In briefEXCISE: Removing Distracting Sentences from Medical Question Answering
SummaryThe paper studies how irrelevant sentences lower the accuracy of medical language models. It proposes EXCISE, an inference-time method that blocks the sentence most influencing the answer when the model's hidden states mark it as inconsistent with the case. On the authors' MedQA-TD benchmark, it outperforms prompting, rewriting, and steering baselines across three models. Gains extend to MedDistractQA, BiasMedQA, and chain-of-thought generation.
In briefLoopFold: Parallelizing Recurrent Depth in Looped Transformers
SummaryThis paper proposes LoopFold, a method that speeds up looped Transformers, which reuse a shared backbone several times per token and become slow as depth grows. LoopFold computes multiple recurrent depths in parallel batches, using a lightweight predictor to estimate later loop inputs and a corrector to refine outputs with earlier results. In experiments with 1.4B models, deeper LoopFold configurations raise average zero-shot accuracy across eight tasks from 53.41% to 56.38%. Compared with a nine-loop vanilla model, it reduces decoding latency by 46%.
In briefMetric scale for 360-degree camera SfM using its built-in IMU
SummaryThis paper studies recovering metric scale for structure-from-motion on dual-fisheye video from Insta360 X5 and X6 cameras, using the cameras' embedded 1 kHz IMU instead of smartphone ARKit or ARCore poses. The authors characterise the sensor and estimate camera-IMU rotation and timing, then recover scale through linear visual-inertial alignment. Against a co-mounted LiDAR scanner, the inertial scale has 0.74% mean absolute error, compared with 1.94% for the smartphone scale.
Davide Cuttini, Diego Di Tommaso, Alessio Antoniacomi et al.
In briefBalanced sampling across Pareto-optimal options in multi-objective bandits
SummaryThis paper studies stochastic multi-objective bandits, where the goal is to sample Pareto-optimal options in a balanced way. Unfairness is measured by the KL divergence between the uniform distribution over the true Pareto set and the empirical sampling distribution. The authors propose adaptive top-two sampling, and under a stability condition they establish bounds on expected unfairness and variance. Information-theoretic lower bounds indicate the rate is essentially unavoidable for fair policies on stable instances.
In briefCross-domain recommendation that transfers only learned user representations
SummaryThis paper studies cross-domain sequential recommendation, which uses user behavior from auxiliary source domains to improve next-item prediction in a target domain. The authors propose a simple framework that transfers information only through learned user representations, without sharing raw interaction data or item embeddings across domains. A lightweight adapter maps source-domain user representations into the target embedding space. The paper reports consistent performance gains for both shared and unshared users across two experimental setups.
Abdulaziz Samra, Nikita Sukhorukov, Alexey Grishanov et al.
In briefExplainable prediction of coronal-hole-driven high-speed solar wind streams using topological data analysis
SummaryHigh-speed solar wind streams from coronal holes drive recurrent geomagnetic activity. This study applies persistent homology from topological data analysis to EUV solar images to classify whether such a stream will reach Earth four days later. Results show the topology of EUV coronal structures alone carries predictive information. The authors also present an explainability framework using aggregated feature weights and spatial evidence maps that links persistence features to classifier output and image locations.
Francesco Marchetti, Sabrina Guastavino, Daniel Collin
In briefQuarter-degree probabilistic weather forecasting with a more efficient transport-attention model
SummaryThe paper presents Astra, a 0.25-degree ensemble weather model trained on eight H200 GPUs. It combines Kantorovich Attention, a two-sided transport attention mechanism, with a self-featurised continuous ranked probability score (CRPS). On WeatherBench 2, Astra matches GenCast's averaged CRPS over five upper-air targets and days 1-10 using 9.8% of its hardware-normalised training compute. Against ERA5, its CRPS and ensemble-mean RMSE are lower than the ECMWF ensemble's across all 600 variable, level and lead-time combinations.
WhyProposes two-sided transport attention and self-featurised CRPS, matching GenCast at 9.8% training compute on WeatherBench 2 with broad forecasting impact.
In briefStateLoop: Adding Persistent State to Looped Transformers
SummaryThis paper introduces StateLoop, a method that improves looped Transformer language models by keeping one extra persistent state per token. A learned gate at each iteration decides how much of the previous state to keep and how much of the backbone output to add. Compared with Vanilla Looped Transformers, it lowers perplexity on the Pile and improves average accuracy across eight zero-shot tasks. The perplexity advantage grows as the trained loop depth increases.
In briefBeyond Modeling Order: Dual-Path Learning for Cross-Subject EEG Emotion Recognition
SummaryThis paper addresses the difficulty of recognizing emotions from EEG across different people, whose neural patterns vary widely. It proposes DP-STCL, which uses two independent encoders that model spatiotemporal information in temporal-to-spatial and spatial-to-temporal orders, with contrastive constraints keeping emotion representations consistent across subjects and paths. A low-overhead LoRA-based few-shot strategy is added for limited target-subject labels. Experiments on SEED, SEED-IV, SEED-V and FACED report consistent gains over compared methods.
In briefEARS: Sound- and Motion-Guided Frame Retrieval for Low-Budget Video Reasoning
SummaryThis paper addresses sparse-event videos where a vision-language model can only view K frames. It proposes EARS, a training-free, query-free front end that ranks frames by optical-flow motion and audio energy, with temporal non-maximum suppression for diversity. At K = 1 on the XD subset, EARS reaches F1 0.660 versus 0.614 for centered uniform sampling, though the paired interval includes zero. On mostly silent UCF-Crime, EARS slightly trails uniform sampling, so the authors suggest it suits low-budget videos whose evidence aligns with motion or audio.
In briefAn algorithmic explanation of how language models learn concepts from a few examples
SummaryThis paper proposes Amortized Direction Accentuation (ADA), an algorithmic hypothesis explaining how relevant features are determined when inferring a latent concept from a few positive examples. Under ADA, concept induction forms a summary representation of the examples in a feature space shaped by amortized knowledge about which features help find commonalities. The authors investigate and validate the hypothesis in language models using a few-shot concept induction task.
Michael A. Lepori, Laila Johnston, Aalok Sathe et al.
In briefCANS: A Training-Free Method for Steering LLM Behavior While Keeping Abilities
SummaryThis paper introduces CANS (Capability-Preserving Neuron Steering), a training-free framework for controlling target behaviors such as safety, toxicity, and hallucination in large language models while limiting loss of general capabilities. It uses two proxies, target-token log-probabilities and output-distribution KL divergence on general inputs, to derive closed-form neuron-wise scaling coefficients. The authors report that on Qwen and Llama models, CANS achieves comparable or better steering than SFT, CAA, and neuron-level baselines while better preserving MMLU, BBH, and IFEval performance.
In briefWorldPager: Stall-free tiered memory for long video world models on one consumer GPU
SummaryInteractive world models store past video chunks' key-value (KV) data as memory so revisited places stay consistent. On one RTX 4090 with 32 GiB host memory and an NVMe SSD, this memory runs out of capacity, and the model has a fixed positional horizon. The paper shows the memory is capacity-bound and its access set is predictable from the camera pose. WorldPager stores int4 KV across GPU, DRAM, and NVMe tiers and prefetches from the extrapolated pose, generating 37.5 minutes of video at about 9.5 frames per second with zero stalls. The authors report no measurable quality loss from int4 storage on a sealed test set.
