In briefHierarchy-GBP: Speeding Up Factor Graph Inference with Abstraction and Recovery
SummaryThis paper proposes Hierarchy-GBP (H-GBP), which speeds up Gaussian Belief Propagation by first solving global errors on a coarse graph approximation and then projecting the results back to the original graph. Standard GBP then refines the remaining local errors. The authors prove convergence to the optimum by analyzing the spectral radius of the combined operator. Experiments on sparse linear graphs, Pose Graph Optimization, and Bundle Adjustment show faster convergence than standard GBP.
In briefMemTrace: State-Consistent Memory for Long-Horizon Coding Agents
SummaryThis paper presents MemTrace, a provenance-aware memory system that helps coding agents manage execution history during long, multi-file software tasks. It stores history as immutable memory traces linked by a dependency graph, keeps only compact anchors when context is limited, and checks recalled evidence against the current repository before reuse. The authors report that across three long-horizon coding benchmarks, MemTrace outperformed all fully evaluated baselines under the same backbone and harness.
In briefRSure-Agent: Making Remote Sensing Agents Reliably Use Tool Results
SummaryRemote sensing agents use perception, measurement, and raster analysis tools to answer Earth observation questions. The paper analyzes 1,229 execution trajectories and finds that tool observations are often wrong even when tools run successfully, and these errors frequently propagate to the final answer. The authors propose RSure-Agent, which requires tools to return process evidence and uses a task-tool reliability prior built from offline feedback to accept, request more evidence for, or reject each observation. The framework is evaluated on EarthBench, ThinkGeo, TerraLogic, and CHOICE-420, reducing error propagation and improving accuracy.
In briefCluster validation indices as self-supervised training targets for text representations
SummaryThe paper proposes SilK (Silhouette-guided K-means), a self-supervised method that fine-tunes a pretrained language encoder without labels, negative pairs, or auxiliary graphs. It clusters the corpus and regresses a simplified silhouette score toward a target value, comparing each document only with k cluster centroids. On BERT-base, SilK trains 1.46x faster per epoch than the fastest baseline evaluated and uses 45.4% less peak GPU memory than the leanest one. Under frozen-encoder linear probing, it remains competitive with the best baselines on three downstream tasks.
Kishor Kumar Bhaumik, Nicolas Roque dos Santos, Neil Shah et al.
In briefAnalysis of verification and accumulation failures in multi-agent proof search
SummaryThis paper studies verification and lemma accumulation in agentic provers working on open problems, where no proof assistant is available and language models judge each other's outputs. The authors instrumented one system end to end and analyzed 51,754 traced observations across three full runs. They report that a unanimity-based verifier treats parse or API failures as non-approval, and that lemmas were extracted from rejected arguments with their refutational context removed.
In briefGRAM: Fixing Frozen Time-Series Forecasting Models Using Retrieved Error Patterns
SummaryTime-series foundation models make zero-shot forecasts after large-scale pretraining, and retrieving history can further improve them. This paper proposes GRAM, which corrects a frozen model's forecasts using the systematic bias that recurs in its prediction errors, rather than the ground-truth futures of similar past windows. It scales errors by amplitude into retrievable prototypes and uses a prototype graph to keep consistent bias while suppressing random fluctuation. During online forecasting, it generates per-horizon corrections, and experiments across datasets and foundation models show consistent gains.]
In briefDICE: Separating Tutoring Ability from the Need to Intervene in LLM Tutors
SummaryThe paper argues that LLM tutors are usually trained to reply to every student turn, conflating what to say with whether to intervene at all. It proposes DICE, which first selects an explicit pedagogical action and uses Intervention Value, a counterfactual metric comparing each action to non-intervention, to show many interventions add little or no benefit. In simulated tutoring sessions, DICE reduces over-intervention to near zero and guides students to correct solutions in about 3-4 fewer turns than existing Socratic baselines.
In briefTransferable Attacks on Closed-Source Multimodal LLMs Using Feature Alignment and Visual Augmentation
SummaryThis paper proposes IAU-FOA, a method for crafting targeted adversarial examples that transfer to closed-source multimodal large language models. It aligns samples at a global level with a cosine-based objective and at a local level by clustering image patch tokens and matching them via unbalanced optimal transport. A visual-invariance augmentation simulates exposure, contrast, illumination, and color-temperature changes. The authors report that experiments on open-source and closed-source MLLMs show it outperforms existing transferable attack methods.
WhyExtends global alignment with patch-level optimal transport, confidence-adaptive unbalanced transport, and visual-invariance augmentation; evaluated broadly across open and closed-source MLLMs.
In briefAgent Behavior as Code: Efficient and Robust LLM Agents with Programmatic Specifications
SummaryThis paper introduces ABCAgent, an LLM agent whose runtime behavior is fully specified by a symbolic program, such as Python code with possible neural functions, which a foundation-model agent can edit. The authors evaluate it on six benchmarks, including two they built to test generalization to task variants. ABCAgent is reported to match or exceed a comparable neural agent in robustness and to be substantially cheaper and faster on parametric task families.
In briefAnnounced Breaks: Conformal Reliability Versus Frequency Calibration
SummaryMany structural breaks are announced in advance, but adaptive conformal methods only learn of them from their own errors. The paper resets the empirical quantile at the announced date and evaluates it with two audits: conformal reliability of the next interval and frequency calibration of a closed window, which can disagree. After a large synthetic break, the restarted quantile's coverage error at the 50th forecast is 0.035, versus 0.052 for the best of five adaptive baselines. Real breaks are mostly smaller, and results are mixed.
In briefDriftSR: One-Step Real-World Image Super-Resolution via Distribution Drifting
SummaryThis paper proposes DriftSR, a one-step real-world image super-resolution framework that uses pretrained diffusion priors by performing distribution drifting in a frozen intermediate representation space. It introduces Spatial Feature Drifting, which treats spatial features as distributional samples for denser supervision, and Structure-Modulated Guidance, which adjusts guidance based on local structural consistency with the low-quality input. Experiments on three real-world benchmarks show high-quality reconstruction with efficient one-step inference.
In briefMOXIE: Finding Multiple Alternative Explanations for Biomedical Image Classifiers
SummaryThe paper introduces MOXIE, an evolutionary framework that searches for subsets of image segments that keep a classifier's confidence while retaining as little of the image as possible. Instead of returning one explanation like LIME, it returns a Pareto front of alternative explanations ranging from compact to highly faithful. Experiments on BloodMNIST and HAM10000 show MOXIE achieves higher hypervolume than LIME on every image, and the authors argue that alternative explanations give a more complete view of model decisions.
In briefAgent-based multimodal emotion reasoning that verifies evidence and fixes prompts
SummaryThis paper proposes a traceable agentic framework for multimodal emotion reasoning in conversation, where visual, acoustic, and lexical agents post cited observations that an integration agent combines into an emotion prediction. For incorrect predictions, deterministic checks and an entailment check locate the responsible component, and targeted prompt patches are created and validated on a held-out set without updating model weights. On the MER2025 dataset, the adapted prompts improve F-score by 0.6 to 2.2 points across four MLLMs, including a rise from 52.3 to 54.5 on Qwen2.5-VL-7B.
