In briefBAM: Bivariate Additive Models for Readable Tabular Prediction
SummaryThis paper proposes Bivariate Additive Models (BAMs), a type of GA²M in which each pairwise shape function is limited to a simple parametric form, an asymmetric bivariate Gaussian. Each component can be shown as a single-peaked contour map, and its feature pair is chosen automatically during training. Experiments on six tabular benchmarks show that BAM's predictive accuracy is comparable to standard GAMs and close to GA²Ms. The paper also examines the trade-off between accuracy and the number of components a user must inspect.
In briefCombining Deep and Classical Methods: A CPU-Efficient Decision-Fusion Model for Modulation Recognition
SummaryThis paper addresses automatic modulation classification, identifying the modulation type of radio signals without cooperation. The authors find that a compact CNN on raw I/Q samples is more robust at low SNR, while a higher-order cumulant classifier is more accurate at high SNR. They propose FuseAMC, which fuses the two models' class posteriors with a logistic-regression combiner. On RadioML 2016.10a it reaches 55.8% overall accuracy and runs on a single CPU core.
In briefDiversity-aware replay sampling using statistical leverage scores
SummaryThis paper addresses reinforcement learning replay buffers that are dominated by routine transitions, so rare situations are seldom sampled. It scores each stored transition by its statistical leverage relative to the buffer's own data and mixes this score with uniform sampling. The authors find that importance-weight correction cancels the benefit, so they omit it, and they extend the approach to relational rare situations via goal-relative trajectory windows. On a three-joint arm pushing an object around an obstacle, rare-task success rises from 0.49 to 0.69.
In briefModality-aware pruning of expert parameters in MoE multimodal language models
SummaryThis paper proposes MAEP, a training-free method for pruning experts in mixture-of-experts multimodal large language models. It scores expert importance separately for vision and text tokens in a single forward pass and ranks experts globally across layers. Experiments on three MoE-based MLLMs show it beats modality-agnostic baselines on most multimodal benchmarks and stays competitive on text-only tasks, cutting BF16 expert weights by about 27 GB.
In briefSplit-KV: Efficient Context Parallelism for Training Diffusion Drafters
SummaryDiffusion drafters speed up speculative decoding by denoising several draft tokens in parallel, and their training uses short draft queries cross-attending to a much longer context. This paper presents Split-KV, an exact context parallelism method that keeps key/value shards local so per-layer communication scales with query size rather than context length. It adds a cost-aware longest-processing-time-first strategy to balance context chunks by their realized attention cost. At a context parallel size of 64, the authors report 85–91% weak-scaling efficiency and up to 3.67× speedup over the fastest existing baseline.
In brief3D Banff Glomerulitis Scoring for Kidney Transplant Biopsies: Volumetric Quantification Feasibility Study
SummaryThis study examines whether Banff glomerulitis scoring for kidney transplant biopsies can be extended from two to three dimensions. The authors propose a framework that aligns serial sections, tracks glomeruli across depth, and integrates inflammatory cell counts under three scoring strategies. Applied to multi-section renal allograft biopsies, the 3D g-scores were more consistent across sections. Under the semi-automatic setting, the Volumetric strategy correlated more strongly with clinical biomarkers than traditional 2D scores.
In briefRoutes of Reasoning: Controlling Overthinking in MoE Models at Test Time
SummaryThis paper examines whether the routers in Mixture-of-Experts models reveal the behaviors a model uses during reasoning, such as planning, verifying, and concluding. The authors label reasoning traces by behavior and show that router signals predict these labels nearly as well as residual-stream probes while using fewer features. They then edit expert router scores during inference to reduce overthinking. Across four models and four benchmarks, the median reduction in generated tokens is 30.4%, with accuracy dropping by no more than 0.5 percentage points.
WhyThe authors use MoE router scores as a training-free handle on reasoning behaviors, reporting a 30.4% median token reduction across four models and four benchmarks.
In briefAutomated data pipelines for merging two Salesforce organizations after an acquisition
SummaryThe paper proposes a reference framework for an autonomous ETL pipeline that consolidates customer data from two Salesforce organizations while both remain in use. It covers bulk extraction, declarative mapping, quality gates, master data matching, constrained loading, and change data capture with a circuit breaker, each mapped to Jitterbit Integration Studio. Evaluated on synthetic scenarios with known ground truth, the hybrid matching reached pairwise F1 scores of 0.918 for accounts and 0.885 for contacts. Validation on a production deployment is left for future work.
In briefDecoupAttn: Training-Free Speed-Up for Video Diffusion via Decoupled Sparse Attention
SummaryThis paper presents DecoupAttn, a decoupled sparse attention method for accelerating few-step video diffusion transformers without training. Semantic clusters compensate every key, while exact attention corrects errors on a few blocks fixed offline by position. The authors also introduce a rate-distortion Distortion Budget for allocating exact computation and Reassign Cache, which updates drifting keys to cached centroids at each step.
In briefBetter approximation methods for capacity-limited cluster editing
SummaryThis paper studies the capacitated minimum-disagreement cluster editing problem, which partitions vertices of a complete edge-signed graph into disjoint cliques of size at most k while minimizing disagreements. The authors give a downsizing procedure built on any approximation algorithm for the uncapacitated version, and a hybrid of cluster-LP sampling and pivoting. These yield randomized approximation algorithms with ratios (1.9+ε) and (1.7537+ε).
In briefAutonomous agents that plan electricity systems under carbon limits
SummaryThis paper presents an autonomous planning framework in which agents propose investment bounds, storage operating conditions and feeder interventions, then interpret numerical feedback to redesign them and decide when to stop. The framework automatically runs annual investment-and-dispatch optimization and representative-feeder checks, and keeps the lowest-cost candidate that passes fixed acceptance criteria. Forty carbon-constrained trajectories were evaluated across ten system-year scenarios in five electricity systems in four countries.
In briefDo Saliency Maps Fail on Dense Breasts? A Chance-Corrected, Confounder-Aware Fairness Audit of Mammography Explanations
SummaryThis paper tests whether saliency maps from mammography classifiers (Grad-CAM, Integrated Gradients, and occlusion) locate cancers less accurately in dense breasts. Using 1,345 malignant CBIS-DDSM views, the authors find large unadjusted gaps, but case mix, mainly more calcifications in dense breasts, explains about half or more of them. Occlusion maps still show a significant average gap. The authors recommend reporting density gaps before and after case-mix adjustment, per classifier, alongside conspicuity and lesion-reliance controls.
In briefDoes a trial simulator understand what the drug does? An intervention-blind test on 7,207 trials
SummaryThe paper proposes a measurement of whether patient world models can truly simulate the effect of an intervention in clinical trials. The authors build 9,996 audited within-trial contrasts from 30,119 phase 2/3 trials with posted results on ClinicalTrials.gov, and compare each simulator against an intervention-blind reference model to score how much its accuracy depends on knowing the tested intervention. A gradient-boosting simulator trained on registry data has negative interventional skill, while a frontier language model reading only the registered protocol scores 0.20, and smaller models sit near the limit of detection.
