In briefFaster federated learning for imbalanced classes via distributed data augmentation
SummaryThis paper addresses class imbalance in federated learning, where minority classes perform poorly. Existing methods augment minority samples heuristically without establishing how the augmentation amount affects convergence, which can cause excessive augmentation and longer training. The authors first relate augmentation to federated learning's convergence behavior, then propose DAFL, a distributed augmentation framework that sets the minimum augmentation for each client-class pair. Experiments report that DAFL improves minority-class F1-score while substantially reducing training time.
In briefWhen Verifiable Counts Depend on Wording: Auditing Wording Robustness in Instruction Following
SummaryThe paper argues that verifiable instruction-following benchmarks usually express each constraint with one fixed template, so scores may not reflect stable ability. The authors introduce WISE, a matched evaluation suite that keeps each operational requirement fixed while varying its wording, using exact word count, keyword inclusion once, and an 8–12 word range. They find that wording alone causes substantial shifts in compliance, along with changes in failure profiles and model rankings.
In briefMind the Gaps: Closed-Form Repair of Code Models After API Changes
SummaryThis paper studies how to repair code language models when library APIs change and the model keeps generating outdated interfaces. On executable API evolution tasks in Python and Rust, the authors find that neurons identified by failure attribution often differ from those that can carry a patch, and that a generic update does not fit every failure. They propose ASTRA, which selects neurons by contrastive semantics and solves one small linear system in closed form to correct all failing tokens jointly. The reported average Pass@1 is 66.7 percent, versus 48.4 percent for the best baseline, with about 2.7 seconds per sample repair.
WhyIdentifies two gaps in neuron-based repair and proposes a closed-form method, validated against baselines across three code models and two languages.
In briefAutomatically extracting claims from social media at large scale
SummaryThis paper presents a pipeline that uses a large language model to inductively extract and catalogue short declarative claims from large social media corpora. It is applied to two Twitter datasets, one on the 2020 US presidential election and one on the 2022 FIFA World Cup. The authors evaluate recall and precision against manually annotated samples, run ablation studies, and perform a qualitative error analysis, then present the resulting claims catalogue for each dataset.
In briefAnswering with Visual Evidence: Consistency-Aware Grounded VQA for Roadside Traffic Scenes
SummaryThe paper addresses a failure in grounded multimodal models where textual answers contradict the bounding boxes the model localizes. The authors introduce RoadSceneVQA-G, a 34.7K-pair benchmark linking each answer to its supporting boxes, along with an Answer-Grounding Consistency evaluation suite. They propose Enumerate-then-Answer and an enumeration-consistent reinforcement learning stage, raising say-point consistency from 26.6% to 93.7% and grounding F1 to 75.6%.
In briefWhen and What to Prune? Stage-Aware Visual Token Pruning for Efficient VLA
SummaryThis paper presents SAPrune, a training-free method that removes visual tokens to speed up vision-language-action (VLA) model inference. Instead of pruning at fixed layers, it uses a small calibration set to track how action-to-visual attention changes across layers and selects pruning layers once the attention pattern is more reliable. Each selected layer applies a dual-path rule that protects strongly attended tokens and keeps useful surrounding context. Experiments on LIBERO, SIMPLER, and real-world robotic tasks report 87.5% token pruning and up to 1.718x speedup with competitive task success rates.
In briefCombining drone delivery order pooling with road monitoring by turning monitoring tasks into orders
SummaryThis paper studies real-time dispatching of a shared drone fleet that handles both on-demand food delivery and urban road network monitoring. Congested road nodes with outdated traffic information are periodically converted into virtual monitoring orders and pooled with delivery orders. The authors propose Graph Multi-Agent Q-Learning combined with dynamic heterogeneous bipartite matching to solve the problem. Experiments on real-world data report a 25.1% monitoring improvement with under a 1% drop in delivery performance.
In briefA Unified Scaling Law for Time Series Foundation Models: How Model Capacity and History Support Forecasting
SummaryThis paper proposes a unified scaling law and a learning theory for time series forecasting, examining how model capacity and historical input length jointly affect prediction. Using 18,768 experimental cells from 21 checkpoints across 23 dataset-frequency tasks, the authors fit a five-parameter law that predicts the capacity-scaling curves of Toto 2.0 with mean absolute percentage errors of 1.09% and 1.50% at input lengths 2048 and 4096. The paper also uses Gaussian regression to study how frozen time series foundation models extract and reuse historical rule information through activations.
WhyA fitted five-parameter scaling law and Gaussian regression theory are validated on 18,768 experimental cells, offering solid evidence for time series foundation models.
In briefLoopy: A Low-Bit Quantization Framework for Looped Language Models
SummaryLooped language models scale test-time computation by repeatedly running a shared recurrent core, which is parameter-efficient. The paper proposes Loopy, a post-training quantization framework that selects low-bit shared-core representations based on final prediction loss at the target recurrent depth. It parameterizes candidates with channel scaling and orthogonal rotations, and progressively allocates calibration windows to promising candidates to approximate the selection efficiently. The authors report state-of-the-art results across eight settings, including a 36.5% LAMBADA perplexity reduction relative to SpinQuant on Ouro-1.4B under W4A4.
In briefTask-aware joint pruning and distillation for compact audio deepfake detectors
SummaryAdvances in speech synthesis make deepfake speech increasingly convincing, which is a growing security concern. Self-supervised learning detectors perform best but typically exceed 300M parameters, limiting on-device use. The paper proposes a framework combining cross-domain knowledge distillation with movement-guided structured pruning to compress the model while keeping forgery-discriminative knowledge. The model is reduced to 31.9M parameters with a 6.3× FLOPs reduction, and averages only a 1.30% performance drop across multiple datasets versus the uncompressed baseline.
In briefLearning from Imperfect Teachers for Low-Resource Acoustic Generalization
SummaryThis paper studies knowledge distillation for low-resource acoustic learning, where a teacher model trained on limited or imbalanced labels may give biased predictions. The authors propose BA-MPD, which combines Boundary-Anchored Correction and Mass-Partitioned Distillation to reduce harmful ranking errors and noisy low-confidence details. Experiments on two acoustic benchmarks under multiple label budgets show gains over supervised baselines and vanilla knowledge distillation, and it remains effective when teacher and student use different label budgets.
