In briefTC-ADA: One-Shot Active Domain Adaptation for Semantic Segmentation
SummaryThis paper addresses the high cost of dense manual annotation when adapting semantic segmentation models to new driving environments. It proposes TC-ADA, a one-shot active domain adaptation method that selects and fully annotates a fixed set of target images in a single round before uninterrupted adaptation. The selection stage combines vision foundation model features with predictions from a fixed unsupervised domain adaptation model, without target labels, while adaptation jointly uses labeled source, labeled target, and unlabeled target data with calibrated supervision. The authors report consistent gains over representative baselines across five transfers and performance within 1.9 mIoU points of full target supervision using only 23 to 140 labeled target images.
In briefMoGround: Measuring and Reducing Modality Distraction in Vision-Language Models
SummaryThe authors release MoGround, a dataset spanning four visual domains where each question's answer is available in exactly one modality. This allows measuring modality distraction, where a model answers correctly from one modality but gets it wrong after irrelevant content from the other is added. Across seven open-source vision-language models, distraction varies by model and scales inversely with grounding strength. A weight-space robustness vector trained on one MoGround split reduces distraction by 9% to 51% across all seven models, costing only 0.1 average points on standard multimodal tasks.
In briefAPEX: An Extensible Agent-Assisted Model for Production Scheduling
SummaryThis paper presents APEX, an extensible production scheduling framework built on a general model and hybrid multiobjective search, with agents assisting data preparation, scenario exploration, and implementation of new constraints and objectives. On 69 public job-shop, flexible job-shop and flow-shop instances, eight APEX configurations were compared with NSGA-II, SPEA2, MOEA/D and SMS-EMOA. Hybrid configurations achieved the best aggregate solution quality, though the leading method varied by problem class and objective. Synthetic workflow and coding evaluations reported that all requested changes passed independent checks.
In briefGraph-Guided Construction of Repository Execution Environments (Graph2Env)
SummaryThis paper presents Graph2Env, an agent-based method centered on DepGraph, a typed dependency graph that represents the environment requirements, their dependencies, and their states for running a code repository. It refines the graph using execution feedback and saves successful repairs as a replayable procedure, then checks reproducibility in a fresh environment. On 200 Python repositories, it reaches an 81.0% Environment Build Success Rate and a 59.3% Environment Setup Success Rate, exceeding the strongest baseline by 9.5 and 9.0 points.
In briefUsing Temporal Graph Learning on Wearable Activity and Sleep Data to Model Adolescent Crystallized Intelligence
SummaryThe study uses about 21 days of daily Fitbit summary data from 6,091 adolescents in the ABCD study (Release 5.1) to predict crystallized intelligence. It proposes SATURN, a temporal graph regression network that represents each participant as 21 daily nodes. In family-controlled cross-validation, SATURN reached R² = 0.2783, outperforming Gradient Boosting and BiLSTM baselines. Light activity, metabolic equivalents, and sleep duration were identified as the dominant predictors.
Md. Tanvir Rahman, Nabil Anan Orka, Asaduzzaman Khan et al.
In briefTeacherGRPO: Closing the Gap Between Teacher and Student in Reasoning Distillation
SummaryThis paper addresses the problem that when a large teacher model is distilled into a smaller student model for reasoning, the gap between their capacities can degrade performance. The authors propose Teacher Alignment, which adapts the teacher toward the student's distribution without discarding data or lowering reasoning quality, and implement it as TeacherGRPO, a reinforcement learning method built on Group Relative Policy Optimization. The method includes curriculum selective alignment and importance-adaptive length regularization, and the authors report significant gains over baselines across diverse reasoning benchmarks and distillation methods.
In briefA lighter method that tunes self-supervised models toward specific tasks before fine-tuning
SummaryThis paper presents BiSSLight, an efficient successor to BiSSL, a method that inserts a bilevel-optimized stage between self-supervised pretraining and fine-tuning. BiSSLight pairs an M-FAC-based implicit gradient approximation with LoRA-based parameter-efficient tuning to make the approach practical for larger models. The authors report consistent downstream gains across several tasks and architectures, with larger gains at bigger model sizes. On a ViT-H backbone, they report computation time cut by more than a factor of ten compared with BiSSL.
