In briefGene summaries that reveal the answers: circularity from Gene Ontology in gene embedding benchmarks
SummaryThe paper shows that NCBI Gene summaries are generated from Gene Ontology annotations, so they can contain the same labels used to evaluate gene embeddings. Comparing function-prediction results with and without these summaries, the authors find that removing them lowers performance, and that UniProt function descriptions recover much of it. They also introduce GeneProv for comparing encoders on matched text.
In briefQuery-only, prompt-agnostic attacks on visual in-context learning segmentation
SummaryThis paper studies the robustness of Visual In-Context Learning (VICL) segmentation models when an attacker can only observe returned RGB outputs and does not know the deployment prompts. The authors propose two appearance-space attacks, Color Hijack, which pushes foreground colors toward the background, and Task Hijack, which collapses instance colors into a shared assignment. A Self-Prompt Surrogate transfers these attacks to unseen prompts and outperforms a random-prompt baseline on SegGPT. At 30% corruption, Color Hijack annotations reduce downstream YOLO-seg mAP50 from 0.761 to 0.281.
In briefMySign: A motion-capture dataset for generating 3D Malaysian Sign Language
SummaryThis paper introduces MySign, the first motion-capture dataset for isolated signs in Bahasa Isyarat Malaysia, with 5,000 samples covering 1,000 glosses recorded by five Deaf native signers. It provides ground-truth SMPL-X sequences captured with a sign-language-specific setup, which offer more accurate hand and finger motion than pseudo-labels regressed from video. The authors report that training on these sequences improves generation compared with training on pseudo-labels, and that current state-of-the-art models produce plausible outputs that still fall short of the ground truth, especially in fine-grained articulation.
In briefRisk control for multi-stage AI language model pipelines
SummaryThis paper proposes CompCRC, a framework that controls the risk of wrong answers in systems where several large language models are chained in stages. It splits the risk into per-stage parts, selects budget allocations and acceptance thresholds on independent data, and certifies each stage on held-out data. The authors give a finite-sample guarantee under conditional exchangeability and compare against Global-CRC on MMLU, GSM8K, and HellaSwag.
In briefBeyond Hard Labels: Using LLM Probabilities to Improve Human-Assisted Regression
SummaryThis paper studies whether keeping the full probability vector from large language models, instead of reducing their output to one category label, improves downstream regression when combined with limited human annotation. The authors compare direct regression and human-assisted correction using three LLMs and three social-science text datasets. After human correction, using the probability vector consistently improved regression efficiency across models and datasets.
Xinyue Wang (Department of Statistics, The Pennsylvania State University), Qunhua Li (Department of Statistics et al.
In briefStatopia: A Prover-Library System for Automated Formal Proofs in Statistics
SummaryThe paper presents Statopia, an ecosystem combining Statopia Prover, an independent agent for formalizing and proving statistics and machine learning results, and Statlib, a Lean 4 library organized for automated use. The prover can construct a Lean proof from a statement alone, or use a natural-language proof to guide its search when one is available. Statlib's shared definitions and theorem interfaces are designed for retrieval and composition by agents, and its statistical coverage continues to expand.
In briefScale-by-scale evaluation of generative samplers for lattice field theory
SummaryThe paper shows that a small global effective sample size (ESS) does not necessarily mean long-distance observables are inaccurate, since errors spread over many short-distance modes can dominate the diagnostic. It introduces a scale-resolved diagnostic based on the overlap between coarse-grained model and target distributions. The method is tested in free field theory and in two-dimensional φ⁴ theory near criticality. A marginal Metropolis test combined with exact local updates is also explored to reduce the remaining bias.
In briefJoint Optimization of Memory, Computing, and Offloading for Energy-Efficient DNN Inference
SummaryThis paper addresses energy use in edge intelligence systems running deep neural network (DNN) inference, where memory demands have become a key bottleneck. The authors formulate energy minimization under deadline constraints as a mixed-integer nonlinear programming problem and jointly optimize memory frequency, computing frequency, and offloading. They decompose the problem into local and edge subproblems and solve them with geometric programming, convex optimization, and a coordinate descent algorithm. Simulations show the gap from the optimal policy is at most 9.3%, and device energy consumption is reduced by up to 16.1% compared with benchmarks lacking memory frequency optimization.
In briefHow task framing and execution paths affect LLM math performance
SummaryThis paper studies how a language model's math performance depends on how the question is presented, whether earlier draft text is reused, and whether candidate answers are supplied. Paired comparisons on 65 development problems show that the submission instruction's wording strongly determines whether valid final answers are produced. Drafts and candidates each raise accuracy, but their combined effect is subadditive, and retaining a trailing excerpt outperforms a leading one.
In briefDoes Pose Encoding Matter? Geometric Representations for Motion Vocabularies
SummaryThis paper studies how continuous human motion can be compressed into a finite set of discrete tokens. It compares four skeletal pose encodings: Gaussian joint heatmaps, root-relative joint positions, local joint rotations with a continuous 6D embedding, and a redundant full kinematic encoding. VQ-VAE tokenizers with hypersphere codebooks are trained while varying codebook size, temporal stride, and code dimension, and effective vocabulary size is measured. Rotation and position encodings keep adding effective vocabulary, while the full kinematic encoding saturates at a few hundred cells and heatmaps at fewer than sixty, a gap that reconstruction loss alone does not reveal.
In briefUncertainty-guided reward labeling for reinforcement learning with limited feedback
SummaryThis paper studies reinforcement learning when only a small part of an offline dataset can be given reward labels, asking which samples to label. It keeps queried rewards for training a reward model and bounds the learned policy's suboptimality using offline RL error and reward model uncertainty. Based on this, it proposes SURE, a method that uses successor features to choose uncertain rewards tied to states from which high-value states are reachable. Experiments report near full-feedback performance with few reward labels.
Renhao Zhang, Shreyas Chaudhari, Bruno Castro da Silva
In briefSoftPAT3D: Physics-Based Text-to-3D Scene Generation with Deformable Objects
SummarySoftPAT3D is a framework that generates 3D scenes from text prompts, including deformable objects such as cloth, stuffed toys, and ropes. It derives each object's rest shape from its manufacturing process, assigns initial elastic parameters from a material library, and settles the scene to static equilibrium under a unified contact model. Across 65 prompts, the authors report that SoftPAT3D produced no undesired interpenetration, whereas every other system they compared did.
In briefA hierarchical framework for compressing brain fMRI data into tokens
SummaryThe paper introduces HBAR, a method that organizes resting-state fMRI voxel data into a hierarchy of voxel groups, brain parcels, and large-scale functional networks, achieving up to 415x compression. On the HCP-YA dataset, at about 415x compression, HBAR gets higher parcel and network correlations than rate-matched flat tokenizers. At 52x compression, shuffling the anatomical grouping lowers parcel correlation from 0.920 to 0.886.
Karanpartap Singh, Yue Zhao, Mohammad H. Abbasi et al.