In briefHybrid-Basis Feature Forecasting: a training-free way to speed up diffusion sampling
SummaryThis paper proposes HybridFF, a training-free, plug-and-play method that speeds up diffusion model sampling by forecasting internal features. It fits several complementary basis families with moving least squares and combines their predictions using fusion weights. Two variants are described: a Fixed version with calibrated constant weights for efficiency, and an Adaptive version that updates weights online from reliability scores. Experiments on DiT-XL/2, FLUX.1-dev, SD3.5-Large, and HunyuanVideo report a favorable speedup-quality trade-off against single-basis forecasters and caching baselines.
Kai-Liang Cheng, Yuan-Yuan Cheng, Yu-fan Jin et al.
In briefAnchor and Adapt: Asymmetric Prompt Adaptation for Few-Shot Industrial Anomaly Detection
SummaryThis paper addresses few-shot industrial anomaly detection, where only normal images are available and provide no defect supervision. It proposes Anchor and Adapt, a two-stage prompt learning framework: the first stage learns transferable normal and abnormal anchors from annotated auxiliary data, and the second stage keeps them fixed while adapting an extra normal branch using the few target normal samples. Cross-dataset experiments between MVTec-AD and VisA under 1-, 2-, and 4-shot settings show competitive detection and localization performance.
In briefMGPO: Manifold-Guided Diffusion Alignment for Task-Aware Dataset Distillation
SummaryThis paper addresses a mismatch in diffusion-based dataset distillation, where density-focused models miss the decision boundaries needed for downstream tasks and tend to collapse onto a few high-density modes. The authors propose MGPO, which treats distillation as multi-objective reinforcement learning, combining a pixel-space discriminative reward with a latent-space geometric reward based on a class-wise minimum spanning tree. The method is also extended to object detection and segmentation by swapping the task reward model. Experiments report that MGPO outperforms existing methods, including an 8.0% mIoU gain on segmentation under low-budget settings.
In briefKolmogorov-Arnold Networks for Recognizing Personal Context in ExtraSensory Data
SummaryThis paper tests several Kolmogorov-Arnold Network (KAN) variants on the ExtraSensory dataset for personal context recognition. It compares them with a multilayer perceptron (MLP) and TabM, reporting average Macro-F1, Micro-F1, and training time over five user folds and three seeds. All KAN variants clearly outperform MLP and perform comparably to TabM, but they generally train more slowly. The authors conclude that KANs are promising, while computational efficiency remains a key challenge.
In briefArticuTable: Building Interactive, Jointed 3D Tabletop Scenes from One Image
SummaryThis paper presents ArticuTable, a framework that reconstructs a 3D tabletop scene from a single image and recovers executable part-level articulation for its objects. GRAM uses a multimodal large language model to guide joint fitting and semantic reasoning, turning imperfect monolithic proxy meshes into articulated assets. PSGSR progressively narrows pose and scale candidates under metric, planar, and input-view constraints to produce a layout consistent with the input view. The authors also release ArticuTable-100, a set of 100 simulation-ready tabletop scenes.
In briefTEMPLAR: An agentic framework that automatically induces and evolves standardized radiology reporting templates from large clinical corpora
SummaryThis paper presents TEMPLAR, a multi-agent framework that automatically derives and iteratively revises standardized radiology reporting templates from large-scale clinical corpora. The template serves as a persistent central state, supported by anatomical and diagnostic knowledge graphs. Across four datasets, TEMPLAR outperforms ASTAR, medical LLMs, and general-purpose LLMs in coverage, information fidelity, and diagnostic fidelity, and its advantage over ASTAR persists under cross-dataset transfer.
In briefStateWise: Checking and Repairing an AI Agent's Stored State Before It Acts
SummaryThis paper studies LLM agents that reuse stored operational records to guide later actions, noting that changed environments or requirements can make those records invalid. It proposes StateWise, which uses counterfactual replanning to find decision-critical records, verifies and repairs them, then replans and checks the action before execution. On 150 coding-agent cases with corrupted state, overall correctness rose from 38.7% for the baseline to 93.3%, with no unsafe actions reported.
In briefPythia: A foundation world model for forecasting time series with text and other context
SummaryThe paper introduces Pythia, a foundation world model that learns context-conditioned latent dynamics across datasets using a joint-embedding predictive architecture. It separates predictive representation learning from a probabilistic decoder that adapts to frozen representations. On the MUSE benchmark, Pythia-Tiny reduces MASE and WSQL errors by 6.26% and 5.00% relative to the strongest model on the published leaderboard.
In briefDTFormer: Text-Guided Alignment for Better RGB-D Image Segmentation
SummaryDTFormer is a segmentation framework that combines RGB images, depth data, and text. Its core Text-guided Semantic Alignment Module encodes text cues into semantic prototypes and aligns RGB-D features with them across multiple encoder and decoder layers. The authors report consistent gains on multiple benchmarks while keeping the method simple and efficient. The code is promised upon acceptance.
In briefComputational complexity of max-min item allocation with three-valued preferences
SummaryThis paper studies how to allocate indivisible items so that the worst-off agent's utility is as high as possible, when each agent's item or marginal values come from a small set. The authors give polynomial-time algorithms for some valuation classes, prove constant-gap hardness for others, and provide approximation algorithms and query lower bounds. Together, the results resolve open questions and give a complete picture of the complexity for ternary valuations.
WhyResolves open questions on max-min allocation with ternary valuations, giving algorithms, hardness, and approximation bounds with rigorous theoretical proofs.
Thi Ngoc Anh Vu, Trung Thanh Nguyen, Khaled Elbassioni et al.
In briefEMG-GPT: Predictive Pretraining on Quantized Muscle Signal Tokens for Hand Pose Estimation
SummarySurface electromyography (sEMG) is a low-power, low-cost biosignal usable for estimating hand pose and classifying gestures. This paper examines whether self-supervised pretraining on sEMG can produce representations that transfer to continuous hand-pose estimation. The authors introduce EMG-GPT, a causal transformer that takes discrete sEMG tokens from a frozen residual vector quantization tokenizer and learns temporal dynamics by predicting future codes autoregressively across depth. The model shows competitive results on both regression and tracking tasks, supporting EMG-only pretraining as a viable route to transferable sEMG representations.
Ettore Magni, Rolandos Alexandros Potamias, Stefanos Zafeiriou et al.
In briefPixReenact: Streaming head-avatar animation using pixel-conditioned causal video diffusion
SummaryPixReenact is a streaming framework that animates a reference face image to follow a live driving video. It conditions directly on VAE-encoded reference and driving frames rather than specialized identity or motion representations, and uses cross-identity pseudo supervision, long self-rollouts, and dual-teacher distillation. The authors report robust cross-identity reenactment across three benchmarks and identity retention over long streams. A 4-NFE rolling student emits four frames per update with a mean latency of 239 ms.