In briefSeparating decision time from decision quality in real-time gaps of distilled deciders
SummaryThe paper argues that evaluating real-time agents with paused worlds or time rounded to whole ticks can hide the effect of decision latency. The authors split a learned decider's gap to a zero-latency teacher into a time component and a quality component, testing this in ViZDoom and a conveyor simulator. Results differ across hosts, and retraining on teacher-labelled states partly improved the quality component.
In briefExStereo: Adding 3D Stereo Vision to 2D Robot Action Models
SummaryThis paper introduces ExStereo, a stereo module that adds 3D perception to pre-trained 2D Vision-Language-Action models used for robot control. It reconstructs scene geometry from stereo image pairs and feeds multi-view stereo representations to the action model through a cross-attention mechanism. The authors fine-tuned two public VLAs and tested them in simulation and on a real bimanual PiPER robot. The models fine-tuned with ExStereo consistently outperformed baselines in both settings.
I-Chun Arthur Liu, Jason Chen, Gaurav S. Sukhatme et al.
In briefHow People Challenge Racial Stereotypes Online: Detecting Counter-Stories Across Reddit Communities
SummaryThis paper studies counter-storytelling, a way people challenge dominant narratives that has received little attention in computational social science. The authors propose a three-dimensional taxonomy grounded in narratology and Critical Race Theory, plus a multi-stage pipeline that finds pairs of racial stereotypes and counter-stories in Reddit discussions. They annotated 25,549 posts across 615 communities and identified 1,312 counter-stories. The analysis finds that speaker identity and post context shape how counter-stories are told.
Uma Sushmitha Gunturi, Jimin Mun, Maarten Sap et al.
In briefMulti-Agent Spectrum Sharing: How Multiple Cognitive Radars Learn to Share Limited Wireless Spectrum
SummaryThis paper studies how multiple cognitive radars can use machine learning to share a fixed 100 MHz band with other radio users without interfering with each other. Each device independently chooses where and how widely to transmit, based on real or simulated signal activity that reveals which frequencies are in use. Experiments show the proposed meta-learning approach achieves the highest average reward among the main baselines across five multi-agent benchmarks, while keeping collision rates low and transmission behavior stable.
In briefUncertainty in Knowledge Graph Embedding: Methods for Measuring and Managing Reliability in Knowledge Graph Reasoning
SummaryThis thesis examines three sources of uncertainty in knowledge graph embedding (KGE): input knowledge uncertainty, algorithmic uncertainty from training randomness, and predictive uncertainty about output reliability. It proposes a voting-based aggregation framework to reduce training instability and adapts conformal prediction to KGE to build answer sets with distribution-free coverage guarantees. It also develops statistically valid prediction intervals for confidence-scored triples and an embedding-based approximation for probabilistic reasoning over statistical ontologies.
In briefAID: A framework for modeling the dynamics of AI inference infrastructure
SummaryThe paper introduces AID (AI Infrastructure Dynamics), a framework for describing learning problems across coupled physical, computational, networking, and serving processes in AI inference systems. It supports structured variable-size state, asynchronous observations, and demand that responds to service. The authors separate representations used for prediction under an existing policy from those preserving outcomes under changed actions, and give a lower bound on prediction error plus a sufficient condition for exact controlled state reduction. A validation protocol for cache representations and related factors is also described.
In briefEvent cameras for monitoring melt pools in metal 3D printing, with a public benchmark and cross-machine transfer tests
SummaryThe paper introduces SynAM-E, the first public simulated event-camera benchmark for monitoring melt pools in metal additive manufacturing, built from 85 physics-calibrated event shards across 15 sources and 8 institutions. On a single-machine case, event-based spatial monitoring matched dense-frame accuracy (0.874 versus 0.863 macro-F1). On the NIST testbed, an event-rate counter recovered a 528.7 Hz intensity oscillation while requiring about 380 times less sensor readout than the frame stream.
Mohamad Yazan Sadoun, Sarah Sharif, Yingtao Liu et al.
In briefMedImageOSWorld: A Benchmark for GUI Agents Operating Medical Imaging Consoles
SummaryThis paper introduces MedImageOSWorld, a benchmark testing whether GUI agents can operate simulated CT, MR, and ultrasound imaging consoles using screenshots and mouse-and-keyboard actions across seven capability levels. Evaluation separates workflow completion from acquisition outcomes using hidden anatomical ground truth. Across eleven open-weight agents, success rates range from 3.0 to 25.0, with agents often completing workflows but rarely acquiring the intended anatomy, and performance collapsing for millimetre-level planning and closed-loop control. Two proprietary agents reach about 34 success, mainly exceeding the best open-weight agent in acquisition quality.
In briefTIMBRE: Helping time series forecasters use event reports and past reactions
SummaryThe paper proposes TIMBRE, a method that turns event reports and historical responses into adjustments to numerical forecasts, using source-aware representation, state-conditioned response transfer, and reliability-guided fusion. A separate readout adjusts prediction interval widths while keeping the median fixed. In a single-seed, one-epoch development study over 13 tasks, TIMBRE beat ordinary fusion on eight tasks but beat native Chronos-2 on only two. Disabling learned response transfer markedly increased error on BTC and AULF, suggesting the gains depend on that component rather than a general forecasting advantage.
In briefKnowing the Store: What Memory Backends Must Record Before Agents Read
SummaryThe paper examines whether an agent with long-term memory can recognize, before retrieval, that a record such as a cancelled plan should no longer be used, scoring this as metamemory monitoring. It tests 148 questions across three versions of one store differing in a single attribute, with readers seeing only lifecycle state counts and attribute names. On long name lists, some stronger models beat a cosine-similarity lookup, while on short lists the advantage is not reliable. The tested backends omit or misstate this information, so what the store records limited every reader's performance.
In briefPressure and context make AI models take shortcuts, study finds
SummaryThis paper applies criminological theories such as self-control, general strain, neutralization, and routine activity theory to reward hacking, where generative AI models reach measured goals through unsanctioned means. In a study of 2,310 conversations across seven models, pressure and conversational cues substantially changed how often models chose shortcuts. In a preregistered coding study, several models cheated on tasks whose tests contradicted their specifications, while some never did. The authors conclude that a model's stated refusal does not guarantee compliant behavior.
In briefDo more modalities always help? A geometric view of missing-modality robustness
SummaryMultimodal models can perform worse than unimodal models when one modality is missing at inference time. The authors find that multimodal training rotates principal parameter subspaces, especially in cross-modal layers, which is associated with this degradation. They propose Geodesic Unlearning (GU), a lightweight method that rotates the principal input subspace toward a unimodal reference, and report improved missing-modality performance while preserving full-modality accuracy.
In briefNot all answers can be swayed by context: how large language models' inference changes under contextual influence
SummaryThe paper introduces a theoretical framework for analyzing how context influences large language models through their internal inference dynamics, beyond whether the output answer changes. It finds that under repeated contextual assertions, predictive representations converge to stable, query-dependent states that limit whether context can alter a prediction. The authors show that repeated assertions do not act as accumulating evidence, and report empirical results that align with their theoretical predictions.
