In briefPrivileged Information Does Not Guarantee Privileged Guidance: Auditing Self-Distillation for Reasoning Models
SummaryThis paper audits on-policy self-distillation (OPSD), where a student learns from next-token targets produced by a frozen teacher copy that also sees a hidden worked solution. The authors test four stages across three Qwen3 model scales on math tasks: whether the teacher uses the solution, whether its targets are useful, whether the student learns from them, and whether the learned behavior appears without the solution. They find that accuracy gains do not show transfer from the solution, and that gains track the teacher's thinking mode, a pattern they call reasoning arbitrage.
In briefVocabulary-free next-event prediction for temporal graphs
SummaryThis paper introduces GTGT, a Generative Temporal Graph Transformer that performs self-supervised next-event prediction on temporal graphs without a node vocabulary. It predicts events for observed destinations, behavioral descriptors for novel destinations, and the time gap before the next event, with targets whose meanings are shared across graphs. This lets one transformer train across disjoint node sets without node-specific parameters. On the Temporal Graph Benchmark (TGB), GTGT achieves link-prediction performance comparable to or better than established baselines at much lower training cost.
Zhishang Luo, Venu Madhav Tammali, Amila Weerasinghe et al.
In briefPothole risk assessment using geometry-grounded vision-language models
SummaryThis paper presents a five-stage framework that assesses the severity of already-detected potholes by combining geometric measurements with vision-language model reasoning, and generates natural-language warnings. It was evaluated on 3,092 potholes across 1,119 dashcam scenes. Agreement of severity classes with human annotators and measured depth is weak, but geometric grounding changes the warnings, and in a blind study raters preferred them.
Abdalmalek Aburaddaha, Zaid A. El Shair, Ali Al-Bustami et al.
In briefREWEAVE: An online method that rewires HNSW vector indexes to match query workloads
SummaryThis paper presents REWEAVE, an online index-maintenance method for HNSW vector search indexes. It learns edge utility from query feedback and periodically rewires the bottom layer while keeping the original per-node degree cap. The authors prove value-estimation and graph-stabilization guarantees under explicit assumptions and report recall and search-cost improvements over HNSW and query-aware baselines.
In briefPyramidal columns as attention heads: a hypothesis linking language cortex to Transformer attention
SummaryThis paper proposes that a suitably organized group of pyramidal columns in posterior temporal language regions (PTLRs) could biologically realize the functional architecture of a Transformer attention head. The authors describe attention heads as networks of simple neural units linked by output-input and output-weight connections, and examine how such processing could be built in neural tissue. They identify candidate mechanisms of signal transmission and modulation within and across pyramidal columns. Whether this organization actually exists in PTLRs is presented as an open empirical question.
In briefSub-actions in Action: Text-Guided Hand-Role Alignment for Sub-Action Recognition
SummaryThis paper aims to improve vision-language models' ability to tell apart the left and right hands' roles in first-person actions. It introduces sub-action detection, which requires recognizing each hand's role and manipulated object within an ongoing high-level action. The authors propose BiScopeNet, which augments existing video-LLMs without changing their encoders or language models, and release BiScopeCap-100K, a large dense-caption dataset. Results show gains over supervised baselines in action, hand-role, and object recognition.
Sadegh Rahmaniboldaji, Filip Rybansky, Quoc Vuong et al.
In briefWayfinder: Learning Which Resources to Search from Agent Citations
SummaryThis paper presents Wayfinder, an adaptive router that helps agents decide which resources, such as chats, issue trackers, email and knowledge bases, to search for information. It uses the documents cited by the downstream agent as feedback to update its internal resource representations, so later searches prioritize likely sources. The authors also construct and release three controlled testbeds derived from NFCorpus, FEVER and Tech Startup, and compare Wayfinder with supervised, tool-retrieval and contextual bandit baselines under a shared closed-loop protocol.
In briefHuman-likeness versus conversational preference for synthesized agent speech in customer-service calls
SummaryThis paper compares two listener judgments of synthesized speech: how human it sounds and how much listeners want to talk to it. Using paired-choice tests on 120 improvised customer-service turns, it evaluates six commercial text-to-speech systems and actors' original recordings. The study finds that systems ranked differently on human-likeness and on conversational preference.
Jeffrey M. Girard, Antony D'Avirro, Jason Z. Zheng et al.
In briefImproved Elastic Weight Consolidation as a Continual Learning Constraint
SummaryThis paper proposes Continual Learning as an Optimization Constraint (CLOC), which requires new training to preserve chosen functionals of previously learned behavior. The authors show that the empirical Fisher matrix in Elastic Weight Consolidation (EWC) is the first-order form of this constraint, but it weights each sample by its loss gradient scale, reducing importance for well-fitted samples. Improved Elastic Weight Consolidation (IEWC) penalizes output changes using a first-order approximation of each old sample's loss level set, removing this weighting. Across classification, segmentation, language, forecasting, and diffusion tasks, IEWC matches or improves on EWC. In a held-out ImageNet-R evaluation, uniform IEWC improves accuracy over matched EWC by 2.00 percentage points (95% CI 0.93-3.07).
In briefContinual event recognition for autonomous optical network monitoring, letting systems learn new disturbances without retraining
SummaryThis paper proposes a continual event recognition framework for monitoring optical networks based on state-of-polarization (SOP), using prototype representations, open-set recognition, and few-shot prototype registration. Experiments on a 2600-km ROADM-based optical testbed report 99.82% closed-set recognition accuracy and adaptation to novel disturbances without encoder retraining. The study also finds that some overlapping disturbances can be recognized zero-shot through prototype composition, while others require explicit few-shot registration.
In briefRobustness and Fairness Trade-offs in Adversarial Classification
SummaryThis paper studies binary classification trained with adversarial examples under a squared soft demographic-parity budget, tested on the Adult Income dataset with logistic regression and a one-hidden-layer MLP. Tighter budgets reduce the demographic-parity difference while keeping balanced accuracy above empirical-risk minimization at the tested settings. The authors also report that mean equal-opportunity difference can reverse sign, and give a rate theorem explaining the required true-positive-rate shift. They conclude that a single fairness metric cannot replace a full group-rate audit.
In briefA study on scaling public input to AI policy using pluralistic AI agents that represent individuals
SummaryThe paper develops and tests pluralistic agents, which are AI representatives that deliberate on behalf of individuals to align AI systems with diverse human views. Experiments show the agents reliably shift policies toward their represented persona's preferences, though their influence varies with the instructions they receive. In a preregistered study with 746 participants, people rated the approach as useful and fair and preferred it over giving input to a single AI model.
In briefA unified, interpretable framework for combining three optical fiber sensing methods when data is limited
SummaryThis paper proposes a unified framework for multi-modal optical fiber sensing under data scarcity. It maps measurements from SOP monitoring, DAS, and OTDR into shared physics-based semantic descriptors such as impulsiveness, persistence, structure, and energy concentration. Inference reasons over semantic class profiles, optionally aided by retrieving similar examples. Experiments on real datasets report accuracies of 95.0%, 91.1%, and 94.6% for the three modalities.