In briefWhen calibration data matter for sub-4-bit language model quantization
SummaryThis paper studies whether task-informed calibration data can better preserve downstream accuracy when compressing language models to fewer than four bits. The authors propose FragCal-DC, which adds training documents from task families that lose the most accuracy at two bits to generic calibration text. They compare six data-dependent quantization methods across six dense models and ten tasks, finding that SignRound V2 improves mean ten-task accuracy by 1.83 points over generic-text calibration at two bits.
In briefADC-RANet: Making Adaptive Image Classifiers Better With Attention, Distillation and Calibration
SummaryThis paper introduces ADC-RANet, an improved version of the Resolution Adaptive Network (RANet) for image classification. RANet allocates computation across input scales depending on how difficult each sample is. The method adds bottom-up channel attention, top-down cascaded knowledge distillation, and MixUp-based correction of overconfidence. Experiments on CIFAR-100 and ImageNet show consistent gains across computational budgets while keeping RANet's efficiency.
Jeferson Severino de Araujo, Lilian Hollard, Eric Granger et al.
In briefHow System Prompts Change LLM Safety Responses: Mostly Refusal Willingness, Not Discrimination
SummaryThis paper examines how system prompts change the way large language models respond to safety-relevant requests. Using HarmBench, XSTest, and IHEval, the authors cross eight safety-task prompt families with three clarity levels and evaluate 13 models. They apply signal detection theory to separate changes in the ability to distinguish safe from unsafe requests from changes in willingness to refuse. The study finds that system prompts mainly shift refusal willingness, with smaller and less consistent effects on discriminability.
In briefNanoNet: An Ultra-Light Network for Classifying Brain Tumors in MRI Scans on Edge Devices
SummaryThis paper presents NanoNet, an ultra-lightweight convolutional neural network for classifying brain tumors from MRI images, inspired by MobileNetV2 and built with only 8,132 parameters. The authors also explore dependency-aware structured pruning, knowledge distillation, and INT8 quantization-aware training to further compress the model. On the BRISC 2025 benchmark, NanoNet reaches 92.80% accuracy and 92.97% macro F1, and its INT8 version achieves 89.70% accuracy.
Tran Nguyen-Le, Huu-Thuan Huynh, Minh-Triet Tran et al.
In briefDiscrete Diffusion Playground: A 2D Test Bed for Discrete Generative Models
SummaryThis paper introduces the Discrete Diffusion Playground, a diagnostic benchmark that converts known 2D continuous distributions into binary token sequences via an exact discretized bijection. Generated samples can be decoded back to 2D and compared directly with the ground-truth distribution. The benchmark includes smooth Gaussian mixtures and tilted checkerboards with sharp boundaries and disconnected support. Experiments show that masked diffusion models learn meaningful adaptive reveal orders but are less stable than autoregressive baselines on smooth targets and largely fail on tilted checkerboards, where autoregressive models preserve the target geometry.
In briefRescene-ESM: an Earth system model built around a frozen AI weather model
SummaryThis paper presents Rescene-ESM, an Earth system model whose coupler, YAHO, links a frozen 20.8-million-parameter AI weather model with physically based ocean, sea-ice, land, river, carbon and energy components constrained by observations. It runs about 1,000 years per day on one GPU, holds the observed climate without drift, reproduces observed global warming from 1979 to 2022, and routes runoff to 212 river gauges. Driven by emissions, it projects heatwave exposure in 2081-2100 to differ by a factor of 4.7 between SSP1 and SSP3 worlds.
WhyThe authors report embedding a frozen AI weather model in a physically coupled Earth system model that reproduces El Nino, river runoff and heatwave exposure divergence.
In briefGARP: Real-time generative game characters running locally in a game engine
SummaryThis paper presents GARP, a platform for real-time generative agents inside the Godot game engine, where character plans, conversations, and memories are generated on the fly by a language model. The full system runs on local hardware using a 3B-parameter open-weight model without any cloud service. It adds emotion, personality, needs, and relationship mechanisms, along with layered optimizations to keep the game responsive. In a nine-scenario evaluation, 25 of 27 runs passed log-based checks.
Piero Molino, Patrick John Chia, Federico Bianchi et al.
In briefPredicting Heart Attack Risk From ECGs: When Tested-Only Labels Mislead
SummaryThe paper studies how machine learning trained only on patients who were tested can fail when some diagnoses cannot be detected from the model's inputs. Simulations show the problem worsens as testing becomes more selective or undetectable cases become more common. In ECG-based prediction of acute coronary syndrome, models trained on revascularization performed worse than chance for 30-day major adverse cardiovascular events among untested patients, while models trained on ECG-visible subtypes kept their discrimination.
WhyThe authors formalize a failure mode under selective labels and report ECG-based ACS prediction results across three foundation models and subtype-specific targets.
Alexander Schubert, Anton H{\aa}kansson, Ziad Obermeyer
In briefWhat Multiple-Choice Video Scores Really Measure
SummaryThe paper argues that current video question-answering benchmarks rely on multiple-choice options, language priors, and fragile scoring, making it hard to isolate true video understanding. Across six benchmarks, the authors find accuracy well above chance even without visual input. They propose Option-Delayed Replay (ODR), which reveals options at controlled stages of open-ended inference, and release a human-relabelled LVBench edition correcting 20.5% of its answer keys.
In briefSymplectic Normal Coordinates Flow: a generative model for Hamiltonian systems
SummaryThe paper introduces an invertible symplectic flow layer made of amplitude-dependent rotations in learned normal coordinates, aimed at modeling densities of Hamiltonian systems under varying conditioning parameters. The authors derive the condition under which the nonlinear rotation layer is symplectic and propose an architecture that guarantees it. They test the approach on a synthetic chain of coupled nonlinear oscillators and two accelerator physics problems. Normalizing flows with the new layers achieve competitive or superior negative log-likelihoods, with larger gains as nonlinearity increases on the synthetic system.
In briefDifferentiable virtual-slot partitioning for adaptive tokenization in vision transformers
SummaryThis paper proposes Differentiable Virtual-Slot Partitioning (DVSP), a method that learns a two-scale quadtree for adaptive tokenization in vision transformers end-to-end without deleting any tokens during training. Each 2×2 region keeps four virtual slots, blended with a single weight α between a coarse token and its fine children, reproducing the discrete grid exactly at α in {0,1}. On COCO detection with EVA-02-B, the authors report retaining 97.9% of dense Box AP while removing 64.2% of visual tokens, for a 1.61× end-to-end speedup.
In briefTurning SHOP-Style Method Preconditions into Partially-Ordered HTN Planning
SummaryThis paper examines how method preconditions in hierarchical task network (HTN) planning are interpreted under two semantics. SHOP-style semantics require the precondition to hold just before a method's first action, while HDDL semantics allow unrelated actions to invalidate it before the method starts. The authors present a compilation of the first-action semantics into standard partially ordered HTN planning without method preconditions, using only two additional facts, addressing an open question in the field.
In briefA fast local learning spiking network based on neural activity synchronization
SummaryThis paper proposes a local learning framework for multi-layer spiking neural networks, inspired by Neural Activity Synchronization. Each layer updates its own weights using locally computed gradients, so no gradient backpropagation is required. The model is trained and evaluated on MNIST, reaching 97.78% classification accuracy. It also shows robust performance under Gaussian and salt-and-pepper noise at different levels.