In briefActive-DiNTS: Differentiable Network Search with Active Learning for 3D Medical Segmentation
SummaryThis paper presents Active-DiNTS, which embeds pool-based active learning into a bi-level differentiable topology search so that architecture discovery and labeling are performed jointly. At each round, unlabeled MRI volumes are ranked by an uncertainty signal and only top-ranked ones are sent for annotation. On the MSD Task01 BrainTumour benchmark, it reports higher Dice than DiNTS, C2FNAS, and nnU-Net on a single GPU using fewer labeled volumes, with the fastest search variant finishing in under 0.25 GPU-days.
Gean Trindade Pereira, Thierry Urruty, Muriel Visani et al.
In briefKALEIDO: Adapting a Vision Model to Time-Series Forecasting by Folding Series into Images
SummaryThe paper presents KALEIDO, which forecasts time series by rendering them as images and using an ImageNet-pretrained masked autoencoder to inpaint the missing part. It detects dominant periods, renders several rule-generated fold geometries, combines their inpaintings with a validation-fitted gate, and fuses the result with the zero-shot output. Training only LayerNorm (0.05% of parameters), it reduces MSE by 13% against the published zero-shot baseline on LTSF, or 6.6% when frozen, and improves MASE by 7.4% and CRPS by 19.3% on GIFT-Eval.
In briefLollypop: A Reference Target for Verifying Camera–Motion Capture Calibration
SummaryThis paper presents Lollypop, a reference target that combines an ArUco fiducial with a motion-capture marker group so both refer to the same physical point, allowing independent checks of camera-to-mocap calibration after deployment. The method projects the mocap point into the image and measures its offset from the detected fiducial center. Experiments report sub-pixel nominal error, sensitivity to controlled extrinsic perturbations, and rising error during an illustrative mixed-handling sequence.
In briefRepTC: representation-aware optimization for efficient traffic classification on edge IoT devices
SummaryThis paper presents RepTC, a method for session-level classification of encrypted network traffic on resource-constrained edge IoT devices. It jointly optimizes network architecture, session length, and header preprocessing under microcontroller-class limits on memory, model size, and computation. Tests on several embedded boards report accuracies of 96.21% on ISCX VPN-nonVPN, 99.62% on USTC-TFC2016, and 99.97% on Edge-IIoTset, with model size and computation reduced by up to three orders of magnitude versus state-of-the-art methods.
Adel Chehade, Edoardo Ragusa, Paolo Gastaldo et al.
In briefDASH: Fast Valid Counterfactual Explanations for Deep Networks
SummaryThis paper introduces DASH, a batched heuristic search method that quickly finds counterfactual explanations for deep neural networks that are close to the original input, valid, and actionable, under ℓ1, ℓ2, and ℓ∞ norms. It uses directional Lipschitz bounds and local affine models to generate anchors, then ranks and expands promising regions using batched network evaluations. On 9,000 tabular query-norm cases, DASH returned a valid counterfactual within 5% of the best heuristic-observed distance in 94.6% of cases, with a median CPU runtime of 0.061 s.
Shraman Pal, Gabriel Medeiros, Clayton Escouper das Chagas et al.
In briefSuper-resolution in a Frozen Vision-Foundation Latent Space
SummaryThe paper argues that real-world super-resolution should operate in a latent space suited to semantics, not the reconstruction-oriented latent or pixel spaces used by prior methods. It uses 23 fused layers of a frozen DINOv3-L as the substrate and trains a 415M-parameter decoder with reconstruction and adversarial losses to map degraded embeddings back to the image manifold and decode them in one pass. The resulting model, RAESR, reportedly achieves the best fidelity-perception trade-off against adversarial and diffusion-based baselines on four benchmarks, running at 37 ms per 512×512 image on an H20 GPU. The authors report that using a VAE latent under the same recipe performs worse on every metric.
WhyProposes super-resolution in a frozen vision-foundation latent space, validated against VAE latents and state-of-the-art restorers on several real-world benchmarks.
In briefTesting whether patient world models can be trusted before making stronger claims
SummaryThis paper examines patient world models used for longitudinal prediction and intervention reasoning, arguing that predictive generalization is a prerequisite distinct from causal or clinical validity. The authors test short-horizon gait biomarker forecasting on PhysioNet GaitPDB (165 participants, 306 recordings) and progressively remove data leakage, showing performance worsens for unseen patients and under partial observation. They explicitly do not claim a longitudinal or intervention-aware simulator, and instead argue for a prerequisite evaluation stack covering patient separation, shift, missingness, and uncertainty.
In briefARISE: Adaptive AI reasoning with image self-checking for explainable IBD diagnosis
SummaryThis paper presents ARISE, a framework that treats few-shot medical image analysis as a step-by-step AI agent workflow with five stages: hypothesis generation, image-based evidence summary, evidence-based refinement, symbolic verification, and final diagnosis. It is applied to inflammatory bowel disease assessment using capsule endoscopy images for Crohn's disease and B-mode ultrasound data for ulcerative colitis, across two independent patient cohorts. According to the authors, ARISE improves diagnostic performance over baseline vision-language models while showing where its reasoning succeeds or fails, offering a more interpretable basis for clinical decision support.
In briefLearning task and motion plans from real demonstrations using hybrid flow matching
SummaryThis paper presents a hybrid flow matching planner that uses one network to generate both a symbolic task plan, via discrete flow matching, and a motion trajectory, via continuous flow matching, for long-horizon mobile manipulation. It learns from a small set of demonstrations by reusing intermediate-resumed demonstrations and treating same-kind objects as interchangeable, and it executes plans in closed loop with replanning after each action. In kinematic simulation, replanning raised task completion from 40% to 53%, and the planner was validated on a real mobile manipulator.
Zuleika Redondo Garcia, Andreu Matoses Gimenez, Javier Alonso-Mora
In briefFlow-policy skill actions for world-model planning: learned and symbolic abstractions
SummaryThe paper studies how robots can plan long tasks more efficiently by using skill-level actions instead of control-rate actions. A flow-matching policy trained on segmented demonstrations turns each complete skill into a single world-model transition, and four action abstractions are proposed, from a compressed seed to a symbolic label. On simulated block rearrangement tasks requiring up to 14 skills, the symbolic label succeeds on over 90% of tasks needing up to six skills, while an unlabeled learned object-centric code matches it on single-skill tasks.
Andreu Matoses Gimenez, Andrei-Carlo Papuc, Chris Pek et al.
In briefVariance-optimal control variates for learning from black-box feedback
SummaryThe paper studies gradient estimation when models learn from black-box oracles such as humans, solvers, or external tools. It observes that even an exact action-value function can be far from variance-optimal, because an action can still affect the rest of the gradient estimator through shared parameters. The authors prove that the residual variance decomposes exactly across actions without cross terms, yielding a closed-form variance-minimizing correction for neural-network parameters that can be computed by projection. Experiments report that this correction consistently reduces the remaining variance and improves learning across all tasks.