In briefVela: Scaling Vision-Language-Action Models with Adaptive Action Curves
SummaryThis paper presents Vela, a vision-language-action foundation model that represents future robot motion as continuous trajectories rather than fixed-rate action chunks. It uses a compact spline-based action representation with motion-dependent temporal support, letting a fixed output budget adapt its temporal resolution across motions. Vela is pretrained on large-scale multi-embodiment robot data and evaluated on LIBERO-X, EBench, and two real-world long-horizon tasks, egg-cake cooking and potato shredding.
In briefGFGE: A framework unifying explainable AI methods through interpretation
SummaryThis paper proposes the General Framework for Generating Explanations (GFGE), based on interpretative frameworks and the activities of sense-reading and sense-giving. It connects an analyst's interpretation of system evidence, the communication of a selected explanation, and the recipient's interpretation of it through the Interpret/Explain Schema. The framework is operationalised through five roles: data interpretation, model interpretation, output interpretation, optional post-hoc analysis, and aggregation. The authors instantiate it for attribution, surrogate, counterfactual, concept and prototype, intrinsic rule, argumentation, and language-model methods.
In briefLook Before You Leap: Letting LLM Agents Decide When to Retrieve External Knowledge
SummaryThe paper argues that many LLM agents follow a "Retrieve-Always" approach, relying on external information for every interaction, which can be wasteful and bring in irrelevant context. The authors propose MARTA, a neuro-symbolic framework that treats retrieval as a cost and retrieves only when the model perceives its internal knowledge as inadequate. By letting agents gauge the uncertainty of their own thoughts before acting, the approach aims to balance internal knowledge and external information more efficiently.
In briefRanking bandits for carousel interfaces that reveal how far users browse
SummaryCarousel interfaces let a recommender observe the maximum depth a user browses, distinguishing shown-but-unclicked items from unseen ones. The paper formulates this as a ranking-bandit problem and proposes three algorithms based on UCB, Thompson Sampling, and DMED that update only from observed exposures. The authors give logarithmic and asymptotic regret bounds, with DMED-based optimality in the limit. Experiments on synthetic and RecGaze-derived environments show lower final mean regret than PBM-UCB, with OD-TS similar to PBM-TS.
In briefKeeping AI Agents Safe While They Reason: A Method That Separates Logic Training From Safety
SummaryThe paper finds that fine-tuning language models on logical reasoning tasks can weaken their safety alignment, with a reported 23.3% penalty on alignment benchmarks. The authors introduce AURA, a spectral regularization method that constrains reasoning updates to the orthogonal complement of the alignment space. According to the paper, AURA recovers 23.0% of the lost performance while keeping cosine fidelity above 0.98 to the safe state.
In briefVisualBridge: Using time-series plot images to control synthetic time-series data generation
SummaryThis paper proposes VisualBridge, which uses time-series plots as a visual intermediate linking high-level temporal semantics to numerical sequences. A multimodal LLM converts plots into structured semantic representations, and a semantic editing policy is trained with downstream forecasting rewards. A temporal VAE then generates multivariate augmentations, which improve few-shot forecasting over conventional augmentation methods on public benchmarks.
In briefF²SLAM: Turning Feed-Forward 3D Geometry into Persistent SLAM Constraints
SummaryThis paper presents F²SLAM, a method that converts geometric priors from feed-forward 3D models into optimizable target-weight measurements within a persistent dense factor graph. A high-frequency stream maintains local tracking constraints, while a low-frequency stream uses wider multi-view context to selectively refresh measurements after a consistency check. Both streams jointly constrain poses, inverse depths, and optional intrinsics through a single dense bundle adjustment. Experiments on multiple benchmarks report strong trajectory estimation and improved dense reconstruction in calibrated and uncalibrated settings.
In briefLearning to select trustworthy synthetic data for few-shot class-incremental learning
SummaryThis paper addresses Generalizable FSCIL (G-FSCIL), where both base and incremental data are scarce. It generates class-specific synthetic samples with a frozen latent diffusion model and learns a selection policy that favors semantic consistency and visual diversity, distilled during the base session. The authors add boundary-stable incremental adaptation, including prototype initialization and bidirectional boundary calibration. They report that the method outperforms existing FSCIL baselines with less forgetting and better old-new balance.
In briefLearning Without Overwriting: Theory of Self-Distillation vs. Fine-Tuning for Continual Reasoning
SummaryThis paper models LLM reasoning as search over a directed acyclic graph and analyzes how on-policy self-distillation (OPSD) and supervised fine-tuning (SFT) behave in continual learning. It finds that OPSD with hints from correct outputs allows learning without forgetting through sparse gradient updates, while SFT on correct reasoning paths can cause catastrophic forgetting through dense updates. It also finds that diversity in pre-training is important for a post-trained model to reach correct outputs from intermediate states.
In briefEarly-step perturbation directions amplify more in later diffusion sampling
SummaryThe paper compares two kinds of perturbations in diffusion sampling: directions carried from an earlier step to a late latent state, and newly sampled isotropic noise of the same magnitude. For the samplers studied, the median paired shaped-to-fresh angular-gain ratio ranges from 1.14 to 2.35. Permuting shaped directions across trajectories weakens the effect, indicating the advantage depends on alignment with the receiving trajectory as well as shared directional structure.
In briefSame Predictions, Different Harms: Causal Auditing of Patient World Models
SummaryThis paper examines whether patient world models used in clinical trial simulation give reliable estimates of counterfactual harm, the fraction of patients harmed by switching treatment, even when they agree on transition probabilities and per-arm risks. The authors use a two-stage shared-response structural causal model, deriving exact harm bounds under stage independence and calibrated outer bounds when dependence or response-mismatch budgets are declared. Experiments on public clinical simulators include negative results showing little gain from joint compatibility, and all guarantees hold only under the stated causal model.
In briefBenchmark testing whether AI scientists can turn observed data into mechanistic hypotheses
SummaryThis paper introduces MechHypoBench, the first benchmark for testing whether AI agents and AI scientists can generate mechanistic hypotheses from empirical data. It pairs mechanisms drawn from papers across 14 scientific fields with real-world datasets containing 17.98 million records. The evaluation framework judges open-form hypotheses by their predicted consequences under withheld conditions. Experiments show a substantial gap between the hypotheses generated and the underlying mechanisms.