WhyThe authors measure counterfactual skill of patient world models via an intervention-blind reference across 9,996 contrasts from 7,207 trials.
In briefCHARM+: A Multistream Attention Method for Faster Cross-Hardware Training
SummaryAs sequences grow longer, sequence-parallel distributed attention such as Ring-Attention has become a mainstream approach for large language models, but its communication dependencies limit overlap between computation and communication. This paper proposes CHARM+, which reorganizes sequential communication into levels for intra-level parallelism and overlap, and separates cross-hardware from intra-hardware communication to avoid synchronization overhead from bandwidth imbalance. Compared with Ring-Attention, its forward and backward algorithms achieve average speedups of 1.3x and 1.7x on a dual-NUMA PCIe architecture, and 2.8x and 2.9x on a dual-node NVLink architecture.
In briefEchoKV: Self-correcting KV cache eviction using iterative query feedback (paper, cs.AI)
SummaryThis paper studies KV cache eviction in large language models, where the workspace used to generate eviction queries can introduce early errors that distort later queries and reinforce wrong eviction decisions. The authors propose EchoKV, which initializes a bootstrap workspace from proposals at future decode positions and repairs it using queries generated in early decoding steps. On LongBench, EchoKV improves over LAQ on tasks requiring dispersed evidence, such as multi-hop cross-document QA and long single-document comprehension.
In briefRetrieval Beats Scale: Predicting Log Levels Using Socio-Technical Context
SummaryThis paper proposes a retrieval method for LLM-based log level prediction that selects in-context examples by code functionality and ownership (code maintained by the same developers) instead of at random. Evaluated on ten popular Java projects with four LLMs, combining both signals yields median AUCs of 0.83 to 0.97. The approach also outperforms the supervised predictor DeepLV and random-example baselines.
Youssef Esseddiq Ouatiti, Mohammed Sayagh, Bram Adams et al.
In briefRisk-Controlled Forecasting of Emotional Departures in Dialogue
SummaryThe paper studies emotion forecasting in conversation and finds that predicting the next emotional state's value is hard, while repeating a speaker's previous state forms a persistence baseline that frozen-encoder corrections rarely beat. The authors recast the task as selective departure forecasting, predicting when emotion leaves persistence, and apply dialogue-level conformal risk control to bound the expected loss of committed predictions. The guarantee holds in all twelve tested configurations, but it is achieved through abstention, and coverage and alarm recall fall sharply on the hardest corpus.
In briefLeakage-safe, information-matched evaluation for predicting emotion shifts in conversations
SummaryThe paper proposes an auditing protocol for forecasting emotion shifts in dialogue that matches the information available to each model and prevents target-turn leakage. It applies the protocol to six causal model families across four text corpora. The authors find that a deployable current-emotion classifier and an oracle gold-label setting can produce different conclusions about learned models versus the transition baseline.
In briefRECAP: context-aware reuse of model predictions to speed up vision-language decoding
SummaryThis paper proposes RECAP, a training-free method that speeds up vision-language models by reusing the target model's own past predictions in speculative decoding. It stores predictions under one-, two-, and three-token keys, keeping contextual memories limited to the current request and token-level memory persistent across requests. Retrieved predictions are fused by entropy-confidence weighting, then arranged into a candidate tree verified in one target pass. The authors report mean speedups of 1.66x on Qwen2.5-VL-7B and 1.56x on LLaVA-1.6-7B across seven single-turn benchmarks, exceeding the Token Recycling baseline.
In briefChronos: Identity-Preserving Video Generation via Temporal Anchoring
SummaryThis paper proposes Chronos, a framework for generating videos that keep a subject's identity consistent from text and visual inputs. It handles both single-segment facial identity preservation and multi-segment sequential action generation. The method combines semantic prompt anchoring, visual frame anchoring, and metric-guided candidate routing. The authors report first place in both Track 1 and Track 2 of the ACM MM 2026 Identity-Preserving Video Generation Challenge.
In briefIs One Query per Template Enough? Moving Learned Cost Models to New Hardware
SummaryA learned database cost model trained on one machine must be adjusted when the database moves to new hardware, where each query's runtime changes by a different factor. The authors time 1,162 TPC-H and TPC-DS query plans in six PostgreSQL environments and test how many measured queries are needed for correction. A correction with separate lines for serial and parallel plans plus a per-template offset, using one measured query per template, approaches the error of a model retrained on the new machine. For templates never measured, no tested correction beats a simple speed factor by more than 2% on average.
In briefCellCanon: An Agentic Cellular World Model for Prediction, Explanation, and Updating
SummaryThis paper introduces CellCanon, an agentic cellular world model that unifies single-cell expression prediction, evidence-grounded explanation, and observation-driven updating. Its predictive core reuses condition-level biological evidence to generate single-cell expression. On 339 held-out SciPlex3 conditions across three cell lines, it outperforms the best baseline on differential-expression AUPRC and Macro-F1 and on direction Macro-F1. Explanation and update capabilities are also evaluated on a source-grounded benchmark and in simulated experimental feedback.
Zhaokang Liang, Shuyang Zhuang, Xiaoran Jiao et al.
In briefDoes Better Confidence Help Audit Knowledge Graphs?
SummaryThis paper studies why better error-prediction confidence in LLM-built knowledge graphs often yields little gain in human audits. It analyses repairability, selection headroom and label validity, modelling non-adaptive auditing as budgeted covering. On two cross-document graphs from Re-DocRED, an error-rate by fan-out queue leaves 3.98 to 4.74 percentage points more errors than a joint model using the same actions.
In briefMIGONet: a graph neural network surrogate trained on solver-free manufactured-solution data
SummaryThis paper presents MIGONet, a multiple-input graph neural operator for surrogate modeling of steady heat conduction, along with a discrete manufactured-solution method that generates training data without solving a PDE. Temperature fields are prescribed first, and source terms and boundary data are computed with the same discrete operators as the finite-difference reference, so each sample satisfies the discrete equations. On a rectangular domain, MIGONet achieves mean relative temperature errors of 10.31% and 10.64% for curved-wire and Gaussian-spot heating, outperforming MIONet and a variant without the complete-source encoder. Data generation for 10,000 samples drops from 486.65 s to 3.15 s. Separately trained models are also evaluated on L-shaped, disc-shaped, and ring-shaped domains.
In briefFeature Puzzles: Learning Multi-View 3D Segmentation From Single Images
SummaryThis paper introduces Feature Puzzles, a framework that learns multi-view-consistent segmentation using only single-image semantic segmentation supervision. It creates pseudo-labels through image cropping and geometric transformations, then combines them with features from pretrained 3D foundation models in a lightweight, plug-and-play network. The approach supports prompts from one or multiple views and can be extended to open-vocabulary 3D segmentation.
