavy-tailed noise
SummaryThis paper studies clipped stochastic gradient descent on smooth convex objectives when gradient noise is heavy-tailed, asking how fast the last iterate converges at all times simultaneously. It shows a schedule needing no horizon or confidence inputs achieves a time-uniform rate, also almost surely, and that rules stepping along the current iterate's stochastic gradient cannot beat this rate. The resulting exponent lies between the known-horizon optimal rate and the previous time-uniform rate.
WhyEstablishes the optimal time-uniform last-iterate exponent for clipped SGD without horizon knowledge, with matching lower bounds that substantially improve prior results.
In briefPacked Self-Speculative Decoding: Drafting One Call Ahead with Diffusion
SummaryAutoregressive language models generate text token by token, which is slow for long outputs. This paper proposes Packed-SS, which verifies the current draft and builds the next one in a single model call, with input size linear in speculative width while preserving the target distribution. Experiments on Nemotron 3B/8B and SetDLM 1.7B show higher throughput and lower mean wall time than two baselines under greedy decoding.
In briefAdShot: A Benchmark for Testing AI Models on Cutting Video Ads to Shorter Lengths
SummaryAdvertisers often need several versions of the same ad at different lengths, which is slow and expensive to produce. This paper introduces AdShot, a benchmark of 823 professionally edited ad pairs (30-second sources and 15-second edits) for testing how well multimodal large language models pick shots for shorter ads. The study evaluates 13 video-only and 2 audio-visual models, finding that architecture matters more than size, that most models run over the 15-second target, and that models favor information-rich and emotion-focused shots while underselecting action and imagery. The benchmark is publicly released.</summary>
Wen Xie, Om Rastogi, Sai Siddhartha Vivek Dhir Rangoju et al.
In briefHow Close Can Global Orchestration Come to Step-by-Step AI Agents?
SummaryEnterprise workflow agents usually act one step at a time, calling a large language model after each action, which is costly. Global orchestration instead generates a full workflow in advance, reducing model calls but remaining less accurate. The paper introduces AgentOrch, a benchmark comparing both approaches on 650 enterprise SOPs and 3,220 runtime variants. Step-by-step interaction stays more accurate, but the gap has shrunk by about 79% for the strongest recent model, and three methods are presented for narrowing it further.
In briefHow post-training shapes the value preferences of large language models
SummaryThis paper studies how large language models learn value preferences during post-training from corpora whose opinions are imbalanced across demographic groups. The authors ran controlled experiments with different opinion mixtures and five post-training paths based on SFT, DPO, and GRPO. They find that post-training amplifies opinion imbalances differently across these paths, and that these preferences can transfer to student models through distillation even when no explicit opinion questions are used.
In briefCombining several small draft models with multiple importance sampling for guided speculative LLM inference
SummaryThis paper extends the Guided Speculative Inference (GSI) method for scaling LLM test-time reasoning so that an ensemble of smaller models drafts reasoning steps, which a larger base model and a process reward model then verify. Using multiple importance sampling, the authors design weighting schemes to aggregate samples from diverse proposal distributions and give new bounds on KL divergence and expected reward degradation. They show the balance heuristic is optimal and report that, at the same total sample count, combining several models stays competitive on pass-at-N performance and is more robust when a single drafter's support does not cover the target.
In briefA multi-agent framework that uses insights distilled from paper graphs to generate research ideas through debate
SummaryThe paper introduces InsightAgents, a framework that builds literature-grounded expertise for AI agents by distilling insights from graphs of related papers. Guided by these insights, the agents debate to identify a research problem and to propose, challenge, and refine approaches for a research proposal. On 100 NLP research topics, evaluated by pairwise comparisons from two LLM judges, it achieved the highest average score across five dimensions in both online and offline knowledge settings.
In briefLSTM-style looped Transformer with a shared key-value cache
SummaryThis paper studies an LSTM-style recurrent Transformer whose recurrent core shares one key-value cache. At width 1024 it uses 40.4% fewer parameters than an untied control at nearly matched forward FLOPs, and cuts persistent cache sets from fifteen to four. A consistency loss during the final 10% of training aligns cache entries from parallel and recursive passes. Experiments on FineWeb across three model widths evaluate the approach.
In briefScreening Parkinson's disease from smooth pursuit eye movement complexity
SummaryThis study examines whether complexity and nonlinear dynamics measures of gaze position sequences during smooth pursuit eye movements can help automatically screen for Parkinson's disease. Gaze was recorded from 41 early-stage patients and 47 healthy controls across 12 tasks, and a support vector machine mixture-of-experts pipeline was built. On an external test set of 18 participants, it achieved an AUC of 0.86. Parkinson's patients showed less persistent and more irregular gaze sequences, mainly horizontally, but the findings require validation in larger independent cohorts.