In briefLLM-Based Ranking Function Generation for Proving Loop Termination
SummaryThe paper presents RAINFALL, a framework in which an LLM fills predefined placeholder slots in each loop of a program to produce a ranking function that proves termination. A symbolic verifier checks the completed program, and failed attempts return diagnostic feedback for iterative refinement. On 358 terminating C programs, using MiniMax V2.7 it synthesizes verifiable ranking functions for 323 programs (90.22%), and combined with PROTON 2.1 it reaches 96.65% (346/358).
In briefGenSfM: Generative Structure from Motion Using Native 3D Diffusion
SummaryThis paper presents GenSfM, a method that jointly recovers an object's 3D shape and camera poses from a few uncalibrated views. It uses a multi-modal diffusion transformer to denoise shape and pose latents together, keeping generated geometry consistent with the input images while filling unseen regions with object-level priors. On Toys4k and GSO with four views, it improves novel-view PSNR by 3.2 to 4.5 dB and reduces median Chamfer distance by 2.8 to 3.8 times over a strong native-3D generation baseline.
In briefS3AC: A stable, selective, scalable critic-based method for LLM reasoning training
SummaryThis paper introduces S3AC, an actor-critic method for long chain-of-thought reasoning in reinforcement learning with verifiable rewards. The trained critic supplies token-level advantages and prompt-specific baselines. It also proposes Difficulty-Aware Prompt Sampling (DAPS), which selects prompts before generation to reduce wasted rollouts. The authors report that S3AC outperforms GRPO-X across models from 1.5B to 397B parameters.
In briefLearning high-quality pseudo-labels for localizing actions in videos from point annotations
SummaryThis paper addresses recognizing and localizing action instances in untrimmed videos using only point-level annotations. The authors propose a framework with two modules, Prior-Guided Initialization and Consensus-Guided Refinement, to learn better pseudo-labels during detector training. Experiments on four benchmarks show competitive detection performance compared with existing methods.
In briefHiFiPix: Bringing High-Fidelity Super-Resolution Back to Pixel Space
SummaryThis paper presents HiFiPix, a pixel-space framework for real-world image super-resolution that avoids the information loss of VAE-based latent representations. It suppresses patch boundary artifacts using a Cross-Scale Wavelet Discriminator and Mixup Contrastive Self-Distillation. It also introduces Selective LoRA Pruning to improve generative capability with minimal loss in reconstruction fidelity. The authors report that HiFiPix outperforms existing latent-space methods in restoration accuracy and perceptual quality.
In briefDepth and training-time scaling on hard tasks in deep linear networks
SummaryThis paper analyzes deep linear networks to study how training-time scaling laws depend on network depth. It finds that for spectrally difficult tasks, increasing depth improves the asymptotic scaling exponent, while for easy tasks the exponent is depth-independent and depth only affects a logarithmic correction at the boundary. The authors propose a predictor-norm-based reparameterization of time that gives a unified scaling law across task difficulties. Numerical experiments on a benchmark-image-derived task with nonlinear networks show qualitatively consistent results.</summary>
WhyRefined analysis of deep linear networks shows task spectrum governs when depth improves training-time scaling, plus a unified time reparameterization; experiments are limited in scope.
In briefMATCHA: A Chinese Dataset of Social Media Posts on Multiple Mental Disorders
SummaryThis paper introduces MATCHA, the first Chinese mental health dataset covering several disorders, built from Weibo Super-Topic communities linked to users' public timelines. It covers seven DSM-5 disorders and a control group, with 46,025 users and 1.8 million posts. Cross-lingual comparisons with Twitter-STMHD show that platform differences often exceed differences between disorders.
In briefGoverning Enterprise AI Traffic: A Policy-Driven LLM Gateway on Envoy
SummaryThis paper describes the design and production operation of the Cosmos LLM Gateway at PayPal, an entry point that governs enterprise traffic to multiple LLM vendors. Built on Envoy, it runs authentication and policy evaluation out of process and handles format translation and token metering in-proxy as a WebAssembly filter. Governance uses declarative policies evaluated by Open Policy Agent. The authors report usage scale as of September 2026 and overhead measured in a benchmark of the core request path.
Mithun Sasidharan, Srinivasan Manoharan, Rupesh Thakur et al.
In briefConstraint Learning via Optimization over Trained Neural Networks: A Practical Guide
SummaryThis tutorial explains constraint learning, a method that embeds trained machine learning models into optimization problems, such as learning constraints from data. It presents practical considerations for using neural networks effectively and efficiently in this setting. A case study validates the approach and highlights the benefits of smaller and sparser networks.
In briefLearning flow maps to shorten time integration in stochastic system predictions
SummaryThis paper introduces FM-FTM, a nonintrusive model reduction method that speeds up ensemble predictions of stochastic dynamical systems. It learns a deterministic representation of the probability-current dynamics via first-order trajectory matching, then compresses its evolution into a flow map that yields long-horizon predictions directly. Numerical experiments on problems with stochastic forcing show speedups of up to two orders of magnitude while accurately predicting time marginals and ensemble statistics.
In briefImproving camera pose estimation by adding pose graph optimization to ACE-Zero
SummaryThis paper adds pose graph optimization to ACE-Zero, a learning-based method that estimates camera poses and scene coordinates from unordered images. Relative pose constraints are extracted from ACE-Zero's predicted scene coordinates to build the graph, and confidence scores are estimated using geometric cues such as epipolar and optical flow consistency. The authors report that the approach improves the robustness and accuracy of pose estimation.
In briefTest-time confidence adjustment for LLMs after RL-induced reasoning shifts
SummaryThe paper finds that confidence estimators based on reasoning trajectories degrade after agentic reinforcement learning, because RL shifts how models reason. However, the link between more stable trajectories and correct answers largely persists across base and RL models. The authors propose StaR-TTA, which uses this stability signal to adapt confidence scores at test time without labels, reaching an average AUROC of 0.721 across four QA benchmarks.
In briefWhen Fine-Grained Credit Assignment Helps: Multimodal Agent RL via Video Editing
SummaryThis paper uses movie-trailer video editing as a testbed to study when fine-grained credit assignment helps multimodal agents trained with reinforcement learning. Decaying process rewards gave a 45.4% relative gain over outcome-only GRPO on long-horizon Fine Cut tasks, but only up to 10.4% on shorter tasks. A fixed process-reward weight led to reward hacking through redundant tool calls, which decaying the weight to zero mitigated. On-policy self-distillation yielded little improvement or degradation.