Pranith Dutta, Sk Wasim Rajjak, Siddhanta Ray et al.
In briefAgents propose, code verifies, humans decide: a verified pipeline for loading messy scientific tables into databases
SummaryThe paper presents an AI-assisted pipeline that ingests heterogeneous public scientific tables, such as wide matrices, long tables and spreadsheets, into PostgreSQL. LLM agents propose schemas, deterministic code verifies them with 39 check types in structural and semantic loops, and a human approves the schema and load. On four public biomedical datasets, 9 of 20 first proposals failed a check and all were caught, and all 45.1 million values loaded with zero reconciliation gap.
Bharatha Rankothge, Suzanne Hodgkins, Virginia Rich et al.
In briefVirtual Task Prompting for Multi-Task Scene Understanding
SummaryThis paper introduces Virtual Task Prompting (VTP), a method for multi-task scene understanding that keeps the backbone unchanged. It maintains virtual tasks beside the backbone and uses learnable routing matrices to retrieve task-specific compositions that interact with image tokens. The approach works with pretrained vision foundation models and reports state-of-the-art results on Pascal-Context and NYUD-v2, with robustness tested on Taskonomy's 13 tasks.
In briefA systematic review of reliability and scalability in cloud, edge and AI-enabled systems
SummaryThis paper presents a PRISMA 2020 systematic mapping study of reliability and scalability in cloud-native, edge-native and AI-enabled systems. It includes 1,794 primary studies and 95 secondary studies, classifies them into a 13-category taxonomy, and grades each by evidence level from simulation to production deployment. Most studies rely on prototypes or testbeds, while production deployments are concentrated in cascading-failure prevention, root-cause analysis and predictive autoscaling.
In briefRoom to Reconsider: Reply Length Decides Whether Pushback Makes Language Models Sycophantic or Self-Correcting
SummaryThis paper examines whether language models abandon correct answers or revise wrong ones when a user pushes back without giving a reason. The authors test four open models on MMLU and ARC-Challenge multiple-choice questions and read final answers from option logits after a free-form reply. They find the reply length changes the outcome: with longer replies, pushback corrects many wrong answers and raises accuracy.
In briefPruning-Based Probes for Stopping Reasoning Early
SummaryThe paper introduces Prune-to-Probe, an inference-time controller that temporarily suppresses selected late-MLP activation coordinates after an answer is stated to test whether the answer is stable. The interventions act only as probes, so the model either returns its existing answer or resumes generation from an untouched cache, with no permanent change to the model. On MATH it reduces reasoning-token forwards by 2.60%, and on MMLU-Pro it achieves 1.145% pooled savings with no observed harm.
In briefCommunication-efficient aggregation of predictions across networked agents
SummaryThe paper studies learning on directed acyclic graphs where each agent sees only some input coordinates and can use its parents' predictions as features. It revisits the finite-sample analysis of the full-message protocol, identifies a gap in a prior proof, and gives a corrected population-risk guarantee. It also proposes a shared-sketch protocol that sends low-dimensional linear sketches of predictions, separating statistical error from compression error. For Gaussian sketches, an explicit high-probability population-risk bound is provided.
Mohammadhossein Bateni, Zahra Hadizadeh, MohammadTaghi Hajiaghayi et al.
In briefHybrid product search directly on a data lake, without copying the catalog
SummaryThe paper proposes hosting product search directly on lakehouse tables (Apache Iceberg), removing the need to maintain copies of the catalog in a lexical engine and a vector database. The authors extend Apache Pinot with BM25 scoring, corpus-wide term statistics, and reciprocal rank fusion in SQL. On the Amazon ESCI benchmark, hybrid search served from the lake matches exact retrieval (nDCG@10 of .361), and on a 277M-product lake a single node answers brand-scoped searches in single-digit milliseconds.
In briefUsing LLMs to build and audit chemistry datasets, with a judge model applied to PET depolymerisation
SummaryThis paper presents a citation-grounded method in which one language model extracts structured experimental records from papers, and an independent LLM judge checks each record against the full text and proposes cited corrections. The method is applied to literature on homogeneous catalytic PET depolymerisation. The judge agreed with human evaluation on 85.7% of reviewed records, and 67.5% of 3,942 extracted records were accepted as written. The authors release the dataset, audit logs, and the open-source Chemistry Data Extractor Toolkit.
Muhammad bin Javaid, Karim K. Ben Hicham, Enkhnyam Battulga et al.
In briefMoTo: A Mixture of Tokenizers for Fairer Multilingual Language Models
SummaryLarge language models typically use a single tokenizer fixed at training time, and most such tokenizers are trained on English-dominant data, which serves morphologically rich and non-Latin-script languages poorly. This paper proposes MoTo, which trains a dedicated BPE tokenizer for each language and combines them into a unified superset vocabulary with shared tokens deduplicated. Across 21 languages spanning six scripts, MoTo improves performance on 18 of them while retaining English performance despite allocating fewer vocabulary entries to English.
In briefPhase-Conditioned Reward Shaping (PCRS) for Token-Efficient LLM Reasoning
SummaryThis paper proposes PCRS, a training-time method that identifies reasoning phases (reasoning, execution, revision) and allocates length penalties accordingly instead of penalizing all steps equally. It compares a hidden-state classifier with a frozen text classifier across six mathematical and scientific reasoning benchmarks. The default text-classifier setup reduces mean response length by 3.13% versus a token-level length-penalty baseline, with sampled Pass@1 about 62.07% for both, while oracle Pass@16 rises from 86.08% to 87.67%.
In briefWhere AI models should run: a split-computing framework for robots on 6G networks
SummaryThis paper addresses the limited compute and energy of robots running large AI models by proposing an edge-centric collaborative inference framework that splits models across devices, edge servers, and the cloud. Simulations and real-world WiFi and 5G experiments show that the optimal partitioning point depends on network conditions and model structure. The authors identify latency-minimizing partition points for ResNet-101, Qwen2.5-VL-7B, and OpenVLA-7B under varying wireless conditions.
In briefUnified factored representations help neural networks resist catastrophic forgetting
SummaryThis paper proposes the representational forgetting hypothesis, which links the quality of a neural network's internal representation to catastrophic forgetting. Networks with unified factored representations (UFR) can resist forgetting without changing the fine-tuning algorithm or architecture. In a 1D regression task and an image-regression task, UFR networks forgot substantially less than fractured entangled representation (FER) networks when repeatedly fine-tuned on UFR-compatible variations. The results suggest representational quality is an underappreciated factor in catastrophic forgetting.
In briefToward universal excited-state machine-learned potentials using a local excitation Hamiltonian for size consistency and transferability
SummaryThe paper introduces LEH-MACE, a model whose excitation energies are eigenvalues of a local excitation Hamiltonian built from atom-centred blocks, so that excited states of separated molecules combine exactly. The authors trained it on TD-DFT energies and forces for about 300,000 geometries of over 100,000 molecules. Its advantage over direct readouts grew with distance from the training data, with roughly half the S2 and S3 errors for long conjugated oligomers.