In briefAI-assisted classification when people ignore overlapping evidence and mistrust AI
SummaryThis paper examines when an AI assistant helps or harms human decisions, arguing that outcomes depend on how much the AI's evidence overlaps with the human's and how the human combines them, not just on AI accuracy. The authors build a behavioral model with two parameters, awareness of overlap and trust in the AI, and derive a phase diagram of complementarity, impairment and automation regions. On ChatBench, AI assistance raises accuracy by 21 points, yet workers use only about a quarter of the Bayes-rational weight on AI information. Calibrating trust recovers about 2.6 points for GPT-4o but almost nothing for Llama-3.1-8b.
WhyThe authors model correlation neglect and trust in human-AI decisions, derive a closed-form phase diagram, and report decomposing accuracy loss on ChatBench.
In briefConvergence Study of Gradient Temporal-Difference Learning with Regularized Corrections and Eligibility Traces
SummaryThis paper studies convergence of temporal-difference (TD) learning, a core reinforcement learning method, for TDRC(λ), a regularized-correction variant. The authors prove convergence under diminishing stepsizes and non-i.i.d. off-policy data, with state-dependent and history-dependent λ parameters. They also propose TDRC(λ)-Huber, which reduces sensitivity to large Bellman errors while keeping the same fixed point. When the mean TD matrix is positive stable, both constrained algorithms converge under sufficiently strong regularization without requiring positive definiteness.
WhyThe authors extend TDRC convergence theory to eligibility traces and non-i.i.d. off-policy data, and propose a Huber-based variant with stability analysis.
In briefAuditable Unlearning for LLMs: Proving Data Deletion from Language Models
SummaryMachine unlearning removes a user's data from a trained model after a deletion request. Existing methods are evaluated on forgetting quality, retained capability, and robustness to membership inference, but not on whether providers can give users evidence that deletion occurred. The paper adds auditability as a fourth axis and proposes a two-stage method: GapWiden uses a hinge loss to widen the audit-statistic gap between forget and retain sets, and GapVeil is an inference-time rejection sampler. Tested on eleven methods across TOFU, MUSE and RWKU, the approach reaches oracle parity on auditability and leakage.
In briefCARTBench: A rubric-based benchmark for evaluating shopping advice on complex, ambiguous requests
SummaryThis paper introduces CARTBench, a benchmark for evaluating LLM shopping assistants on complex and ambiguous multi-turn purchase requests with timelines. It contains timestamped dialogues across 17 retail domains paired with 1,603 product records, and uses case-specific rubrics with 1,328 expert-annotated criteria covering Core, Excellence, and Safety. Seven commercial API models and open-weight models were evaluated, with three of ten models exceeding 90% on Core but none exceeding 45% on Excellence. The authors identify an understanding-to-execution gap between required behavior and high-quality purchase advice.
In briefGenerative AI for urban planning: DualControlNet jointly synthesizes land-use maps and satellite imagery
SummaryThe paper introduces DualControlNet, a generative land-use model that jointly produces high-resolution land-use maps and satellite imagery. Trained on over 40,000 paired tiles from Orlando and Chicago, it is evaluated with land-use, spatial consistency, and image quality metrics. Compared with unimodal baselines, it shows gains across these metrics and responds to text prompts such as "high-density development" for scenario-based planning.
In briefRisk control bounds error rates, not error sizes: auditing selective fetal biometry
SummaryThe paper audits a tool-grounded pipeline that measures fetal head circumference and refuses when the image cannot support a measurement. Calibrating the contour-completeness threshold with exact Clopper-Pearson risk control gives a guarantee no different from a hand-set floor, yet the pipeline still outputs a 271.3 mm error. Raising the lower bound of the admissible range from 50 to 55 mm cuts the worst-case error to about 11 mm at a small coverage cost.
In briefA study of how a low-rank decoder limits the reliability of perturbation predictions
SummaryThis paper audits the state-transition decoder used in STATE across three encoders and four CRISPRi cell lines. It finds that the decoder's first attention layer has a low-rank output, and that projecting onto 4 to 32 fixed directions preserves mean per-perturbation Pearson delta correlation. The authors propose predicted-NN-distance, the cosine distance from a predicted response to its nearest training response, as a per-prediction confidence score, which has a mean absolute Spearman correlation of 0.81 with that correlation. The score's thresholds transfer across encoders and cell lines and remain predictive for drug perturbations.
In briefTime-to-Goal: A learned temporal distance metric for long-horizon planning in JEPA world models
SummaryThe paper addresses a weakness of planning with Joint-Embedding Predictive Architectures (JEPAs), where Euclidean goal costs can fail to distinguish states at different temporal distances from the goal. The authors propose Time-to-Goal (T2G), a lightweight metric trained on latent triples from offline trajectories with a margin ranking loss so that temporally nearer states get smaller distances. With the JEPA encoder and predictor kept frozen, T2G replaces the Euclidean planning cost. Across four navigation and manipulation benchmarks, average planning success rises from 32.7% to 63.0%.
In briefDual Flow: a joint image and measurement flow model for recovering images from indirect, noisy measurements.
SummaryThe paper proposes Dual Flow, a model that solves linear inverse problems by jointly estimating a clean image and its noise-free measurement using coupled flows. The final estimates are guaranteed to satisfy the forward model. Experiments across several linear inverse problems and image datasets report state-of-the-art distortion and perceptual quality, along with substantially shorter reconstruction times than competing methods.
In briefPEARL: A Learning Method for Selective Forgetting Under Partial Preference Drift
SummaryThis paper studies personalized recommendation and model selection when user preferences change over time. It proposes PEARL, which reverses confidence accumulation only along feature directions where predictions are badly wrong, leaving other directions intact. On the PRISM benchmark of 1,500 users and 21 LLMs, PEARL outperforms greedy and neural bandit baselines, as well as sliding-window and discounted forgetting methods.
In briefVision Paper: Why Distributed and Decentralized LLM Post-Training Is Now Essential
SummaryThis vision paper argues that distributed learning, including federated and decentralized approaches, should become a core paradigm for LLM post-training rather than just a privacy add-on. It contends that methods such as RLHF and DPO assume preference data must be centralized, a premise the authors judge scientifically, practically, and ethically questionable. The paper builds its case on privacy and data sovereignty, pluralism and representational equity, and robustness and scalability, and it outlines technical challenges, algorithmic directions, and a research agenda.
In briefRACE: A rollout-free checkpoint selection score for robot policies
SummaryThis paper introduces RACE, a metric for ranking training checkpoints of robot policies without running costly real-world trials. It calibrates action deviations against thresholds derived from action differences at similar states in a fixed set of held-out successful trajectories, and aggregates them through a running product. Experiments in simulation and on real robots report that its rankings align better with deployment success than baselines.