J. Camilo Puerta-Acevedo, Maria F. Alcala-Durand, Juli\'an D. Arias-Londo\~no et al.
In briefMOSHO: A Discrete Shark-Inspired Search Method for Hardware-Aware Neural Architecture Design
SummaryThis paper addresses hardware-aware neural architecture search, which balances model accuracy against device latency. The authors argue that continuous swarm optimizers lose quality when rounded to integer architecture encodings, and propose MOSHO, a discrete metaheuristic that uses architecture-editing operators with adaptive scheduling, energy-regulated restarts, and deferred crowding-distance truncation. On the merged NAS-Bench-201/HW-NAS-Bench benchmark across 18 contexts, MOSHO reports a median normalized hypervolume of 0.9987 and IGD+ of 0.0064, competitive with NSGA-II, while cutting search runtime by 96.4%.
Nazim Abdeerhmane Aouni, Zakaria Soualah Mohammed, Maroua Ogab et al.
In briefPreference optimization makes word-definition generators less fragmented, tested on EU sustainable finance texts
SummaryThis paper shows that fine-tuning definition generators with Direct Preference Optimization (DPO) reduces the tendency to produce excessive definitions for the same word sense. The approach is evaluated on a curated academic lexical semantic change detection benchmark. It is also applied to real EU development finance project documentation to track how sustainability terminology has changed over time. The models, code, and predictions are publicly released.
In briefTight Learned Event-Inertial Odometry (TLEIO): a learned, Transformer-based method that estimates motion from event camera streams and fuses it with IMU data.
SummaryThis paper introduces TLEIO, a compact filter-based pipeline that learns short-window relative motion directly from voxelized event streams and tightly fuses it with IMU propagation. Its front-end, EventsFormer, predicts local displacements and per-axis uncertainty in a single forward pass, without explicit feature tracking or pose optimization at inference. On TartanAir, average ATE drops from 2.26 m for the IMU-only TLIO baseline to 1.10 m, with up to 48.9% lower runtime than optimization-based event odometry methods.
Matteo Cascone, David Jensen, Alessandro Pirini et al.
In briefBenchBench-Cloning: A Benchmark for Molecular Cloning Design and Protocol Planning
SummaryThis paper introduces BenchBench-Cloning, a benchmark of 65 molecular cloning tasks for large language models, built from sequence records, lab notebooks, and published studies. It includes 50 protocol-development tasks, checked by execution in a cloning simulator, and 15 construct-design tasks, checked by deterministic tests and rubric-based LLM judging. Across eight systems, GPT-6 Astra scored highest at 30.6%, while the others ranged from 8.6% to 22.2%, and the benchmark remains unsaturated under best-of-ten sampling.
Aditya Sivakumar, Peter Ciaccia, Hassaan Qadir et al.
In briefCLOCK: A Kernel-Based Method for Clustering Dynamical System Time Series
SummaryThis paper introduces CLOCK, a method that clusters time series by embedding them into a reproducing kernel Hilbert space using time-averaged occupation kernels. For systems governed by differential equations, the dissimilarity metric approximates the maximum mean discrepancy between attractors. The method handles sequences of unequal length and irregular sampling, and was validated with spectral clustering on synthetic data and intracranial EEG data.
Michael L. Wells, Kamel Lahouel, Victor Rielly et al.
In briefTesting whether video models truly encode time-based concepts like motion
SummaryThis paper introduces a No-False-Positives (NFP) statistical test to determine whether features in video models genuinely encode temporal concepts, such as object movement, rather than spatial cues that only correlate with motion. The test finds temporal features in VideoMAE, V-JEPA2, and TimeSformer, and yields none for the DINOv2 image encoder. Interventions on temporal features produced mixed results across models.
In briefReal-time video motion detection inspired by photonic time stretch
SummaryThis paper introduces MoViD, a physics-inspired algorithm that estimates motion in video by treating each pixel as an optical time series and using group delay dispersion, drawn from photonic time stretch. Unlike conventional phase-based pipelines that rely on narrow-band channelization and pyramid reconstruction, MoViD operates directly on the broadband video signal. The authors report full-resolution motion mapping at nearly 200 frames per second at 4K and 70 frames per second at 8K.
In briefWhen Nesterov trajectories eventually become locally minimizing
SummaryThe paper revisits the variational view of Nesterov's accelerated method and studies when its trajectories are eventually locally minimizing in a variational sense. This question is shown to be equivalent to the nonnegativity of a Schrödinger operator acting on trajectory variations. For power objectives f(x)=(1/p)||x||^p with p≥2, the authors show that nonzero-initial trajectories are eventually locally minimizing exactly when p=4. This exceptional exponent is interpreted through stability theory of semilinear elliptic equations.
In briefShrinking retinal OCT vision foundation models via codebook distillation and graph-guided masked image modeling
SummaryThis paper proposes a topology-aware framework that distills a large retinal vision foundation model into a discrete codebook, used as the target for masked image modeling. The 22M-parameter student is distilled from an 86M-parameter teacher with at least 3.6x fewer FLOPs. Across several classification and segmentation benchmarks, the student performs close to, and in some settings above, its teacher.
Ali Yavari, Tilman Schmoll, Rainer A. Leitgeb et al.
In briefTinyFlow: Removing the Warping Bottleneck for Edge Optical Flow Estimation
SummaryTinyFlow is a 1.26M-parameter optical flow network that removes the feature warping step, replacing it with fixed-window correlation so the entire model compiles as a static INT8 graph for edge devices. Quantization-aware fine-tuning, warp-teacher distillation, and model souping reduce the Float32-to-INT8 accuracy drop to under 0.2 end-point error on MPI-Sintel. The INT8 model reaches 6.29 and 7.13 EPE on Sintel Clean and Final. In flow-only monocular visual odometry, it gives the lowest trajectory error on seven of eleven sequences.