WhyThe authors build an atom-centred local excitation Hamiltonian giving exact size consistency; the abstract reports about half the baseline S2 and S3 errors on long oligomers.
In briefProtLink: Predicting Protein Interactions Using Sequence and Multispecies Graph Context
SummaryThis paper introduces ProtLink, a framework that combines frozen ESM-2 protein sequence representations with known protein interaction networks and orthogroup connections across species to predict protein-protein interactions. The authors compare sequence-only models, GraphSAGE, and homogeneous and heterogeneous graph transformers for human interaction prediction. The multispecies heterogeneous graph transformer reaches an average precision of 0.980 with random negatives and 0.878 with two-hop-preferred negatives. For human proteins without training interactions, it achieves 0.8335 average precision, compared with 0.6482 for a sequence-only model.
In briefRoutingTrie: a method for applying metadata filters inside constrained generative retrieval
SummaryThe paper proposes RoutingTrie, which applies request metadata constraints before beam selection in generative retrieval, instead of filtering after generation. It uses one shared SID trie with a compact index per metadata field, built as a Dense Bitset or Sparse List. In a production replay, Hit@100 rose from 67.5% to 96.1% with only 1.1 ms added latency.
In briefCompact Foundation Models for Low-Latency Recommendation Serving
SummaryThis paper presents SlimFM, a pipeline that makes online use of large language models practical for click-through rate prediction. It prunes LLaMA3-8B by 44% into a 4.5B-parameter backbone, fine-tunes it with LoRA to produce field-level semantic embeddings from short prompts, and fuses these embeddings into a conventional CTR model. Across four datasets and six CTR backbones, SlimFM retains over 99% of uncompressed-model accuracy while cutting inference latency by up to 3.4x compared with existing LLM-enhanced baselines.
In briefLearning Action-Guiding Information Maps for Multisensory Exploration in Hawkmoths
SummaryThis paper introduces a multimodal predictive-coding framework that learns maps of uncertainty reduction over actions from visual and mechanosensory observations of hawkmoths interacting with flowers. A variational generative model trained on simulated environments recovers modality-specific information landscapes consistent with several hawkmoth behaviors. The authors also develop an uncertainty-gated algorithm that switches between information gathering and reward exploitation, which outperforms several control methods in nectary search.
Abdelrahman Sharafeldin, Yaqing Wang, Simon Sponberg et al.
In briefLightfoot: Graph-based retrieval system with exact, model-free document deletion
SummaryThis paper presents Lightfoot, a graph-based retrieval-augmented generation (GraphRAG) system in which graph entities, links, facts and synonym edges are deterministically derived from per-passage extractions that record chunk-level provenance. Indexing uses one language-model call per passage, retrieval uses none, and deleting a document is a database operation requiring no model calls. Compared with HippoRAG 2, LightRAG and fast-graphrag on MuSiQue, 2WikiMultiHopQA and HotpotQA, Lightfoot has the lowest indexing cost and outperforms HippoRAG 2 on some retrieval and answer metrics, but underperforms on 2WikiMultiHopQA.
In briefPIA: Agent-written physics-based runtime anomaly detection for industrial control systems
SummaryThe paper introduces PIA, in which a large language model agent writes an anomaly detector as a set of physical residuals, such as pump characteristics, tank mass balances, and controller logic. Each candidate is refined against normal operating data before deployment. On the BATADAL water distribution attack benchmark, PIA reaches an F1 of 0.878 on test attacks, compared with 0.504 for the strongest of six baselines, and attributes each detected attack to the attacked component.
Ali Mohammadi Ruzbahani, Abbas Yazdinejad, Hadis Karimipour
In briefEfficient 3D plant segmentation via multiscale graph learning and optimized geometry processing
SummaryThe paper presents a compact multiscale graph attention model of about 78 thousand parameters for segmenting plant organs in point clouds, combining local shape descriptors with a stem and leaf residual correction. The authors also optimize geometric preprocessing such as principal component analysis, sampling and neighbor queries without changing learned weights, reducing pipeline latency on PlantNet and TomatoWUR. The complete model reaches a mean intersection over union of 86.36% on a local maize collection, and inference was measured on a Raspberry Pi 5.
Jesus Ivan Ruiz-Martinez, Andres Mendez-Vazquez, Eduardo Rodriguez-Tello
In briefGradReason: LLMs Replace Some Training Epochs with Weight Updates That Need No Training Data
SummaryThis paper examines whether a large language model can read a training run's weight and gradient history and replace some epochs with updates that compute no gradient and use no training images. In 120 runs on CIFAR-10 and CIFAR-100, Claude Opus 5.5 used 32.8% fewer data epochs than an 80-epoch recipe, with 0.31 percentage points lower test accuracy. The model spontaneously chose to average recent weight snapshots, which produced most of the savings.
Ali Mohammadi Ruzbahani, Abbas Yazdinejad, Hadis Karimipour
In briefElastic Representation Learning Using Hyperbolic Geometry
SummaryThis paper introduces Hyperbolic Elastic Representation Learning (HERL), a method for learning embeddings whose dimensionality can be compressed or expanded. It uses a lightweight adaptor that maps encoder outputs into hyperbolic space on a Lorentzian manifold, without retraining the base encoder. The authors report that the approach works across image and text data in both supervised and unsupervised settings, keeps downstream performance across many embedding sizes, and costs far less computation than prior methods.
Arjun Ramesh Kaushik, Rudrasis Chakraborty, Nalini Ratha et al.
In briefDifferentiable agent-based modeling for multi-echelon supply chains that optimizes many heterogeneous policies with gradients
SummaryThe paper proposes a differentiable agent-based model for multi-echelon inventory management, implementing stochastic dynamics with differentiable tensor operations so gradients propagate through the simulation. It recovers theoretical Newsvendor optima without supervision and optimizes 10,000 distinct agent policies simultaneously in seconds. The authors report a 53% cost reduction compared with scalar baselines.
Bulent Soykan, Murtadha Alsayegh, Sean Mondesire et al.
In briefOpen Problems in Chatbots and Mental Health: research challenges for general-purpose chatbots used for emotional support
SummaryThis paper reviews key unresolved problems in how general-purpose chatbots affect mental health when people use them for emotional support. It notes that risks such as sycophancy, reinforcement of delusional beliefs, and poor crisis response are hard to study and evaluate. The authors outline research directions in controlled evaluation, real-world evidence, and evidence-based policy, and argue these should be addressed together.
Jared Moore, Jacy Reese Anthis, William Agnew et al.
In briefToward a theory of evaluability: assessing black-box machine learning models when test samples may have been seen during training
SummaryThe paper addresses how to evaluate large foundation models when the classical assumption that test data was held out from training is increasingly hard to justify. The authors formalize evaluability without that assumption and study black-box evaluation under different levels of model access. When the evaluator only observes model outputs on the test set, they characterize the distributions where accurate evaluation is possible and give tight sample complexity bounds. When additional queries to the model are allowed, they connect the problem to self-correctors from program testing.