In briefLearning from Imperfect Demonstrations: Robust Offline-to-Online Imitation from Mixed-Skill Demonstrators
SummaryThis paper studies how an AI agent can learn a control policy from demonstrations given by people or algorithms with differing and unknown skill levels. The authors propose LEOS-IL, which jointly learns a model that scores how optimal each state-action pair is and a latent expertise level for each demonstrator from unlabeled suboptimal offline data. During online training, the agent maximizes these scores while using its own rollouts to refine the scoring model. Experiments on continuous control benchmarks report that the method outperforms existing baselines in highly heterogeneous, low-quality demonstration settings.
In briefTheoretical runtime analysis of Cartesian Genetic Programming on the Max problem, showing exponential speedup
SummaryThis paper studies Cartesian Genetic Programming (CGP), an evolutionary method that encodes candidate programs as graphs, by analyzing its runtime on the Max problem. The authors prove that a (1+1) CGP using common mutation operators finds an equivalent optimal program in expected time at most O(log^9 n), where n is the optimal program length. This is an exponential speedup over previously proven superlinear bounds for tree-based genetic programming. It is the first proven performance guarantee for CGP on an established benchmark, and experiments support the theoretical results.
In briefTopology-aware 4D Gaussian splatting for reconstructing changing everyday scenes from first-person video
SummaryThis paper addresses reconstructing dynamic scenes from monocular first-person video when everyday actions change object topology, such as cutting, assembly, opening, or closing. The authors note that existing 4D Gaussian Splatting methods, which deform a fixed representation, produce severe artifacts in these cases. They introduce TopoEgo, a benchmark covering fission, fusion, open, and close events, and Topo-4DGS, a method that assigns independent rigid transformations to parts and activates or deactivates Gaussians as surfaces appear or separate. The experiments reportedly show Topo-4DGS significantly outperforms state-of-the-art approaches, especially on post-event frames.
Leonardo Vanni, Stefano Gasperini, Federico Tombari et al.
In briefIris-Z: Scaling One Generative Model Beyond Specialist Geometry Models
SummaryThe paper presents Iris-Z, a general image generator that produces appearance and geometry (depth) together using a single backbone. The same weights generate images with depth, estimate depth from photographs, and paint images onto a given depth. The authors report that at sufficient scale it surpasses state-of-the-art specialist models on established depth benchmarks while retaining the backbone's generative abilities without specialized layers. Two models are released to the community.
In briefJustRUA: A Controlled Study of GPT-6 Robot-Use Agents
SummaryThis paper dissects recent robot-use agents built on GPT-6 Astra and tests their design choices, including the harness, action interface, experience reuse, and scene geometry, within one framework. It introduces JustRUA, a lean harness that keeps the full episode history and stops each motion early. On the RoboDojo simulation benchmark, this simple harness raises the score from 29.7 to 50.3.
In briefFit-LRP: fitting linear relaxations to polytopes for neural network verification
SummaryThe paper proposes Fit-LRP, a framework for verifying neural networks over non-convex regions represented as unions of polytopes, rather than axis-aligned boxes. It optimizes affine bounds exactly by solving a sequence of small, low-dimensional linear programs, and enhances the final step of LiRPA and α,β-CROWN branch-and-bound methods. On ARCH-COMP and neural Lyapunov benchmarks, it reports a 47× reduction in branching on the hardest region and a 15× reduction in worst-case verification time.
Brian Yarbrough, Jesse Wayment, Jingbo Wang et al.
In briefPredicting Needles in Petabyte-Scale Satellite Imagery: Open-Vocabulary Event Anticipation
SummaryThis paper represents change as the vector difference between vision-language model embeddings from different time points, enabling automated global event discovery in satellite imagery without manual annotation. The authors build a dataset of over 21,000 locations, more than 64 event types, and about 2 million PlanetScope and Sentinel-2 images across six continents, and use it to train a global event anticipation model. According to the paper, the model reaches a maximum F1 score of 93.5% for change detection, outperforms baselines 3.6 times on predicted event retrieval, and forecasts completion dates with a 3-day median error up to nine months before completion.
WhyThe authors represent change as embedding differences, build a 2M-image global event dataset, and report 93.5% max F1 and 3.6x retrieval gains.
Shreelekha Revankar, Mikhail Klassen, Creon Levit et al.
In briefAdaptive distance and instance detection methods for spotting enemies in a shooter game
SummaryThis paper studies object detection and semantic segmentation models for detecting enemies in the first-person shooter game Counter-Strike: Condition Zero. The authors trained an object detection model with YOLOv5s and compared inference times with semantic segmentation. Object detection was slightly faster, while semantic segmentation was slower but better at capturing exact object shapes. The paper proposes Adaptive distance detection, which combines both algorithms, and Adaptive instance segmentation, which combines semantic and panoptic segmentation.
In briefHow training settings shape flat minima in priorless variational learning
SummaryThis paper studies variational deep learning trained only on the expected loss, without an explicit divergence penalty toward a prior, relying instead on implicit regularization from the optimizer. The authors show, theoretically and empirically, that larger learning rates and fewer parameter samples lead to flatter loss minima, which are associated with better generalization. They describe training as operating at a modified edge of stability, set by the signal-to-noise ratio of gradient estimates and the number of Monte Carlo samples. Experiments on out-of-distribution benchmarks are reported to support these findings.
In briefDo System Prompts Really Assign the Roles They Claim? Validating and Steering Occupational Role Vectors in LLMs
SummaryThis paper introduces a framework for extracting and injecting occupational role vectors, activation-level directions that capture how someone in a given role would distinctively respond to a task. The authors validate these vectors on Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct using interpretability and control lenses. The extracted vectors are stable and separable, but injecting them alone only matches a role prompt at best, and adding them to prompts did not slow role decline over 30 turns.[Summary limited]
In briefTRACE: Training-free method to raise video generation resolution by syncing trajectories
SummaryThis paper introduces TRACE, a training-free method that increases the resolution of text-to-video generation models. It synchronizes low- and high-resolution generation trajectories during early denoising and uses local window attention to refine fine details. It also checks for duplicate subjects against the low-resolution layout and corrects them. Experiments report better perceptual quality and text-video alignment than FreeSwim, with 2.0-2.9x end-to-end speedups.
In briefDexRemo: A Force-Sensing Robot Hand and Task Suite for Contact-Rich Removal
SummaryThe paper presents DexRemo, a three-finger robotic hand with high-force actuation, swappable fingertips, and three-axis fingertip force sensing. It also provides a suite of nine real-robot scenarios covering adhesive detachment, surface cleaning, and retrieving occluded objects. Fingertip forces are expressed in a shared wrist-centered frame, and an Action Chunking Transformer fuses vision, proprioception, fingertip pose, and force inputs to predict actions. Ablations show that wrist-frame forces, especially combined with fingertip pose, improve stability and generalization over local-frame force inputs and no-force baselines.