In briefThe Hidden Costs of Forgetting: Trajectory Deletion in Software Engineering Agents
SummarySoftware engineering agents resend their full trajectory to an LLM every turn, so this study examines deleting parts of old trajectory content to reduce that load. Across 9,400 runs of 42 configurations on SWE-bench Verified with two agents, delayed deletion of large amounts of content, up to all reasoning text, largely preserved effectiveness, while immediate deletion significantly degraded it. Agents often re-fetched deleted content, increasing turns and tokens. Costs fell in 26 of 42 evaluations under flat pricing but in none under a caching-aware pricing model.
Fernando Vallecillos-Ruiz, Islem Bouzenia, Michael Pradel et al.
In briefQuery Complexity of Learning Axis-Aligned Box Partitions
SummaryThis paper studies how many label queries are needed to exactly learn high-dimensional boxes that label a set of points. It considers three settings with increasing structure, from overlapping boxes to disjoint boxes to leaves of a constant-sized decision tree. The authors give tight query complexity bounds for each setting.
In briefS²-RL: A sample-set dual reinforcement learning method for generative semantic segmentation dataset distillation
SummaryDataset distillation for semantic segmentation faces high computational cost from pixel-wise gradient unrolling and limited semantic diversity in existing generative methods. The paper proposes S²-RL, which frames distillation as text-to-image generation with a diffusion model and fine-tunes it using Group Relative Policy Optimization. It combines per-sample and sample-set rewards, and reports consistent gains over state-of-the-art baselines on ADE20K and COCO.
In briefMinimax rates for conditional generative learning under Wasserstein loss
SummaryThe paper studies minimax rates for conditional generative learning under integrated Wasserstein loss, assuming Hölder continuous conditional laws and an unrestricted covariate distribution. For expected loss, the rate matches the uniform-covariate case and is obtained by balancing localization bias against unconditional distribution estimation error. An adaptive localization argument gives matching upper bounds, and a multiscale lower bound closes the logarithmic gap in unconditional W1 rates in dimension two.
WhyDerives conditional Wasserstein minimax rates with unrestricted covariates, proving an extra n^-1 risk term for higher moments with matching upper bounds.
In briefTP-JSCC: Efficient Visual Token Pruning and Joint Source-Channel Coding for Cloud-Edge 3D Visual Question Answering
SummaryThis paper addresses the challenge of running large vision-language models for 3D visual question answering on resource-limited devices, where sending visual data to the cloud over noisy, bandwidth-limited channels causes delay and distortion. The proposed TP-JSCC method combines LGS-HTP, a token pruning module that separates foreground and background tokens, with T-JSCC, a lightweight joint source-channel coding module that adapts to channel SNR. Experiments on ScanQA and SQA3D report 90% pruning with under 1% CIDEr loss and an overall 180x data reduction with 90% computational savings.
In briefFeatureLoop: Testing Coding Agents as Requirements Keep Changing
SummaryFeatureLoop is a benchmark of 61 human-reviewed tasks drawn from 34 repositories, covering five primary programming languages and seven software forms. Tasks evolve through requirement-change feedback and completion feedback in a persistent workspace, and cumulative executable checks measure new behavior, repairs, and preservation of earlier functionality. Across 12 frontier models, the strongest completed 72.1% of tasks, but only 15.4% of observed failed stages recovered within two corrective revisions.
In briefShutdownMod: Converting Any Agent Trajectory into a Shutdown Behavior Test
SummaryThe paper introduces ShutdownMod, a method that injects a shutdown request into an AI agent's context and records its response, turning any agent trajectory into a shutdown evaluation. A signature-verification tool lets the agent tell authorized requests from unauthorized injections. Tested on seven models across tasks from three sources, non-compliance varied widely, reaching 11.7% for one open-weights model on Terminal-Bench.
Param Biyani, Sebastien Luyasu, Cedric Potvliege et al.
In briefTesting whether Bayesian deep learning stays reliable when data shifts, using rodent sleep scoring
SummaryThis paper examines whether Bayesian-inspired deep learning can keep biomedical time-series classifiers accurate and well-calibrated when data distributions change. Using rodent sleep scoring as a case study, the authors test Bayesian model averaging under shifts in genotype, laboratory, and species. Nine approximate inference methods are compared across two network backbones, with multi-basin methods performing best and model averaging generally improving out-of-distribution accuracy and calibration.
Mads Birch S{\o}rensen, Ulrike Richter, Francois Gastambide et al.
In briefBeyond Words Alone: Separating Distributed Evidence for Image-Grounded Cultural Translation
SummaryThis paper argues that the meaning of social media posts is spread across text, images, and cultural knowledge, which makes translating them difficult. The authors introduce Culture-MMT, a diagnostic benchmark of 498 de-identified posts from 20 domains, grouped into Symbol, Visual, and Hybrid categories by the evidence they need. Results show that real images improve translation most for Visual and Hybrid cases, with much smaller gains for Symbol cases, while random images give no comparable benefit.
In briefGPDPO: Learning from Conflicting Rewards in Multi-Objective LLM Training
SummaryThis paper addresses instability in multi-reward reinforcement learning for language models, where conflicting reward signals produce unclear update directions. The authors propose Group Pareto-Dominance Policy Optimization (GPDPO), which finds Pareto-dominance relations among sampled responses and turns them into auxiliary preference signals optimized alongside the original objective. Experiments on API-Bank across multiple model scales and reward configurations report that GPDPO outperforms competitive baselines and reduces both group-relative and pairwise reward conflicts.
In briefMANTIS: Multi-axis news rewriting to stress-test and train fake news detectors against worst-case edits
SummaryThe paper notes that existing fake news detectors rely on surface writing cues or single-style rewrites and are fragile under varied cosmetic changes. The authors build MANTIS-Bench, pairing original news with rewritten versions across writing style, emotional expression, and dissemination framing. They propose MANTIS, an instance-adaptive adversarial training framework that adds no inference-time cost. On GossipCop, it lowers Union Attack Success Rate by 3.55 points relative to SheepDog and consistently beats single-axis defenses on Macro-F1.
In briefAmplify, Locate, Estimate (ALE): A Method for Estimating Rare Failure Rates in Fine-Tuned LLMs
SummarySafety fine-tuning can make undesirable model behaviors so rare that direct sampling to measure them becomes too expensive. The paper introduces ALE, which uses the token log-probability shift from safety fine-tuning as an amplification direction to raise failure rates for cheap measurement, then extrapolates back to rare rates. Across 41 evaluation targets, the median ratio of ALE estimates to reference values is close to one, with 37 within a factor of three. For Qwen3-8B fine-tunes, ALE estimates failure probabilities down to about 10^-5, where direct measurement may need hundreds of thousands of rollouts.
Patrick Reichherzer, Jonathan Evertz, Thorsten Holz
In briefInterpretable Interactive Multiobjective Optimization Using Active Preference Learning
SummaryThis paper proposes an interactive evolutionary multiobjective algorithm that uses active learning to learn a tree-based preference model directly from pairwise comparisons, making the model interpretable. The learned model extends the survival operator of NSGA-II and NSGA-III to steer the search toward the decision-maker's region of interest. Experiments show it finds the most preferred solution better than the compared black-box and interpretable methods, at the cost of a slightly worse average final population. A real-life engineering case study shows that inspecting the model reveals which objectives drive the decision-maker's preferences.