In briefExplainable deep learning maps brain network biomarkers of adolescent intelligence from resting-state scans
SummaryThis paper presents an explainable deep learning framework that predicts fluid, crystallized, and total intelligence from resting-state fMRI of 5,285 adolescents in the ABCD study. Reported Pearson correlations of 0.44, 0.58, and 0.56 represent 6 to 9 percent improvements over existing approaches. Three explainability methods produced near-identical feature rankings, and consensus maps show primary predictive hubs within canonical systems while the strongest global pathways often bypass them through distributed long-range connections.
Md. Tanvir Rahman, Nabil Anan Orka, Asaduzzaman Khan et al.
In briefRecovering a Measure's Lower-Dimensional Semialgebraic Support from Moments
SummaryThis paper addresses recovering the compact, semialgebraic support of a measure from its moments when the support has codimension at least one and the measure need not be finitely atomic. The authors combine moment-matrix kernel information with the Christoffel–Darboux kernel to obtain a discrete approximation of the support. They validate the method on analytically computed moments and pseudo-moments from polynomial optimization problems without unique global minimizers.
Ruben Karapetyan, Shenyuan Ma, Ales Wodecki et al.
In briefTT-VidT: A video pretraining method that separates the time axis for efficient motion learning
SummaryThis paper compares video self-supervised learning methods under matched conditions, testing 24 combinations of architecture and objective at roughly 170M to 190M encoder scale. It proposes TT-VidT, which pairs a DINOv3-initialized ViT-B/16 per-frame spatial path with a compact Temporal Transfer Layer trained by Diff Compression. The authors report leading results on several motion-focused fine-tuning benchmarks while using fewer encoder FLOPs than DisMo, VideoMAE, and V-JEPA2. Results on HMDB51, IARD, and EPIC-Kitchens are noted as bounding the claims.
Shih-Ying Yeh, Daniel Z. Kaplan, Xuehai Wang et al.
In briefMemory summaries tuned to future query usefulness: the MemSuit method
SummaryExisting summarizers for memory systems are usually tuned for human-facing qualities like faithfulness, which does not match their real goal of keeping evidence useful for future queries. The paper proposes MemSuit, a self-distillation framework where a teacher summarizer, conditioned on query-answer pairs, writes utility-aware memory entries that a student learns to reproduce from raw conversation alone. The authors report that MemSuit consistently outperforms state-of-the-art baselines across a diverse set of conversational query types.
In briefEffect of k-NN Graph Preprocessing on Fair Graph Neural Networks
SummaryThis paper proposes a method for designing fair graph convolutional networks, where graphs are built with k-nearest neighbor preprocessing and fairness is measured by equalized odds. A fairness-driven loss term is added to the optimization objective. Experiments on three real-world tabular datasets with known biases examine how the neighborhood size k and fairness constraints affect classification performance and disparities in false positive and false negative rates across protected groups.
Nikolaos Zafeiropoulos, Emmanouil Mavrikos, George E. Tsekouras
In briefTTRSD: Test-time reinforcement learning with self-distillation for vision-language models
SummaryThis paper proposes TTRSD, a framework that lets vision-language models adapt at test time using only unlabeled inputs. It combines multi-view answer-level self-distillation with selection of visually sensitive tokens, so learning feedback targets perception bottlenecks. With only 20 unlabeled samples, the method improves results across seven benchmarks and three models, raising InternVL3-2B's MMMU accuracy from 35.79% to 49.32%.
In briefResolving State-Representation Mismatch: State-Space Visual Reasoning for Open-Loop VLA Planning
SummaryThis paper addresses a state-representation mismatch in open-loop planning for vision-language-action (VLA) models, which must simulate action-driven state changes from only an initial observation. The authors propose State-Space Visual Reasoning (SSVR), which separates static visual context, language constraints, and a recurrent latent state updated by an action-conditioned GRU. Using Qwen2.5-VL, SSVR substantially outperforms prior methods on FrozenLake, Maze, and MiniBehavior.