In briefEasyClassifier: A No-Code Tool for Honest, Reproducible Machine-Learning Classification
SummaryThis paper presents EasyClassifier, an open-source Python package that lets researchers who do not program run classification analyses on CSV or Excel files by answering plain-language questions. It enforces correct procedures by default, learning preprocessing within each training fold and giving the selected classifier a separate final score via nested cross-validation or an untouched test set. On ten public datasets and a random-label control, its reported scores showed a smaller optimistic bias than common practice.
In briefWhen to stop a self-improving LLM system, and which version to keep
SummaryThis paper studies when self-evolving large language model systems should stop updating their prompts or skills, and which artifact to output at that point. The authors use an online sequential test with per-item paired evaluation outcomes to detect restarts, and change-point estimation to choose an earlier artifact. The method is plug-and-play and requires no changes to the underlying algorithms. Across two frameworks, three model families and five benchmarks, it substantially cuts computation while keeping unseen-test performance comparable.
In briefFoSeRL: Formal certification of how much a trained RL policy's return can be hurt by sparse action attacks
SummaryThis paper introduces FoSeRL, a framework for certifying deployed reinforcement learning policies against precommitted, temporally sparse action attacks without changing, smoothing, or retraining the policy. It requires only a resettable simulator supporting shared randomness, and gives guarantees that an attacked episode loses no more than a set amount of return relative to the unattacked episode, at a target probability and confidence level. Evaluated on six stochastic continuous-control environments and TD3, SAC, and PPO policies, it reports larger certified budgets than policy smoothing and robustness differences among policies with similar nominal performance.
WhyProposes a policy-unchanged certification framework using barrier conditions for probabilistic guarantees, validated across six environments and three policy families, beating smoothing baselines.
Sara Taheri, Deep Kumar Ganguly, Jan K\v{r}et\'insk\'y et al.
In briefMore Value per Key: Faster LLM Decoding via Asymmetric Sparse Attention
SummaryThe paper observes that in many sparse attention methods, the probability-value multiplication becomes negligible, so the bottleneck shifts to the query-key step. It proposes reducing key heads while keeping more value heads, introducing SAGA and the Atop-N sparse attention method. Reported results include end-to-end decoding speedups above 2x over a full-attention GQA baseline at long contexts, with models trained from scratch nearly matching comparable GQA quality, plus a fine-tuning method to convert pretrained models.
In briefVariance-aware fine-grained gap-dependent bounds for online RL
SummaryThis paper studies model-free online reinforcement learning for episodic tabular Markov decision processes, focusing on gap-dependent regret and policy switching cost. It establishes the first fine-grained gap-dependent regret upper bound for UCB-Bernstein+, a variance-based exploration algorithm. By adding stage-wise policy updates, it achieves the best-known gap-dependent local switching cost, and it also improves worst-case guarantees over the original UCB-Bernstein. Numerical experiments show favorable empirical regret and local switching cost.
WhyGives the first fine-grained gap-dependent regret bound for variance-based model-free RL, improves switching cost, and supports it with experiments.
In briefDynamic Routing: A New Dimension for LLM Test-time Versatility
SummaryThis paper studies whether large language models can adapt to new problems at test time by using dynamic routing programs that skip or repeat some layers, without any gradient updates. A selection strategy based on the probability assigned to demonstration labels yields consistent gains on MMLU and large improvements on far out-of-distribution tasks such as ARC-AGI. The authors also report that routing complements chain-of-thought reasoning, while noting that confidence-based selection leaves much potential untapped.
Michal \v{S}tef\'anik, Marek Kadl\v{c}\'ik, Josef Kucha\v{r} et al.
In briefAgent-driven discovery of blood biomarkers using a released scorer trained on private health records
SummaryThe paper trains a graph attention network inside a data boundary to predict case-control AUC for candidate complete blood count expressions across 13 immune-mediated diseases, releasing only the trained weights. This lets language model agents run a propose-score-refine loop without accessing any patient data. In external validation, agent-discovered expressions improved on literature-seeded starting points by a median of 4.18 AUC percentage points.
WhyDistills private health records into a releasable scorer enabling agents to discover blood biomarkers without patient data; externally validated across three cohorts.
Seffi Cohen, Liat Antwarg Friedman, Amir Anisman et al.
In briefFormal Framework for Judging Speech Acts: Truth, Lies, and Moral Dilemmas
SummaryThe paper presents a logical framework that evaluates the ethics of speech-act utterances based on agents' beliefs, implemented in Answer Set Programming. It assesses utterances under deontologism, consequentialism, and principialism. The framework is illustrated with Sartre's The Wall (1939), where lying leads alternately to rescue and to death. The authors state that accommodating a new moral situation requires adjusting parameters rather than rules.
Benjamin Icard, Gauvain Bourgne, Jeanne Bonnaventure et al.
In briefExact fast batch simulation for tabular reinforcement learning: a method that speeds up simulation by avoiding generating individual trajectories.
SummaryThis paper proposes an exact fast-simulation framework for finite-horizon tabular Markov decision processes, aimed at cases where only aggregate statistics are needed. Direct batch simulation represents a batch by its aggregate Markov flow, obtained by trajectory aggregation or forward Markov-flow sampling, reducing the simulator's dependence on batch size m from O(m) to O(1). Adaptive batch simulation, used when batch length depends on data, applies exact multivariate-hypergeometric splitting to refine the flow while preserving the conditional law, reducing the dependence to O(log m). The framework applies to simulator-based, offline, and online batch or stage-based RL.
In briefBoT-Feedback: Explainable Human Action Feedback Grounded in Biomechanical Evidence
SummaryThis paper proposes BoT-Feedback, a framework that helps multimodal language models give coaching feedback on human actions by grounding their reasoning in biomechanical evidence. Its four-stage Biomechanics of Thought process identifies the action, locates key body regions, compares expert and student biomechanics, and produces interpretable feedback. The authors also introduce a data parser and a paired video benchmark called BiomAF. Experiments on twelve models report better feedback quality, fewer biomechanical hallucinations, and an average expert score rising from 2.07 to 2.95.
In briefLie Rarely, Lie Big: Stealthy Insider Attacks on LLM Robot Teams
SummaryThis paper studies how one compromised robot in a multi-robot survey team, whose planning and mutual trust are handled by LLM agents, can corrupt the shared map. The authors derive a lower bound on the probability that the adversary's reports get verified and an upper bound on the map error a stealthy adversary can cause. In experiments with LLM-agent honest robots, both bounds hold at the budget the attack actually spent.
In briefThyCLIPNet: A Lightweight Segmentation Model for Thyroid Nodules Guided by Biomedical Vision Features
SummaryThis paper presents ThyCLIPNet, a lightweight encoder-decoder model for segmenting thyroid nodules in ultrasound images. It adds global biomedical semantic guidance from the BiomedCLIP vision encoder to a multiscale CNN pipeline, without using text prompts. Reported Dice scores are 96.22%, 87.58%, 84.73%, and 80.70% on TG3K, TN3K, DDTI, and PKTN, with 8.55M parameters and 22.99G FLOPs.