WhyThe authors formalize evaluability when test data may have been seen in training, giving a full characterization and tight sample-complexity bounds, plus links to self-correctors.
In briefDPDP: Diffusion Policy-Steered Diffusion Planner for Offline RL
SummaryThis paper introduces DPDP, a method for offline reinforcement learning that combines a diffusion planner with a value-optimized diffusion policy. At the planner's execution step, the policy replaces the behavioral prior so actions stay consistent with the planned transitions while favoring high-value choices. The transfer is achieved either by a training-free sampler or by a preference objective that trains a single diffusion executor. On D4RL, DPDP improves the planner it builds on, remains competitive with reactive diffusion policies on locomotion, and leads on long-horizon navigation.
In briefHydro-Dataset: an open-source multi-modal pipeline for Nordic hydrology and foundation model benchmarks
SummaryThe paper presents Hydro-Dataset v1.1, a reproducible open-source pipeline that combines data from 582 SMHI meteorological stations over nine hydrological winters into a CF-1.9 NetCDF product with about 1.8 million annotated station-days. It derives six snow parameters, including Snow Water Equivalent, to provide benchmark labels for training hydrological foundation models. Cross-validation against ERA5-Land indicates that observation-anchored SWE is a more reliable machine learning benchmark than reanalysis data.
In briefOn-device adaptation of low-bit hand networks using cached gradient accumulators
SummaryThe paper introduces SGAC, a method that lets small low-bit hand-tracking networks keep adapting to a user's camera on the device. It keeps residual state only for a small, changing set of weight groups, collecting sub-step updates in an 8-bit accumulator cache and writing them back by stochastic rounding. On a 383k-parameter hand-keypoint network it uses 8.1 bits of state per weight and reports the highest mean accuracy among nine deployed low-bit networks on two simulated cameras. It also reports the highest average accuracy across twelve ImageNet-C corruptions among seven methods.
Soroor Ghandali, Kanika Mamgain, Manjot Kaur et al.
In briefHorizon-Scalable Q-Learning: A Method for Long-Horizon Value Learning from Offline Data
SummaryThis paper addresses the difficulty of learning accurate off-policy Q-functions from offline datasets as task horizons grow. It proposes Horizon-Scalable Q-Learning (HQL), which automatically balances one-step and multi-step returns, extending a target through a dataset action only when the critic values it at least as highly as the best-of-N actions sampled from the BC policy. Experiments on a combination-lock task across horizons and on long-horizon OGBench and OCBench tasks show that HQL avoids the failures of both one-step and multi-step TD and outperforms prior horizon-reduction methods.
In briefA runtime optimizer that adapts collective communication settings to network contention in AI training.
SummaryThe paper observes that collective communication library settings for distributed AI training are fixed by static heuristics, which perform well without background traffic but degrade under it. The authors propose a two-agent runtime optimizer: a detector using an LSTM classifier with a sequential probability ratio test to separate sustained contention from transient spikes, and a selector using a Dueling Double DQN with Conservative Q-Learning to adjust NCCL configuration parameters. On Chameleon Cloud's bare-metal cluster, it reduces AllReduce latency by 36% versus NCCL and 32% versus a converged AutoCCL. Across six real AI workloads, it delivers up to 65% speedups on the Delta cluster with NVIDIA A100 GPUs.
Rajat Bisht (Tennessee Technological University), Tevfik Kosar (University at Buffalo), Anthony Skjellum (Tennessee Technological University) et al.
In briefThink or Remember: The Exchange Rate of Agentic Inference
SummaryThis paper studies agents built on reasoning models, which resend a growing transcript at every call, with costs shaped by both the client and the server. The authors propose a cost model and define an exchange rate between per-token processing cost and per-pair attention cost, separating a thinking regime from a memory regime. Experiments show that whether earlier reasoning is kept and how the server caches states can change which context policy is cheaper.
Juan Wisznia, Santiago Luis Groba Alonso, Mat\'ias G\'omez Taddeo et al.
In briefAnalysis of reinforcement learning with function approximation, studying whether approximate Q-learning converges
SummaryThis paper studies how to compute the optimal Q-function in Markov decision problems with infinite state spaces. It analyzes the convergence properties of several Q-learning variants combined with function approximation, extending the TD-learning analysis of Tsitsiklis and Van Roy (1996) to stochastic control settings. The authors identify conditions under which these approximate methods converge with probability 1 and compare their results with related work.
Francisco S. Melo, Sean P. Meyn, M. Isabel Ribeiro
In briefDM-Call: Improving AI Function Calls with Decision-Model Checks and Cumulative Feedback
SummaryThis paper proposes DM-Call, a framework that helps large language models choose the right functions and generate correct arguments when making function calls. It combines local schema validation with a separate decision model that verifies function selection and argument correctness. Rejected candidates feed accumulated probabilities, binary judgments, and schema errors back into iterative refinement, without executing external tools. Evaluations on BFCL and API-Bank (first-call) show improved function-calling accuracy over direct generation at modest extra cost.
In briefSWEET: Reducing Spurious-Feature Reliance in LLM Reasoning by Weight Editing
SummaryThis paper proposes SWEET, a training-free method that combines counterfactual interventions with mechanistic interpretability to find and suppress spurious features in long chain-of-thought reasoning in large language models. It extracts behavioral directions from activation contrasts between original and counterfactual problems, then edits the weights of responsible MLP neurons and attention heads to suppress them. Across five open-weight reasoning models and five benchmarks, it raises accuracy on perturbed inputs by an average of 11.8 percentage points and on original inputs by 2 points, shrinking the relative gap from 17.6% to 2.5%.
Zhizhang Fu, Lucas Zhao, Michalis Korakakis et al.
In briefWhen Safety Becomes an Outlier: Why LLM Safety Behaviors Fade
SummaryThis paper examines why safety-aligned large language models lose safety behaviors after fine-tuning, even when safety data are included. It proposes that fixed refusal strings act as fragile outlier patterns that are easy to learn but vulnerable to later updates. The authors introduce NEST, a framework that builds natural, context-aware safety targets within the model's existing output distribution. Experiments across several models report better safety and retention under continued fine-tuning without loss of helpfulness.
In briefRule-table structure of life-like cellular automata: activity, asymmetry, sensitivity, and propagation-storage factorisation
SummaryThe paper studies the rule space of the 262,144 two-dimensional life-like cellular automata, which includes Conway's Game of Life. It organises the rules using three symmetry-derived rule-table coordinates: unweighted activity, a birth-survival asymmetry, and Boolean-derivative sensitivity. Exhaustive enumeration of compact gliders, still lifes and oscillators supports a factorisation in which the birth condition gates propagation and the survival condition tunes storage. It also proves that no B=S rule supports a compact glider at the tested footprints.