In briefMetaBench-Harness: a dual-loop search method that optimizes the benchmark generation workflow end to end
SummaryStatic benchmarks for large language models saturate quickly. This paper proposes MetaBench-Harness, a dual-loop framework in which an inner loop uses a benchmark harness to generate a new benchmark each round, while an outer meta-harness layer iteratively refines harness implementations based on historical evolution trajectories. Tests on CodeContests and AIME-2024 show the evolved benchmarks are challenging and discriminative for frontier models, with improvements across rounds in evolution reasonableness, benchmark competency, and evaluator robustness.
In briefFoldAttention: Softmax attention with a pre-declared reference for faster decoding and deterministic backprop
SummaryThis paper addresses the high cost of attention when language models decode long contexts, which repeatedly read a growing KV cache. It proposes FoldAttention, which fixes a reference value before scanning keys so each weight is final when computed and partial results can be added without rescaling. On H100 the authors report faster decoding than the fastest BF16 baseline on several models, and a faster deterministic backward pass.
In brief"Dense Is Not Enough": a paper on allocating teacher guidance more effectively when distilling AI agents on long tasks.
SummaryThis paper studies on-policy distillation, where a smaller model learns from a larger teacher's feedback on its own rollouts, in long-horizon agentic tasks. The authors argue that uniform token-level supervision allocates guidance poorly and propose LENS-OPD, a coarse-to-fine framework with Locate, Validate, and Refine stages. Experiments across multiple long-horizon agent benchmarks and student-teacher setups report consistent gains over vanilla OPD and several curriculum- and selection-based baselines.
In briefStarBOA: real-time Mamba state-space unrolling for reconstructing sparse radar micro-Doppler signals in integrated sensing and communications networks
SummaryThis paper addresses radar sensing in Integrated Sensing and Communications (ISAC) under chirp subsampling with up to 90% missing data. It proposes StarBOA, which replaces an attention-based model with a causal Mamba state-space model that updates incrementally per window using a persistent state. The authors report that StarBOA outperforms the baseline's published results at all sparsity levels, processing each window in 1.53 ms with zero lookahead.
Mustafa Bora \c{C}elik, Ceren \c{C}elik, Orhan Gazi
In briefANCHOR: Searching Crystal Compositions Around a Frozen Structure Predictor
SummaryThis paper proposes ANCHOR, which keeps a crystal structure prediction (CSP) model frozen as a physical prior and trains a GRPO composition policy with multi-objective rewards, including a novelty score (CAN). On the same CSP backbone, replacing DNG compositions with ANCHOR's policy raises MSUN from 11.4% to 47.6% and SUN from 1.1% to 22.1%. The authors note that the on-hull rate depends on the potential used for evaluation.
Emma Lei Hovmand, Jonas Elsborg, Melih Kandemir et al.
In briefSeparating the two roles each token plays when training language models
SummaryThis paper studies how a single token affects autoregressive pretraining, separating its role as a prediction target from its role as context. Using controlled corruption, the authors find a reversal: making a noisy token easier to predict reduces its damage as a target but increases it as context. The same decoupling is used to help explain how language models generate text.
In briefDataMagic: Creating Data Videos from Raw Data with Declarative Multi-Agent Orchestration
SummaryThis paper presents DataMagic, a system that creates data videos, which combine charts, voice narration, and animations, directly from raw tabular data. It uses a declarative specification, DVSpec, to link charts, narration, and animations to the data and keep them synchronized, plus a multi-agent approach that generates scenes in parallel and then refines narrative coherence. On 109 real-world samples, DataMagic raised quality from 2.13 to 3.89 out of 5 for GPT-5 and kept execution success above 95%. A user study reported a 79.7% reduction in task time compared with a conversational LLM workflow.
In briefVaME: Variational Latent Reasoning for Multimodal Embeddings
SummaryThis paper proposes VaME, a framework that learns compact embeddings for universal multimodal retrieval by modeling latent reasoning as a learnable distribution over trajectories. It combines Variational Latent Reasoning with reinforcement learning guided by a Semantic Decoding Reward. On the 78-task MMEB-V2 benchmark, VaME outperforms most explicit CoT-based models and all latent-reasoning baselines, with at least a 4.25x inference speedup over deterministic latent autoregressive baselines.