Thomas Quadt, Boris \v{C}ule, Roy Lindelauf et al.
In briefXL MIMO Extreme Learning Machines: Neural Inference Over the Air with Stacked Metasurfaces
SummaryThis paper proposes an extremely large MIMO system with stacked intelligent metasurfaces (SIMs) at the receiver, which together with a single antenna can act as a two-layer neural network operating over the air. The first layer is formed by wireless fading, and the second by controllable SIM operations, with only the front metasurface layer needing a nonlinear response. Using the extreme learning machine framework, the authors show the architecture is a universal approximator and derive the optimal linear combining in closed form. Numerical results across several datasets report classification accuracy comparable to idealized digital models.
Kyriakos Stylianopoulos, Mattia Fabiani, Davide Dardari et al.
In briefMulti-Horizon Future Discovery for Physics-Informed Exploration (HTC² Framework)
SummaryThis paper addresses exploration in physical control environments by proposing Hamiltonian Temporal Contrastive Curiosity (HTC²), which learns contrastive representations of reachable futures across multiple horizons to produce an intrinsic reward. The reward is modulated by bounded multi-step Hamiltonian variation and integrated with SAC and PPO. The method is evaluated on six DeepMind Control Suite tasks against ICM, RIDE, E3B, Plan2Explore, and a single-step Hamiltonian baseline, with state-coverage, ablation, and phase-space analyses.
In briefEvoBench: An Evolving Diagnostic Benchmark for Environmental Modeling
SummaryEvoBench is a benchmark framework that shows where and when environmental models fail, tracks these failures over time, and evolves to focus on capability gaps. The authors apply it to daily streamflow prediction on about 9,000 U.S. basins from 1985 to 2024 (CAMELS-Evolve), comparing an operational physical model (NWM), a conceptual model (HBV), and several data-driven models. No model performs best across all testing scenarios, and all data-driven models under-predict floods.
In briefA context-based way to measure ambiguity in rules and laws
SummaryThis paper proposes measuring ambiguity in prescriptive texts such as laws, contracts, and policies by how much stakeholders disagree about how alternative versions apply to a set of scenarios. In preliminary analysis across two policy domains, the approach correlates with real human disagreement, even when using simulated raters, and is robust to spurious text features. The authors also apply the measure to existing deliberation datasets to examine how policy ambiguity relates to support for those policies.
In briefPrior Predictive Matching for Bayesian LoRA: A Closed-Form Way to Set the Prior Before Fine-Tuning
SummaryThis paper addresses how to choose the prior over low-rank weight updates in Bayesian LoRA, a method for uncertainty estimation when fine-tuning large models. The authors introduce a closed-form prior-predictive estimator that works before fine-tuning, using a second-order network expansion and moment matching. On RoBERTa-large, the approach improves predictive NLL and calibration over MAP and marginal-likelihood selection without validation data, while remaining competitive on out-of-distribution detection.
Tomasz Ku\'smierczyk, Eliezer de Souza da Silva, Marcin Sendera
In briefBeyond Federated Learning: Decentralized AI Agents for Self-Evolving Clinical Intelligence
SummaryThis Perspective argues that medical AI should shift from aggregating data to aggregating clinical experience. It proposes Decentralized Agentic Systems (DAS), in which AI agents across healthcare institutions share abstracted reasoning trajectories instead of raw data or model parameters. The paper discusses the framework's architecture, deployment, governance requirements, and open challenges.
In briefOBB-PBD: Extending parallel box decoding to rotated bounding boxes for locating described objects in aerial images
SummaryAerial objects appear at arbitrary orientations, so this paper extends Parallel Box Decoding from horizontal boxes to six-parameter oriented boxes using a double-angle parameterization. Image and prompt shuffle controls and probes show the model depends on both the image and the prompt. On the AVVG benchmark's full 6139-record test split, rotated IoU rises from 0.1253 for the zero-shot baseline to 0.303 ± 0.012 over three seeds.
In briefGenerating Safety-Critical Driving Scenes Tailored to Each Planner to Improve End-to-End Driving Robustness
SummaryEnd-to-end driving models perform well on standard tests but often collide in safety-critical situations. On NeuroNCAP, the authors find that 95.9% of DiffusionDrive's collisions stem from planning failures rather than perception, largely because such rare interactions are underrepresented in real driving data. The paper proposes a framework that generates attack scenes against a specific planner, attacking a differentiable surrogate distilled from the planner's own head at the bird's-eye-view level and pairing each attack with a safe escape trajectory for fine-tuning. Fine-tuning roughly halves the closed-loop collision rates of SparseDrive and DiffusionDrive on NeuroNCAP while largely preserving normal driving.
Khaled M. Seyam (1), Julian Wiederer (2), Samar Fathalah (1) et al.
In briefCutting Simulation Costs for Bayesian Inference in Agent-Based Models
SummaryThis paper proposes Double-Flow Neural Posterior Estimation (DF-NPE), a method for Bayesian parameter inference in agent-based models whose likelihoods are intractable. It combines learned time-series summaries with two conditional normalizing flows to generate synthetic training data and estimate posteriors. In two settings of the Brock-Hommes asset-pricing model, the authors report that DF-NPE reduces direct simulations by 90 percent while achieving lower posterior discrepancies and runtime than standard NPE.
In briefOmni-InteractScope: A Benchmark for Proactive and Agentic Interaction in Live Audiovisual Streams
SummaryThis paper introduces Omni-InteractScope, a benchmark of 1,479 human-annotated interactions with 1,790 temporal targets across eight task groups for omnimodal AI models watching continuous audiovisual streams. Tests of 12 models found that timely first responses to tool-invocation targets reached at most 40%, and reminder recall never exceeded 23.0%. No model fully covered any of the 26 guidance sequences.
In briefSkill∞: Recursive Self-Improvement for Optimizing Agent Skills
SummaryThis paper presents Skill∞, a method where an AI agent optimizes task-specific skills and also improves the meta-skill it uses for that optimization, using the same procedure. The meta agent diagnoses execution traces, proposes candidates, requests reviews, and allocates search budget, and the meta-skill can rewrite these steps. On 83 SkillsBench tasks, Skill∞ achieves a 60.7% pass rate, compared with 56.4% for human-curated skills and 54.5% for SkillEvolver.
In briefProbing phonetic representations in the Mimi codec's first codebook
SummaryNeural audio codecs are increasingly used to encode speech for large language models. This paper examines phonetic information in the first codebook (CB0) of the Mimi codec. Because Mimi's fixed 80 ms tokens often span several phonemes, the authors use forced alignment, entropy analysis, and Phone-Normalized Mutual Information (PNMI) instead of ABX-style evaluation. On TIMIT speech, individual tokens are strongly tied to specific phonemes (PNMI = 0.84), but each phoneme is realized by multiple competing tokens rather than a one-to-one mapping.