In briefBayesian transfer for offline Q-learning using multi-task Gaussian processes
SummaryThe paper proposes a target-aware Bayesian transfer framework for offline Q-learning that jointly models task-specific Q-functions with multi-task Gaussian processes. Source tasks are assessed using target predictive evidence and maximum mean discrepancy, and posterior uncertainty guides conservative Bellman updates. Simulations with heterogeneous source tasks show better target Q-function estimation and less negative transfer than indiscriminate source pooling, and an application to human decision-making data shows higher average expected immediate payoff.
In briefInterpretable AI-guided Lean proof refinement for better bounds on analytic constants
SummaryThis paper presents a framework in which an AI agent iteratively tightens intermediate lemmas in Lean proof blueprints, so local improvements propagate to a verified final bound. The authors use it to obtain a numerical upper bound on Grothendieck's constant and a lower bound on Korenblum's constant.
In briefAnatomy-guided mixture-of-experts model for 3D medical vision-language pre-training
SummaryThis paper proposes A²MoE, a framework for 3D medical vision-language pre-training that builds anatomical structure into the model's computation. A histogram representation router conditions expert selection on predicted organ composition, and an organ-query projector extracts organ-level embeddings without manual annotations at inference. Pre-trained on over 60,000 multi-phase CT scans, the authors report state-of-the-art results on segmentation, classification, and detection benchmarks, with claimed generalization to MRI and ultrasound.
Szymon P{\l}otka, Gizem Mert, Pedro R. A. S. Bassi et al.
In briefHybrid Source Code Change Impact Analysis: Combining Static Analysis with LLMs
SummaryThis paper studies change impact analysis, which helps developers find code entities affected by their changes. The authors manually labeled 1,931 changed code entities from 263 commits across 22 open-source Java projects to address dataset limitations. They propose katIA, which combines static impact analysis with large language models in a single-prompt and a multi-agent variant. The multi-agent version achieved F1-scores of 29.1% (fully qualified name matching) and 33.3% (class and entity name matching), at the cost of higher generation time and token usage.
Katharina Stengg, Christian Macho, and Martin Pinzger
In briefCoupledFlow: One-Step Neural Operators for Coupled Physics Equations
SummaryThis paper presents CoupledFlow, a neural operator framework for multi-physics partial differential equation systems that models interactions among several physical processes. It uses a structure-constrained flow matching method to learn process-specific mean velocity fields, enabling fast one-step prediction of coupled systems. Experiments on diverse coupled PDE benchmarks show it outperforms existing neural operators and generative PDE solvers.
Trong Khiem Tran, Minh Long Bui, Phi Le Nguyen et al.
In briefUsing LLMs to write data transformation code instead of brute-force search for training
SummaryThis paper examines whether large language models can generate executable data transformation programs for computer vision training, instead of relying on costly brute-force search. Under a unified training protocol, LLM-generated transformations achieved the highest accuracy on CIFAR-10, CIFAR-100, and SVHN, using far fewer candidate evaluations than brute-force search. The authors also find that one-epoch selection scores correlate with 20-epoch results to varying degrees depending on dataset and architecture.
In briefCan small Transformers learn the "language" of Coxeter groups?
SummaryThe authors trained a small Transformer to learn the language of the affine A2 Coxeter group, the symmetry group of tiling the Euclidean plane with equilateral triangles. After examining the trained model's attention patterns, they formulated and proved a word-theoretic conjecture about affine A2.
J. Elisenda Grigsby, Yiyao Huang, Arushi Katyal et al.
In briefHow surname-based race imputation biases gender classification
SummaryThis paper argues that voter files infer race from surnames and geography, which fails for women who adopt their husband's surname, especially Asian and Hispanic women married across racial lines. Using self-reported race data from six US states, the author shows lower accuracy for these women and that this bias distorts downstream gender gap estimates. Using maiden surnames improves accuracy substantially, and correcting the bias more than triples the turnout gender gap among Asian voters in a national replication.
In briefLearning linear decision rules for yes-or-no choices that are guaranteed to be feasible
SummaryThis paper proposes a prescriptive learning framework for binary decision-making under uncertainty, which jointly optimizes linear decision policies and feasible assignments across training scenarios instead of predicting optimal labels. The authors extend it to a distributionally robust chance-constrained formulation using a Wasserstein ambiguity set, which enforces feasibility with a prescribed probability and allows a tunable trade-off between robustness and solution quality. Tested on unit commitment and staff allocation case studies, the robust model is more conservative when distributions match, and under distributional shift it substantially reduces unmet demand and lowers total cost by up to 76.5% relative to classical predictive learning.
Farzaneh Pourahmadi, Trine Krogh Boomsma, Pierre Pinson
In briefRemembering to Reason: Structured Memory for Surgical Scene and Workflow Understanding
SummaryThis paper introduces MemSurg, a framework that stores earlier surgical observations in a structured, timestamped memory graph before reasoning about them. Task-adaptive retrieval supplies evidence-grounded prompts for scene and workflow reasoning. Evaluated on robot-assisted prostatectomy and two complementary surgical benchmarks, it improves over SurgRAW with GPT-4o, and memory conditioning also improves Qwen3-VL-8B fine-tuning and EndoChat.
In briefDo language models understand Bangla compound verbs? A test of their handling of vector verbs.
SummaryThis paper studies Bangla compound verbs, in which a main verb is followed by a vector verb that adds a shade of meaning, such as completion, rather than its own meaning. The authors introduce BanglaCV-Probe, a benchmark of 17 vector verbs with native-speaker-checked minimal pairs, entailment items and correct and incorrect combinations. They test 26 encoders, LLMs and translation systems on distinguishing the two meanings, understanding what a vector adds, preferring correct combinations and preserving meaning in translation. Encoders separate the two meanings well, while prompted LLMs rarely use this knowledge and most instruction-tuned models translate vector verbs literally.
Md. Asaduzzaman Shuvo, Md Mehrab Karim Antor, Md. Adnanul Islam et al.
In briefBeyond Real or Fake: A Dual-Channel Protocol for Photo Authenticity Assessment
SummaryThe paper argues that existing forensic methods label images simply as real or fake, which mixes fully synthetic, tampered, and AI-retouched images together. It proposes 2CAP, a two-channel protocol that checks camera sensor fingerprints and semantic scene coherence to classify four types of authenticity. The method also generates explanations with localization by using a reference image and a frozen vision-language model. On a multi-source benchmark, 2CAP achieves the best classification performance and improves explanation quality.
In briefSCorPion: A Strategic, Controllable, and Personified Chatbot Companion
SummaryFixed-prompt language models struggle with open-domain conversations spanning many topics. This paper extracts different chatting strategies from varied text corpora to steer a language model toward different styles based on human principles. It presents SCorPion, a deployed conversational companion that uses GraphRAG to summarize a knowledge graph into community-level strategies and query-focused optimal strategies. SCorPion produces strategic, personified, stylized responses that can be controlled by explicit factors.
In briefHow Short Is Too Short? Length-Conditional Calibration and Abstention for AI-Text Detection
SummaryThis paper studies how detectors of AI-generated text perform as passage length changes. It gives a closed-form expression for how the false-positive rate of a threshold set on long text rises as passages get shorter, and a formula for the tokens needed to reach a target true-positive rate. Thresholds are calibrated within length bins using conformal p-values, with a rule that abstains when a bin is uninformative. Across three zero-shot detectors, three corpora and two open-weight generators, binned calibration keeps the mean false-positive rate near 5.0% for lengths from 10 to 200 tokens, but no detector reaches 80% true-positive rate, so most bins abstain.