In briefBetter boundary prediction does not automatically yield better structural representations in language models
SummaryThis paper tests whether adding punctuation restoration to language model training improves representations of argument, semantic, and syntactic structure. The authors examine English across decoder-only, encoder-only, and encoder-decoder models below 1B parameters. They find that optimising for boundary prediction does not necessarily produce higher-quality structural representations, and performance on downstream tasks requiring structural understanding did not improve.
Michael Zhou, Markus Frohmann, Junghyun Min et al.
In briefDuoJEPA: Translating Between Two Unaligned Self-Supervised Embedding Spaces
SummaryThe paper proposes DuoJEPA, which trains one self-supervised encoder per modality and appends a transformer layer that maps each embedding into the other encoder's space, using a contrastive loss computed only in that target space. The two original embedding spaces are never compared directly by any loss term. The method is evaluated on chest radiographs with reports from CheXpert Plus and on Sentinel-2 optical and Sentinel-1 radar imagery.
Mahmut S. Gokmen, Mitchell A. Klusty, Evan W. Damron et al.
In briefTesting Deep Ground Reaction Force Estimators on Unseen Subjects Across Two Labs
SummaryThe study compares three sequence models on two independent datasets of locomotion (6 subjects) and occupational patient handling (10 subjects) for estimating ground reaction forces and centers of pressure. Trial-level splits overstated accuracy compared with subject-held-out evaluation, and transfer between the two laboratories performed below chance in both directions. The authors conclude that such models should be validated with subjects held out, checked for coordinate-frame consistency before training, and tested across laboratories before deployment.
In briefPhysICL: In-context learning of physical fields on unstructured meshes using procedural priors
SummaryThis paper presents PhysICL, a model pretrained solely on procedurally generated tasks that predicts physical fields directly on unstructured query meshes from labeled simulation examples, without updating its weights. On five PLAID datasets not seen during pretraining, prediction error decreases as the number of context examples grows. It performs comparably to per-dataset reduced-order models on two of the datasets.
WhyThe authors transfer in-context learning to physical field prediction on unstructured meshes, pretrained only on procedural priors and tested on five unseen PLAID datasets.
In briefWELP: Weighted Ensemble in the Loop for Learning Committor Functions
SummaryThis paper addresses the difficulty of sampling rare molecular transitions needed to learn the committor, a probability that describes how a system moves between states. It proposes a framework that alternates between weighted ensemble simulation, which generates transition data, and training a neural network committor, which then serves as an improved collective variable for further simulation. The method outperforms greedy baselines on a two-channel 2D system, alanine dipeptide, and chignolin, and the authors release a public benchmark of empirical committor values for alanine dipeptide.
In briefIs Spoken Hungarian Low-resource? A Quantitative Survey of Hungarian Speech Datasets
SummaryThis paper surveys existing spoken Hungarian speech datasets, grouping them into five categories: monolingual, the Hungarian portion of multilingual collections, pathological, child-related, and dialectal corpora. The authors estimate the available data at about 2800 hours across roughly 7500 speakers, reflecting rich linguistic diversity. However, they find the data distribution and its fit to tasks such as speech recognition far from optimal, calling for larger-scale natural speech datasets. The survey also aims to serve as a starting point for a unified foundational speech model for Hungarian.
P\'eter Mihajlik, Katalin M\'ady, Anna Koh\'ari et al.
In briefRiordan Arrays as Structured Operators in AI
SummaryRiordan arrays are traditionally studied as infinite lower-triangular matrices that encode generating-function transformations. This note proposes using them as structured operators in artificial intelligence and data analysis, mapping a generating function A(z) to g(z)A(f(z)). This yields a parameterized family of causal linear operators combining convolution, substitution, and multiscale mixing. The authors suggest possible uses in neural-network layers and sequence models, and argue that the algebraic structure can provide a useful inductive bias in certain settings.
Dennis Davenport, Dewayne A. Dixon, Moussa Doumbia
In briefStarscream: Rethinking LLM sharding on AMD Instinct GPUs using GPU partitioning
SummaryThe paper observes that when a chiplet-based GPU is partitioned into multiple logical devices, parallelization strategies designed for homogeneous GPUs cause performance degradation due to memory redundancy, collapsed arithmetic intensity, and communication overhead from flat collectives. The authors present Starscream, which adds intra-GPU sharding along an orthogonal dimension, guided by principles of kernel shape preservation, memory redundancy elimination, and communication overhead minimization. Using tensor parallelism as a case study on AMD MI300X and MI355X, they report up to 4x speedup for attention and 1.43x for fused GEMMs over chiplet-oblivious baselines.
Sumanth Gudaparthi, Smeet Chheda, and Ganesh Dasika
In briefReinforcement learning for inventory control when demand regimes switch
SummaryThis paper studies inventory control when demand switches between hidden regimes. The authors build a Gymnasium environment where daily demand follows a two-state hidden Markov process with Poisson emissions, and compare an optimised (s,S) heuristic, an MLP PPO agent, and an LSTM Recurrent PPO agent across lead times of 0 to 10 days. The recurrent policy performs best at every lead time, and its advantage over the MLP grows with lead time, reaching 2.17 times at lead time 10. Principal component analysis suggests the LSTM shifts from encoding a binary regime indicator to encoding multi-step pipeline state as lead time increases.
In briefLang-SVG: Language-Guided Hierarchical Image Vectorization
SummaryThis paper presents Lang-SVG, a method that converts raster images into editable SVG vector graphics organized as a tree of semantically labeled, parent-child related shapes. It uses segmentation models to initialize object-part hierarchies and multimodal large language models to label nodes, along with techniques for simplifying shapes, pruning redundant nodes, and recovering missing regions. The authors also propose an evaluation protocol that measures structural quality and semantic part alignment, and report state-of-the-art results in rendering fidelity and structural quality.
In briefMOSAIC: A Modular Framework for Robot Manipulation in Cluttered Scenes
SummaryThe paper presents MOSAIC, a modular framework that integrates with pretrained Vision-Language-Action (VLA) models to improve scene understanding and obstacle awareness in robotic manipulation. It adds a semantic grounding task predicting masks for the target, destination, and obstacles, and uses depth-derived geometric tokens along with obstacle-aware attention supervision. Evaluated on the RoboPRO benchmark and real-world tasks, MOSAIC improves success rates by 5.1% in cluttered scenes and 9.7% under visual perturbations.
Richard Shi, Mohammad Mahdavian, Mehrdad Momeni Zadeh et al.
In briefShadowBench: Testing Whether LLMs Really Forget Named Entities
SummaryThis paper introduces ShadowBench, a benchmark that tests whether language models can recall facts about entities when their names are removed. Across model scales, including GPT-5.4 and Claude-Sonnet-4.6, performance drops by over 20% without explicit name tokens. The authors also find that current machine unlearning methods mostly hide entities on direct queries, with unlearned models reconstructing popular forgotten entities in over 74% of reasoning traces.
In briefSpec-Driven Property-Based Testing: Deriving Test Properties from Specifications for AI-Written Code
SummaryThis paper studies using a software specification, rather than the implementation, as the source of properties for property-based testing of code written by AI agents. The proposed Spec-Driven PBT agent first states invariants in natural language from the specification, then checks the implementation to turn them into concrete tests. Across 25 applications from ViBench, it writes 2.4x fewer property tests than the code-driven baseline, and the two approaches find largely non-overlapping bugs, suggesting complementary value.