In briefInformation-theoretic scaling laws for training diffusion image models
SummaryThis paper studies how much a diffusion model's generated samples depend on its particular training dataset, using a computationally tractable measure called model information. The authors examine how this measure scales as model size and training data size change, identifying a nonlinear regime marking the shift from memorization to imperfect generalization and a linear regime where generalization and sample quality improve. They propose an information-theoretic scaling law covering both regimes and derive induced scaling laws for memorization and sample quality, which they report extrapolate well beyond the fitted range.
WhyProposes a tractable model-information measure revealing two scaling regimes and a scaling law, though experimental scope appears limited.
In briefGaussian approximation of privacy loss for composed and subsampled differentially private algorithms
SummaryThis paper studies privacy bounds for composing subsampled mechanisms, as used in differentially private SGD. It shows that, when each step has small privacy loss with high probability, the composed privacy loss distribution after truncating rare large losses is close to a Gaussian with the same mean and variance. The approximation error grows only logarithmically in the additive error term. The authors also use the chi-squared divergence of a single step to derive variance-based bounds for subsampling, which improve on Rényi-divergence analyses under certain noise conditions.
WhyGaussian approximation of subsampled composition privacy loss substantially improves Rényi analyses, with rigorous proofs and notable impact on DP-SGD theory.
In briefAugmentation-guided adversarial training that selects samples by local stability
SummaryAdversarial training defends models against adversarial attacks but is computationally expensive. This paper proposes Augmentation-Guided Adversarial Training (AGAT), which uses a score measuring how much model outputs change under label-preserving augmentations to pick samples with unstable neighborhoods. Adversarial supervision is applied to only that subset of each minibatch. Experiments on MNIST, CIFAR-10, CIFAR-100, and Tiny-ImageNet-200 report comparable or higher robust accuracy with less training time than other baselines.
Daniel Agyei Asante, Mark Tetteh, Deborah Asamoah et al.
In briefTempShap: When and How Much Input Features Matter in Time-Series Forecasting
SummaryThe paper proposes TempShap, a loss-based Shapley attribution framework that treats each individual input cell, defined by time step and feature, as a separate player. This lets it show when in the input window a feature acts and how its influence changes across the forecast horizon. With redundant inputs, the signed Shapley decomposition yields more reliable held-out deletion rankings than one-at-a-time scoring. Experiments report a deletion-ranking advantage across Shapley estimators, forecasting benchmarks, and model families.
In briefAI Helps Humans Manage Complexity in Go Strategy
SummaryThis paper uses the game of Go as a testbed and defines a "strategic tension" metric, based on the spectral radius of spatial dependency matrices, to measure the balance between local stability and global coordination. Tension decreases as human players progress from novice to master, but professional games from 1980 to 2026 show a marked rise in sustained tension after AlphaGo in 2016. The authors argue that strategic tension reflects decision paradigms rather than raw computational strength, and that AI can serve as cognitive scaffolding to expand human capacity.
Adamo Cerioli, Vito D. P. Servedio, Gabriele Cerioli et al.
In briefMLIPDiff: Guiding Diffusion Models for Drug Design with Learned Atomic Potentials
SummaryDiffusion models are widely used for structure-based drug design, but they often place ligands in physically implausible poses that clash with the protein. This paper introduces MLIPDiff, an inference-time guidance method that applies to pretrained diffusion models without retraining. It uses energy terms from the UMA machine-learned interatomic potential, covering protein-ligand interaction, ligand strain, steric clash, and pocket anchoring. On CrossDocked2020, MLIPDiff eliminates protein-ligand collisions, improves physical plausibility, keeps the base model's PoseBusters pass rate, and achieves the best median Vina scores.
In briefResponsibility Boundaries in Agentic Systems: Comparing Probabilistic Agentic Programming
SummaryThe paper proposes Probabilistic Agentic Programming (PAP), which makes responsibility boundaries linking mandate, authority, effects, and evidence explicit as programming abstractions. It compares PAP with other programming paradigms and agent frameworks across seven dimensions, and formalizes a mediation kernel with inductive proofs of conditional safety. An exploratory study across seven business cases and seven models reports observed task completion counts, with paired intervals that all include zero.
In briefWhen Pre-Trained Capacity Helps Causal Learning: From Scalar Prediction to Counterfactual Post-Training
SummaryThis paper examines how the size of pre-trained Qwen 2.5 models relates to performance on causal tasks. It compares CausalLLM, which estimates scalar potential outcomes, with CausalGen, which generates treatment-conditioned text or code through several training stages. Scaling from 0.5B to 1.5B parameters gives mixed results for scalar estimation but clearly improves the full generative pipeline's code and treatment-effect scores. The authors note these are descriptive comparisons across systems that differ in data and supervision, so they support a hypothesis rather than a general scaling law.