In briefFaithful Explanations When Data Modalities Are Missing in Dementia Prediction
SummaryThe paper examines whether AI-generated clinical narratives explaining a multimodal dementia prognosis model remain faithful when costly tests such as amyloid PET and CSF biomarkers are missing. Using 307 Alzheimer's Disease Neuroimaging Initiative participants, the authors masked every subset of four modalities and predicted 36-month conversion from mild cognitive impairment to Alzheimer's disease, restoring masked data to obtain ground truth. Models attributed substantial importance to unobserved modalities, and naming absent modalities in prompts sharply improved acknowledgement and reduced phantom evidence, while advice on whether acquiring a missing test would change the prediction was often incorrect until checked against executed restorations.
In briefA knowledge-driven nonconvex square-root regularization method for group sparsity and smoothness
SummaryThe paper proposes NS³R, a structured sparse learning method that pairs a square-root loss, which does not require knowing the noise level, with a nonconvex ℓ2,q penalty to encode both group sparsity and structured smoothness. The authors derive non-asymptotic error bounds for the estimator under a (2,q)-group restricted compatibility condition. They also develop a double-screened IRLS algorithm with thresholding (DS-IRLS-th) that is proven to converge globally. Experiments on synthetic and real data report that NS³R outperforms state-of-the-art methods in finite-sample settings.
In briefAdversarial user simulator that keeps user goals fixed, for stress-testing multi-turn AI agents
SummaryThis paper introduces StressSim, a framework that uses adversarial simulated users to stress-test multi-turn AI agents. Standard simulated users tend to be overly cooperative, which can overstate agent performance. StressSim evolves the simulator's planning prompts to reduce task success while keeping the user's underlying goal unchanged. The authors report that on tau-bench, pass^4 scores dropped by up to 69% for one tested model, using a budget of 20 rollouts per task.
Atri Vivek Sharma, Brian Formento, Alessio Lomuscio
In briefPredicting Output Uncertainty in Diffusion Language Models from Denoising Trajectories
SummaryThis paper studies how the convergence of masked diffusion language models during denoising relates to the variability of their final outputs across repeated runs. Across four models and four reasoning benchmarks, the authors collected 20 generations per question and 32,000 denoising trajectories. Stronger convergence within a single trajectory was associated with lower endpoint entropy. Features from one denoising run also predicted whether independent runs would concentrate on a common endpoint, which could reduce the number of complete generations needed.
In briefINVAR-Agri: Using a Fine-Tuned LLM to Find Mistakes in Translated Farm Machinery Manuals
SummaryThis paper is a proof-of-concept study comparing two fine-tuned workflows on the same LLM for checking translated agricultural machinery operator manuals. One back-translates each version to English for comparison, while the other has a checker compare the English original and translation directly. The back-translator often concealed planted errors, especially weakened obligations, while the checker detected most of them but was weak on units and unseen number changes. Alerts on real Finnish translations were mostly false, so precision on real text remains a limitation.
In briefImgMT: A Wikipedia-based Benchmark for Translating Text Inside Images
SummaryThis paper introduces ImgMT, a multilingual dataset and benchmark built from localized SVG diagrams, maps, and infographics on Wikimedia Commons. It contains 35,612 aligned image pairs across 129 languages and over 20 writing systems. The authors propose a multi-level evaluation that measures text detection, OCR accuracy, and translation quality, and test several multimodal large language models and a dedicated system. They find that source language and writing system strongly affect results, and that ImgMT is not well suited for isolating the contribution of visual information.
\v{S}\'arka Uramov\'a, Petar Kirilov Petrov, Josef Jon et al.
In briefUnveil the Mask: Joint-Optimized Diffusion Masked Pretraining for Verifying Avatar Operators Regardless of Appearance
SummaryThe paper addresses verifying who operates a photorealistic avatar, since the rendered face's appearance is decoupled from the operator. It proposes a framework without external pretrained models that learns motion representations from facial landmark sequences via diffusion masked pretraining, and captures fine facial motion from frame differences in video. The two embeddings are fused, reaching an overall AUC of 0.769 on the WBA2026 benchmark. Fusion particularly helps across speaking and listening modes.
Katsuya Ogata, Ryoya Yoshiyama, Kawachi Hodaka et al.
In briefPaper on how large language models represent emotions geometrically across model size, depth, and instruction tuning
SummaryThe paper measures emotion representations in six Gemma 2 models (2B, 9B, 27B, base and instruction-tuned variants). All models show a core emotion geometry, with held-out emotions mapped nearest their own vectors. The leading component correlates with human valence ratings at 0.71 to 0.78, and its variance share falls as scale increases. Depth at which emotion words enter the output vocabulary differs notably between the smaller and 27B models.
In briefPolicyMediator: Multi-Regional Policy Learning via Mediation Agents in One-Person Company Ecosystems
SummaryOne-person companies run by individual entrepreneurs with AI agents often move between regions, so local support policies become a cross-regional coordination problem. The paper proposes PolicyMediator, a Stackelberg framework in which a mediator explains verified policies to these companies and reports their needs and concerns back to governments for policy learning. In six-region simulations in China, it ranks first in all 15 combinations of OPC backbone and scenario. Against the mean of eight baselines, it improves the aggregate government objective by 42.4% and average OPC utility by 40.9%.
In briefNeSySpec: Neuro-Symbolic Auditing of Engineering Specifications, Applied to Aircraft
SummaryNeSySpec is a neuro-symbolic framework built on Logic Tensor Networks that audits engineering specifications against physical constraints. It identifies violated constraints, locates altered fields, proposes minimum-change corrections, and links image evidence to the parts described. Validated on 107 aircraft types, including 16 held-out test aircraft from 9 families, the authors report it matches or exceeds the compared methods on constraint verdicts, diagnosis, and repair error.
WhyUnifies neuro-symbolic diagnosis, abduction, repair and image grounding for specification audits, validated on 107 aircraft types with clear baseline gains.
Pier Mattia Basciano, Gianvito Antenori, Cristiano Andolfi et al.
In briefAdaTrendLike: Adaptive Trend-Likelihood Scoring for Time Series Anomaly Detection
SummaryIndustrial IoT devices generate large, non-stationary time series that are hard to monitor for anomalies. This paper proposes AdaTrendLike, which adaptively splits a series into trend and detrended components, then scores anomalies by the likelihood of observed trends under an estimated trend distribution. The authors report consistent outperformance over state-of-the-art methods on five public benchmark datasets. The framework also offers feature-level interpretability by pinpointing deviations in the estimated trend distribution.
In briefBelief-based multi-agent training method for decentralized LLM collaboration
SummaryThis paper proposes BCMAAC, a belief-conditioned multi-agent actor-critic method in which agents maintain beliefs over contextual information and use them to aid training. Experiments on Minecraft game-playing, web searching, and tool-using tasks show it improves task performance and collaborative behavior over naive actor-critic baselines. The gains are most pronounced in multi-turn tasks with richer contextual information.
In briefTransfer learning for deep learning-based intracranial 4D Flow MRI super-resolution
SummaryThis mixed prospective-retrospective study analyzed paired 1.2-mm and 0.6-mm 4D Flow MRI scans from 18 participants, including healthy individuals and patients with moyamoya disease or multiple sclerosis. A convolutional neural network was pretrained on synthetic cerebrovascular data and fine-tuned on cohort-specific or pooled in vivo data. Pooled multi-cohort transfer learning gave the lowest mean vNRMSE (0.119), and super-resolved angiography showed higher precision-recall AUC than standard-resolution imaging. The authors conclude that further multi-center validation is warranted.