In briefPrivate exceptions to a shared decision rule, learned as typed, gated, withdrawable weight edits
SummaryThe paper addresses organisations that apply the same published decision rule but hold private evidence that it is wrong for their cases and will not pool it. Each party compiles the rule into a small ReLU network as a prior and searches a small typed space of declarations, proposing an edit only when its own records pass a selection-aware evidence gate. On synthetic data, the approach was reported as non-inferior to seven tuned baselines at 30 records per party and better than the unchanged rule.
In briefSmart model routing to cut LLM inference costs by an order of magnitude
SummaryThis paper presents a cost-quality optimization framework for routing queries among large language models, sending each query to the cheapest model expected to answer it adequately. It derives an optimal threshold-policy structure, a cascade breakeven condition, and a sensitivity analysis over price ratio and traffic skew. A numerical example under stated pricing and traffic assumptions shows roughly an order-of-magnitude cost reduction, and the paper discusses practical risks affecting production savings.
In briefHeriKGQA: A knowledge-graph question-answering method for heritage sensor data
SummaryThis paper presents HeriKGQA, a knowledge-graph question-answering method for querying cultural heritage sensor data in natural language. The large language model only interprets questions and phrases answers, while computation is handled by SPARQL queries and deterministic code. On 500 questions from five real heritage sites, it reaches 99.0% accuracy, outperforming the strongest of three baselines using the same model. Removing the grounding layer reduces accuracy by 12.2 points.
Lingxiao Kong, Vasileios Arampatzakis, Vasileios Sevetlidis et al.
In briefHypergradient Skewness and the Antithetic Neumann Estimator
SummaryThe paper shows that estimating hypergradients in bilevel optimization with a truncated Neumann series introduces a third type of error, skewness, beyond bias and variance. It proposes the Antithetic Neumann Estimator (ANE), which pairs each Hessian sample with a mirrored counterpart around a reference Hessian, removing odd-order noise terms when the reference is exact. Across robust regression, data cleaning, offline reinforcement learning, and imbalanced classification, ANE-based methods over 20 seeds lower loss by 4 to 25% relative to the best baseline.
In briefReward-driven collider physics challenges for self-improving agentic AI via reinforcement learning
SummaryThe paper introduces an open-ended optimization task format for high-energy physics problems that frontier models still struggle to solve, with continuous rewards evaluated automatically. The authors benchmark frontier reasoning models on several such tasks. They demonstrate a proof of concept in which a small language model improves through reinforcement learning using automated feedback, without human intervention. The goal is to move agentic AI from reproducing established workflows toward autonomously discovering better scientific solutions.
In briefOpen Vision Banana: A Fully Open Method for Generative Visual Perception
SummaryThis paper presents Open Vision Banana (OVB), a fully open and reproducible method for turning pretrained image generators into unified multi-task visual perception systems. It analyzes design choices such as color encoding, prompt design, parameter adaptation, and task balancing, and trains 4B and 9B parameter open-weight base models on public datasets. The authors also introduce legend ensembling, a test-time method that fuses segmentation results under several class-color legends and improves mIoU by up to 8.7 points on Cityscapes and 15.0 points on COCO-Object without retraining.
WhyThe authors convert pretrained image generators into unified perception models and add legend ensembling, reporting mIoU gains on Cityscapes and COCO-Object.
Arpit Jadon, Valentinos Pariza, Jona Ruthardt et al.
In briefFood Critic: An Auditable Agentic Visualization System for Restaurant-Safety Reasoning
SummaryThis paper presents Food Critic, an agentic visualization system that helps diners choose restaurants by combining public health inspection records with crowd reviews. It runs its AI agent under three grounding conditions: a database-only agent, an agent that retrieves food-safety regulations, and an ontology-guided agent producing structured hypotheses with typed evidence. The system displays the agent's reasoning as inspectable visual evidence. The authors evaluate several large language models with use cases under each grounding condition.
In briefFewer Expert Transfers Without Changing Routing: Expert Eviction for MoE Offloading
SummaryThis paper studies offloading for mixture-of-experts (MoE) language models, where most expert weights stay in CPU memory and are copied to the GPU when the router requests them. Using real CPU-to-GPU copies on OLMoE-1B-7B, the authors find that policies holding a fixed expert set and decoding faster than LRU raise next-token loss by at least 0.067 nats. They propose a rule that leaves every routing decision unchanged: when an expert must be evicted, remove the one whose next use is farthest in the future. This copies 2.1x fewer experts than least-recently-used (LRU) with no loss in quality, and comes within 3.5% of the fewest copies possible.
Pierre-Luc Bacon, Diego Calanzone, Jean-Maxime Larouche
In briefA competitive tournament testing rule-based, search, optimization, learned, and language-model negotiation agents under scarce supply or demand
SummaryThe paper pits nine negotiation architectures against each other in a Gale-Shapley matching market, covering rule-based, search, optimization, imitation-learning, and reinforcement-learning approaches. The comparison runs under a balanced market and two scarcity conditions across all 36 exclusion pairs. Nine large language models from 1 to 72 billion parameters were also tested, and they lost significantly to the two strongest engines in 32 of 36 comparisons, winning none.
Eduardo de Avila-Armenta, Ra\'ul Alejandro Aquino-Ibarra, Jorge Octavio Nicol\'as-D\'iaz et al.
In briefBEATS Dataset: Studying How Music Affects the Body in Everyday Life
SummaryThe BEATS dataset compiles psycho-physiological data from 30 participants who listened to curated well-being playlists in real-world, free-movement settings rather than a lab. It includes wearable sensor recordings, listening records with acoustic features and self-reports, individual traits such as demographics, music preference, and personality, and psychological measures taken before and after participation over several weeks. The dataset is intended to support research in well-being, music psychology, affective computing, and signal processing.
Benjamin Gutierrez Serafin, Tanaya Guha, Fahim Kawsar
In briefControlling the Chance of Undiscovered Failures Before AI Deployment
SummaryThis paper addresses the problem of unknown failure modes in AI systems, such as AI agents, that have not appeared during pre-deployment testing. It presents an algorithm that certifies when the probability of encountering such an unseen failure is sufficiently small. The method builds an anytime-valid upper confidence sequence, allowing developers to test repeatedly, check the bound, and stop once it falls below a chosen threshold. The approach is evaluated on SWE-Chat, a dataset of real-world coding agent traces.
In briefWhere to Generate Touch: Whether Image Latents Keep Material Recognition
SummaryThe study tests whether a material classifier trained on real GelSight touch images still works on reconstructions from pretrained image models, to choose a latent space for generating touch. On ObjectFolder Real, reconstructions from three 16x discrete tokenizers cause the classifier to lose accuracy, while SD-VAE reconstructions cause only a small drop. SD-VAE reconstructions also keep object identity at 76.7% matching accuracy, close to 78.9% for real touch. The authors therefore select SD-VAE latents and show a photo-to-touch generator trained in that space.