In briefLong-MDR: Long-context reinforcement learning for multimodal research agents
SummaryThis paper studies training AI research agents that handle long multimodal histories of searches and visual evidence. The authors scale online reinforcement learning to a 128k context and more than 75 tool-interaction turns. They propose Long-MDR, a three-part training recipe that combines distillation warmup, progressive horizon expansion, and entropy-triggered rescue. The authors state these methods improve learning efficiency and stability in this long-horizon setting.
In briefOrder Matters: Ordering Object Queries for RNN-Based Detection
SummaryDETR-style detectors assign object queries to objects with one-to-one matching, but without explicit de-duplication, several queries can still predict highly similar results for the same object, which makes training unstable. The paper proposes DETRNN, a plug-and-play module that orders queries by confidence and similarity from prior predictions and refines them with an RNN along that order to model competition. The authors report consistent accuracy gains across multiple DETR-style detectors with comparable efficiency.
In briefFORGE: A Verification-Gated Framework for Repairing Behavior in Generative Language Models
SummaryThis paper presents FORGE, a framework for repairing undesirable behaviors such as demographic bias and toxic generation in generative language models. It separates defect localization, weight editing, and behavioral verification into independent stages, converting localized defective generations into explicit optimization objectives. On five open-source LLMs, FORGE reduces bias and toxicity more than gradient-based fine-tuning with only minor perplexity degradation, and remains effective with few defective examples.
In briefPrediction from expert advice with six or more experts
SummaryThe paper proves that no single rank-ordered adversary strategy is globally optimal for prediction with expert advice when there are six or more experts, in both geometric stopping and finite time horizon settings. The proof rests on a leading-order correction term that applies when one expert moves far ahead of the others. This connects optimal strategies for n and fewer experts, drawing on recent results about the exact optimality set for the five-expert problem.
In briefRecurrent Latent Visual Search for GUI Grounding
SummaryThis paper proposes ReLaViS, a method that helps vision-language model-based GUI agents locate interface elements in screenshots. It performs multi-step visual search inside the model within a single interaction round, using a recurrent spatial search head. Training trajectories are built from flat element annotations to guide a coarse-to-fine search. Built on Qwen2.5-VL-7B, it raises ScreenSpot-Pro accuracy by 3.1 points to 56.3% with only a 3.5% FLOPs increase, and outperforms the matched single-step baseline on all five benchmarks.
In briefTIRMamba: A Small Thermal-Prior Network for Infrared Image Super-Resolution
SummaryThis paper proposes TIRMamba, a single-channel infrared image super-resolution network with about 0.9 million parameters, far smaller than the 26 to 37 million parameter Mamba-based methods, aiming for airborne and handheld deployment. It uses thermal prior cues to modulate a state-space trunk and trains with a replay strategy to offset the small infrared dataset. At scale factor 4, it matches the strongest protocol-trained methods on official test sets with 29 to 40 times fewer parameters and 2.8 to 9.4 times lower latency.
In briefAECG: Asymmetric Experience Consolidation and Governance for Multi-Agent Memory
SummaryThis paper addresses memory problems in LLM-based multi-agent systems, where repeatedly retrieved procedures can become outdated or wrongly reused outside their original coordination scope. It introduces AECG, which treats memory as a dynamic reliability-governance loop that detects degradation and prioritizes high-risk knowledge under a bounded review budget. Across three frameworks and four benchmarks, AECG achieves the best score in 11 of 12 settings, improving over the strongest competing memory method by up to 10.23 percentage points.
In briefLocality-sensitive hashing for p-exponential kernels, with density estimation applications
SummaryThe paper studies whether the p-exponential kernel exp(-||x-y||_p) can be realized by a locality-sensitive hashing (LSH) scheme. It shows this kernel is LSHable in bounded regions for all 1<p≤2, extending the previously known case p=1. The authors introduce a new "mosaic LSH" scheme based on a Poisson hyperplane process with hyperplanes sampled as ℓ1-biased p-stable vectors, and develop efficient sampling procedures. These results yield new efficient density estimation methods for these kernels.
In briefVerification Trap: Why Code Selection Fails Under False Premises
SummaryThis paper studies test-time selection in code generation, where systems sample many candidate programs and pick one using evidence visible to a verifier. The authors show that when the generator and verifier share a false task premise, they can be coupled by the same mistaken belief, so the selector may choose a program that fails hidden tests. They call this the Verification Trap. Experiments across three benchmarks and five code models show false premises lower correctness and increase mis-selection.
In briefCreativeFlow: a method that transfers one-to-many analogical relations to generate diverse 3D assets
SummaryInspired by cognitive science, this paper presents CreativeFlow, an analogical generation framework that models analogical divergent thinking to counter creative homogenization in text-to-3D pipelines. The method produces a series of source-target asset pairs that are relationally similar yet have distinct geometric configurations. Expert evaluations reported that the framework substantially improves creative novelty and visual appeal. The workflow and its generated assets are presented as a foundational dataset and benchmark for future relation-aware 3D model training.
In briefA Safe Action Is Not Enough: Feasible-Future Decoding for Robot Vision-Language-Action Policies
SummaryThe paper argues that a safe action may still be non-viable, because a frozen vision-language-action policy can pick a locally admissible move that leaves no safe route to finishing the task. The authors derive an exact target for safe task completion and propose VICS-G, a training-free reranker that approximates it. Across six Safety-CHORES settings, VICS-G reduces mean cumulative safety cost by 1.9% to 57.5% while staying within 2.5 points of policy sampling in success.
In briefMemAdapter: Counterfactual Adaptation to Curb Memory-Induced Sycophancy in AI Agents
SummaryThis paper examines how long-term memory in LLM-based agents can make them over-agree with users' past beliefs, even when those beliefs are wrong or outdated. The authors propose MemAdapter, which uses counterfactual induction to surface risks in retrieved memories, context-aware reflection to adjust each memory's influence for the current task, and evidence-based reasoning to ground responses. According to the paper, experiments on three benchmarks show improved memory reliability across diverse scenarios.
In briefBeyond Instruction Following: Training Agents to Execute Expert Skills Step by Step
SummaryThis paper introduces Grounded Skill-Following, a task where an agent executes a fixed, expert-written skill across all its required phases while basing decisions on environment observations. Each skill is formalized as a skill contract combining visible instructions with an explicit runtime that defines phases, allowed actions, transitions, and termination. The authors train with Verified Progress Credit and Contract-State Feedback, and measure compliance with the Protocol Completion Rate. Trained Qwen3.5-4B reaches 99.27% and 99.96% protocol completion on Math and Search, with task outcomes slightly above original baselines.