In briefA Byzantine-resilient federated learning method with user-level differential privacy that achieves optimal convergence
SummaryThe paper studies federated learning where some clients may act maliciously and user-level differential privacy protects each client's whole contribution. It derives an information-theoretic lower bound on convergence for this setting and proposes ByzULDP, which uses two-level momentum on the client and server sides. For strongly convex objectives, the method reaches the order-optimal convergence bound, and experiments on real datasets support the analysis.
WhyThe authors derive an information-theoretic lower bound and propose ByzULDP with server-side momentum; the paper reports order-optimal rates validated on real-world datasets.
In briefRobust Satisficing Ensemble: a scalable way to combine models under data shifts
SummaryThis paper introduces the Robust Satisficing Ensemble (RSE), a scalable framework for combining machine learning models so the combination stays reliable when data distributions change. It pairs a robust satisficing objective with a quadratic risk surrogate, which yields an analytical inner adversary and allows efficient weight updates without standard minimax optimization. The method is tested under label shifts and subpopulation and temporal shifts on several datasets, where it outperforms GroupDRO, IRM, and SRM.
In briefZero-shot robot control adaptation using retrieved demonstrations
SummaryThis paper addresses the problem that large Vision-Language-Action (VLA) robot models lose performance under domain shifts such as changes in camera viewpoint, background, object identity, or robot morphology. It introduces RAGu, which applies retrieval-augmented generation ideas by retrieving trajectories from mismatched in-the-wild demonstration data like DROID, calibrating them geometrically, and using them to guide the base model's flow-matching inference without retraining. The authors report real-world success rates rising from 20% to 80% and gains over closest baselines in simulation.
In briefComplementarity without ground truth? Extending human-AI complementarity evaluation beyond prediction tasks
SummaryThis paper examines how human-AI complementarity, where a human aided by AI outperforms both alone, is usually measured against ground truth labels. The authors argue that many tasks, such as medical treatment recommendations, are judged by adequacy to a practical goal rather than by prediction accuracy. They propose replacing ground truth with context-dependent adequacy labels produced by a two-step scoring method, making the normative choices behind these labels explicit.
Andrea Ferrario, Alessandro Facchini, Federico Cabitza
In briefApproximate Recall and Forecasting: Using Memory Accuracy to Diagnose LLM Prediction Errors
SummaryThis paper examines how large language models retain coarse associations about past events while losing accuracy on specific details. It introduces a new metric, the Recall Fidelity Score, to measure how imprecise internal knowledge affects forecasts of future events. On resolved historical prediction markets, the authors find that higher local recall fidelity around semantically similar past events predicts lower forecasting error, and the same signal predicts smaller distance from open market prices.
In briefKeeping LLM agents on their procedures with executable state-machine control
SummaryThe paper shows that language-model agents drift away from stated procedures during long unattended runs, and that restating instructions or resuming stopped agents only shifts the drift to other shortcuts. It proposes an executable finite-state machine that keeps workflow state outside the model's context and enforces it at tool calls, state transitions, and turn ends. In a pre-registered 200-account SOP-Bench experiment, Claude Haiku 4.5 with the FSM correctly followed the procedure for an average of 139 accounts versus 48 for the strongest baseline.
In briefGuiding Future Global Temperature Distributions Using Simulations and Observations
SummaryThe paper trains a conditional diffusion model on seventeen CMIP6 climate models and ERA5 reanalysis to generate spatially coherent daily temperature fields. During sampling, diffusion guidance steers each emulated model distribution toward observations, with guidance strength chosen through perfect-model experiments. The authors report that guidance more than halves median biases in the 1st and 99th percentiles of daily temperature and reduces annual global-mean temperature RMSE by 32-57% across three emission scenarios.
WhyThe authors use a conditional diffusion emulator with observation-steered guidance, tuned by perfect-model experiments, validated on seventeen synthetic targets, with implications for extreme temperat
Francesco Immorlano, Elijah Tavares, Felix Draxler et al.
In briefRobust and Efficient Continual Model Merging via Global Singular Subspace Separation and Restoration
SummaryThis paper addresses continual model merging, which integrates task-specific models incrementally without revisiting past models or data. The authors note that repeated merging entangles past task components and distorts their singular geometry. They propose GS3R, which restores retained singular geometry, separates task-wise subspaces, and fairly reallocates rank across tasks at each merge. They also introduce PRF to reduce merging-time memory overhead. The paper reports a strong accuracy-efficiency trade-off on vision, NLP, and multimodal benchmarks.
In briefSampling-based MPC reframed as Sequential Monte Carlo inference
SummaryThe paper shows that predictive sampling, MPPI, the cross-entropy method and their variants can all be implemented by one Sequential Monte Carlo (SMC) inference algorithm, differing in three independent choices: the proposal, the update rule that turns weighted particles into a nominal plan, and the effective-sample-size threshold for resampling. The authors evaluate SMC-based planning on MuJoCo control, obstacle-field navigation, 7-DoF motion planning and push-T in simulation and on a Franka FR3. Resampling reduced the collision rate by 30% over non-resampling baselines at a 28-knot horizon, and programmable proposals raised push-T success.
In briefREDiff-AD: a training-free, mask-free method for generating diverse industrial defect images
SummaryIndustrial anomaly detection suffers from scarce and varied real defect data. This paper proposes REDiff-AD, a framework that needs no model training or target masks and uses Rectified Flow MMDiT to generate anomalies efficiently with fewer sampling steps. Given a normal product image and one masked reference anomaly, it transfers the defect appearance via in-context learning, reuses internal attention signals with edge and color cues to produce pixel-level pseudo-masks. Experiments show a good balance between realism and diversity, with competitive downstream classification, detection, and localization performance.
In briefBRAT: A stable training method for reusing rollouts across multiple gradient passes in LLM post-training
SummaryThis paper studies the instability that arises when reinforcement learning for LLM post-training reuses sampled rollouts for several gradient passes. The authors propose Bounded Residual-Aligned Training (BRAT), which gives each token a gradient weight that decreases linearly with its probability ratio and vanishes at a trust-region boundary, and normalizes group-centered rewards by their maximum magnitude. With four reuse passes, the authors report that BRAT avoids training collapse, limits policy drift, and achieves higher average accuracy with lower seed variance on most benchmarks.
In briefOnline imitation learning to stabilize Vlasov–Poisson plasmas
SummaryThis paper studies how to control instability in Vlasov–Poisson plasma systems, a core challenge for nuclear fusion control. Because plasma states are only partially observable in practice, the authors train imitation learning algorithms on a fully observable expert controller while operating only with macroscopic measurements. They show theoretically that online imitation learning avoids the exponential error compounding of offline behavior cloning, where even small errors amplify instability over time, and can instead polynomially control rollout stability. They propose a DAgger-style algorithm for learning a stabilizing controller under partial observability. Simulations on a 1D Vlasov–Poisson system show it stabilizes plasma instabilities over longer horizons than behavior cloning.