In briefNIBench: A benchmark testing AI agents on neuroimaging execution, skill learning, and scientific judgment
SummaryThis paper presents NIBench, a neuroimaging benchmark with three evaluations of autonomous AI agents. It tests execution and workflow reconstruction on 56 tasks across seven tool ecosystems, examines whether agents can evolve procedural skills by distilling execution trajectories, and assesses scientific judgment using 36 claims from published Alzheimer's disease studies. Nineteen language models were evaluated in a shared framework, finding that leading agents execute supplied protocols nearly perfectly while performance gaps widen without procedural guidance. Distilled skills improved acquisition tasks and transferred between models, though benefits on unseen tasks were inconsistent, and agents sometimes reached conclusions that differed from expert answers.
In briefTraining calibrated conditional generative models using a loss that needs only samples
SummaryConditional generative models are widely used to sample posteriors in scientific inverse problems, but they are often overconfident. The paper introduces a differentiable loss based on expected coverage probability, estimated from samples alone, so it works with any conditional generative model without tractable densities or architecture changes. The authors prove it is a strict divergence that vanishes only when the model equals the true posterior.
Benjamin Remy, Raphael Rossellini, Chihway Chang et al.
In briefTwo random forest methods using window-specific functional principal components for curve data
SummaryThe paper addresses functional random forests that apply one principal component decomposition to the whole curve, which may dilute features occupying short intervals. It proposes the global-local functional random forest (GL-FRF), a late-fusion stack of a full-curve forest and per-window forests with penalised convex weights, and the window-concatenated functional random forest (WC-FRF), an early-fusion method training one forest on concatenated window scores. On twenty UCR datasets, GL-FRF outperformed a tuned full-curve forest with the same number of trees on 16 datasets.
In briefLarge Payments Model: A Reusable Foundation Model for Payment Prediction
SummaryThis paper presents the Large Payments Model (LPM), a domain-specific foundation model trained from scratch on large-scale structured transaction data to support payment tasks such as fraud detection, authorization prediction, and payment optimization. It uses two complementary architectures, and optional distillation enables deployment under resource constraints. The authors pretrain the model on up to 400 billion tokens and report that it outperforms strong task-specific baselines across multiple prediction tasks in Merchant and Treasury Services, along with production deployment.
Natraj Raman, Abhenandhan Manoharan, Vishwasreddy Thuniki et al.
In briefSparse PCA with overlapping supports: hardness and an approximation scheme
SummaryThis paper studies sparse principal component analysis with several components, where each component must be sparse and all components must be pairwise orthogonal, and their supports may overlap arbitrarily. The authors show the problem is NP-hard even with two components and rank two, but admits a fully polynomial-time approximation scheme when the rank and number of components are fixed. The returned solution has exactly orthonormal columns. The approach reduces the problem to a sparse orthogonal Procrustes problem, combining approximate dynamic programming over rows with two-column updates.
WhyThe authors show NP-hardness for general overlapping supports and give an FPTAS for fixed r and d, extending known disjoint and common-support results.
In briefDetecting and correcting hallucinations in aerial image captions
SummaryAutomatic captions for drone aerial images can reduce manual inspection work, but multimodal large language models often produce text that does not match the image. This paper proposes a lightweight, reference-free framework that uses InternVL3 and Qwen2.5-VL to detect hallucinated content in captions and then corrects it with a refinement module. Experiments on the Labelled Illusion Dataset (LID) show that the framework significantly reduces hallucination rates while remaining practical for real-time use in resource-constrained settings.
Debolena Basak, Somsubhra De, Maunendra Sankar Desarkar
In briefComputational explanations of semantic universals: a look at why languages share deep meaning features.
SummaryLanguages around the world share many deep features in how they express meaning, which likely reflect common aspects of human cognition, communication, and social structure. The paper lays out a framework for computationally explaining these semantic universals. It applies principles of ease of learning and efficient communication to several domains, including quantifiers, color terms, modals, and indefinites. The authors close by discussing future directions for strengthening the explanatory power of computational methods in semantic typology.
In briefRUMI: A Handheld Interface Matched to a Robotiq Gripper, with AR and Haptics for Robot Learning
SummaryThe paper presents RUMI, a handheld manipulation interface matched to the Robotiq 2F-85 gripper, with exchangeable fingertip sensors that allow natural input before contact and force measurement after contact. It integrates tracking into a virtual reality headset to give real-time augmented reality feedback on motion feasibility for the target robot, and adds a haptic mechanism for force-aware teleoperation. A user study found fewer episodes with joint-limit violations with AR and pose initialization (3/50) than without AR (16/50). The authors also train imitation policies from demonstrations collected with two fingertip types and evaluate the haptic interface for collecting force-aware teleoperation data in simulation.
In briefVibePrevisBench: Testing AI Coding Agents at Building 4D Scenes in Blender
SummaryThe paper introduces VibePrevisBench, a benchmark where coding agents turn a brief into an executable 4D scene in Blender and render a proxy video. It contains 242 tasks across four input types, including scripts, photos, continuations, and reference videos. Outputs are scored by a rule-based scene check against hidden ground-truth facts and by a VLM judge on whether those facts appear in the video. Tests on 11 frontier coding agents show differences in spatial and camera understanding, success on basic camera work but failure on challenging requests, and an inability to correct errors after inspecting their renders.
In briefHonesty as a Race: Predicting When Verifiable-Reward RL Teaches Models to Bluff
SummaryThis paper studies whether reinforcement learning with verifiable rewards makes a model honest or teaches it to bluff. It sets up payments for unverifiable verdicts and for checker-accepted certificates that reveal bugs, and derives an approximate condition under which a GRPO step makes the model more honest. Experiments on propositional Datalog, two code models, and abstention on grade-school math show that runs above a payoff-ratio threshold tend to collapse.
In briefFrom Link Prediction to Linear Community Detection: Finite-Sample Guarantees for Graph Pretraining
SummaryThis paper explains why link prediction, a common pretraining objective for graph neural networks, yields node embeddings usable by linear classifiers for node classification. In the finite-sample stochastic block model, the authors prove that small population held-out link prediction excess risk implies small community classification error for a linear classifier on the learned embeddings. They also propose a nearest-centroid few-label classifier and validate the theoretical chain with synthetic block-model experiments.
In briefSO-Arena: A Benchmark for Testing the Range and Reliability of Fast "System One" AI Models
SummaryThis paper introduces SO-Arena, a benchmark that evaluates the range of abilities and decision reliability of System One models across 11 domains and five capability families. It uses verified reference labels, domain-specific difficulty criteria, interactive sandbox tasks, and comparisons with reasoning-intensive System Two approaches. Beyond accuracy, it measures decision stability when options change and whether models recognize when no valid answer exists. Tests on 21 models show that strong overall performance does not guarantee reliability, as the best-performing System One model ranks only seventh in reliability.