In briefStreaming-capable detection of action boundaries in untrimmed video without class labels
SummaryThis paper addresses action boundary detection, which splits untrimmed video into intervals without labeling their actions. The authors propose an adapter that works on pre-extracted video features, learning temporal representations with groupwise selective scans and decoding them into boundary timestamps. Its class-agnostic boundary F1 scores are 0.457 on Breakfast, 0.622 on GTEA, and 0.611 on 50Salads. A stateful variant supports streaming inference with zero neural look-ahead and bounded memory.
In briefSciGen-Verifier: A Multimodal Reasoner for Explainable Scientific Image Verification
SummaryThe paper addresses the difficulty of verifying scientific images such as circuits, geometric constructions, and function plots, where existing verifiers mostly target natural images and output only scalar scores. The authors introduce SciGen-Verify, a benchmark with a three-tier protocol covering binary judgement, explanation, and corrective editing instructions. They also present SciGen-Verifier, a reasoning-driven multimodal verifier trained with cold-start supervised fine-tuning and two-stage reinforcement learning, which reportedly performs competitively against much larger proprietary models and can serve as an online critic for iterative image correction.
In briefCOEVO: Jointly Evolving Context and Parameters for Recursive Self-Improvement
SummaryThis paper studies recursive self-improvement in large language models, arguing that parameter updates and context optimization are usually studied separately despite interacting with each other. It proposes COEVO, a framework that updates model parameters from on-policy experience while adapting the contextual guidance based on the evolving policy's state, using policy entropy and prompt-conditioned attention as signals. According to the experiments, COEVO consistently improves task performance over fixed-context reinforcement learning and yields policies more robust to changes in system prompts.
In briefCoViST: Compressing Image Tokens via Composable Visual States
SummaryThis paper proposes CoViST, a training-free framework that compresses visual tokens for vision-language models by representing each compressed image as a composable visual state. The state records representative features, original positions, contribution weights, and selection metadata. On seven LLaVA-1.5-7B benchmarks, CoViST-Fixed retains 99.9%, 99.5%, and 98.1% of uncompressed performance at 192, 128, and 64 tokens. Code will be released publicly.
In briefKeeping Word Stems Intact: Morphology-Guided Pre-Tokenization for Nepali
SummaryThis paper tests whether splitting Nepali words into stems and affixes before byte-level BPE tokenization helps. The authors present Papaya, a pre-tokenizer built on a finite-state transducer from a published Nepali grammar, which reaches 0.96 boundary F1 on 607 annotated words. In a 17M-parameter language model it lowers bits per byte by about 1% at equal training steps, while downstream gains are small.
In briefAudio-visual cross-modal translation for detecting video deepfakes via latent denoising
SummaryThis paper proposes CTCLD, a framework for detecting video deepfakes using both audio and visual signals. It performs bidirectional latent denoising, with each modality conditioned on the other, to transfer information across the two domains. The authors report that this captures subtle inconsistencies in manipulated content and achieves competitive detection performance.
In briefAutoHGNN: Robust and Efficient Architecture Search for Hypergraph Neural Networks
SummaryThis paper presents AutoHGNN, a neural architecture search framework that automatically designs hypergraph neural networks instead of relying on manual crafting. It adds a Hyper-Interaction Module to the search space to handle the mismatch between standard graph designs and hypergraph data, and uses Hypergraph Stable Topological Distance as a criterion for selecting architectures. Experiments on benchmark datasets report that AutoHGNN outperforms manually designed and automatically searched baselines in classification accuracy and time efficiency.
In briefBeyond Timestamps: Aligning Decisions for Long-Horizon Agent Training
SummaryThis paper addresses a mismatch in training long-horizon agents, where privileged guidance may not line up with the student's decision timing. The authors propose AlignOPSD, which first rectifies supervision in functionally matched contexts and then assigns credit over variable-duration decision spans. Evaluated with Qwen2.5-3B and 7B on ALFWorld, WebShop, and Search-QA, it outperforms GRPO and StepOPSD in all eight comparisons.