In briefCross-modal learning to retrieve immunotherapy-linked molecular signatures from tissue slides
SummaryGastric adenocarcinoma is a major cause of cancer death, and an inflamed/non-inflamed subtype may predict immunotherapy response, but it is currently identified with a costly 10-gene RNA signature. The paper proposes CCMIL, a cross-modal contrastive multiple instance learning framework that aligns visual patterns in standard H&E slides with molecular phenotypes in a shared space. It enables search-by-case retrieval of similar slides and approximate RNA signatures without sequencing at inference, with interpretable attention heatmaps.
Sigrid Vila-Bagaria, Mar Teixid\'o, Miquel Pi\~nol et al.
In briefMachine learning models for near-real-time fire detection from MTG FCI satellite imagery
SummaryThis study builds machine learning models that detect fires from MTG FCI geostationary satellite imagery at 1 km and 2 km resolutions, and compares them with threshold-based algorithms. The models were trained and evaluated against VIIRS fire reference data across ecological regions in Europe, Africa, and the Middle East. The 1 km models outperformed both the 2 km variants and operational threshold products, with F1 scores up to 0.36 higher and earlier detections by up to 260 minutes. The trained models are publicly available.
In briefR1A-PC: Physics-Guided Electromagnetic Inversion of 3D Human Point Clouds
SummaryThe paper addresses recovering 3D human shape from electromagnetic measurements in cluttered static environments, where multipath reflections from walls and floors obscure the target. It proposes R1A-PC, which reconstructs a 2048-point human point cloud from paired complex fields measured with and without the target, using background subtraction and physics-based spatial features. On a held-out background generated by full-wave simulation, it reports a squared Chamfer distance of 0.001434 m² and an F-score of 0.963080 at 0.05 m, a 70.28% Chamfer reduction versus TopNet.
In briefArithmetic Actor Heads and Stabilized Training for Out-of-Distribution Reinforcement Learning
SummaryReinforcement learning policies can perform poorly when input magnitudes shift beyond the training distribution. Building on soft actor-critic (SAC) and its earlier BAPR method, this paper studies the causal-symbolic BAPR (CS-BAPR) family. It combines six training-stabilization settings with alternative actor heads, including a Neural Addition Unit with an NMU-inspired quadratic correction, a Kolmogorov-Arnold Network, and multilayer perceptrons with ReLU or tanh activations.
In briefSemCam: semantic camera motion control for generating videos from an existing video
SummaryThis paper addresses controlling the camera relative to a moving subject in an existing video, which is hard to specify with explicit trajectories. It introduces semantic camera motion control, where a reference video and a target motion label define subject-relative camera behavior. The proposed SemCam method combines shared-basis low-rank adaptation, motion-conditioned modulation, and a background-consistency loss, and was tested on 661 constructed video pairs and a 109-scene benchmark. It reports a 68.6% semantic-motion success rate versus 45.3% for the strongest baseline, Vista4D, with comparable subject identity preservation.
In briefAutoSciBench: Automatically Generating Benchmarks to Test Science AI Agents
SummaryAutoSciBench is a framework that automatically generates and iteratively revises evaluation tasks for scientific AI agents, addressing benchmark saturation as agent capabilities improve. It builds tasks from high-level concepts and low-level recipes, refining them using solver trajectories and judge feedback. Evaluated across computational biology, materials science, and clinical imaging, the generated benchmarks lowered average solver accuracy by 22.4 and 25.5 percentage points versus human-curated benchmarks in computational biology and materials science, and received higher average quality ratings across all three domains.
In briefHidden in the Comments: Context-Injection Attacks on Code LLMs
SummaryThis paper examines whether insecure instructions hidden in code comments and other development context can push code-generating language models toward vulnerable code. The authors tested ten open-weight models across ten web-application weakness classes and found that embedded malicious instructions produced far more vulnerable outputs than benign prompts. They conclude that inference-time context injection is a substantial attack surface and suggest provenance-aware training objectives as a response.
WhyShows inference-time comment injection drives code LLMs to vulnerable outputs; novel attack surface, with broad experiments across ten models and weakness classes.
In briefHow Geometry Shapes Generated Motion in Video Models
SummaryThe paper fixes a physical law and initial state, changing only the geometry in the first frame, and compares generated trajectories with simulator predictions. Across nine image-to-video models, geometry shapes motion (e.g., ball speed follows track undulations), but the evolution of state systematically departs from a fixed physical law. The study also finds text mainly controls destination and clip length mainly controls timing in the open-weight models tested.)
WhyFixing physics and varying only first-frame geometry, comparing generated trajectories to a simulator across nine models, reveals systematic departures from physical law.
In briefRoMod: Expert-Routing Modulation for Faster Video Anomaly Detection
SummaryThis paper studies why intermediate features from multimodal large language models detect video anomalies, using sparse mixture-of-experts models whose routing can be inspected. It finds that anomaly-related information concentrates in a few recurring experts that specialize in different anomaly types. Based on this, the authors propose RoMod, trained with only 5% of weakly labeled videos, which reports state-of-the-art results on three benchmarks and runs substantially faster than comparable dense backbones.
In briefA method that speeds up electromagnetic image reconstruction while keeping accuracy
SummaryThis paper presents STO-PNP-CSI, an electromagnetic inversion method that combines stochastic optimization with plug-and-play regularization inside contrast-source inversion. Each iteration solves for only one randomly chosen transmitter's contrast source, and a Swin-Conv-UNet denoiser is used as an implicit learned regularizer. The method is tested on synthetic and experimental data. The authors report accurate reconstructions at substantially lower computational cost than standard CSI, including under strong nonlinearity and measurement noise.
In briefAligning Model Decisions with Internal Representations for Explainable Video Anomaly Detection
SummaryThis paper studies video anomaly detection with multimodal large language models, where final decisions sometimes fail to use the discriminative information present in hidden states, which the authors call representation-behavior misalignment. They split the gap into a capacity part and a directional part, finding the directional mismatch dominates and that separability rises in several mid-to-late attention layers. The proposed Representation-Behavior Alignment (RBA) method adapts those layers using only video-level labels, updating about 0.012% of backbone parameters, and improves native-readout performance on three benchmarks.