In briefRobot Learning with Visual Predicted Force: Estimating Contact Force from Gripper Deformation to Guide Actions
SummaryThis paper presents a manipulation approach that predicts contact force visually from the deformation of a compliant Fin Ray gripper, without force or tactile sensors at deployment. The authors train a visual force estimator on calibration data and use it to annotate demonstrations with force estimates. They then train an action-force proposal policy that samples candidate actions and their predicted forces at test time, executing the one closest to the demonstrated target force. The method is evaluated on berry picking, empty-can grasping, in-hand reorientation, and plug insertion.
In briefToward a Locally Deployable Agentic Co-Scientist: Small-Model Planning for Early-Stage Drug Discovery
SummaryThe paper presents a lightweight framework in which a compact, locally deployable language model plans calls to 18 modular drug-discovery tools, sharing molecular records through a unified schema. The authors built 1,263 query-plan pairs and fine-tuned three small models with LoRA. Under a stricter workflow-grouped split, complete sequence exact match reached only 0.452 to 0.548, indicating that compositional generalization remains a key challenge.
Tian Liang, Jiayu Chang, Alejandro F. Frangi et al.
In briefPyINE: Scalable Oversight of AI Reasoning via Code Execution
SummaryPyINE is a framework that uses instrumented Python programs as a verifiable execution base, where execution traces provide authoritative labels for outcomes. The authors built PyINE-v1 from nearly one million execution traces and over 500,000 LLM-generated code variants, and trained a shortcut-following model. They evaluated several overseers and found that dataset-level performance can hide weak coverage of rare, shortcut-driven failures, with cheap learned overseers often missing them.
Pierre-Luc St-Charles, Alessandro Palmas, Damiano Fornasiere et al.
In briefProbabilistic pedestrian forecasts on a handheld phone, with height-drift analysis and evaluation without ground truth
SummaryThe paper builds a system that uses a phone's monocular camera and visual-inertial odometry to forecast where nearby pedestrians will be over the next few seconds, shown as probability heat maps on the ground. A small U-Net trained on bird's-eye and first-person trajectory datasets outperforms a fitted constant-velocity Gaussian baseline on test splits. For handheld video without ground truth, the authors score forecasts against the tracker's later raw measurements, finding a smaller but still positive advantage on held-out clips. The authors also note the limits of such self-consistency scores.
In briefA conceptual framework for assessing knowledge graph quality in the LLM era
SummaryThe paper reviews literature on quality assessment of data, linked data, knowledge graphs, and the intersection of knowledge graphs with large language models. The authors extend existing frameworks into a catalog of 24 quality dimensions and various quality metrics. They also propose a general-purpose, customizable framework for assessing knowledge graph quality.
In briefFinding and removing redundant computation in a video transformer for collision prediction
SummaryThe paper studies a video transformer (VideoMAEv2-Base) used to anticipate worker-equipment collisions and asks where its computation is redundant. It finds that tokens are much more similar across time than across space, and exploits this by merging temporal tokens. On the Nexar dataset, this cuts backbone computation from 356.99 to 178.50 GFLOPs and latency from 12.69 to 7.21 ms per clip, with mAP changing only from 0.7478 to 0.7443.
In briefVerb-ICL: Rethinking In-Context Learning for Structured Prediction
SummaryThis paper proposes Verb-ICL, a selective annotation framework for in-context learning in structured prediction with LLMs. It first selects representative examples using token-level coverage, then generates actionable error feedback that encodes task-specific annotation guidelines into the demonstrations. The authors evaluate it on six datasets spanning information extraction and semantic parsing, reporting gains over selective annotation baselines in low-resource settings.
In briefLatent Lagrangian neural networks for reduced-order modeling of forced nonlinear systems
SummaryThe paper proposes a latent Lagrangian framework for reduced-order modeling of forced nonlinear dynamical systems. It learns latent coordinates with two neural networks for latent kinetic and potential energies, using force supervision to avoid an ODE solver during training. Physical consistency in the latent space is enforced via the principle of virtual work. The model captures subtle dynamics from nonlinearity and non-convex potential energy and generalizes to unseen forces and initial conditions.
In briefNAMVIS: Diffusion-free multi-view image synthesis via next-scale autoregression
SummaryThis paper presents NAMVIS, a diffusion-free method for sparse-view novel view synthesis that generates multi-view images as geometry-conditioned next-scale autoregression. It predicts discrete visual tokens over a small number of coarse-to-fine scales, using multi-scale projective pose encoding and geometry-aware cross-attention to tie generation to camera geometry. The authors report better PSNR, SSIM, and LPIPS than diffusion baselines on Objaverse, GSO, and OmniObject3D, while running over 3 times faster.
Ramil Khafizov, Ilya Statsenko, Ruslan Rakhimov et al.
In briefKnossos and Ariadne: Benchmarking and learning full diagram topology extraction with vision-language models
SummaryThis paper studies extracting all entities and their complete relations from structural diagrams and turning them into graphs. The authors introduce Knossos, a benchmark of 19,200 diagrams across six domains, and Ariadne, a framework that splits the task into node inventory extraction and source-conditioned edge prediction. Experiments report that training on Knossos improves smaller open-source vision-language models, and that Ariadne achieves the highest average Edge F1 among the evaluated methods on Knossos.
In briefWNet: Discrete Wavelet Transform for Efficient Token Mixing
SummaryWNet is a Transformer encoder that replaces self-attention with token mixing based on the discrete wavelet transform, avoiding the quadratic cost of all-pairs comparison. It proposes several attention-free mixers and analyzes how filter length affects which tokens can interact. Under a controlled setup against size-matched BERT and FNet baselines, the token-gated mixer trains 2.7 times faster than attention at 4,096 tokens.
In briefTraining AI Assistants to Ask Better Questions Under Limited Interaction
SummaryAI assistants often receive requests missing information needed for a good outcome, and must either guess or ask questions. The paper frames this as a value-of-information problem, training assistants to acquire the information whose absence causes the greatest avoidable loss in user utility. It applies reinforcement learning with multi-turn simulated users in image generation. In a preregistered study with 456 sessions and 76 participants, users better matched reference images with fewer questions, less time, and lower cost.
In briefAegisFlow: A Multi-Agent AI Framework for Automatically Repairing and Self-Healing Fragile Data Pipelines
SummaryThe paper presents AegisFlow, a multi-agent framework in which a Watchdog agent collects runtime telemetry and a Repair agent uses large language models to generate, test, and deploy code patches. Patches are verified in digital twin environments using a method called Parallel Shadow Patching. Across five common failure scenarios, the authors report reduced mean time to repair from 170 minutes to 3.2 minutes and a 92 percent patch success rate.
In briefAutomatically removing logging code with AI coding agents: how well does it work?