In briefFaster text generation for diffusion language models via better token selection
SummaryDiffusion language models can predict several tokens in parallel to speed up generation, but these parallel predictions do not fully account for the tokens chosen earlier in the same block. The paper introduces BRISK-DLM, which trains on self-generated sequences using risk-reward weighting by position and adds a lightweight prefix-conditioned corrector that reranks candidates at inference. The authors report end-to-end throughput gains of up to 37.4% while preserving task quality.
In briefFrom Grey-Box to Green-Box: Can Physics-Informed Machine Learning Cut Carbon Footprints in Structural Health Monitoring?
SummaryThis paper examines whether physics-informed machine learning, or grey-box models, can reduce the environmental cost of training in structural health monitoring. Four approaches spanning Gaussian processes and neural networks are compared: residual modelling, input augmentation, hybrid modelling, and constrained learning. In most cases, the physics-informed models produced lower training emissions to reach the same error threshold, with input augmentation as an exception. The authors note a trade-off between added model complexity and reduced training data needs.
Daisy R. Bradley, Nathan A. Hinchliffe, Daniel J. Pitchforth et al.
In briefSivia: A System That Learns to Draw and to Learn, Generating Paper Overview Figures
SummaryA paper's overview figure is usually the first thing readers see, so it must explain the method accurately and look polished. Sivia reads a paper, plans the relations among its components, adapts visual layouts from recent top-conference papers, and composes a design for image generation. It improves through human-guided recursive self-improvement at two levels: its drawing skill and the updater that proposes revisions to that skill. On 48 test papers, learning both levels reached 93.54% required-relation accuracy, and readability and visual quality rose to 4.70 and 4.78 on five-point scales.
In briefOLED-MoE: An expert offloading system that speeds up MoE-based diffusion LLM inference by reusing experts across denoising steps
SummaryThis paper addresses semi-autoregressive diffusion large language models (dLLMs) that use mixture-of-experts (MoE) layers, whose expert parameters exceed GPU memory. Instead of prefetching experts within each iteration, the proposed system OLED-MoE keeps high-value experts in GPU memory across adjacent denoising iterations, guided by token confidence, and handles cache misses through CPU-GPU cooperative execution. The authors report 1.23x to 7.93x lower time per output token and 1.44x to 4.23x higher expert cache utilization than state-of-the-art offloading systems. The source code is publicly available on GitHub.
In briefPulseQuant: Propagation-Guided 4-Bit Quantization for Video Diffusion Transformers
SummaryThis paper introduces PulseQuant, a 4-bit post-training quantization method for video diffusion transformers that uses trajectory sensitivity and activation geometry to guide offline calibration. The authors show that short-horizon propagated error predicts final latent error more reliably than immediate block-output error. Evaluations on Wan models, Self Forcing, and MiniMax-H3 report improved key consistency and dense-reference metrics while remaining competitive on other attributes.
In briefLocal LMO Turns Out to Be a Projection Method
SummaryThe paper shows that the Local Linear Minimization Oracle (Local LMO), which avoids computing projections, equals a Euclidean projection onto the feasible set intersected with a separating half-space when its radius does not exceed the Polyak radius, and this projection is computable using only a linear oracle. The authors place Local LMO in a broader family of projection methods indexed by the depth of the half-space. For objectives with ϑ-Hölder continuous gradients, sufficiently deep members of this family achieve the universal rate O(K^{-(1+ϑ)/2}) over the first K iterates.
In briefNLPG: Self-Evolving Language Agents via Natural-Language Policy Gradients
SummaryThis paper proposes Natural-Language Policy Gradients (NLPG), an external policy-memory method that improves a fixed language agent without changing its model parameters or program structure. It analyzes execution traces, propagates downstream feedback backward through the module graph, and turns recurring failures into route-local natural-language corrections aggregated into bounded policy updates. Across six benchmarks covering memory, reasoning, instruction following, and evidence verification, NLPG outperforms the strongest listed baseline on each by 8.71 percentage points on average.