In briefBudget-aware routing of LLM calls that learns tail costs online
SummaryThis paper studies how to route tasks among large language models under a fixed budget, deciding which model to call and whether to retry failures, with the goal of solving as many tasks as possible before the budget runs out. The authors propose a contextual bandit policy, ADP-HT, which learns each model's task-specific success probability and builds conservative cost estimates from coarsened response-length labels on a quantile grid. At the tightest tested budgets, it solves more than twice as many MATH500 and GPQA tasks as the best baseline, and it achieves the highest mean cumulative score at ten of fourteen RouterBench budgets.
In briefAutonomy and coordination in smart manufacturing: a recursive agent framework
SummaryThe paper views intelligence in human production organizations through a complex adaptive systems lens, identifying individual autonomy and collective coordination as coupled mechanisms improved by learning. It proposes the Recursive Autonomous Agent (RAA) framework, combining goal-driven agents with a recursive organizational structure governed by goal and capability contracts. Common interfaces let equipment, production lines, and factories join larger coordinated organizations.
In briefTandemBench: A benchmark for testing decision models in multi-agent teamwork
SummaryThis paper introduces TandemBench, a set of 1,200 questions set in multi-agent cooking scenarios, where models must pick the best team plan from given options. Some variants also ask whether better arrangements exist beyond those options. The authors evaluate 40 native decision models and 10 general-purpose language models. The best native model reaches 62.50% decision accuracy, while native models perform near chance on optimality verification, and some report better alternatives even when an optimal plan is already present.
In briefDualCast: Dual-image diffusion transformers for digital human e-commerce videos
SummaryDualCast is a framework for generating digital human e-commerce videos that keep the product's appearance, the presenter's identity, and the overall style consistent. It encodes product and human images separately and places them at the front and rear of the transformer sequence to separate their roles. A mask-weighted loss emphasizes product, face, and overlap regions, and condition dropout lets one model handle dual-image, product-only, or human-only generation. The authors also built a curated dataset of 10K e-commerce videos and report strong experimental results.
In briefSurvey of ship navigation safety in poor visibility, from vision to multi-sensor fusion.
SummaryThis survey reviews intelligent vision methods for marine surface vehicles operating in restricted visibility. It covers shipboard video stabilization, visual perception enhancement, object detection and tracking, and navigable area segmentation. It also examines vision-centered multi-sensor fusion approaches, including infrared-visible, radar-vision, LiDAR-vision, and AIS-vision fusion. The paper summarizes open-source datasets and safety-critical applications, and outlines open challenges and future research directions.
Ryan Wen Liu, Fengchuan Song, Tianxiang Wang et al.
In briefLearning When Human Input Matters: Executor-Relative Consultation Value for LLM Agents
SummaryThis paper studies when an LLM agent should consult a human under a limited consultation budget. The authors propose Consultation Value Distillation (CVD), which estimates consultation values offline via Monte Carlo executor rollouts and distills them into lightweight predictors that need no rollouts at deployment. The method is reported to outperform direct prompting and Value of Information baselines across several batch and sequential settings.
In briefReTailor: A Real-Time Residual Feature Tailoring Framework for Cross-Sensor Pansharpening
SummaryPansharpening models trained on one satellite sensor often perform poorly on unseen sensors. The paper proposes ReTailor, a real-time framework that rapidly optimizes only a lightweight residual feature tailor under self-supervised, physics-aware constraints to correct cross-sensor shifts. It keeps the learned fusion prior unchanged and adds content-aware patch selection and rolling residual bootstrap for efficiency and quality. Experiments show it mitigates cross-sensor degradation effectively and efficiently, and it is also evaluated on super-resolution pansharpening and image segmentation.
In briefAligning Inference-Time Expert Deactivation With the Training Weight Rule
SummaryThis paper studies how deactivating experts in mixture-of-experts language models at inference time affects performance. It compares four implementations (removal, freezing, remaining-expert renormalization, and unnormalized re-routing) across five instruction-tuned checkpoints and a base OLMoE checkpoint. The authors propose a weight-sum model linking the loss of expert contributions and branch scale changes to the weight rule used in training. Perplexity rises by 1.4 to 73 times when the training rule is kept, and by 14 to 28,740 times under a rule mismatch.
Hao Cheng (Zhejiang Gongshang University), Zhiyu Yang (Chang'an University), Bo Ma (Zhejiang Gongshang University) et al.
In briefPhonological Knowledge of Chinese Characters in Language Models: A Benchmark and Analysis
SummaryThis paper introduces PinyinBench, a benchmark covering homophones, initials, finals, tones, and polyphonic characters, to test how well 17 text-only and multimodal language models know the sounds of Chinese characters. It combines behavioral evaluation, layer-wise probing, and activation steering. The models know initials and finals well, but judging characters that differ only in tone is much harder, and tonal information is harder to access and manipulate.
In briefGREP: a principle-based framework for judging AI-generated furnished rooms
SummaryThis paper introduces GREP, a framework for evaluating the overall quality of furnished room scenes generated by LLM-driven selection and arrangement of assets. It uses a three-tier set of principles: basic physical validity, visual coherence, and human-use logic, guiding an evaluation agent that identifies issues, gathers evidence, and produces scores and reports. Experiments report 87% agreement with human majority preferences and quality gains when scenes are refined using GREP reports.
In briefWayPOP: A Panoramic Open-Set Panoptic Tracking Benchmark
SummaryThis paper introduces WayPOP, a benchmark for testing whether autonomous-driving systems can segment and track objects outside a fixed set of categories across multiple cameras and time. It is built on real Waymo sequences with about 670 Objaverse 3D assets inserted into the scenes, yielding 980 unknown-object tracks and 53,452 mask annotations. Evaluation of representative open-set and open-vocabulary methods shows substantial limitations in truly open-set settings.
Rohit Mohan, Maximilian Luz, Swarnava Chowdhury et al.
In briefItem Representations Matter: Bridging Hard Semantic IDs and Soft Prototype Mixtures for Generative Retrieval
SummaryThis paper proposes CAP-Mix, a flow-based recommender for generative retrieval that represents each item as a soft mixture of shared prototypes plus an item-specific vector. The assignments, prototypes, and item vectors are trained jointly, and the soft representation is kept at inference. Across three datasets, it improves over a flat lookup table on sparse Amazon datasets and reports higher HR@5 and NDCG@5 averaged over six benchmarks.
In briefConsistency Is Not Understanding: Auditing Neural Field Weight Readers
SummaryThe paper studies how well readers can recognize visual content from neural network weights while ignoring arbitrary parameter coordinates. It tests readers on 480 shape fields and 1,000 MNIST fields under function-preserving transformations such as hidden-neuron permutations and positive ReLU rescalings. Results show a statistics reader's shape accuracy drops under joint transformations while a normalized set reader keeps unchanged predictions. The authors also report that canonicalizing prompts does not establish recognition or guarantee identical generated responses.
In briefDARE: A Framework for Analyzing and Improving Repair Task Difficulty
SummaryThis paper introduces DARE, a framework that measures and predicts how difficult code repair tasks are, using the observed failure rate across multiple coding agents as the difficulty measure. It trains a difficulty model on 151 static features taken from benchmark artifacts and analyzes seven repair-task feature groups to choose task-specific agent support. Tested on four benchmark partitions with 2,594 tasks, the predictions reach Spearman correlations of 0.476 to 0.621 on held-out data, and matching support to each task's salient feature group improves repair success.