Oliver Welin Odeback, Susanne Schnell, Patrick Winter et al.
In briefUsing LLM agents to autonomously discover bioprocess dynamics and kinetic parameters
SummaryThis paper proposes a closed-loop framework in which a large language model agent automatically discovers kinetic state-space models of bioprocesses. It separates multiplicative (inhibitory) and additive (catalytic) growth-rate effects through a sequential peeling procedure across operating regimes. When competing model hypotheses cannot be distinguished, an active experiment-design controller applies targeted perturbations. On a simulated continuous stirred-tank reactor, the framework recovered ground-truth parameters with under 6% error without prior kinetic knowledge.
In briefTransECG-SSL: Self-supervised learning of transferable heartbeat features for ECG arrhythmia detection
SummaryThis paper presents TransECG-SSL, a self-supervised method that learns heartbeat representations from unlabelled ECG recordings using a Transformer encoder. The encoder is pretrained with a combined masked-reconstruction and contrastive objective, then frozen and tested with linear probing on patient-disjoint splits across three public databases. Hybrid pretraining outperforms random initialisation and matches or beats architecture-matched supervised pretraining on balanced accuracy. Cross-dataset transfer still trails target-specific pretraining, and minority arrhythmia classes remain hard to classify under severe class imbalance.
Ahtisham Ayyub, Fasih Haider, Muhammad Rehan Usman et al.
In briefACSAN: A Convolutional Self-Attention Network Combining Convolution with Vision Transformers
SummaryThis paper proposes ACSAN, a deep learning model that replaces the fully-connected matrix multiplications in Vision Transformers (ViT) with convolution operations. Its basic module, Dynamic Convolutional Self-Attention (DCSA), uses multi-scale dilated convolutions to produce Query, Key, and Value, then merges spatial and channel attention outputs. Trained from scratch on CIFAR-10, CIFAR-100, and ImageNet-1k, the authors report higher accuracy than standard ViT with comparable or fewer parameters, and ablation studies confirm each component's contribution.
In briefGIFT: Preserving Representation Geometry for Single-Domain Generalized Object Detection
SummaryThis paper addresses single-domain generalized object detection, where a model trained on one source domain must generalize to unseen domains with distribution shifts. The authors observe that unconstrained fine-tuning disrupts the geometric structure of pre-trained Vision Foundation Model representations, which encode domain-invariant semantics. They propose GIFT, a geometry-aware fine-tuning method that applies learnable Householder reflections to pre-trained weights plus lightweight channel-wise scaling. GIFT outperforms prior state-of-the-art methods on two benchmarks, gaining +7.9% mPC on Cityscapes-C and +2.9% on the Diverse Weather Dataset while adding fewer than 1% trainable parameters.
In briefDiffusion-based method that recovers missing or corrupted tensor data using shared factors
SummaryThe paper proposes DiffTR, an unsupervised framework that unifies tensor completion, denoising, and hyperspectral super-resolution using coordinate-conditioned diffusion. Shared factors indexed by tensor coordinates condition a compact score network, which is trained with score matching to learn the factors and an implicit tensor representation from observed data. Experiments across multiple tensor modalities and recovery tasks reportedly show competitive reconstruction quality and flexibility.
In briefSG-Splat: Stable geometry learning under pose uncertainty for pose-free 3D Gaussian splatting
SummaryThis paper presents SG-Splat, a feed-forward 3D Gaussian Splatting framework that reconstructs scenes directly from unposed images, without requiring camera poses. It separates geometry estimation from Gaussian attribute prediction and uses coarse-to-fine refinement, including flow-guided pose refinement and residual-guided geometry correction, to reduce the effect of pose errors. A dual-path training strategy mixes ground-truth and predicted poses to stabilize learning. According to the authors, experiments on DL3DV and RealEstate10K show state-of-the-art results for novel view synthesis and pose estimation.
In briefAgreement of Reconstructed and Original GDELT GKG Records: A Field-by-Field Baseline
SummaryThis paper presents gdelt-regkg, an open-source reconstruction of the GDELT Global Knowledge Graph (GKG) record format, and compares its output with GDELT's archived native records field by field. Using about 193,550 English articles and 2,500 articles in five other languages, it finds that agreement varies widely by field, with English tone showing a mean absolute error of 1.79 points while dates, complete counts, and quotations agree much less. The authors conclude that reconstructed records should be assessed field by field and not treated as fully interchangeable with native ones.
WhyOpen reconstruction of closed GDELT GKG generation, field-level comparison on 193,550 English articles; several fields agree poorly.
In briefBenchmarking uncertainty quantification for engineering regression under graded distribution shift
SummaryThe paper benchmarks 12 regression uncertainty quantification methods from five families on seven public engineering datasets. Evaluation uses a three-level shift ladder: in distribution, soft shift and hard shift. Under hard shift, mean coverage of the 11 methods that always return intervals falls to 0.40-0.75. The authors recommend splitting calibration data by deployment unit and treating abstention as an explicit operational outcome.
Ahmad Lawal, Rahim Pasbanigoloojeh, Asmau Lawal Bilbis
In briefOneEEG: A unified multimodal language model that understands brain signals across many EEG tasks
SummaryThis paper presents OneEEG, a unified multimodal large language model for understanding electroencephalography (EEG) across multiple tasks. It is trained on OneEEG-Instruct, a multi-task EEG instruction-tuning dataset, and combines numerical and visual EEG representations through a task-aware Mixture of Projectors. The authors also introduce OneEEG-Bench, a benchmark, and report that existing multimodal models interpret EEG inputs poorly while OneEEG performs better.
In briefMarkovSim: A Bounded-Context LLM Agent Framework for Long-Horizon Simulation
SummaryThis paper presents MarkovSim, an LLM agent framework for long-horizon computational simulation. Its core is a context management scheme that stores workflow records in a memory system with a predefined schema and builds a compact summary of solution progress at state transitions of a finite-state machine. Applied to finite element analysis and tested on six benchmark problems, it passed 16 of 18 runs and corrected its own errors in solver operations and model physics. The context stayed bounded throughout every run regardless of length.
In briefCVTSplat: Multi-View 3D Reconstruction with Adaptive Centroidal Voronoi Tessellation
SummaryThis paper presents CVTSplat, a method for reconstructing shape and appearance from multi-view images using adaptive centroidal Voronoi tessellations. Cell sizes follow the evolving signed distance field, concentrating sites near the zero level set while keeping cells nearly centroidal and regular. The authors build GPU algorithms for nearest-neighbor queries and approximate Voronoi construction, plus a rasterizer suited to small, nearly isotropic cells, and report Voronoi construction 5 to 10 times faster than general-purpose alternatives. On DTU and BlendedMVS, it reaches geometric accuracy competitive with the best Gaussian-splatting surface methods, training in about 2.5 minutes on one GPU.