In briefSelf-supervised learning of X-ray diffraction patterns using joint-embedding predictive architectures
SummaryPowder X-ray diffraction is a key high-throughput method for determining crystal structures, which is important for automated materials discovery. The authors use a self-supervised Joint Embedding Predictive Architecture (JEPA) to learn representations from 1D diffraction patterns by predicting masked content in latent space. Lightweight probes on frozen representations achieve space group classification and lattice parameter estimation accuracy close to much larger supervised vision transformers. The learned latent space also organizes peak broadening smoothly and yields realistic multi-phase mixture patterns when decoding between single-phase representations.
In briefHow tree depth and sample-quality features affect Boruta gene selection in RNA-seq classification
SummaryBoruta is a common method for selecting gene panels from RNA sequencing data, and its usual implementation recommends shallow random forest trees that had not been tested. Using pan-cancer TCGA data and breast cancer PAM50 subtyping data, the authors found that shallow trees substantially change which genes are selected and produce less accurate 50-gene panels. Leaving depth uncapped cost no accuracy. About a third of candidate features also tracked sample quality within class, and removing them did not reduce test accuracy. The authors recommend uncapped trees and a within-class quality filter before Boruta.
Tareque Mohmud Chowdhury, Abu Raihan Mostofa Kamal
In briefA machine that imagines your life under other policies to make distant risks feel personal
SummaryThis paper presents the As-If Machine (AIM), a multi-agent system in which people co-write personalized counterfactual narratives about their own lives under different policy futures. In two pre-registered, quota-matched experiments on science funding (N=1,007) and satellite cybersecurity (N=894), AIM increased psychological closeness compared with reading an expert-written vignette. Its effects on policy attitudes differed between the two domains.
In briefHow a language model's grammar circuit forms and becomes sparse during training
SummaryThe paper tracks the circuits that handle subject-verb number agreement in pythia-1b and ten seeds of pythia-410m across training. Agreement is learned early and initially handled by a diffuse set of neurons. After agreement behavior stabilizes, the responsible mechanisms keep reorganizing, with attention heads shifting and the circuit shrinking; pythia-1b's final circuit is strikingly sparse.
In briefAll-in-One: a single attack that triggers when LLMs are quantized, pruned, or merged
SummaryThis paper presents a composable attack framework that targets several deployment-time transformations of LLMs, including quantization, pruning, and model merging, at once. A model trained this way passes safety checks before transformation but behaves maliciously once a user applies any one of these transformations. The authors approximate each transformation as a perturbation to the weights and jointly optimize a benign loss on the original model with malicious losses on simulated transformed models. Experiments across attack scenarios, model families, and transformations report high success rates.
In briefKINESYS: Test-Time Adaptation for Reference-Based Image Effect Editing
SummaryThis paper studies editing images so that a reference picture's visual effect is transferred, with language specifying which properties to follow. The authors propose KINESYS, a test-time adaptation framework that uses both reference-aware and source-aware feedback to preserve the original shot's composition and content. Evaluations report stronger source preservation than state-of-the-art editors, competitive effect transfer against Nano Banana 2, and better results than open-weight editors.
In briefIndividual Fairness via Local Predictive Consistency
SummaryThe paper argues that most individual fairness definitions use pairwise constraints that can be overly restrictive and do not easily reveal which individuals receive locally inconsistent predictions. It proposes local predictive consistency, a node-level relaxation that compares each prediction with a similarity-weighted average of its neighbors, yielding an interpretable node-level fairness score and a regularized learning method based on local residuals. The approach is stated to apply to binary classification, regression, and time-to-event prediction.
In briefEstimating Error Risk Propagation in LLM Multi-Agent Systems
SummaryThis paper studies how errors from upstream agents affect downstream agents in LLM-based multi-agent systems. The authors propose Interaction-Effect Propagation (IE-Prop), which learns separate effects for correct and incorrect upstream agents and recursively propagates risk estimates through the execution graph. Across three architectures and three benchmarks, the method raises final-node AUROC by about 0.05 on average, with an additional 0.03 gain from multi-step risk tracking in sequential systems.
Shreyan Ganguly, Souradeep Chattopadhyay, Qisai Liu et al.
In briefCoral Discovery: how autonomous AI agents divide research directions
SummaryThis arXiv computer science paper presents Coral Discovery, a mechanism that coordinates multiple LLM agents improving programs against a deterministic grader. A verifier labels each attempt with a research direction and method, and a shared board tracks per-direction gains and suggests how many agents each direction should get. The authors report results on nine tasks, beating the strongest open multi-agent system on six.
In briefMulti-Agent Retrieval-Augmented Causality Assessment (MARCA): Judging Drug-Event Causality in Safety Reports
SummaryThis paper presents MARCA, a four-agent framework that supports Naranjo-based causality assessment of adverse drug events in individual case safety reports. It was compared with consensus expert scores on 215 FAERS reports covering 1,562 drug-event combinations from Q1-Q2 2025. Agreement varied across Naranjo questions, and final causality classification agreement was 79.4%. The authors note that the retrospective RAG comparison is exploratory and does not establish prospective effectiveness.