In briefZipDrift: Generating multimodal short-horizon MPC costs with one-step drifting models
SummaryThis paper proposes ZipDrift, a method for correcting the cost of short-horizon model predictive control (MPC), which is fast but short-sighted. A one-step drifting generator is trained over cost corrections, favoring low closed-loop cost and learned only from cost evaluations of candidate corrections. Generated candidates are ranked by a learned critic, and model-based MPC then enforces dynamics and constraints. In a kinematic overtaking benchmark, ZipDrift was collision-free in all 600 test episodes.
In briefFlyHash for LLM Agent Memory: Empirical Validation and New Use
SummaryThis paper applies FlyHash, a hashing method inspired by the fruit fly's olfactory circuit, to memory retrieval for LLM agents, comparing it with SimHash and HNSW. Across multiple datasets and random seeds, FlyHash preserves neighbor structure better than SimHash at matched bit counts. However, at equal bits its codes are larger and slower, and HNSW outperforms both hashing methods at ordinary scale. The quality advantage also holds on non-circular agent-memory relevance labels.
In briefUnsupervised neural operators that speed up modeling of fast-spinning neutron star interiors
SummaryThis paper proposes physics-informed neural operators that solve the axisymmetric general relativistic equations governing the interiors of rotating neutron stars. The solvers are trained in a fully unsupervised way, using only the governing partial differential equations as the training objective. The authors report a 10^3-fold speed-up with negligible loss of accuracy, which makes equation-of-state inference and population synthesis that account for rotation computationally feasible.
In briefOptimizing Recourse When Multiple Models Disagree
SummaryAlgorithmic recourse finds changes that turn an unfavorable model prediction into a favorable one. The paper addresses predictive multiplicity, where similarly performing models disagree on individual predictions, and formulates robust recourse as minimizing the worst-case validity loss across competing models under a budget constraint. It proposes an oracle-efficient algorithm, ORPM, with proven sufficient conditions for convergence. The method is evaluated on one synthetic and five real datasets using linear, neural, mixed, and decision-tree model sets.
In briefHow Well Can LLMs Impersonate Individual Authors? A Comparative Evaluation
SummaryThis study compares six large language models on their ability to imitate an individual author's writing style, using short product reviews with consistent personal styles. Two prompt types were tested, a simple one and one with explicit stylometric guidance, and mimicry was measured with a cosine-distance metric. Gemini 3 Pro scored highest, but it copied sentences and phrases from training texts at levels far above human authors. For most models, generated texts were largely separated from the target author's human clusters in embedding space.
In briefME^2FDiff: Multiscale Expert-Enhanced Feature Diffusion for Brain Tumor Segmentation
SummaryThis paper proposes ME^2FDiff, a single-model network for segmenting post-treatment glioma in MRI scans. It refines the bottleneck features with an attention-conditioned spectral diffusion block and fuses intermediate encoder scales through a multiscale expert-enhanced feature block. On the BraTS 2024 post-treatment glioma dataset, it achieves the best results on every tumor sub-region among single-model baselines, reaching 91.23% Dice on whole-tumor segmentation.
In briefBobFlow: Dataflow-Based Human-Agent Collaboration for Data Science
SummaryThis paper presents BobFlow, a system that combines the dataflow abstraction with the ReAct paradigm so that agents can iteratively build and run dataflows for data science tasks. Analysts can inspect the analysis and give feedback during this process. A user study found that BobFlow helps analysts understand, inspect, and verify agent-generated analyses. Experiments across proprietary and open-weight large language models show accuracy gains of up to 18.6 points at just over one-third of the inference cost of a script-based ReAct agent.
In briefPAYWORTHY: Can LLMs Predict Which Freelance Jobs an AI Agent Will Complete?
SummaryThis paper tests whether language models can predict, before any attempt, which freelance job postings an AI agent will complete successfully. The authors built a dataset of 1,494 real postings from a freelancing platform, where an agent attempts each job from the posting alone and an LLM judge acting as client decides acceptance. Eighteen models forecast success likelihood, and their rankings are scored by AUC. Forecasts rank jobs well (AUC 0.66 to 0.86) when the agent's limitations are described, but perform near random (0.45 to 0.56) when it is described as a generic agent with a browser and code execution.
Mohammad Reza Rezaei, Dingxian Wang, Brett Levert et al.
In briefMechanical Engineering Skills: an open library of AI agent skills for CAD, simulation and manufacturing
SummaryThis paper presents an open library of 33 agent skills that lets mechanical engineers use CAD, simulation, and manufacturing tools through a single text or voice conversation. Each skill combines operating instructions, executable tools, and verification procedures, with a language model coordinating skills and deterministic tools producing checked results. Five case studies with independent reruns report exact agreement for lattice homogenization and rocket apogee predictions within 4% of OpenRocket references.
Achuth Chandrasekhar, Jesse Barkley, Peter Pak et al.
In briefDecomposed Distillation for Single-Pass Diffusion Guidance
SummaryThe paper presents CtrlFlow, a single-pass guidance method for conditional image generation with diffusion models. It keeps the separable base flow and guidance direction of Classifier-Free Guidance by attaching two lightweight LoRA adapters (2.5% extra parameters each) to a frozen backbone. A unified multi-scale distillation loss aligns training with inference, and scale-adaptive modulation and time-grid warping reduce residual errors at extreme guidance scales.
In briefA layered knowledge design that keeps organizational interpretations revisable
SummaryThe paper proposes a four-layer architecture (Signals, Domains, Claims, Dimensions) that stores organizational interpretations together with their justifications and standing. A truth-maintenance network revises that standing as evidence changes, and validated, versioned Dimensions place interpretations on named axes. The authors call the result a shared organizational cortex and report retrieval gains, including a 0.081 nDCG@10 improvement over Signals-only retrieval.
Andrew Miracle, Phyl Georgiou, Christine Okoth et al.
In briefFeeding GPUs Data on Time: A Pipeline Approach for LLM Training on AMD Hardware
SummaryThis paper addresses idle GPU compute time during large language model training, which occurs when data is not ready when needed. It uses AMD GPU features for explicit data movement, including the Tensor Data Mover, transpose-load instructions, and SDMA and RDMA for inter-GPU transfers, to keep a steady data pipeline. Reported results show average kernel speedups of 1.88x for TDM and 1.19x for transpose-load, and average speedups of 1.16x for SDMA and 1.33x for RDMA in overlapped computation and communication.
Gengyu Rao, Ghadeer Ahmed H Alabandi, Su Ann Chong et al.