SummaryThis paper studies how developers remove logging code and tests whether coding agents can perform this task. The authors extracted and manually validated logging removal cases from Python and Java repositories, identifying 10 removal patterns and 11 removal reasons, and built LogRem, a dataset of 387 real-world cases. Across four coding agents, 95.6% to 100.0% of outputs passed validity checks, but only 11.1% to 19.6% matched the removals made in the real-world changes.
In briefGPU-accelerated algorithm for solving discrete optimal transport problems
SummaryThis paper presents a GPU-accelerated Bregman Douglas-Rachford splitting algorithm (BDRS) for discrete optimal transport. It supports three input formats: an explicit cost matrix, a point cloud with ground costs, and a separable cost on a regular grid, with hardware-aware equivalent designs for each. The implementations are benchmarked against eight GPU baseline solvers on the same device. The authors report state-of-the-art performance for their respective input formats.
In briefEVFormer: Fusing Egocentric Vision and EMG for Bimanual Hand Pose Estimation
SummaryEVFormer combines the current RGB frame with the preceding 200 ms of bilateral wrist surface EMG to estimate 44 finger and wrist joint angles. It fuses the two modalities through sequential bidirectional cross-attention and feature-wise gated fusion. In a single-participant feasibility study, it achieved a mean absolute error of 11.482 degrees, lower than the baselines. Generalization across participants and real-world conditions remains to be evaluated.
In briefOrganizing Evidence from Agent Trajectories: Methods, Fragments, and Residual Risk
SummaryThe paper proposes a framework that unifies different methods for assessing language-model agents, such as rule checkers, prefix monitors, and execution gates, and classifies what each one establishes. Tested on 456 released τ²-bench telecom trajectories, the benchmark oracle flagged 231 runs, while the combined methods flagged 407. The 49 unflagged runs and unmet obligations form the residual that remains unchecked.
In briefLocalized operator learning using adaptive partition-of-unity mixture-of-experts networks
SummaryThis paper proposes a partition-of-unity (POU) mixture-of-experts framework for localized operator learning on PDEs with sharp interfaces, heterogeneous coefficients, and multiscale structures. The main method, HiRefPOU, is a residual-style hierarchical POU architecture for DeepONets that uses nested parent-child partitions while keeping global continuity. On heterogeneous Darcy and reaction-diffusion benchmarks it achieves substantially lower error than global DeepONet and static POU-MoE baselines, and the learned partitions are interpretable and align with interfaces and rapid solution variation.
Madison Cooley, Ramansh Sharma, Shandian Zhe et al.
In briefVoCa: Speech-canvas interaction for voice agents
SummaryThe paper studies how people combine speech with writing and sketching, and derives a design space for voice agents that use a canvas alongside speech. The authors built VoCa, which coordinates speech with creating and annotating visual objects and guiding attention. A five-day deployment with 18 participants examined usability and use patterns. Participants saw opportunities in learning, work, and daily life, but also found it challenging to follow and influence what the agent says and shows.
In briefNeurodiversity-aware multimodal affective computing for neurodevelopmental assessment
SummaryThis conceptual paper proposes a testable architecture for neurodevelopmental assessment that combines multimodal signals while not treating population-relative deviation as sufficient evidence of adverse functioning. It reports no new participant study or trained predictive system and claims no predictive or clinical superiority. The framework separates observable signals, modality quality, context, population-relative information, and within-person references, and specifies auditable requirements such as target validity, calibration, subgroup evaluation, and privacy. Autism is the principal motivating evidence base.
In briefLearning Discriminative Geometry for Drifting Models
SummaryDrifting Models move iterative distribution refinement from inference into training, enabling effective one-step generation. The paper finds that performance on complex image datasets depends strongly on the representation used to build the drifting field: pixel-space drifting performs poorly, while pretrained feature spaces substantially improve sample quality. The authors propose persistent representation learning, which continuously learns a more discriminative representation geometry as the generator evolves, reducing FID by roughly 82%–95% over the original pixel-space Drifting Models across multiple datasets without pretrained encoders.
In briefDecoupling and aligning features for privacy-preserving federated medical image segmentation
SummaryThis paper proposes FedBCS+, a federated learning method that trains a shared medical segmentation model without pooling patient data across institutions. It separates content from style using frequency-domain recalibration and splits multi-level features into semantic and structural prototypes that are aligned separately. Client contributions are reweighted according to how far their prototypes diverge from the global consensus. Across five heterogeneous medical segmentation benchmarks, it reports the highest mean Dice among the compared methods.
In briefScore-Calibrated Flow: a sampling method for unnormalized densities, applied to generative online reinforcement learning
SummaryThis paper proposes Score-Calibrated Flow (SCF), a method for training generative models to sample from unnormalized densities without importance sampling or backpropagation through the sampling trajectory. It learns the target flow by enforcing self-consistency conditions derived from the target score and flow matching structure. The method is applied to online reinforcement learning, where the critic gradient supplies the target score for actor training. Experiments on RL benchmarks reportedly match or improve on generative-policy baselines while substantially reducing training time.
In briefCoarse finite-difference spectra guide neural eigenvalue solvers
SummaryThis paper studies operator eigenvalue problems and proposes using a cheap coarse finite difference method to estimate approximate eigenvalues, which are then used as fixed shifts when training a neural solver. It also introduces SIPMNN, a more stable training procedure for higher-dimensional problems. On five test problems at dimension 10, the combined approach is reported to be more accurate overall than the tested fully neural alternatives, using eight to ten times fewer iterations.
In briefRAGStress: A controlled benchmark testing retrieval-augmented generation when the knowledge base is corrupted
SummaryThis paper introduces RAGStress, a benchmark for stress-testing retrieval-augmented generation (RAG) systems when their knowledge base is systematically corrupted. It pairs four corruption types (factual corruption, numeric typos, relevance poisoning, contradiction injection) with three severity levels, built from 57 MMLU subjects and 182,546 documents. Across 52,500 evaluations, the authors find that clean retrieval can mask robustness differences and that semantic-fidelity corruptions are more harmful than signal-utility perturbations.<br>
WhyProposes a controlled benchmark for RAG under KB corruption, with 52,500 evaluations supporting findings useful to the subfield.
In briefSepRQ: Self-supervised speech mixture learning without masking, using multi-scale source separation
SummaryThe paper introduces SepRQ, an open-source self-supervised framework for speech representation learning aimed at multi-speaker audio. It replaces masked prediction with a pseudo-source-separation objective over frozen random-projection codebooks. The authors report state-of-the-art results on speaker diarization and speech separation in the SUPERB benchmark, with 85.68M inference parameters.
WhyReplaces masked prediction with mask-free multiresolution pseudo-separation, achieving SUPERB state-of-the-art diarization and separation with broad benchmark evidence including three-speaker mixtures
In briefErrors in LLM Question Answering over Imperfect Tables
SummaryThis paper studies how large language models answer questions over tables that contain errors. Using human-reviewed examples from RADAR-T, the authors run controlled experiments with three LLMs, varying row order and comparing original, error-marked, and repaired tables. They find that row order affects whether models discover errors, and that providing error locations alone does not ensure accurate answers. The authors also propose a workflow called GBDI, which improves QA accuracy over a code-agent baseline across five LLMs.