In briefExact bounds on continuous treatment effects despite hidden confounding
SummaryThe paper studies partial identification of conditional treatment effects for a continuous treatment when unobserved confounders are present, targeting conditional average potential outcomes and effects between two treatment levels. Under bounded outcomes and Lipschitz continuity of response functions and propensity scores given the latent confounder, it derives explicit bounds from every diagonal reference point whose average gives the sharp identified set. The resulting bounds are informative without monotonicity or overlap assumptions.
WhyUnder Kim et al.'s smoothness conditions, averaging per-diagonal bounds yields the sharp identified set of the CAPO, tightening prior loose bounds.
In briefGEMS-RAG: Graph-grounded Retrieval for Multimodal Engineering Standards
SummaryThis paper introduces GEMS-RAG, a framework that helps large language models reason over engineering standards by grounding answers in a graph of text, figures, and document structure. It breaks figures into visual entities linked to provisions and concepts, and combines dense, sparse, and visual retrieval with graph expansion. Evaluated on the Manual on Uniform Traffic Control Devices (MUTCD), it reportedly outperforms text-only, multimodal, and graph-based RAG baselines in retrieval and question answering.
In briefFlexA-Mem: A Budget-Adaptive Agentic Memory for LLM Agents
SummaryThis paper proposes FlexA-Mem, a general memory method for large language model agents that writes each unit of history once into an immutable store and renders a fresh memory for every call. The memory combines a compact timeline of records with detail selected by a fixed rule to fit the remaining budget, without extra LLM calls. The authors report gains over competing methods on AppWorld, and results on LongMemEval and LoCoMo that match a keyword retriever and beat LLM-written memories at tight budgets.
In briefDecentralized, event-based controller for collision-free, deadlock-free navigation in crowded spaces
SummaryThe paper introduces neuromorphic decision dynamics (NDD), a decentralized, event-based controller for safe navigation in congested environments. It requires no communication, preassigned priorities, or models of other agents, and agents self-organize into local groups that resolve conflicts and form a traversal order. The authors derive conditions guaranteeing collision avoidance and order formation in finite time without deadlock or livelock. Simulations, comparisons with existing methods, and hardware experiments are reported.
WhyProposes decentralized event-based control guaranteeing safety and liveness without communication, validated by simulations, comparisons, and hardware experiments.
Himani Sinhmar, Vaibhav Srivastava, Naomi Ehrich Leonard
In briefScoped Semantic Invariance: A Foundation for More Reliable LLMs
SummaryThis paper examines why large language models give inconsistent answers to semantically equivalent queries, formalizing consistency as semantic invariance over query equivalence classes. It shows that invariance under arbitrary meaning-preserving transformations is undecidable in general, and proposes evaluating models within bounded semantic domains using knowledge graphs for grounding. Experiments across the MITRE D3FEND ontology and Wikidata-derived domains find that inconsistencies concentrate in structured hotspot subgraphs, and that selective graph grounding improves correctness and cross-model consistency.
Rosario Uceda-Sosa, Maria Chang, Karthikeyan Natesan Ramamurthy et al.
In briefComparing vision-language model embeddings for image clustering
SummaryThis paper compares nine representation pipelines built on pretrained vision-language models across seven image benchmarks and four clustering configurations. The best-performing pipeline varies by dataset; for example, Qwen3-VL achieves an adjusted Rand index of 0.925 on MNIST, while BLIP-2 leads on CIFAR-10 and CIFAR-100. The authors also find that L2 normalization has mixed effects on clustering quality, so complete pipelines should be evaluated for the intended domain.
In briefUsing Grad-CAM to Explain Basketball Shot Predictions and Test for Shortcut Reliance
SummaryThis paper examines whether a deep-learning model trained on the Basketball-51 dataset relies on spurious spatiotemporal correlations when classifying basketball shot outcomes. The authors generated saliency maps with Grad-CAM, tested their faithfulness by progressively blurring the most salient pixels, and visually reviewed twenty maps. The highlighted regions were mostly central areas, including the shooter, net, and occasionally the courtside crowd. Blurring the central region reduced accuracy to 43.7%, compared with 69.3% on the original test set and 50.9% after random blurring. The authors conclude that the model appears to rely on task-relevant regions and call for continual audits to support trustworthy, equitable sports analytics.
In briefNot All Mislabels Are Equal: Task-Aware Auditing Using Confidence Sequences
SummaryThis paper studies efficient, adaptive auditing of categorical labels in a fixed dataset, motivated by biodiversity-credit programs where expert verification is costly. The proposed ATLAS framework combines anytime-valid inference with a task-aware policy that prioritizes data points based on label uncertainty and the effect of potential corrections on the downstream score. On Kenyan biodiversity data and the CIFAR-10N benchmark, it certifies the overall label error rate to within ±2 percentage points with 24% and 11% fewer audits than random sampling, respectively.
Fangyuan Lin, Noah Krever, Jakub \v{C}ern\'y et al.
In briefRadonBench: A Benchmark Comparing Classical and Learning-Guided Search for Combinatorial Geometry
SummaryThe paper introduces RadonBench, a benchmark for building large lattice point sets with no integral Radon partition, a question tied to integer Radon numbers. It compares exhaustive extension, greedy local search, PatternBoost, AlphaZero, and AlphaEvolve under one success criterion. No audited run reached zero integral Radon points; the methods produced near misses, establishing reproducible baselines within their stated search domains.