In briefSAME: A semantic watermarking method to protect open-source large pretrained models
SummaryThis paper proposes SAME, a semantic watermark embedding method for protecting the intellectual property of open-source large pretrained models against unauthorised copying and modification. SAME adds a fixed private vocabulary that pairs each native token with a private embedding, so private tokens can be combined into flexible semantic triggers rather than fixed trigger-response pairs. The authors state that because the watermark is grounded in the model's native semantics, it persists through downstream fine-tuning. Experiments reportedly show reliable and robust watermarks across models of varying scales and architectures.
In briefPatient-derived organoid drug screening data for training and evaluating AI drug response prediction models
SummaryThis paper presents four drug screening studies using patient-derived organoids (PDOs), generating PDO multi-omics profiles, drug feature data, and treatment response measurements. The authors retrained and evaluated several public drug response prediction models under three scenarios: cell-line-only training, combined cell-line and PDO training, and cell-line pretraining followed by PDO fine-tuning, comparing prediction performance on PDOs. They also assessed the consistency of viability data generated by different laboratories using different assay technologies or protocols.
In briefPrior-as-Skill: Teaching visual agents to store step order outside the model
SummaryThis paper studies self-evolving visual agents that learn task steps and their order as skills, trained into a vision-language model with on-policy self-distillation. The authors propose PriorSplit, which keeps the step-order information outside the model and uses it to fill in frames that are missing at deployment, without training beyond the standard procedure. Experiments on synthetic skills and on web, cooking and egocentric tasks show it outperforms compared methods when frames are missing, and stays close to the baseline on unusual step orders.
In briefAIFR: Adaptive Identity Feature Recovery for Diffusion-Based Makeup and Anime Portraits
SummaryThis paper addresses the problem that makeup and anime portraits generated by diffusion models can hinder face recognition. It proposes AIFR, a framework that inserts a lightweight recovery module into an intermediate layer of a frozen face recognition model to correct identity features, so recovered embeddings can match existing gallery templates without re-encoding. With only 1.22M trainable parameters, the authors report that average TAR at FAR=10^-4 improves from 66.1% to 90.2% over the original ArcFace. The recovered embeddings can also condition a pretrained Arc2Face model to generate faces with better identity preservation.
In briefImage deblending with diffusion models (splitting a blended image back into its source images)
SummaryThis paper treats image deblending as a joint inverse problem, using diffusion posterior sampling to reconstruct two source images at once from a single blended image. It proposes a training-free two-stage framework that uses text prompts to guide semantic separation in latent diffusion, followed by pixel-space refinement for fine details. The authors report state-of-the-art results on unseen domains, extension to spatially-varying alpha maps, and competitive performance with supervised baselines on standard scalar blending benchmarks.
In briefPhysics-based simulation for scalable, generalizable MRI reconstruction
SummaryThis paper introduces the Scalable Inverse Model (SIM), which trains on physics-simulated data and co-designs its architecture and learning methods around scalability to improve MRI reconstruction across unseen conditions. It was evaluated on 14 public and in-house datasets spanning seven centers, five vendors, field strengths from 0.3 to 7 T, and many anatomical regions and protocols. The authors report that SIM recovers fine anatomical details and suppresses aliasing artifacts on zero-shot cases where conventional deep-learning methods fail.
WhyThe authors co-design physics-simulated training data, architecture, and learning for zero-shot MRI reconstruction, validated across 14 datasets and a blinded radiologist study.
In briefHow well can frozen vision-language models handle event camera data? A training-free study on EventBench
SummaryThis paper presents a training-free system for the General Understanding track of the EBMV @ ECCV 2026 EventBench challenge, which tests vision-language models on five event-camera understanding tasks. The authors use no challenge training data and rely only on representation and prompting with RGB-pretrained models. The system reaches 62.28% overall accuracy on 2,084 test questions, ranking 9th of 27. The paper also analyzes weaknesses, including bounding-box cropping that left event OCR text at reduced resolution and biased answer choices on detailed-understanding questions.
In briefMULTILOOP: Automatically Evolving Multi-Agent Harnesses for Complex Search Tasks
SummaryThis paper studies whether a coding agent can automatically optimize the harness (prompts, tools, and orchestration logic) of multi-agent LLM systems. The authors propose MULTILOOP, an iterative loop in which a coding agent analyzes, evaluates, and edits the harness, and test it on complex information-seeking tasks from DeepSearchQA. With only 10 training samples, the evolved harness improves accuracy by 4.6% over the base harness and 3.2% over the best single-agent system. A failure-mode taxonomy shows that orchestration failures characteristic of multi-agent systems largely persist after optimization.
In briefA method for building faster surrogate models that learns which inputs matter most
SummaryThe paper proposes spike-and-slab random feature expansions (SAS-RFE), a surrogate modeling method that learns how to sample random features using sensitivity analysis. Inclusion probabilities for each input are set from estimated total Sobol' sensitivity indices, and the sampling distribution is updated iteratively. Numerical experiments on six benchmark problems with up to 100 input dimensions show improved surrogate approximation and sensitivity estimation.
In briefLatent imagination method (HyLI) for parameterized action spaces
SummaryThis paper addresses parameterized-action MDPs, where an agent must choose a discrete action along with continuous parameters. The authors propose HyLI, a model-based agent that learns a hybrid policy in the latent imagination of a world model and acts with a single forward pass. It adds counterfactual mode evaluation and success-window replay to support learning. On seven standard benchmarks, the authors report competitive or better performance, improved sample efficiency over the strongest model-free baseline, and much faster decisions than the model-based planner DLPA.
Atahan Cilan, \.Ihsan Mert Muhac{\i}ro\u{g}lu, \"Umit Can Bekar
In briefWhere Uncertainty Is Observed Inside LLMs: A Mechanistic View of Uncertainty Quantification
SummaryThis paper proposes four mechanistic observation lenses for studying uncertainty in large language models: structural, representational, distributional, and semantical, which observe component updates, hidden states, token probabilities, and meaning distributions. It links these lenses through the residual stream and orders them by the information they carry, offering a principle for choosing uncertainty quantification methods for different targets. The authors survey existing methods through these lenses and distinguish evaluation targets, criteria, and metrics.
Xiaoyu Liang, Virginia Ceccatelli, Aliai Eusebi et al.
In briefDISCERN: A Persistent Memory Framework for LLM Agents
SummaryThis paper argues that persistent memory for enterprise LLM agents is mainly a problem of resolution rather than retrieval, since facts conflict, get superseded, and similarly named entities get confused across sessions. It proposes DISCERN, which follows the rule of preserving information at write time and resolving it at read time, using an additive bi-temporal provenance graph and an agentic loop that commits to one of four gated actions. On the ENT-310 benchmark, DISCERN leads on abstention calibration and macro answer correctness, driven by entity resolution. It also leads macro answer correctness on the full LoCoMo corpus with no tuning.
Chirag Ahuja, Vladislav Tanov, Ramcharan Kakarla et al.
In briefReliable Near-Field Source Counting and Positioning via Two-Stage MUSIC
SummaryThis paper studies reliable source enumeration and positioning for near-field signals in coherent multipath environments. It proposes a conformal risk control scheme that calibrates the penalty coefficient of an eigenvalue-based criterion to keep the expected miss-detection ratio at a set level. It also introduces MUSIC-Net, a model-based deep learning framework that learns line-of-sight subspaces and surrogate distances to position sources directly, plus a split conformal prediction method that builds angle-range regions with finite-sample coverage guarantees. Numerical results indicate better control of source underestimation, higher positioning accuracy, and tighter predicted regions than benchmarks at the same coverage levels.