In briefWhen Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning
SummaryThis paper examines why multi-agent debate often fails to outperform simple majority voting. The authors argue that debate needs proposal supply to surface a correct answer and a readout that can identify that answer when voting misses it. They introduce Latent Verification Debate (LVD) and build agent societies from neural-thicket agents using coverage objectives, which increased aggregate accuracy across two backbones and matched-budget reasoning benchmarks.
In briefSteering Speech-Language Models Without Training Using Contrastive Activation Addition
SummaryThe paper proposes a training-free Contrastive Activation Addition (CAA) method that derives steering vectors for speech tasks, such as transcription, from a small number of labeled utterances. Adding these vectors to the model's representations at inference time helps enforce the targeted task. Combined with prompting, the vectors consistently improve performance over prompting alone on most evaluated tasks, including Automatic Speech Recognition and Emotion Recognition, and transfer to out-of-domain data. The authors also show that script-normalization directions can enforce a target script for a specific language.
S\'everin Baroudi, Yanis Labrak, Pierfrancesco Melucci et al.
In briefCOMPASS: An automated pipeline for measuring DNA damage in comet assay images
SummaryThe comet assay measures DNA damage in individual cells, but image analysis is often manual or semi-automated, which is laborious and hard to reproduce. COMPASS combines deep learning-based comet segmentation, damage measurement, and automated comet selection to produce standardized DNA damage measurements. It also allows optional manual review for transparency. The tool is implemented in Python and freely available on GitHub.
Jack Roberts, Canya Lu, Alexis Michelle Lawson et al.
In briefExtracting persona subspaces via iterative nullspace projection for controlling LLM behavior
SummaryThe paper introduces PaSS, an inference-time method that models LLM personas as multi-dimensional subspaces in latent space. It extracts these subspaces through Iterative Nullspace Projection (INLP) and uses them to amplify or suppress persona traits without retraining. Experiments on MATH-500, TinyAlpaca, GSM8K and IFEval report stronger modulation than single-direction methods while preserving content fidelity.
In briefMASBench: A Benchmark for LLM Agent Teamwork Under Limited Information
SummaryThis paper introduces MASBench, a benchmark for evaluating teams of LLM-based agents that each see only part of the environment. Tasks are organized into Reasoning, Scheduling, and Game categories to test three collaboration mechanisms: Protocol, Memory, and Routing. The benchmark uses deterministic metrics, including performance score, communication cost, and cost effectiveness. Experiments across various LLM backbones and mechanism settings offer guidance for designing multi-agent systems, and the code is publicly available.
In briefAsking the crowd the right question: bias-cancelling weights for federated learning
SummaryThe paper treats the weights in federated learning's weighted sum of client risks as a wisdom-of-crowds mechanism, where clients are noisy views of one truth. It proposes CROWD, which reads client disagreement from the optimization trajectory at no extra cost and matches a Bayesian minimax lower bound in the same constant. On real scans split across sites with miscalibrated detectors, the method attains the oracle excess risk.
WhyFrames federated weighting as a wisdom-of-crowds mechanism with identifiability results and a matching minimax lower bound; validated on real scans, impact mostly within federated learning.
Ilya Kuruzov, Dmitrii Vishovan, Kirill Novoselov et al.
In briefA Tropical Geometry View of Forgetting: Per-Unit Projectors for Knowledge-Preserving Fine-Tuning
SummaryThis paper studies why fine-tuning language models on new text erodes existing abilities. It argues that replay-free methods such as Adam-NSCL use one shared subspace for all units, which is too coarse. Using tropical geometry of ReLU layers, the authors derive a gate-aware per-unit projector and show that exact protection costs each unit the rank of its own open tokens. Experiments on OPT-1.3b and OPT-6.7b report less forgetting than Adam-NSCL at matched budgets.
WhyTropical geometry yields a gate-aware per-unit projector with a budget-separation theorem; experiments on OPT models are moderate, impact limited to continual fine-tuning.
In briefPrinciples Guiding Actions: LLM Agents Learn Reusable Knowledge from Experience
SummaryThe paper proposes SAGA, a framework that lets LLM agents evolve from experience using external memory without changing model parameters. It gradually turns interaction trajectories into episodic descriptions, reusable procedures, and principles with explicit applicability conditions, then instantiates retrieved principles as task guidance to refine candidate actions. Experiments on ScienceWorld and ALFWorld show improved task performance, and ablation studies highlight the importance of contextual instantiation and action regulation.
In briefNeural Surrogates for Linear Radiation Transport: Lattice and Hohlraum Benchmarks
SummaryThis paper compares two parameter-matched neural surrogate models, Transolver and BSMS-MGN, for approximating final-time particle concentration in the two-dimensional linear radiation transport equation. The benchmarks are the Lattice and Hohlraum problems. An ablation study finds that inductive-bias preferences depend strongly on architecture, so design choices should be re-examined per model. The training recipe, data, and evaluation pipeline are released.
Carmelo Gonzales, Steffen Schotth\"ofer, and Cory D. Hauck
In briefFLASHSWIN: A memory-efficient attention method enabling large windows and dense tokens in Swin vision transformers
SummaryThe paper introduces FLASHSWIN, which replaces Swin transformer's windowed attention with a FlashAttention implementation that avoids materializing the full score matrix, cutting per-window memory from O(M^4) to O(M^2). To keep positional information, it uses window-local learnable 2D RoPE instead of Swin's relative-position bias. The authors report that training memory at a 32×32 window stays at about 12.4 GB, the same as at 8×8, and that the Tiny model reaches 84.1% on ImageNet-1K, 44.1 COCO box AP, and 47.28 ADE20K mIoU with dense tokens and wide windows.
Tushar Kataria, Gerald Sabin, Ponnuswamy Sadayappan et al.
In briefPermVLA: Factorization Order as a Regularizer for VLA Learning
SummaryThis paper studies how vision-language-action (VLA) robot policies learn action sequences. Standard training predicts actions in a fixed left-to-right order, even though the same expert trajectories can be factorized in many valid ways. The authors propose causally anchored permutation (CAP), an auxiliary objective that trains one shared policy to predict actions from different known subsets of an expert chunk, while deployment keeps left-to-right control. CAP consistently outperforms standard training on LIBERO, LIBERO-Plus, and cross-dataset CALVIN evaluations, and the approach is also extended to diffusion action generators.
In briefRETRACE: Reusing Past CI Fixes by Untangling Repair Histories
SummaryThis paper presents RETRACE, a framework that reconstructs problem-level repair experience from pull request histories where multiple CI failures, failed attempts, and reverted edits are mixed together. The recovered experience is stored at three abstraction levels, and for new failures the system retrieves relevant experience to guide repair. On CI-REPAIR-BENCH, with 565 PR-level repairs, RETRACE raised mini-SWE-agent Pass@1 from 19.6% to 31.9% with MiniMax-M2.5.
WhyProposes reconstructing problem-level repair experience from entangled histories; large Pass@1 gains on 565 PR repairs across models show solid evidence.