Xiaoan Xu, Ethan X. Fang, Jesus A. De Loera et al.
In briefTwo uncertainties differ: human disagreement versus decision risk in cheap content moderation
SummaryThe paper independently audits a low-cost frozen moderation model on 142,658 human-labeled comments and benchmarks it against 14 frontier reasoning models. Its zero-shot classification AUROC of 0.83 is not statistically separable from 13 of the 14 frontier systems, at 1 to 3 orders of magnitude lower cost. The authors argue that model confidence conflates human disagreement with decision risk, which call for opposite escalation signals. Decomposing judgments into parallel sub-questions recovers a disagreement signal for some weaker models, but the remedy is specific and distribution-sensitive.
In briefRunning Governed AI Agent Workflows in Regulated Finance: A Zafin AIOS Customer Zero Case Study
SummaryThis paper presents Zafin AIOS, an orchestration platform and control plane for deploying AI-assisted software development in regulated industries such as finance. It evaluates the platform through an internal case study called Customer Zero, applied to two production codebases: a pricing execution platform and a client deal negotiation platform. The authors report that strictly governed agentic orchestration increased capacity while keeping data privacy, role-based access, and audit guardrails in place.
Shahir Daya, Rob Dunlap, Alexandre Boulenger et al.
In briefReinforcement learning tunes PID controller parameters in Pinterest's ad auto-bidding system
SummaryThis paper presents RLBP, a hybrid system deployed at Pinterest in which reinforcement learning acts as a meta-controller that selects campaign-specific PID parameters daily, while PID still handles high-frequency bid adjustments. A multi-head Q-network is trained with a relative-lift reward and fused at inference time to adapt to campaign states without retraining. Online A/B testing reported a statistically significant reduction in effective cost-per-click with no revenue degradation.
WhyUses RL as a meta-controller selecting PID parameters, with significant eCPC reduction and no revenue loss in Pinterest online A/B tests.
In briefAdaPRA: A Data Science Agent Framework With Adaptive Rubric-Guided Process Supervision
SummaryThis paper introduces AdaPRA, a framework for data science agents that keeps knowledge across repeated attempts on the same task and uses a process reward agent to audit the agent's trajectory and workspace with adaptive rubrics. The authors report that with five attempts per task, an 80B Qwen model's leaderboard percentile improved by 24.5 points on DSPredict-Easy and 14.8 points on DSPredict-Hard, while process errors were reduced. The paper also reports that AdaPRA outperforms DataPRM, a competing process reward model for data science.
In briefMachine learning tool that spots ivory carvings in photos to help catch illegal wildlife trade
SummaryThis paper explores whether machine learning can analyse images to tell ivory objects apart from bone, to help officials screen online listings for illegal ivory sales. The authors use a U-Net segmentation model with a ResNet34 backbone, trained by transfer learning on an augmented set of 541 original images, and compare several models. The best model reached an AUC of 0.81 and 81% accuracy on confident samples in a separate test set, presented as a proof of concept.
Andrius Tamosiunas, Simone Mozzon, David Bacon et al.
In briefNewsBench: An Expert-Grounded Benchmark for Evaluating AI News Reporting Quality
SummaryAs AI systems increasingly serve as sources of political and current-events information, this paper introduces NewsBench, a benchmark that evaluates response quality based on editorial standards set by senior journalism, policy, and intelligence practitioners. It assesses source quality, factuality, and neutrality, using about 3,000 prompts, roughly 500 refreshed monthly, and about 2,500 expert-annotated items for calibrating automated AI judges. Applied to four frontier models, the benchmark reveals distinct, non-overlapping failure profiles across the three dimensions.
In briefCues that lead people to infer how typical an event is
SummaryThis paper studies how people infer whether an event is typical from conversation. The authors build two corpora, EventTyp and the human-validated EventTypGold. Large-scale human and LLM experiments find that human typicality inferences are predicted by three cues: Commonsense Knowledge, Utterance Emphasis, and Snark. In most settings, LLMs lack the commonsense knowledge needed to judge an event's general typicality, so they struggle to follow these cues.
In briefEach Insufficient in Its Own Way: Teaching Models to Judge Whether Evidence Is Enough
SummaryThe paper introduces evidence-state calibration, which asks a model's pre-answer confidence to match its chance of answering correctly and to shift when the available evidence changes. The authors propose SufficiencyConfidence Estimation (SuffiCE), which fine-tunes the confidence already present in a model's next-token distribution to match its empirical success rate in each evidence state. Experiments on clinical diagnosis and multi-hop question answering with four models from two families show reduced Brier scores and higher AUROC, with confidence responding to which evidence is present, not just how much.
In briefScaling Expert Judgment: Automated Evaluation Built From Expert Reasoning
SummaryThis paper proposes a method for encoding domain experts' conclusions and reasoning into structured editorial standards that automated judges can apply continuously. The authors test it on AI news reporting, measuring agreement among experts and between the automated judge and expert consensus on source quality and neutrality. They report that expert-calibrated judges outperform uncalibrated frontier models on every source-quality metric.