In briefExpert critiques plus LLM refinement for building AI evaluation datasets
SummaryThe paper presents SME-C, a human-AI workflow in which subject matter experts give lightweight critiques of candidate responses, and LLM-driven analysis and refinement iteratively curate response datasets and evaluation rubrics. Validated on financial business briefs using 3,853 critiques from 411 finance professionals, both refinement strategies significantly outperformed the experts' top-ranked baseline responses in a blind survey. Critique-based refinement produced more domain-grounded improvements, while vote-based refinement was better at removing irrelevant content.
In briefText-only domain adaptation for Hausa medical speech recognition using language model fusion
SummaryThis paper studies how to improve medical speech recognition in Hausa, a low-resource language, using only text data and no new domain-specific speech. It compares two decoding-time language-model fusion methods, Shallow Fusion and the Density Ratio Approach, applied to a fixed acoustic model. Both methods reduce word error rate compared with greedy decoding, but they do not differ significantly from each other. A general-domain language model alone achieves nearly the same improvement, so the medical language model adds only a modest benefit.
Oche David Ankeli, Hassan Adelani Luqman, Theodora Ngozi Egbunike et al.
In briefLearning smarter image patch splitting to speed up Vision Transformers
SummaryThis paper proposes SPLIT, a method that learns how to allocate token resolution in Vision Transformers based on what the model finds meaningful, rather than on pixel-level heuristics like local entropy. This produces shorter token sequences matched to task-relevant content. SPLIT matches full-resolution ViT classification accuracy while using 30% fewer tokens, and achieves competitive results on COCO and ADE20K dense prediction tasks.
Aditya Gupta, Jean Sebastien Dandurand, Kai Qiu et al.
In briefPersonaCodeBench: A benchmark testing coding agents in multi-turn sessions shaped by user personas
SummaryPersonaCodeBench is a benchmark for evaluating coding agents in multi-turn coding sessions driven by user personas. It turns user roles, technical abilities, interaction preferences, and practical constraints into verifiable requirements that are revealed progressively during the interaction. A completion verifier checks code correctness and a personalization verifier checks whether persona requirements are met. The authors evaluate major frontier agents and publish a leaderboard.
In briefOdorSpaces: Visual-Olfactory Embodied Navigation in 3D Environments
SummaryThis paper introduces OdorSpaces, a physics-based simulation platform and benchmark for navigation that uses both vision and smell. It extends Habitat-Sim with 3D odor transport and dispersion modeled by Fire Dynamics Simulator (FDS), so agents can navigate indoor scenes using visual and olfactory observations. The benchmark covers two source-seeking tasks: locating an invisible gas leak and locating an early-stage fire. The authors also develop a multimodal reinforcement learning policy that combines visual input with olfactory measurements over time, which they report outperforms a range of baselines.
In briefC-WFDM: A ControlNet-guided diffusion model that fills in missing brain MRI scan types.
SummaryThis paper presents C-WFDM, a ControlNet extension of the Wavelet-Fusion Diffusion Model, for completing missing MRI contrasts in individual brain scans. Because pooled neuroimaging datasets often lack certain contrasts, the model synthesizes T2-weighted, FLAIR, and gadolinium-enhanced T1-weighted images from the scans a participant does have. Trained on the BrainScape collection of 157 public datasets and 45,880 scans, it is compared against ALDM, cWDM, and Pix2Pix. C-WFDM achieves the best T2w and FLAIR results among the evaluated models, with SSIM values of 0.9232 and 0.8904.
Muhammad Nabi Yasinzai, Remika Mito, Mangor Pedersen
In briefSelf-supervised learning of shared latent spaces for neural operators (a reusable function representation trained without labels)
SummaryThe paper pre-trains a functional autoencoder on synthesized analytic functions, with no solver or labels, to represent functions on arbitrary point clouds. The resulting latent space is frozen and reused across tasks, so neural operators train on cached embeddings. Across 28 tasks in fluid dynamics, solid mechanics, acoustics, Helmholtz and Poisson equations, the authors report accuracy that can exceed end-to-end architectures, with higher throughput and lower memory use.
WhyThe authors decouple representation and operator learning via a self-supervised frozen latent space, reporting gains over end-to-end models on 28 PDE tasks with large memory savings.
Sepehr Mousavi, Levi Lingsch, Siddhartha Mishra et al.
In briefReducing spurious invariance in contrastive learning with probabilistic self-supervised learning for medical images
SummaryThis paper addresses a limitation of contrastive self-supervised learning on medical images, where data augmentations can suppress subtle but critical features and produce spurious invariance. The authors encode feature embeddings as probability distributions instead of fixed vectors to capture richer information, and propose a unified probabilistic framework that integrates multiple self-supervised approaches. The method outperforms state-of-the-art self-supervised methods on downstream medical image segmentation and classification across multiple datasets. The paper also analyzes robustness to stronger augmentations, label efficiency with limited fine-tuning labels, and computational trade-offs.
In briefProspective Memory for Proactive LLM Agents: a memory framework that helps AI agents remember future intentions and act at the right time.
SummaryThis paper observes that most LLM agent memory systems look backward at past information, whereas proactive agents also need to retain future intentions and carry them out at suitable moments. The authors propose Pro-Memory, a framework inspired by human prospective memory theory that separates maintaining intentions from deciding when and how to realize them. Experiments on two benchmarks report better proactive task performance across several LLM backbones while reducing unnecessary interventions.
In briefMale abstraction bias: how abstract concepts link to men in internet culture and AI language models
SummaryThis paper examines tens of thousands of single and multi-word expressions across internet text and images, finding that more abstract concepts are more associated with men. The pattern replicates across languages and extends to grammar, with verbs and adverbs being more abstract and more male-coded. The bias also appears in free association responses from both humans and AI, suggesting it is both internalized and reproduced algorithmically.
WhyThe authors report that abstractness and male association link across concept space, grammar, and AI free-association, replicated across languages using large internet corpora and images.
In briefRESHPart: Feedback-Reshaped Spectral Embedding for Hypergraph Partitioning
SummaryThis paper studies spectral hypergraph partitioning by analyzing the geometry of spectral embeddings, identifying a ray-like structure that helps explain why projection-based partitioning works. The authors propose FRESHPart, a framework that refines spectral embeddings using Geometry Feedback and Search Feedback so they align better with projection-based partitioning. On the Titan23 benchmark, it reports average cut reductions over hMETIS of 13.09% at ε=2% and 25.21% at ε=20%, finishing every benchmark in under 120 s on a single GPU.
In briefGenerative Modeling of Cascading Extremes with Autoregressive Transformers
SummaryThe paper introduces a categorical token representation for multivariate extreme events, where a quiet token marks no threshold exceedance and otherwise the token names the component with the largest marginal probability-scale score. A decoder-only transformer estimates the history-dependent next-token kernel, and recursive sampling generates future token sequences. The model is pretrained on simulated process families and evaluated against exact probabilities and particle filter references. Applied to ten-second returns of five semiconductor stocks, it transfers zero-shot, and brief adaptation improves agreement in token frequencies and transition patterns.
Miguel de Carvalho, Myung Won Lee, Jordan Richards et al.