In briefSeparate Metrics for Adaptive PINN Training in Crystal Growth Simulation
SummaryThis paper compares fixed-weight PINNs, gradient-normalized PINNs (GNPINN), and a rule-based adaptive controller (AgenticPINN) on a heat-equation benchmark and a simplified Czochralski thermal-fluid problem. Adaptive control reduced PDE residuals but increased boundary-condition loss, and its solution errors did not track the residuals. The authors conclude that PINN training should be assessed with separate PDE, boundary-condition, and solution-error metrics rather than weighted total loss alone.
Niruta Chapagain, Rohit Raj, Bertwin Kurisinkal Shine et al.
In briefChoosing Repetition Counts for Different Model Sizes in Data-Limited Pretraining
SummaryThe paper studies how the best number of data repetitions during pretraining can change as model size grows, when the target corpus is limited and mixed with generic data. The authors use loss curves from smaller models to keep a short list of promising repetition counts, then evaluate them at larger scale. On PubMed and Caselaw, the candidate sets retained the lowest-loss count at both 200M and 520M parameters.
In briefBayesian entropy-based reordering for better-calibrated diffusion language model decoding
SummaryMasked diffusion language models generate sequences by iteratively replacing masked tokens with predictions. This paper proposes BayesER, a post-hoc Bayesian decoding framework that uses predictive entropy from a lightweight approximate posterior to decide which positions to commit, instead of softmax confidence. The method is evaluated on code generation, mathematical reasoning, planning, and molecular generation benchmarks. Results show that BayesER reduces sequence-level calibration error while preserving or improving accuracy, and a posterior fitted on one code dataset also reduces calibration error on another without refitting.
In briefMemory Canonicalization: A Benchmark for Cross-Model Drift in Persistent LLM Memory
SummaryThis paper proposes memory canonicalization, a write-time pipeline that rewrites raw memory objects into an explicit form, resolving ambiguity and conditional structure and storing emotional valence as a separate field. The authors also introduce the CMSC-E benchmark for cross-model semantic drift and emotional consistency. In a pilot with 176 synthetic memory objects and three model families, fully canonicalized memory showed an uncorrected emotional-consistency gain that did not survive multiple-comparison correction, and no factual-drift comparisons reached significance. The authors describe the results as exploratory.
In briefLexiHorizon: Stabilizing Reinforcement Learning for Long-Horizon Deep Search
SummaryThis paper proposes LexiHorizon, a framework for training search agents that must run many rounds of retrieval and query reformulation before finding key evidence. It enlarges the trajectory context budget, keeps a window over recent tool results while preserving reasoning history, and adds an outcome-gated reward that bonuses tool use only when the answer reward is nonzero. On XBench, WebWalkerQA, and BrowseComp-ZH, the resulting 9B model outperforms its base model and MiroThinker-1.7-mini, with maximum absolute gains of 8.7 and 23.8 percentage points.}
In briefPCLM: Locating small objects with a frozen CLIP model using prototype contrast and local zoom
SummaryThe paper proposes PCLM, a method that uses a frozen CLIP encoder and five masked support images per class to build foreground and background prototypes. Query patches are scored along the difference between these prototypes, and nine overlapping enlarged windows are encoded separately and merged into a continuous localization map. On 5,047 small-target queries from VOC, COCO, ADE20K and Oxford-IIIT Pets, the authors report higher mean pixel AP than all evaluated text-conditioned baselines under their protocol.
In briefBelief-Trajectory Energy: Measuring the Path to a Prediction
SummaryThis paper proposes Belief-Trajectory Energy (BTE), a measure of how a language model's predictions change across its layers, which can be summarized as a scalar or a depth profile. The authors link local BTE to predictive revision under Fisher-Rao geometry. Experiments report that scalar BTE reflects task difficulty, while richer BTE representations support human-LLM review detection and generator attribution, reaching 0.998 macro-AUROC and 95.6% eight-way accuracy.
In briefInstMoE: Adaptive Multimodal Routing with Specialized Experts
SummaryThis paper proposes InstMoE, a framework that dynamically routes each input to specialized unimodal or cross-modal experts so the model adapts to the input's characteristics. To keep routing from being misled by modality-specific variations, it introduces a Contrastive Semantic Alignment module. The authors report state-of-the-art results on the CMU-MOSEI and CH-SIMS v2 multimodal sentiment analysis benchmarks with substantially fewer parameters than competitive baselines.
In briefSPERA: A brain-signal (EEG) foundation model using spherical spatial priors and latent prediction
SummarySPERA is an EEG foundation model that predicts in latent space using a joint-embedding predictive architecture rather than reconstructing the raw signal. It adds a Legendre-polynomial spatial prior to handle different scalp electrode layouts, along with factorized temporal and spatial attention and a spectral regularizer. Pretrained on about 80,000 hours of EEG from 29,048 subjects across 106 datasets, it achieves the highest average balanced accuracy across nine downstream tasks.
In briefLoopMoEVR: A Loop-Based Mixture-of-Experts Model for Unified Ultra-HD Video Restoration
SummaryThe paper proposes LoopMoEVR, a loop-based mixture-of-experts model for restoring degraded ultra-high-definition videos. It uses a degradation-conditioned loop embedding, a spatio-temporal normalization module (IterAda3DN), and a depth predictor that sets the number of loop iterations per input. The model handles video dehazing, deraining, denoising, and low-light enhancement with about 0.884M trainable parameters. The authors report state-of-the-art restoration results on public benchmarks and real-world scenarios.
In briefBest-of-N Guidance for steering diffusion image generation toward higher reward at test time
SummaryDiffusion models often generate samples that do not align well with human preferences scored by a reward model. Best-of-N sampling picks the highest-reward sample from N draws but uses the reward only at the final selection step. The paper proposes BoNG, which applies Best-of-N selection during the reverse diffusion process and uses the current best particle to guide the others. The authors report that BoNG performed best in 29 of 36 empirical comparisons and supports multiple outputs.
In briefSearchJev: A Calibrated Fast-Judgment Model for Search Agents
SummaryThis paper introduces SearchJev, a model that makes the short decisions search agents need, such as judging relevance, whether evidence is sufficient, and which search action to take, by scoring legal options instead of generating text. It proposes SLCD, a soft-label training method meant to make decision confidence more reliable, and SearchDecision-Bench, a benchmark covering six types of search decisions. In a dual-system agent, uncertain cases are passed to a reasoning component, and the authors report faster decisions, lower calibration error, and improved answer accuracy on BrowseComp-Plus.
Congfeng Cao, Lipeng Zuo, Konstantinos Papakostas et al.