In briefReusable Intervention Guides for Robot Manipulation via Recursive Harness Distillation
SummaryThis paper proposes Recursive Harness Distillation, in which a strong agent turns its experience of interacting with vision-language-action manipulation policies into a playbook for a lighter agent. The playbook is then refined recursively using the light agent's execution feedback, without updating model parameters. In real-world manipulation, success rose from 37.3% to 64.0%. On SimplerEnv Bridge, the light agent with the playbook reached 66.7%, compared with 41.7% for the GR00T-only baseline.
In briefOptimal Transport Dropout: learning structured uncertainty for neural predictions
SummaryThis paper proposes Optimal Transport Dropout (OTD), a method that lets neural networks and neural operators output structured predictive uncertainty instead of only point predictions. It starts from a simple independent reference distribution and learns a flow that transports latent perturbations, trained with the Energy Score and a kinetic regularizer. On a PDE surrogate task, predictive dispersion aligns with spatial error patterns, and OTD gives more accurate and better-calibrated intervals than Monte Carlo dropout.
In briefSpectral Embedding Method for Detecting Auto Insurance Fraud with Machine Learning
SummaryThis paper studies fraud detection in automobile insurance, comparing classifiers such as Decision Tree, Random Forest, XGBoost, LightGBM and Gradient Boosting Machine. It uses MWMOTE and TGAN to address data imbalance and Topological Node Feature (TNF) based spectral embedding for low-dimensional representation, compared with MDS, Isomap and t-SNE. Across 65 model combinations, the authors report that MWMOTE, TNFSE2 and Random Forest performed best on the dataset.
Rohan Yashraj Gupta, Lalith Srikanth Chintalapati, Satya Sai Mudigonda et al.
In briefWhich Features Hold Up When Speech Codecs Change: Pooled No-Vocals Residuals for Speech Deepfake Detection
SummaryThe paper compares 12 acoustic representations with a shared linear classifier, finding that hierarchical XLS-R performs best on the pooled test set while pooled no-vocals residual statistics perform best on unseen codecs. Based on this, the authors propose MN-P, a dual-view detector that fuses these features through adaptive gating. MN-P reduces EER by 54.2% overall and by 60.9% on the codec-unseen condition relative to the best retrained state-of-the-art system.
In briefIdentifying a semimartingale's predictable drift from marginal laws
SummaryThe paper studies how to recover the predictable finite-variation drift of a semimartingale when it is observed only through repeated cross-sections. It treats the drift as the output of a hidden linear dynamical system and formulates the joint estimation as a bilinear quadratically constrained program, solved to certified global optimality by spatial branch and bound. The authors also derive estimator convergence rates at a fixed mesh.
Jakub Marecek, Enrico Biffis, Abigail Langbridge et al.
In briefEvaluation Rules Can Change Which Graph Anomaly Detector Wins
SummaryThis paper audits open-set graph anomaly detection, where a few labeled anomalies from one class are used to find unlabeled anomaly classes. It examines common reporting practices, including selecting the best epoch on the test set, copying baseline numbers, and treating minority classes as anomalies. The authors re-run DEMO, NSReg, and their own OUTPOST under one protocol and score each run with both a best-epoch rule and a validation rule. They find that the scoring rule changes which method leads.
In briefDrafTS: Time-aware decomposition with residual correction for more accurate time series modeling
SummaryThis paper proposes DrafTS, a model-agnostic framework for time series modeling. It uses features from instantaneous amplitude and frequency to decompose a series into a primary component meant to capture underlying dynamics. A task-specific backbone models that component, and a lightweight correction module uses residual information to refine the backbone's output. The authors report that DrafTS improves six diverse backbones across four time series modeling tasks.
In briefPreSO: Isolating Privileged Representations for On-Policy Self-Distillation
SummaryThe paper proposes PreSO, a method for on-policy self-distillation that uses compact hints to limit solution-specific privileged detail. It estimates a low-rank subspace from representation shifts pooled across related problems and adds the projected correction to the student's unprivileged representation as a distillation target. On five competition-level math benchmarks, PreSO outperforms strong baselines on Qwen3-1.7B, 4B, and 8B, improving average accuracy over the base model by 6.8, 9.0, and 10.0 points.