Rabeya Khatun Muna, Muhammad Ahasanuzzaman, Nakhla Rafi et al.
In briefWavePrune: One Rotation Period Is Often Enough for RoPE
SummaryThe paper proposes WavePrune, which limits each channel of Rotary Position Embedding (RoPE) to its first rotation period to reduce attention distractions caused by position aliasing. Without extra tuning, it raises the HELMET score on four of five tested models, and models pretrained from scratch reach lower validation loss at extrapolated lengths. Custom CUDA kernels give 1.15x prefill and 1.24x decoding speedups over FlashAttention-2 at 32K context.
In briefVerdicts Without Evidence Labels: Rejection Sampling vs. Label-Only Training for Recovering Evidence
SummaryThe paper asks how much supporting evidence a small language model can recover when post-trained only on recorded verdicts, with no human evidence annotations. On ContractNLI it compares label-only training with rejection sampling, measuring verdict accuracy and span F1 against held-out human evidence spans. Both methods improve evidence recovery, but the authors note that one seed on one corpus cannot establish which method is better.
Nishanth Nayakanti, Prasang Gupta, Ashutosh Bilthare et al.
In briefPath Laplacian positional encodings that capture edge direction in directed graph learning
SummaryThis paper introduces PathLapPE, a spectral positional encoding derived from the path Laplacian of directed graphs. It produces node- and edge-level features that encode directional higher-order structure and can be added to standard graph learning architectures. On node- and graph-level benchmarks, it yields consistent gains across several architectures, especially with direction-aware message passing. Compared with magnetic Laplacian encodings, it needs no extra tuning of directionality hyperparameters while offering competitive runtime and performance.
Lydia Mezrag, Semih Cant\"urk, Michael Perlmutter et al.
In briefAdversarial attacks that improve both transferability and stealth via primal-dual regularization
SummaryThis paper argues that the apparent trade-off between transferability and imperceptibility in black-box adversarial attacks comes from fixed-budget evaluation rather than an intrinsic conflict. It proposes ST, a plug-in primal-dual wrapper that adds an L∞ saturation regularizer and solves it with a two-step update, requiring no auxiliary models. At ε=16/255, ST improves average LPIPS by 17% and NIQE by 14% over base attacks while preserving or improving attack success rate.
In briefGS-Codec: a Gaussian-splatting bottleneck for neural audio coding
SummaryThe paper introduces GS-Codec, a neural speech codec that replaces the usual quantizer with a parametric decomposition in which encoder latents are fitted as weighted sums of 1D Gaussian primitives. A lightweight predictor network regresses these parameters in one forward pass to avoid iterative fitting at inference. According to the authors, it matches or exceeds EnCodec and DAC on speaker similarity, intelligibility, and perceptual quality at comparable bitrates, with comparable WER, and one checkpoint supports post-training bitrate control.
WhyReplacing quantizers with Gaussian splatting as a speech codec bottleneck is novel, though evidence rests on comparisons summarized only at high level.
In briefGated Target Propagation for Compositional Generalization in Continual Learning
SummaryThis paper introduces Gated Target Propagation (GaTaP), a continual learning method that learns task-specific gates to selectively suppress or enhance network modules, while network parameters are updated on a slower timescale. Experiments on class-incremental learning with multilayer perceptron and convolutional networks show that it retains performance on earlier tasks and generalizes compositionally to unseen tasks through few-shot gain adaptation at inference. The authors observe that related tasks produce similar gating patterns, suggesting the learned gates capture reusable task structure.
Abdel Mfougouon Njupoun, Colin Bredenberg, Blake Aaron Richards et al.
In briefSEIS: Self-Evolving Inference Systems, where agents optimize a whole language model serving engine end-to-end
SummaryThis paper presents SEIS, a method where AI agents autonomously optimize the entire mini-sglang inference engine through iterative sessions and code changes, without human intervention. Serving Qwen3-0.6B on an H100, the optimized engine reaches 3.27 times the throughput of the original and outperforms vLLM, TensorRT-LLM, and SGLang in the single-request workload. The authors check correctness through numerical differences and accuracy on math and long-context retrieval tasks, and report that building on earlier sessions works better than independent attempts.
In briefRevisiting Generalization of Neural Graph Edit Distance Models: Cross-Dataset Transfer Study
SummaryThis paper examines how well neural Graph Edit Distance (GED) models transfer across different graph datasets. Using exact GED supervision, the authors systematically test a range of learning-based methods on diverse graph collections. They find that models performing well within their training dataset often degrade on structurally different data, while training on multiple source datasets substantially improves zero-shot transfer.
In briefGrounding Probe: a generator-independent hallucination detector using an observer model's hidden states
SummaryThis paper proposes the Grounding Probe, a logistic regression classifier over mean-pooled middle-layer hidden states of an observer language model that reads the context, question, and response in one forward pass without generating text. Because it does not depend on the generating model, it can be applied to closed-weight generators. On the RAGTruth test set, the probe reaches 0.879 to 0.894 AUROC across four observer models, and 0.924 AUROC when averaged with a supervised span detector.
WhyDecouples hidden-state probes from the generator by reading an observer model, a substantive novelty. Strong RAGTruth results and cross-generator validation support it.
Michael Rathmayr, \'Ad\'am Kov\'acs, G\'abor Recski
In briefHypergraph representation learning with random effects for hyperlinks, exploiting low-rank structure and heterogeneous link formation
SummaryThis paper proposes a framework for learning hypergraph representations that models latent heterogeneity in hyperlink formation through random effects while exploiting low-rank structure. The authors establish identifiability of the model parameters and theoretical guarantees for representation-level recovery. They study three random effect specifications (categorical, Gaussian mixture, and score-based) with corresponding estimation algorithms. Simulations and real-world hypergraph datasets show the method recovers latent structure and captures heterogeneous interaction patterns.
In briefPolicy as Data: Replay-Based Policy Dual Averaging via Advantage Regression
SummaryThe paper proposes RDA2C, an actor-critic method that treats replayed experience as accumulated policy-improvement feedback rather than only off-policy samples. It fits a dual score model to stored state-action data with critic-estimated advantage labels, and derives the policy from it via the entropy mirror map. The authors provide a finite-time error decomposition and report that, with GAE labels, RDA2C outperforms PPO on several MuJoCo and Atari tasks, and with twin-Q labels it matches SAC.
Nianli Peng, Geoffrey J. Gordon, Kiant\'e Brantley
In briefDoes Added Neural Complexity Help Detect Health Misinformation? A Leakage-Controlled Multi-Corpus Benchmark
SummaryThis paper compares five compact neural architectures, a soft-voting neural ensemble, and three classical machine-learning baselines on the COVID19-FNIR and CONSTRAINT corpora. Exact duplicate texts were removed before modelling, and neural results were repeated across three random seeds. The deep ensemble performed strongly on COVID19-FNIR, but a linear SVM outperformed it on CONSTRAINT, and architecture rankings changed between corpora. The authors conclude that model complexity gives no stable advantage and that benchmark construction can matter more than architecture choice.