In briefCan language models tell which constraints should be grouped together?
SummaryThe paper studies how to choose which groups of constraint families to convexify in a mixed-integer program, where some families strengthen the linear relaxation only when applied together. The authors prove that methods observing only bounds need many evaluations to find such hidden groups, and propose JointDW, which combines language-model predictions with a staged search and a one-pass lookahead. On MIPLIB-NL models, the strongest language models identify these groups about as often as the lookahead, and reasoning roughly doubles accuracy.
WhyThe authors prove a covering-number lower bound for finding hidden constraint groups and test a staged search with language-model predictions on MIPLIB-NL models.
Junbo Jacob Lian, Kaixuan Chen, Hanzhang Qin et al.
In briefParallel Reconsideration: Giving Speculative Decoding's Drafter a Second Guess
SummaryThis paper addresses how a small drafter model in speculative decoding can be made more accurate for accelerating large language model inference. The proposed Parallel Reconsideration method lets the drafter, guided by its own uncertainty, re-examine hesitant guesses in extra passes that run alongside the single-pass draft, and adds these refined candidates to the draft tree. Across models and benchmarks, the method improves acceptance length by 5.3% to 10.2% and speedup by 3.5% to 9.2% on average over baselines.
Hoang Anh Duy Le, Huibing Dong, Alireza Khadem et al.
In briefCognitive Biases in Search-Equipped LLM Agents: A Stage-by-Stage Analysis
SummaryThis paper studies how large language model agents with search tools are affected by cognitive biases. The authors build a controlled search setup that keeps questions and document collections fixed while varying result order, prior work records, and initial hypotheses. Tests across 10 models and 1,000 synthetic search worlds track queries, retrieved documents, document openings, and final answers. The results indicate that final answers depend more on initial evidence order during active search, and that dead-end histories reduce strategy switching beyond what context length alone explains.
In briefSame Method, Different Task: Limits of the Bitter Lesson in Social Simulation
SummaryThe paper argues that the bitter lesson's successful cases all scaled general methods on the same task as the domain engineering they replaced. Social simulation instead swapped hand-built agents for models optimized to predict text, a change of task the authors say the lesson does not support. They contend that prediction involves no consequences for the predictor, whereas cooperation and norm enforcement require interested parties, so what is missing is stakes rather than emergence. A pre-registered pilot with 540 agents found that persistent consequences altered strategies in two of three model families but did not produce human-like heterogeneity.
In briefCoReFlow: A consistency-relaxed flow map model for probabilistic traffic forecasting
SummaryThe paper introduces CoReFlow, a model for probabilistic traffic forecasting that models uncertainty in road networks. It uses the one-step generation ability of flow maps and a training objective that supervises interpolation endpoints over finite time intervals, restricting consistency with the conditional mean. Experiments on four real-world datasets report state-of-the-art forecasting accuracy with lower inference cost than existing generative baselines.
In briefReducing labelling needs for smart-glasses facial expression recognition via self-supervised pretraining
SummaryThis paper studies how self-supervised learning can reduce the labelled data needed for facial expression recognition on OCOsense smart glasses, which sense expressions without cameras. The authors pretrain a Vision Transformer with a masked-autoencoder objective on unlabelled recordings, then finetune it for five-class recognition under person-independent cross-validation. Pretraining improves performance at every label budget from 5% to 100%, and the model pretrained with only 5% of labels outperforms a model trained from scratch on all labels.
Zhengtao Gao, Francesco Bombassei De Bona, Leonardo Alchieri et al.
In briefAgents Should Think Only When It Helps: Testing Thought Supervision in LLM Agent Distillation
SummaryThis paper examines whether the thoughts in teacher trajectories are critical for student LLM agents' action performance during distillation. The authors independently perturb teacher thoughts and actions and evaluate on widely used agent benchmarks across model scales. Altering or discarding thoughts has little effect on student performance, but thought supervision matters when the thought's information is not disclosed in the corresponding action.
In briefNeural networks and games: an exact equivalence between Hopfield networks and games
SummaryThis paper proves an exact equivalence between symmetric binary Hopfield neural networks and binary-action games that admit a quadratic potential. Stable network configurations correspond to pure-strategy Nash equilibria, and asynchronous neural updating corresponds to best-response dynamics. The author further argues that storing and retrieving memories in neural networks correspond to designing and selecting equilibria in games.
In briefFrom Type to Token: Petri Nets and Individualized Causal Effects
SummaryThe paper argues that standard causal models such as Structural Causal Models and Bayesian Networks describe population-level average effects, which lose meaning when individual treatment effects vary widely. The authors propose Petri Nets under the Individual Token Philosophy as a token-level representation of causal structure and give a formal translation from Bayesian Networks to Petri Nets. A conformal prediction layer is added to quantify individual causal effects with finite-sample coverage guarantees.
Fernando Delbianco, Fernando Tohm\'e, Gustavo Bodanza
In briefTACT: A Practical AI Council Method for Qualitative Analysis
SummaryThis paper presents TACT (Traceable AI Council for Thematic analysis), a procedure in which multiple language models jointly perform qualitative coding while teams keep control of interpretation and judgment. It has two variants, TACT-Lite and TACT-Full, the latter adding critic models from other model families and a decision model that flags high-uncertainty code assignments. Evaluated on 256 Stack Overflow threads coded by other researchers, TACT reached F1 scores of 0.71-0.77, and it costs $13.77 to $33.59 per 100 threads in model calls.
Rudrajit Choudhuri, Christian Bird, Carmen Badea et al.
In briefGeoDIVE: Best-first search for finding objects in satellite images
SummaryGeoDIVE is a model for locating specific objects, structures, and spatial patterns across large, multiscale satellite image collections. It treats the image gallery as a tree and uses best-first traversal to decide which tiles to encode, with a lightweight model estimating the best match among each node's descendants. For image queries it matches exhaustive search at 10% of the encoding cost and outperforms all tested baselines at every budget in both image and text query settings.
In brief3D CT segmentation from three orthogonal annotated slices using anchored propagation
SummaryThe paper proposes propagating segmentation masks from three orthogonal annotated slices, correcting each propagation chain where it crosses another annotated plane and weighting chains by their agreement at fusion. It compares two ways of transporting boundary geometry: a truncated signed-distance field and a learned contour tracker. On about 403 CT scans spanning 17 abdominal structures, anchoring raises mean Dice from 0.737 to 0.769 and improves every structure.
Tushar Kataria, Md Rakibul Haque, Shireen Y. Elhabian
In briefTRUST-Stereo: A Fast Stereo Matching Method Using Teacher Assistance and Uncertainty-Guided Search
SummaryThis paper presents TRUST-Stereo, a fast stereo matching method that addresses errors in propagated disparity in coarse-to-fine approaches, which can place correct correspondences outside the local search range. Its Uncertainty-Guided Residual Search Module estimates the reliability of propagated disparity and predicts residual offsets to adaptively relocate local cost-volume centers. The method also uses a two-stage Senior-Junior Teacher Assistant distillation strategy to transfer monocular knowledge from a large foundation model to a lightweight backbone. Trained on only 1.4M synthetic stereo pairs, it achieves state-of-the-art results on several benchmarks at about 50 ms per pair.