In briefTerms of Care: Designing Data Governance That Disability Communities Can Join
SummaryThis paper examines how disability communities can take part in governing datasets used in AI-generated media, since open sharing of such data risks misuse, decontextualization, and exploitation. The authors worked with three disability advocacy organizations through interviews, an educational webinar, and a modular licensing activity. The modular approach let communities adapt Creative Commons clauses to state permitted uses, sector restrictions, and commitments to respectful representation. Concerns about monitoring and enforcing appropriate data use remain.
In briefORBIT: A Benchmark for Testing LLMs on IT Operations Reasoning with Sourced Evidence
SummaryThis paper introduces ORBIT, a benchmark of IT operations questions in multiple-choice, open-ended, and troubleshooting formats, with each question paired with a reference answer and recorded source. An audit found template defects in about a tenth of generated items, which were removed, and the troubleshooting format was rebuilt from real practitioner threads with accepted answers. Under an independent judge, five LLMs answered most multiple-choice items correctly, between a third and just over half of open-ended items, and only a minority of troubleshooting items.
Sathwik Narkedimilli, Rakshit Verma, Rahul Soni et al.
In briefBeyond the Reward: Capability Drift in Multilingual Post-Training
SummaryThis paper studies two GRPO post-training branches started from the same supervised checkpoint in a multilingual mixture-of-experts model. The released branch gains 36.07 points on MATH but loses 19.32 F1 points on XQuAD-Indic, while XQuAD-English improves. A separately trained sibling branch, adding Indic instruction following, recovers the Indic task scores and preserves MATH, but does not restore script compliance. The authors associate most of the measured MATH gain with more boxed-answer emission, so benchmark scores should be read alongside the response forms they reward.
Sutanshu Saksena Raj, Purusharth Malik, Pasunuti Prasanjith et al.
In briefA study on whether explanations from AI tools for choosing freight transport modes can be trusted
SummaryThis paper examines explainable machine learning that classifies shipments as air or truck/ocean, using a PEPFAR supply chain dataset of 5,964 records. The authors compare several models and evaluate the stability and faithfulness of SHAP and LIME explanations. Tree ensemble models achieved the highest accuracy, but models with similar accuracy disagree on which features matter most. The authors conclude that explanation stability and faithfulness should be assessed before such explanations support freight decisions.
In briefCuriosity-0826-4B: A Medical Event Model for Children and Adults
SummaryThis paper presents Curiosity-0826-4B, an updated generative medical event foundation model trained on Epic Cosmos that simulates patient timelines to make zero-shot clinical predictions. On tasks both models support, it generally matches or exceeds its predecessor, Curiosity-0226-1B, with laboratory value prediction error 18–46% lower. The model extends predictions to patients aged 2–17 and adds adult mortality prediction tasks based on documented death.
In briefFaster federated Bayesian optimization via shared neural network surrogates
SummaryThis paper introduces GLOBAFED, a framework that lets agents transfer knowledge between related but non-identical optimization objectives without sharing raw data. It uses a probabilistic neural network surrogate with a shared feature backbone trained by federated learning and agent-specific variational Bayesian output heads. Experiments reported in the paper show faster convergence and better final performance than state-of-the-art federated Bayesian optimization baselines, along with the ability to warm-start from learned global features.
Alexander Gr\"afe, Max van Gemmeren, Paul Brunzema et al.
In briefPredictive multiplicity in ranking and its effects on fairness
SummaryThis paper studies predictive multiplicity, where several comparably accurate models give conflicting outputs, in score-based ranking problems. It gives formal definitions for multiplicity in ranking and supporting theoretical guarantees. The authors analyze how this multiplicity can cause unfairness and how exploring it can help find less discriminatory rankings, and they validate these findings on synthetic and real-world datasets.
In briefAskALCF: A Self-Improving AI System for HPC User Support
SummaryThis paper presents AskALCF, an agentic pipeline that analyzes logs from an HPC user-support AI system. It evaluates production interactions in detail, identifies whether failures stem from knowledge, retrieval, prompting, model behavior, or tool execution, and routes each to a targeted fix. Across 2,174 support turns, unattributable failures fell from 41.4% to 14.2%, and expert agreement rose from 108 to 123 of 150 turns.
Jingyan Jiang, Akhilesh Bondapalli, Krishna Rauniyar et al.
SummaryThis paper studies learning a mixture of two bounded-support exponential-family distributions that share an interaction parameter matrix, with each component having its own linear parameter vector. The component labels belong to a known structured class. The authors propose a convex-loss estimator that avoids the log-partition function, so estimation is convex for each fixed label assignment. They derive finite-sample bounds on the interaction and linear-parameter errors and show component and label recovery under additional conditions.
In briefCGOcc: Learning Dependencies Between Gaussians for 3D Semantic Occupancy Prediction
SummaryThis paper presents CGOcc, a supervised multi-camera framework for predicting 3D semantic occupancy. It explicitly models dependencies between overlapping Gaussian primitives using a diagonal-plus-low-rank representation, then uses class-conditional projection to refine voxel class probabilities. Experiments on nuScenes report state-of-the-art performance, with larger gains when Gaussians overlap heavily and better predictions near boundaries.
In briefStreaming Speech Recognition for Low-Resource Mixed-Language Speech with Morphology-Aware Evaluation
SummaryThis paper studies streaming automatic speech recognition for low-resource, agglutinative languages with intra-sentential Bangla-English code-switching. It proposes a Dynamic Block-Online framework aligned with voice activity detection, using global bidirectional attention to preserve morphology. To address limited data, it introduces Script-Anchored Loanword Injection, which augments training text with English loanwords, and defines the CS-WER metric for switch, root, and suffix errors. The authors also report successful transfer of the approach to menstrual-medical domain speech.
Kazi Rafat, Afifa Imran, Md. Ismail Hossain et al.
In briefSpaceMCP: Giving Vision-Language Models Spatial Reasoning via Structured Tool Calls
SummaryThis paper presents SpaceMCP, a framework that lets vision-language models make structured function calls to query a scene graph for spatial information such as distances, object positions, and camera properties. The scene graph is built in real time using feed-forward 3D reconstruction for metric geometry and open-vocabulary segmentation for object localization. Without any training, the method reports state-of-the-art results on VSI-Bench, MMSI-Bench, and VSTI-Bench, improving its base model by 15 to 22 points per benchmark.
In briefMeasuring the noise floor when comparing quantized and BF16 language models
SummaryThe paper measures how much decisions change from non-quantization sources, including engine swaps, server restarts, batch-size changes and kernel selection, on 2,400 public decisions for two Cloudflare models, Clef-flash (9B) and Clef (27B). It reports quantization effects as excess over those sources. For both models, every 4-bit interval excludes all non-quantization sources, and two of five pre-registered predictions failed.