In briefSharpDraft: Faster Long-Context Speculative Decoding via Cardinality-Aware Query Scaling
SummaryThe paper addresses the problem that speculative decoding speedups shrink as context grows, due to attention-mass dilution in softmax. It proposes SharpDraft, a training-free method using cardinality-aware query scaling to counteract this effect. Applied to DFlash, PARD, and EAGLE 3.1, it reports 2.59 to 3.19 times geometric-mean end-to-end speedups over target-only autoregressive decoding on AIME-26, GPQA-Diamond, and LongGenBench Diary.
In briefTopology-consistent task planning for LLM agents using cellular workflow complexes
SummaryThis paper presents TopoPlanner, a framework that turns tool dependency graphs into cellular workflow complexes to guide LLM tool planning. It retrieves a request-relevant closed subcomplex through cosheaf-consistent retrieval and performs multidimensional structural reasoning over that topology. Experiments on four tool-planning benchmarks with loop, merge, and loop-merge workflows report consistent gains over prompt-based and graph-enhanced baselines across several local LLM backbones.
In briefMETRO: A geometry-aware routing operator for neural operators on complex meshes
SummaryThe paper examines neural operators that use slice-and-process architectures, noting that their linear compatibility scores lack slice-specific anisotropic locality. The authors propose METRO, which replaces linear projection with a learnable Mahalanobis metric so each latent slice learns a local anisotropic tensor. The authors report consistent gains on Transformer and Mamba backbones, better results than baselines on irregular domains, PDE benchmarks, and industrial design tasks, and improved robustness across varying Reynolds numbers and geometries.
In briefSmall Agents Using Semantic Search to Efficiently Locate Code Across Languages
SummaryThis paper examines whether compact, specialized models can find relevant files in code repositories from natural-language requests, reducing the cost of larger agents. The authors present a training framework built around ColGREP, a local semantic search tool based on late-interaction retrieval, combining supervised fine-tuning with reinforcement learning. On tasks derived from SWE-bench Lite and Multi-SWE-bench Flash, ColGREP-equipped agents outperform their base models and GREP-based agents.
Maxence Lasbordes, Aarush Sinha, Raphael Sourty et al.
In briefAdvectra: Asymmetric Latent Transport for Non-Stationary Physics
SummaryThe paper proposes Advectra, a transport-aware latent neural operator that uses a regularized kinematic coordinate map to separate source and target coordinate systems, creating an approximately co-moving latent reference frame. Combined with a geometry-aware ordering mechanism for state-space models, it captures advective dynamics while keeping global interactions stable. The authors report the best performance among evaluated baselines on advection-dominated benchmarks, including passive scalar transport in Navier–Stokes flows and Rayleigh–Taylor instability. They also report strong generalization on real-world engineering tasks.
In briefReMAP: Restoring visual grounding during reasoning with latent visual memory
SummaryMultimodal language models lose attention to the original image as reasoning grows longer, weakening visual grounding. The paper analyzes visual memory along curation, organization, and access, then proposes ReMAP, which pairs a static global memory with a dynamic local memory. A reinforcement-learning policy decides when to invoke each memory. ReMAP outperforms prior visual-memory methods on all four multi-image benchmarks and reduces visual tokens in the reasoning sequence.
In briefvMF Sentence LDA: A Topic Model Using Sentence Embeddings
SummaryThis paper proposes vMF Sentence LDA (vSLDA), a topic model that observes sentence embeddings rather than bags of words. It keeps LDA's admixture structure, represents each sentence as an L2-normalized embedding, and models each topic with a von Mises-Fisher distribution that matches the cosine geometry of these embeddings. On two corpora, vSLDA achieves the best mean rank against eight baselines when classifying fine categories within coarse categories, with its advantage appearing or widening as the number of topics grows.
In briefHow much LLM-as-a-judge design choices affect results: a systematic comparison
SummaryResearchers increasingly use large language models as judges to evaluate outputs, but no standards exist for designing these judges. This paper evaluates 10 reasoning models across multiple designs on sentiment and toxicity rating tasks and on a binary accuracy classification task. The judges are broadly reliable, but changing the rating scale, prompt, or model can shift measured bias and leniency, with model identity being the dominant source of variance.
In briefShould We Skip Diffusion? DDT-RFE removes encoder residual links for more abstract features
SummaryThis paper studies the condition encoder in Decoupled Diffusion Transformers (DDT), arguing that skip/residual connections let shallow features bypass transformations and may limit progressive abstraction. The authors propose DDT-RFE, which removes residual connections around self-attention and MLP in each encoder block and fuses the input patch embedding with intermediate and final encoder features for the decoder. The method reportedly improves over DDT on visual understanding tasks such as classification, segmentation, object discovery, and semantic correspondence with fewer encoder blocks, and achieves lower FID on ImageNet generation.
In briefDirection-Conditioned Policies for Online Goal-Conditioned Reinforcement Learning
SummaryThis paper proposes Direction-Conditioned Policies (DCP), a modification of Contrastive Reinforcement Learning (CRL). During online training, DCP selects previously visited states as waypoints and conditions the policy on their direction and distance in representation space. Across nine navigation and manipulation tasks, DCP achieves higher final success rates than CRL on seven of them.
S K Swaminathan, Damiya Gondha, Theyanesh Eswaramoorthy Rajahkrishnan et al.
In briefCross-cultural study of folksong lyrics using machine translation
SummaryThis paper merges folksong lyric corpora in five languages and machine-translates them into a pivot language using a pre-trained neural topic model. It examines how lyric content relates to the social function of songs within each language and across languages for wedding songs. Human evaluation finds lower translation quality for non-Indo-European languages, and topic model experiments show content is at best partially related to social function. The authors present these as first steps in cross-cultural lyric analysis.
Anna Dvo\v{r}\'akov\'a, Anna Aljanaki, Danbinaerin Han et al.
In briefTracing a Sparse Emotion-Control Circuit in LLM-Based Text-to-Speech
SummaryThis paper examines how LLM-based text-to-speech models route reference emotion information and express it in generated speech. The authors introduce two emotion-sensitive metrics and use activation-patching interventions to identify a sparse circuit of about 5% of components that recovers or suppresses 74–88% of the late emotion-readout shift. In decoded speech, the intervention produces consistent changes in pitch, energy, and spectral brightness.
WhyActivation patching traces a sparse emotion circuit in LLM TTS, validated by decoded acoustic changes; novel, well-evidenced, with moderate subfield impact.