Pouya Sohrabipour, Yihong Feng, Chaitanya Kumar Reddy Pallerla et al.
In briefOptimal settings for classifying tree species from bark images, tested without data leakage
SummaryThis study systematically compares design choices for deep-learning tree species classification from bark photos, using the public BarkNet 1.0 dataset with 20 species and ConvNeXt-V2 backbones. Patch size has the largest effect, with image-level accuracy plateauing near 94.9% around 512 px. Learned aggregation and backbone capacity have only secondary effects. The authors recommend large patches with a nano backbone as a reproducible reference configuration.
In briefHow AI agents make decisions from visual data
SummaryThis paper studies how artificial agents powered by vision-language models perceive uncertainty and make decisions based on visual data. The authors ran two complementary studies: one on how agents perceive and quantify uncertainty shown in visualizations, and one on decision-making from images. They compared perception accuracy and decision performance across models and visualization types. GPT-4o showed the highest accuracy in perceiving uncertainty in pie and bar charts, and agents aligned best with rational decision theory when uncertainty was shown as circle charts, icon arrays, or text.
In briefRevisiting how critics are estimated in Soft Actor-Critic, a reinforcement learning method
SummaryThis paper examines design choices for the critic in Soft Actor-Critic (SAC), a deep reinforcement learning algorithm. It compares six critic variants, including learning soft advantages, using multi-step return targets, and training on replayed trajectory segments, across fourteen tasks with random hyperparameter searches. The authors find that learning soft advantages does not help and that soft GAE offers no gain over the simpler λ-return critic. Training a λ-return critic on replayed trajectory segments performs best, most clearly on continuous control tasks.
In briefCombining relevant expert knowledge with mixed feedback fuzzy cognitive maps
SummaryThis paper proposes a neuro-symbolic method that combines only the expert answers relevant to a given query. It extends a legal-evidence definition of relevance, requiring answers to be both probative and material, and applies it to mixing the feedback fuzzy cognitive maps of multiple experts. The probative status of each answer is estimated by comparing expert predictions with those of the mixed map, and a Bayes-like theorem de-mixes the knowledge into posterior-like maps that yield query-specific relevancy weights. The approach is demonstrated on synthetic maps generated from a 13-node expert map of ischemic-stroke risk.
In briefGrounded but Not Immune: Confirmation Bias in Agent-as-a-Judge
SummaryThis paper studies whether an Agent-as-a-Judge, an evaluator that uses tools and reviews an agent's full trajectory, is still susceptible to confirmation bias. A single false pre-screening note attributed to an authority overturned up to 40% of correct verdicts across four judges and five datasets, even when the judge gathered decisive evidence. The authors classify the resulting behavioral changes into three failure modes: over-trust, hallucination, and under-investigation.
Drake Benjamin Brown, Khushbu Pahwa, Yubin Ge et al.
In briefAdaptive domain decomposition for neural PDE solvers that learn their own partitions
SummaryThis paper proposes ADD-FBPINNs, which combine Partition of Unity Networks (POUNets) with finite basis physics-informed neural networks (FBPINNs). The method adapts the domain decomposition to solution features, learning partitions that are stable and geometrically interpretable. It is tested on the two-dimensional Helmholtz equation, a two-dimensional Poisson equation with multiscale features, and the viscous Burgers equation. The authors report improved accuracy over FBPINNs with fixed decomposition.
Sarah Helfert, Wanxin Chen, Amanda A. Howard et al.
In briefBeyond Accuracy: Information-Theoretic Evaluation of Brain Decoders
SummaryThis paper examines how brain-signal decoders for assistive brain-to-robot control should be evaluated, arguing that single-trial accuracy is only a partial proxy for sequential planning performance. The authors model control as a partially observable Markov decision process and identify a regret phase transition governed by the worst-pair Chernoff information. They derive a closed-form regret predictor computable from any confusion matrix, and report that on real EEG data two decoders with equal 35.1% validation accuracy differ by 9.6 times in held-out regret, which the predictor captures.
In briefWhen is an SAE latent the same feature across two dictionaries?
SummaryThe paper proposes a framework for comparing sparse autoencoder (SAE) dictionaries using an energy-weighted optimal-transport plan that matches latents and quantifies dictionary-level discrepancy. For fixed trained SAEs, it derives joint asymptotic inference for transport costs and fractions and uses a joint document bootstrap to build simultaneous confidence bands. Experiments across four language models locate cross-seed discrepancy and separate context-dependent changes.
Alexandra Suvorikova, Ilgam Latypov, Anton Trukhan et al.
In briefMeasuring how EEG models encode stable traits versus changing brain states
SummaryThis paper examines how EEG embeddings from signal processing methods and foundation models encode two kinds of information: stable traits such as subject identity, age, and sex, and states that respond to stimuli. Using a hierarchy metric and kNN purity on brain-computer interface trials, the authors find that most EEG features form trait-level clusters with micro-states inside. Linear probe results indicate this structure reduces state distinguishability, and normalization that reduces trait information improves state classification balanced accuracy by about 10%.
In briefProving Agent Failure Causes Requires Executed Counterfactuals: A Reporting Standard
SummaryThis paper argues that claims linking an agent's behavior to its history should be backed by intervention-verified evidence. The proposed method freezes an executable mid-trajectory state, alters only the history the agent sees, and continues execution to observe the outcome. The authors note that several existing methods execute interventions from checkpoints or recorded states, but none reports a measured replay certificate. They also present two cautionary measurements, including one showing that restoring a random stream's position made snapshot results match the live run.
In briefNitroFusion2: Adding Adversarial Loss to the Text Stream of Multimodal Diffusion Transformers
SummaryThis paper proposes Text-Stream Adversarial Distillation (TSAD), which uses intermediate text-token states of multimodal diffusion transformers as the representation for adversarial discrimination. It also introduces Aspect-Cued Discrimination (ACD), which conditions the discriminator on aspects such as typography, objects, or style. Combining these in NitroFusion2, the authors report reducing few-step distillation from 4 to 8 steps down to 1 to 2 steps at comparable quality.
In briefARIDA: Semi-automated annotation of railway infrastructure images using neural network predictions
SummaryThis paper presents ARIDA, a semi-automated pipeline for annotating railway images. It converts instance predictions from a mask scoring segmentation network into bounding boxes, polygons, polylines, and paired polylines. Automatic labels are accepted using confidence thresholds and a contour-agreement quality score, while an active learning loop ranks remaining images by uncertainty so human effort targets hard cases. Experiments on disjoint and time-sequential imagery quantify effort reduction and accuracy trade-offs.
Maximilian Dornik, Christian Sieberichs, Simon Geerkens et al.