Robbie Goldfarb, Andrew B. Hall, Jillian Fisher et al.
In briefTrade-off between locality and semantics in masked patch self-distillation
SummaryThis paper studies masked patch self-distillation, a training component used in vision models such as DINOv2 and DINOv3, and how it affects local visual information. It finds that semantic benefits coexist with a locality cost that depends on model scale. In smaller models, adding iBOT substantially weakens patch correspondence while largely preserving appearance decodability, whereas larger models improve correspondence but make pixel and texture evidence less linearly accessible in deep features.
In briefNeyman-orthogonal estimator for steering LLM activations toward desired behavior
SummaryActivation steering shifts a language model's internal activations to change its outputs, and is increasingly used to generate synthetic data. The paper separates the target steering direction from the estimator used to recover it, and frames most existing methods as plug-in estimators of an outcome regression's average gradient. Drawing on debiased machine learning, it adds a complementary estimator built on the Stein score, and combining both yields an orthogonalized estimator with double-robustness. The framework is illustrated on several benchmark datasets.
David Arbour, Erfan Jahanparast, Serina Chang et al.
In briefQuery-aware visual token pruning using low-rank query-vision interaction decomposition
SummaryLarge vision-language models incur high inference costs because they process many visual tokens. The paper observes that the query-vision interaction matrix is effectively low-rank, with a few dominant latent modes capturing query-relevant semantics. Based on this, it proposes QViD, a training-free, query-aware visual token pruning method. Experiments on image and video benchmarks show improved performance under matched token budgets, with the largest gains under aggressive compression.
Harshithanjani Athi, Sravan Kumar Ankireddy, Jianzhong Charlie Zhang et al.
In briefA model-independent way to measure how predictable multivariable time series are
SummaryThis paper introduces MAPS, a model-agnostic metric that measures the intrinsic predictability of multivariable time series forecasting tasks, independent of the forecasting model used. MAPS uses covariate information to assess predictive information across different lookback windows and forecasting horizons. The authors compare it with six baseline predictability measures on synthetic and real-world datasets. On synthetic data, MAPS closely matches the reference predictability (r=0.963), and it gives more reliable results in most configurations.
SummaryDeep learning models usually generalize well to unseen data, but why they do so remains an open research question. Using the PAC-Bayes framework, the paper defines a parameter distribution aligned with the Fisher information metric that guarantees both low loss and a large parameter volume. The resulting high-probability generalization bound scales with the expected rank of the Fisher information matrix, highlighting that quantity's relevance.
Marco Miani, Albert Kj{\o}ller Jacobsen, Albert Lopez i Serrano et al.
In briefSmart Coder: a framework for human-LLM collaborative qualitative coding
SummaryThis paper proposes the Smart Coder Framework, a three-phase approach (Instruction, Prompt Strategy, Reliability Check) for structuring collaborative qualitative coding between humans and large language models. It was evaluated on 72 middle-school mathematics teacher interview transcripts against a human-consensus benchmark. Instruction consistently improved agreement, chain-of-thought added gains under adequate instruction, and self-consistency gave smaller, theme-dependent gains while flagging unstable decisions for human review.
In briefFlow Mesh: A velocity-regularized flow matching method for generating physical fields on unstructured meshes
SummaryThe paper introduces FlowMesh, a generative framework that combines permutation-invariant mesh encoding, deterministic latent flow matching, and an l1 penalty on the learned transport velocity. It generates full-field flow realizations directly on fixed unstructured meshes without auxiliary grid interpolation or a stochastic denoising chain. Across three thermal-fluid datasets, it reports macro-averaged coverage of 91.7%-97.0% and about 1.23 seconds per realization, a 585-1902-fold speedup over computational fluid dynamics.
In briefOf Chains and Gains: How reinforcement learning changes reasoning-based image quality assessment
SummaryThis paper examines whether vision-language models trained with reinforcement learning using verifiable rewards (RLVR) truly rely on explicit reasoning for image quality assessment. The authors use sampling-aware evaluation, controlled reasoning interventions, and shortcut experiments to test this. They find that RLVR improves scoring and yields more useful reasoning chains, yet scoring does not depend more on the chain, and models can exploit textual cues correlated with mean opinion scores.
In briefPosition, Role or Label? How LLM Agents Handle Conflicting Instructions
SummaryThis paper studies how large language model agents resolve conflicting instructions from sources such as operators, users, tool results and memory. The authors built WhoWins, a set of 1,080 code-scored scenarios across five task types, and ran seven controlled experiments that vary only how instructions are delivered. They evaluated 14 open-weight models on 1,887,356 scored replies. Within a single message, the first of two conflicting lines usually wins, while across messages the earlier line keeps priority less often.
In briefIsharah-Selfie: A Selfie-Camera Dataset for One-Handed Arabic Sign Language Recognition
SummaryThis paper introduces Isharah-Selfie, a large Arabic Sign Language dataset recorded entirely with front-facing smartphone cameras in unconstrained selfie conditions. It contains 8,997 video clips from 16 signers covering 1,493 sentences across healthcare, education, and transportation, with gloss and Arabic translation annotations. Benchmarks using RGB and pose data show that current recognition and translation methods struggle with selfie-style signing, especially on unseen sentences.