WhyThe authors separate non-quantization noise from quantization effects, measuring restarts, engines, batching, and kernels on two decision models.
In briefBeyond Reweighting: Certified Gains from New Transitions in Discrete MCMC
SummaryThis paper studies how to design the set of allowed transitions in reversible Markov chain Monte Carlo (MCMC), considering both the transitions and their stationary flows. It gives a necessary and sufficient dual condition under which adding transitions strictly lowers the optimal asymptotic variance, along with computable primal-dual certificates. On identical augmented Ising supports, full weight optimization certified strict improvements in 24 of 24 conditions, while degree-based Metropolis-Hastings weighting certified none.
WhyThe authors give a dual necessary and sufficient condition for support augmentation to strictly improve asymptotic variance, construct a family with Ω(n) gains from one edge, and report experiments on
In briefALMeD: A Medieval French Gold-Standard Corpus
SummaryALMeD is an annotated, semantically rich corpus (still a work in progress) of medical and surgical treatises written in medieval French, including Old French, Middle French and Anglo-Norman. Its semantically disambiguated gold standard focuses on medical terminology. It also accounts for morphological change, evolving syntax and the absence of spelling normalisation.
In briefCan AI Agents Recall Past Emotional Experiences? A Diagnostic Study of Event-Affect Decoupling
SummaryThis paper introduces Affective-Episodic Retrospection (AER), a task requiring agents to recall past events, their associated emotions, and the correspondence between them. It defines Event-Affect Decoupling as a failure where the correct emotion is recalled but linked to the wrong episode. The authors propose EM-TRACE, a framework that splits AER into retrieval, affect recovery, and realization stages for diagnosis. On a GoEmotions-based benchmark, emotion-only retrieval recovered affect in 28.65% of cases but retrieved the target episode for only 4.44% of queries.
In briefMechanistic interpretability study of AION-1, an astronomy AI model
SummaryThis paper presents a causal mechanistic interpretability study of AION-1, an astronomy foundation model, focusing on its multimodal advantage in redshift prediction, where combining imaging with photometry outperforms photometry alone. The advantage is largest for spatially extended and dust-attenuated galaxies and is mostly localized to the encoder's first block. A single attention head is necessary for most of the advantage, but none of the tested astrophysical operations explains what it computes.
WhyThe authors report the first causal interpretability study of astronomy model AION-1, localizing its multimodal redshift advantage to one block and a single attention head.
Rachel C. Zhang, Alberto Bietti, Wes Gurnee et al.
In briefFrequency-Lipschitz curriculum for neural fields in sparse-view CT reconstruction
SummaryThis paper proposes FLIC, a neural field framework for reconstructing CT images from sparse projection views while controlling how much high-frequency detail the network can express. It pairs random Fourier feature encoding with spectrally normalized MLP decoding under scheduled Lipschitz budgets, moving reconstruction from coarse anatomy to fine detail. The authors report state-of-the-art unsupervised reconstruction quality on four public CT datasets under noise-free and low-dose settings.
In briefUltrasound Elastography: Low-Cost Ways to Excite Tissue
SummaryThis paper addresses ultrasound elastography, a method that measures tissue stiffness to help diagnose some diseases. It proposes using ordinary smartphones as the source of mechanical excitation instead of costly dedicated devices. The resulting displacements are used to solve the elasticity inverse problem with a finite element method. Preliminary reconstructions on a tissue-mimicking phantom are similar to those obtained with a dedicated actuator, suggesting the approach is feasible.
In briefMemory Consolidation for Conversational Agents: What the RRR Method Explains and Doesn't
SummaryThis engineering study introduces Memory-TCG, an extension of Conversation ESAA inspired by ringing, resonance and renormalization (RRR), for keeping current decisions retrievable across agent sessions. On synthetic data with 480 facts and a 1,024-token budget, full RRR recalled 0.44 of current decisions, compared with 0.06 for chronological retrieval. The authors note that fixtures are repetitive, comparators simplified, and results not independently replicated, so the work is an engineering analysis rather than a claim of superiority.
In briefDifferential Backpack language models: letting context suppress word senses
SummaryBackpack language models predict by linearly combining context-free per-word sense vectors with non-negative context weights. This paper computes those weights with differential attention, letting context actively inhibit a sense, which the authors call a differential Backpack. At equal parameter counts it reduces the perplexity gap to a Transformer by 31% on WikiText-103 and keeps an advantage up to 350M parameters on OpenWebText. Its sense representations also improve zero-shot sentence similarity by up to 33%.
WhyThe authors replace non-negative softmax weights with differential attention so context can inhibit senses, reporting perplexity gains on WikiText-103 and OpenWebText with ablations.
Konstantinos Skianis, Giannis Nikolentzos, Polykarpos Meladianos et al.
In briefContour-Assisted Evolutionary Computation: Peak Estimation from Solution Contours
SummaryEvolutionary algorithms usually use their evaluated candidate solutions only to rank individuals. This paper treats these solutions as elevation points, draws contour lines through them, and splits the search region into peaks, with a peak memory directing the search toward new ones. It proposes CA-EA and CA-SAEA, which are tested on two-dimensional benchmarks and the CEC 2015 expensive suite. Their performance relative to competing methods varies across settings.
WhyThe authors treat evaluated solutions as contour maps to decompose peaks and extrapolate their locations, reporting results on 2-D benchmarks and CEC 2015 expensive problems.
In briefMemInStream: Testing whether streaming video models truly remember or just recognize
SummaryThis paper argues that question-answering accuracy alone cannot reliably measure memory in streaming video understanding, since errors may come from visual misunderstanding and correct answers may rely on language priors. The authors introduce MemInStream, a benchmark that links each question to temporally localized evidence intervals. Its validation protocol keeps only questions answerable from the annotated evidence but not from recent observations, post-evidence observations, or text alone. The benchmark evaluates memory across four categories: Episodic Retrieval, State Tracking & Updating, Multi-Evidence Integration, and Contextual Memory.
In briefChJGBench: A Diagnostic Benchmark for Generating Chinese Court Judgments
SummaryThis paper introduces ChJGBench, a benchmark for testing whether language models can generate the reasoning and dispositive sections of Chinese court judgments. It is built from 3,462 real judgments covering civil, criminal, and administrative cases, and uses a diagnostic framework with 13 metric types. The experiments find that legal-domain models do not uniformly outperform general-purpose models, with strengths varying by dimension. Performance is generally better on criminal cases, while civil and administrative cases remain difficult.
In briefData-free metrics for neural network layers may not reflect model quality
SummaryThis paper examines data-free metrics from random matrix theory (RMT), which are used to judge the quality of neural network layers and pre-trained models. The authors show that functionally preserving reparametrisations, made by reinitialising robust layers, can give these metrics arbitrary values across several architectures and scales. They also find that relationships among RMT metrics are spuriously linked to generalisation, and that whether models land in a 'good' metric range depends partly on initialisation.
WhyThe authors show RMT-based data-free metrics can be arbitrary via functionally preserving reparameterisations, across several architectures and scales.
Gabryel Mason-Williams, Israel Mason-Williams, Fredrik Dahlqvist