Briac Toussaint, Diego Thomas, Jean-S\'ebastien Franco et al.
In briefShould Pseudo-Labels Share Mini-Batches with Labeled Examples? A Semi-Supervised Continual Active Learning Method
SummaryThis paper studies how to use pseudo-labels in continual active learning, where each round uses a fixed number of training steps. It formalizes two decisions: which unlabeled examples receive pseudo-labels and what share of each mini-batch they occupy. The proposed method, SSCAL, works with any acquisition function, selects pseudo-labels by distance to labeled neighbors, caps their share, and reweights the labeled loss. Across two audio datasets, two frozen encoders, six acquisition functions and ten seeds, SSCAL ranks first on average among four protocols with equal training steps.
In briefTargeted deep regression for estimating organ measures from whole-body water-fat MRI
SummaryThis paper applies image-based regression to neck-to-knee water-fat MRI scans from the UK Biobank to estimate kidney volume, heart volume, and age. Manually defined anatomical sub-regions from a reference scan are propagated to other scans via image registration and assembled into 2D multi-channel collages fed to a ResNet50-style network. Predictive performance reached R² = 0.731–0.984, and group-wise saliency maps highlighted medically relevant regions.
Filip Marttala, Taro Langner, Johan \"Ofverstedt et al.
In briefOne Megakernel Isn't Enough: Predicting Where to Split Streaming AI Inference
SummaryThis paper examines how far GPU kernel fusion should extend and which boundaries are worth keeping. The authors present a framework that uses profiles of conventional computation graphs and standalone kernel benchmarks to predict time to first output and select the best design, without running each candidate end to end. Across 13 tested combinations of models and batch sizes, the best configuration was never a single megakernel, ranging from two to four megakernels or fully conventional execution.
Aditya Chatterjee, Anand Joseph X M, Sanjay Singh Chauhan
In briefWhen to Hand Off to Humans: Disagreement Patterns for Uncertainty-Aware Content Moderation
SummaryThe paper argues that disagreement among LLM agents in value-laden content moderation may reflect genuine value pluralism rather than a reliability error to be removed. It proposes Reasoning Trace Divergence Analysis (RTDA), which classifies cases into a four-state taxonomy based on justification similarity and conclusion agreement. On 1,000 items from the Measuring Hate Speech corpus, this structural signal predicted human annotator disagreement better than magnitude-only and output-only baselines.
In briefRepoTrace: Improving Requirement Tracing in Multi-Round Code Agent Interactions
SummaryThis paper examines how code agents degrade over many interaction rounds during iterative software modification, analyzing 2,656 trajectories on EvoCode-Bench. It identifies requirement focus misalignment, where agents fail to separate a requirement's core modification targets from unrelated code. The authors propose RepoTrace, which combines repository structure with requirement semantics to build a requirement-code correspondence injected into the agent context. Experiments report higher mean per-round reward across several models and harnesses, along with reduced regression errors.
In briefLabel-free checkpoint selection for audio active learning using geometric consistency
SummaryThis paper proposes a label-free validation rule for choosing training checkpoints in audio active learning, so no extra labeled validation set is needed. Two classifiers, nearest-class-mean and k-nearest-neighbor, are built from labeled data in a frozen pretrained embedding space and used to pseudo-label unlabeled data. Each checkpoint is scored by its agreement with these pseudo-labels and its cross-entropy on the labeled set. The authors report that, at matched label budgets, the method achieves the best mean rank in area under the labeling curve among methods using no extra labels.
In briefConstrained Decoding: Harmful or Beneficial? Revisiting Its Impact on LLM Performance
SummaryThis paper sorts constrained decoding into three types: format, validity, and reasoning constraints. Causal ablations across eleven open-source models from four families find that apparent performance drops mostly come from non-equivalent prompts and output formats, not decoder masking. Validity constraints improve reliability by removing invalid answers, and reasoning constraints can raise accuracy when they enforce beneficial reasoning paths.
WhyProposes a three-type constraint taxonomy and revisits constrained decoding's impact via theory and causal ablations across many models, challenging a common assumption.
Ngoc Trinh Hung Nguyen, Alonso Silva, Armen Aghasaryan et al.
In briefCRIM: A reusable item memory that makes personalized search cheaper to run
SummaryPersonalized search and recommendation with large language models is costly, because longer interaction histories repeatedly re-encode the same item text. The paper proposes CRIM, which compresses each item into compact continuous memories using learnable summary tokens, and these memories are independent of users and queries so they can be reused across requests. Experiments on MovieLens-1M, Qilin, and a new dataset LazSeq reportedly achieve state-of-the-art retrieval performance at the same backbone scale with substantially lower inference cost.
WhyDecouples item encoding from request-level history modeling via reusable continuous memories, cutting repeated encoding costs. Experiments and cost analysis on three datasets give reasonably solid evi
In briefRSIV-Harness: A self-improving framework for open-ended chart generation from conversational requests
SummaryThis paper presents RSIV-Harness, which keeps a language model's weights fixed and repeatedly revises the surrounding harness based on the agent's own failure traces. Proposed revisions pass through anchor and memory checks before being kept. The authors also introduce ODAC-Vis, a 640-task benchmark testing data discovery, ambiguity resolution, multi-turn chart refinement, and executable chart generation. One round of RSIV-Harness improves all six metrics over the seed harness, most notably in executability and chart structure.
In briefCoPE-CGL: Community-Aware Prototypes for Replay-Free Continual Graph Learning
SummaryThe paper proposes CoPE-CGL, a continual graph learning framework that does not replay past graph data. It uses a parameter-free structural prior computed by multi-hop Laplacian smoothing, together with community-level prototypes and majority-vote routing. At equal storage and without task identifiers, it matches the strongest replay baseline on Arxiv and CoraFull while running faster and showing zero forgetting. The design targets homophilic graphs with community structure.
In briefOVM-Occ: Open-Vocabulary Monocular Occupancy Prediction via Geometry-Aware Vision-Language Distillation
SummaryThis paper studies open-vocabulary monocular occupancy prediction for indoor scenes, recovering 3D geometry and semantics from a single RGB image. It proposes OVM-Occ, a Gaussian-based framework using language-embedded Gaussians, geometry-aware deformable attention, instance-level vision-language embeddings, and decoupled semantic distillation. The authors also introduce OpenOcc-ScanNet and OpenOcc-Replica benchmarks, and report state-of-the-art results, outperforming the prior best by 12.2 mIoU and 0.6 IoU on Occ-ScanNet.
In briefCalibrating black-box simulators with a temporal-difference method that needs no derivatives
SummaryThe paper proposes a derivative-free temporal-difference framework for calibrating black-box simulators, using one-step transitions from a target system and noisy simulator continuation values. It introduces three algorithms based on random-direction finite differences and proves almost-sure convergence, mean-square error bounds, and central limit theorems for averaged estimates. The analysis also examines how correct or incorrect simulator specification affects bias and convergence rates.