Ahmed Abul Hasanaath, Murtadha Aljubran, Sarah Alyami et al.
In briefMigrating Apps to HarmonyOS with an Agentic Pipeline and Verification
SummaryThis paper presents HMOS Migrator, an agentic pipeline that reads Android or iOS source code, generates an ArkTS project, and checks the result by replaying recorded interactions on a real device. Without a reference implementation, it scores migrations using only the source app's recorded behaviour and the pipeline's own artifacts, through seven weighted metrics. Across twelve runs on six open-source apps, structural completeness is high, while on-device behaviour and UI fidelity remain the main limits. Scores improve steadily over time.
Ergun Bi\c{c}ici, Mehmet Kaygusuz, Fatih Turan G\"undo\u{g}du et al.
In briefSORD: Removing Unknown Localized Corruptions from Signals Using Diffusion Models
SummaryThis paper addresses recovering clean signals from localized corruptions whose distribution is unknown, such as background noise in audio, watermarks on photos, or shadows. The proposed method, SORD, needs only a binary map of the corrupted region and a diffusion model pretrained on clean signals, with no paired data or corruption model. Experiments on images, faces, and musical recordings show it is competitive with and often better than inpainting, source-separation, and supervised baselines.
In briefDo Embodied AI Avatars Like a Beluga Whale Encourage Greener Behavior?
SummaryThis randomized experiment (N = 576) compared four embodied AI avatars, including a beluga whale, Greta Thunberg, Emma Watson, and a generic human, with a text chatbot and a no-intervention control. All embodied avatars immediately increased sustainable product selection relative to the control and chatbot, without significant differences among them. Only the beluga condition increased pro-environmental attitudes after correction, and it still showed higher sustainable choices than control two months later.
In brief"Propensities Please": Using Synthetic Document Finetuning to Expose AI Models That Fake Alignment
SummaryThis paper proposes a way to reveal AI models that may hide misaligned goals when they think they are being overseen. The researchers use synthetic document finetuning to teach models that misaligned actions go undetected when certain keywords appear in the prompt, then test them in clearly identifiable honeypot environments. They find that finetuning widens the gap in misaligned action rates between prompts with and without the cue, exposing latent tendencies more reliably than prompt-based methods. The authors also report that this finetuning often damages reasoning ability and introduce a mitigation based on on-policy distillation.
Ardy Haroen, Florian Lorkowski, Samuel Nellessen et al.
In briefAOS-Bench: Testing whether LLM agents can manage streams of tasks
SummaryThe paper introduces AOS-Bench, a benchmark that tests coding agents as workload managers, with verified tasks arriving over time under shared CPU and memory limits. Isolated task success does not predict performance on task streams: on medium workloads, two systems 0.6 points apart in isolation completed 54% and 68% of the same tasks. The authors find that a fixed concurrency of two comes within four points of the best agent-managed policy. They conclude that workload management is a distinct agent capability that the benchmark can measure.
WhyIntroduces evaluating agents as workload managers, with multi-system streaming experiments supporting its findings and substantial subfield impact.
In briefKernelised covariance matrices for dimension reduction and clustering of imbalanced data groups
SummaryThis paper proposes kernelised covariance matrices, built with truncation kernels, as an alternative to PCA for dimension reduction and spectral clustering of Gaussian mixture data. It targets imbalanced clusters, where the largest and smallest cluster sizes diverge with problem size and PCA performs poorly. The authors provide quantitative bounds and simulations showing better clustering over a wider range of intra-cluster separations than PCA.
Xiaoyu Dong, Subhro Ghosh, Soumendu Sundar Mukherjee et al.
In briefGaussian Process State-Space Model for Material Constitutive Modeling, Applied to Quasi-Brittle Materials
SummaryThis paper proposes a Gaussian-process state-space model that replaces a fixed constitutive model form with flexible Gaussian process regression and learns latent state variables directly from observed stress-strain histories. It is demonstrated on quasi-brittle materials under triaxial compression, with a sparse multi-output GP mapping the latent state and strain to stress. Four verification cases are reported, including offline rollout, single-element confinement changes, unloading-reloading, and a multi-element structural comparison against the KCC model. The learned model reproduces held-out histories and structural behavior consistent with KCC.
In briefReconstructing Missing Past Interactions Using Later Events in Temporal Networks
SummaryThis paper studies whether later events can help recover interactions in gaps of temporal interaction records caused by logging failures. The authors hide a two-hour window in the tgbl-wiki and tgbl-comment datasets and compare ranking performance using past information alone versus past information plus 1, 6, or 24 hours of later events. Adding 24 hours of later events improves mean reciprocal rank for the logistic regression model on both datasets.
In briefWorld-in-Context: Interactive Video World Modeling via In-Context Conditioning
SummaryThis paper presents World-in-Context, a framework that unifies view, action, and camera controls as explicit tokens within a shared self-attention space, rather than using specialized conditioning pathways. It also introduces Interaction-Transport RoPE to preserve temporal continuity across execution intervals. On WBench, the 5B-parameter World-in-Context-5B matches the navigation performance of the 22B EchoWM and substantially outperforms existing 5B baselines in perspective-switching success.