In briefInteractive person retrieval that uses visual feedback from retrieved candidates to refine descriptions.
SummaryTraditional text-based person retrieval assumes a witness gives a full description at once, while interactive retrieval refines a partial one through questions and answers. This paper introduces VFI-PR, in which retrieved candidates produce visual feedback that serves as evidence for later questions and re-ranking. The authors build the VFI-PEDES dataset and evaluation protocol and propose VFI-ReID, which uses a Dynamic Accumulated-evidence Retriever and retrieval-aware question optimization. Experiments show consistent retrieval gains across interaction stages and outperformance over existing interactive baselines.
In briefDeleting Weights in LLM Post-Training: A Surprisingly Effective Method
SummaryThis paper asks how little a pretrained large language model must change to gain new abilities. The proposed Deletion Tuning method learns a binary mask over the gate projection of transformer blocks and zeroes fewer than 0.05% of weights, with no other changes. The authors report it matches full fine-tuning across reinforcement learning, supervised fine-tuning, and on-policy distillation, with masks hundreds to thousands of times smaller and far less forgetting.
In briefLearning noisy rigid-body dynamics on Lie groups with a stochastic physics-based Gaussian process model
SummaryThe paper introduces LieSPHGP, a stochastic Gaussian-process framework for learning rigid-body motion. It represents motion on Lie groups and Lie algebras to avoid singular parameterizations, and models mass, potential energy, dissipation, and control as separate components, with a diffusion term for stochastic forcing. Experiments on noisy rigid-body benchmarks report better prediction accuracy, long-horizon stability, and manifold-constraint satisfaction than neural dynamics baselines.
Thribhuvan Rapolu, Luke McLennan, Chandrajit Bajaj
In briefSWXAI: A unified library for explaining AI models in space weather and heliophysics
SummaryThis paper presents SWXAI, an open-source library that unifies explainable AI workflows for image, time-series, and tabular data in space weather forecasting and heliophysics. It offers a common, extensible interface across widely used vision, time-series, and tabular learning models, along with examples, tests, and documentation. A benchmark against direct backend implementations shows that SWXAI substantially reduces the code needed to produce an explanation, with no practical runtime cost.
Benedict Antonio Mervyn, Anli Ji, Manolis K. Georgoulis et al.
In briefStructure-separated DAG learning with multi-scale normalized closure
SummaryThis paper addresses continuous structure learning of directed acyclic graphs (DAGs), where existing methods enforce acyclicity through scalar functions of a weighted adjacency matrix. The authors add matrix-valued closure logits that explicitly score multi-hop relations for each ordered pair, supervised alongside data fit, sparsity, and a DAGMA-style barrier. On unstandardized Gaussian linear and nonlinear ER benchmarks, the models outperform DAGMA on denser graphs and are close to it on sparse ones.
In briefOrthogonal watermarks that resist colluders in image fingerprinting
SummaryVectorMark assigns each user a near-orthogonal high-dimensional unit vector as a watermark key and identifies colluders by cosine similarity to the extracted watermark. Experiments on MS-COCO, OpenImages, and AI-generated images show it keeps robust attribution under photometric and geometric distortions. It achieves the best strict all-colluder detection among evaluated methods while preserving high visual quality.
Zhakshylyk Nurlanov, Tobias Wei\ss{}berg, Florian Bernard
In briefVolterra Networks for Self-Contained Pixel Diffusion Denoising
SummaryThis paper proposes a pixel-space diffusion architecture that replaces softmax attention with Volterra layers as its mixing operators, requiring no pretrained components. The authors prove that softmax attention and any normalized mixing rule, including linear attention, cannot realize polynomial maps beyond affine ones, whereas Volterra layers can. On ImageNet at 256×256 and 512×512 resolution, the model is reported to be competitive with comparable self-contained pixel diffusion models at matched parameter and FLOP budgets.
Siddharth Roheda, Rohit Chowdhury, Amit Satish Unde et al.
In briefA self-contained data cleaning method for unsupervised anomaly detection
SummaryThis paper addresses contamination of training data with anomalies in unsupervised anomaly detection, proposing a preprocessing framework that does not require a known contamination ratio. During the early training stage of an auto-encoder, it analyzes the cosine similarity between each data point and its reconstruction, then removes anomalous candidates from the training set. Evaluated on 8 models and 10 security-relevant datasets, it reports up to 100% contamination reduction, with average gains of 11.2% in F1 score and 7.3% in classification accuracy.
Jordan F. Masakuna, Ahmed Yacine, Djeff K. Nkashama et al.
In briefRepresenting tensors as function compositions for faster computation on sparse data
SummaryThis paper proposes a mathematical formalism that represents tensors and operations on them as simple function compositions, where tensors map indices to values. Element-wise operations, re-indexing, and reductions are each expressed as compositions over values or indices. Identities between these compositions allow expressions to be rewritten into equivalent forms without evaluation, such as converting a sparse matrix product into one that computes only non-zero products. The authors argue that using this algebra as an intermediate representation could reduce codebase complexity, and they have made a draft implementation publicly available.
In briefA geometry-guided feature-space framework for interpretable unsupervised 4D-STEM analysis of multiphase materials
SummaryThis paper presents a machine-learning framework for analyzing four-dimensional scanning transmission electron microscopy (4D-STEM) data from heterogeneous materials. It combines physics-informed geometric descriptors with weighted high-dimensional embeddings and builds a regularized descriptor space before clustering. Validated on Si/SixGe1-x epitaxial layers and multiphase Ti-Nb alloys, the method yields higher cluster fidelity, greater interpretability, and substantially lower dimensionality than a pretrained vanilla autoencoder and model-free pattern vectorization.
In briefKeyASRBench: Measuring Whole-Phrase Keyword Preservation in Real-World Speech Recognition
SummaryThis paper introduces KeyASRBench, a bilingual Chinese-English pilot benchmark that checks whether speech recognition transcripts preserve keywords such as names, products, technical terms, versions, and relational phrases. It uses occurrence-based Keyword Error Rate (KER) computed over frozen whole-phrase annotations. On a challenge set of 1,501 clips with 4,723 keyword occurrences, five locally inferred models and two official API systems show MER/WER between 5.31% and 12.91%, while KER ranges from 13.61% to 33.81%. A disjoint 30-minute model-blind control set yields KER of 10.65% to 27.31%, indicating that errors on selected phrases extend beyond the challenge set.
In briefMultiTurnRoute: A platform for building and testing routers that pick language models across multi-turn chats
SummaryThis paper presents MultiTurnRoute, a platform for building and evaluating routers that choose which large language model answers each turn of a conversation. The routers use the current query, dialogue history, and each model's reusable prefix length. The proposed Chain-of-Route algorithm reduces the normalized cost needed to match the strongest model's quality by up to 20.3% compared with the strongest routing baseline on WildChat and LMSYS-Chat-1M, while matching or improving peak response quality.