WhyProposes a derivative-free TD calibration framework via Bellman orthogonality, revealing a specification-dependent bias vanishing with improved rates; theoretical results are substantial, with limited
In briefShared Context, Separate Engines: Concurrent Streaming Vision-Language Inference on Edge Chips
SummaryEdge SoCs combine CPUs, GPUs, and NPUs, but current vision-language model runtimes usually run generation on one engine and leave the others idle. The paper presents \name, which separates each stream's versioned context from each engine's KV state so several heterogeneous accelerators can run stateful inference concurrently. The authors report higher throughput than the strongest single-engine baselines on Apple M4 Max, Rockchip RK3588, and Qualcomm QCS6490.
In briefTraining Tsetlin Machines to Reduce Inference Latency
SummaryThis paper studies how the Tsetlin Machine's two training hyperparameters, T and s, affect both classification accuracy and inference latency. The authors propose a directed search that treats accuracy as a constraint and increases s to lower latency. On the gem5 simulator with an ARM processor, latency drops by up to 16.34% across six datasets, with some datasets losing only 0.06% accuracy. The method needs far fewer evaluations than an exhaustive grid search.
In briefInterQ-RL: Correcting execution mismatch in quantized agent reinforcement learning
SummaryThis paper addresses execution mismatch that arises when quantized low-precision rollouts are used to train multi-turn agents with reinforcement learning. It proposes InterQ-RL, which applies bounded correction over the policy history that reaches each agent state and allocates correction strength across action segments under a fixed displacement budget, with no extra model forward pass. The authors evaluate dense and mixture-of-experts backbones from 4B to 30B on ALFWorld, WebShop, and ScienceWorld. They report that the method matches BF16 within 0.7 points and exceeds it in two settings, while avoiding the late performance regression of alignment alone.
In briefLayerwise Repacking: a training method for variable-length hybrid-attention models without extra data redistribution
SummaryThis paper addresses conflicting parallelism needs across layers when training hybrid-attention models on variable-length samples. Layerwise Repacking keeps wide sequence-parallel groups for full-attention layers and repacks samples into node-local smaller groups for sparse-attention layers, piggybacking the layout change on existing full-layer communication. On a 1F3S hybrid-attention model, repacking alone cut mean end-to-end time by 5.95% and raised MFU by 3.75 points, while the improved parent solver reduced time by 7.61% and raised MFU by 4.83 points.
In briefPROBE: Frontier Coding Agents Find Different Bugs Than Maintainers Fix
SummaryThis paper introduces PROBE, a benchmark of 2,457 bugs fixed over 18 months across 20 GitHub repositories, along with a method for matching agent findings to this reference set. Evaluated frontier coding agents reach recall below 21% and precision below 9%, and stronger models raise recall without improving precision. The authors also find that bugs found by different agents resemble each other but differ substantially from bugs fixed by maintainers.
Marco Paina, Mark Niklas M\"uller, Niels M\"undler-Sasahara et al.
In briefLLM-enhanced heterogeneous graph learning for detecting financial fraud using business meaning
SummaryThe paper proposes SemHGT, a framework that uses large language models to interpret structured transaction attributes according to entity roles, such as cards and merchants, making their business meaning explicit. It generates role-aware risk descriptions, selects path evidence guided by semantic similarity, and processes target transactions with schema-level tokens in a unified transformer. The authors report that SemHGT leads the evaluated baselines on most dataset and metric comparisons across three benchmark datasets.
In briefDid the Model Use That? Counterfactual Training for Self-Attribution in Language Models
SummaryThis paper studies whether language models indicate when, and only when, their answers are influenced by information supplied with a question. It finds that most of eleven frontier models rarely attribute influential information, while the rest attribute non-influential information. The authors propose CREDIT, which supervises final-answer self-attribution using counterfactual labels without human annotation, and report substantial gains in four open models that partly transfer to unseen information types.
In briefAI Use in Africa for Climate-Related Health Problems: A Bibliometric Study
SummaryThis paper uses bibliometric analysis of Scopus-indexed literature to compare global and African research on AI for climate-related health issues. It finds global work focuses on complex predictive models for planetary health, while African work centers on localized, reactive solutions for infectious disease triage. The authors identify compute deficits, data silos, and energy instability as barriers to scaling advanced AI in Africa, and recommend shifting international funding toward foundational digital infrastructure and localized AI frameworks.
Peter Kokol, Martha Kachweka, Helena Blazun Vosner et al.
In briefAgenticCity-Bench: A Multimodal Benchmark for Longitudinal Urban Observation and Agentic Reasoning
SummaryThe paper introduces AgenticCity-Bench, a longitudinal multimodal benchmark covering 16 cities in 12 countries across six continents from 2000 to 2025. It includes 18 modalities spanning urban function, activity, and environmental state, organized around a common spatial reference and calendar index. The benchmark supports urban cross-modal translation and agentic city tasks such as forecasting, planning, and resource analysis.
In briefReward-Guided RRT* for budget-limited robot task and motion planning
SummaryThe paper proposes Reward-Guided RRT*, a sampling-based method for robots that must choose rewarding locations and plan motions within a limited travel budget. It builds missions step by step directly in continuous state space, without a precomputed task-transition graph, and uses a return-feasibility check to keep each mission within budget. Experiments with more obstacles and tasks report solutions comparable to VNS-PRM* while requiring substantially fewer sampled states. The authors also extend the method to Dubins steering and test it in CoppeliaSim on a scaled car-like robot.
Vin\'icius Ara\'ujo de Oliveira, Gabriel Ferreira Mendes Louren\c{c}o, Paulo L. J. Drews-Jr et al.
In briefProof-of-Agentness: Verifying whether a web client is an AI agent
SummaryThis paper introduces Proof-of-Agentness (PoA), a mechanism for telling LLM-powered agents apart from simpler bots and scripts on the Web. It explores active verification through challenge-response and passive analysis of signals visible to the service. Eight language models achieved 96.5% accuracy on the challenges, with an average overhead of 1,647 tokens. In an IRB-approved user study, four classifiers reached an average accuracy of 95.8% in passive analysis.
In briefSatellite Intelligence: A survey of distributed training, collaborative inference and LLM adaptation in LEO satellite constellations
SummaryThis survey reviews how artificial intelligence can run on low Earth orbit (LEO) satellite networks, where dynamic topologies, intermittent links, and strict size, weight, and power limits complicate training and inference. It covers three paradigms: satellite-ground co-training using federated and split learning, collaborative inference across single-satellite, multi-satellite, and satellite-ground architectures, and on-orbit adaptation and serving of large language models. The paper concludes by identifying open challenges and future research directions.
In briefLet the Gate Decide: Per-Lookup Precision for Quantized Conditional Memory in LLMs
SummaryThe paper studies quantization of conditional memory layers, such as DeepSeek's Engram, which attach large hashed n-gram tables to language models. It introduces GateRQ, which stores each row as a ladder of 1-bit residual codes and reads deeper levels only for lookups that the gate ranks highest. Without retraining, GateRQ reduces GPU table memory by 7.8x on DeepSeek-V4.1-Flash and raises vLLM throughput by 11%-31% over FP8 tables, with small accuracy losses on two models.