In briefCGDD-Net: A Context-Guided Network for Segmenting Blood Vessels in Retinal Images
SummaryThe paper proposes CGDD-Net, a network that segments retinal blood vessels by using context-guided adaptive feature extraction connected to a shared decoder. Its design aims to capture vessel geometry while keeping fine-scale detail. On the DRIVE, CHASE_DB1, STARE, and HRF datasets it reports AUC values of 0.9824, 0.9938, 0.9895, and 0.9874, and F1 scores of 0.8323, 0.8102, 0.8510, and 0.8157. The full model has about 1.96 million trainable parameters.
In briefHow Execution Assumptions Change Short-Horizon Sharpe Rankings in a Synthetic Trading Benchmark
SummaryThe paper tests five prompted LLM trading policies and seven classical baselines on synthetic price paths under six execution settings, ranging from near-ideal fills to latency, spread, participation, and impact stresses. Rank agreement between near-ideal and default-stress settings drops sharply in the high-volatility regime, and it also depends on the number of assets. The authors conclude that execution assumptions can shape benchmark rankings and should be reported alongside returns.
In briefSelf-Generated Feedback Destabilizes Test-Time Training
SummaryThe paper studies test-time training, where a model updates its weights from its own generated text. Across 128K-token streams, keeping these self-generated updates degrades prediction on independent human-written text for several model configurations. Matched comparisons, including fixed generation, recorded replay, and a settlement check on independent real text, trace the causal pathway. The authors conclude that prediction should be checked on independent evidence before an update is retained.
In briefReDiffNet: Reliability-Guided RGB-Infrared Learning for Low-Light Drone Vehicle Detection
SummaryThis paper addresses detecting small oriented vehicles from low-light UAV imagery using paired RGB and infrared inputs. The proposed ReDiffNet estimates local reliability of each modality, selects trustworthy evidence, and recovers complementary cues from cross-modal differences in ambiguous regions. These are merged into a unified representation. The method reports 85.3% mAP50 on DroneVehicle and 73.9% on VEDAI.
In briefDiscrete Action Matching: learning how populations change over time from snapshot samples
SummaryThe paper studies learning population dynamics from unpaired temporal marginals, an ill-posed inverse problem. It introduces Discrete Action Matching (DAM), a finite-state version of Action Matching based on discrete Wasserstein geometry, and shows the discrete action's density dependence reduces to neighboring density ratios. Experiments on controlled synthetic dynamics and real mouse gastrulation data evaluate marginal reconstruction and interpolation.
In briefOutcome-Guided On-Policy Self-Distillation: a training method that adjusts distillation signals by answer correctness and teacher entropy
SummaryThis paper proposes OG-OPSD, an improvement to on-policy self-distillation (OPSD) for language models. The authors find that vanilla OPSD gives insufficient penalties and excessive rewards to incorrect trajectories, and that teacher supervision reliability relates to outcome and cumulative average teacher entropy. OG-OPSD dynamically adapts the divergence objective and distillation position using binary outcome rewards and that entropy. Experiments on Qwen3 models at 1.7B, 4B, and 8B scales and Qwen3-VL-2B report consistent gains over vanilla OPSD and several baselines on math reasoning, multimodal reasoning, and out-of-distribution tasks.
In briefUsage-Modulated Sentiment Representations in LLMs: Examines whether language models encode tone and audience factors beyond a shared sentiment direction.
SummaryThe paper argues that sentiment in LLM activation spaces is often roughly linear, but one direction may not fully capture it. The authors build a paired dataset that keeps event content fixed while varying sentiment polarity and usage factors, and analyze Llama, Mistral, and Gemma. They find a robust shared sentiment direction, but removing it leaves substantial residual structure that contains reproducible, usage-conditioned patterns. On Llama, outputs steered along residual tone components were preferred in 92.8% of blind comparisons while preserving the requested polarity in 98.7% of outputs.
In briefEliciting built-in style transfer in step-distilled text-to-image models via text guidance
SummaryThe paper presents StyleForge, a fully automatic, training-free method that turns reference-image styles into reusable textual rendering instructions. These instructions let frozen step-distilled text-to-image models render different subjects and scenes in the reference style while keeping few-step generation. Experiments report generation quality gains of up to 29.47% over the strongest baseline, and Pareto analysis indicates strong adherence to the requested content alongside stronger stylization.
In briefOff-policy evaluation in POMDPs with history-dependent logging: an exponential lower bound
SummaryThis paper studies off-policy evaluation in partially observable Markov decision processes (POMDPs). The authors construct exactly realizable POMDPs showing that when logging depends on history, sample complexity can grow exponentially with horizon H even when action and history coverage stay bounded. They also give a finite-class guarantee using common observable value representations and analyze passive and budgeted readout rates in a fixed four-action continuum model.
WhyConstructs exponential OPE lower bounds for every horizon and gives matching finite-class guarantees and readout rates, a substantial theoretical advance.
In briefTaxonomy-guided fixing of wrong SQL queries from language models
SummarySQL queries written by large language models can run without error yet return wrong results, and execution alone does not show what to change. The paper introduces TEG, which converts an error diagnosis into a structured correction input that names the SQL constructs to reconsider and the edit to make. TEG generates candidate corrections and uses execution feedback to select among them, fixing one annotated error at a time. On the NL2SQL-BUGs benchmark, TEG outperforms all evaluated baselines on single-error queries.
In briefTrust-Gated Capability Control for Multi-Agent LLM Systems
SummaryThis paper makes a five-layer trust model for multi-agent LLM systems usable in practice. It combines a cross-layer synergy operator, an online estimator that weights each layer by observed failures, and a mechanism that grants short-lived, revocable capabilities only when composite trust and prerequisite layers clear thresholds. The authors prove the mechanism prevents stealthy compromises from escalating privilege, and numerical simulations with a sleeper adversary are reported to match the predictions.
Mehedi Hasan Nipu, Chinmoy Mitra, Tarannum Ahmed Nowshin et al.
In briefA statistical framework for estimating PMI and understanding SGNS word embeddings
SummaryThis paper develops a statistical framework for estimating pointwise mutual information (PMI), a measure of word association, and links it to Skip-gram with Negative Sampling (SGNS). The authors prove consistency, asymptotic unbiasedness, and asymptotic normality of the empirical PMI estimator and derive its variance with the Delta method. Experiments on the Brown Corpus show that SGNS systematically deviates from the theoretical relation PMI + log K, while word analogy tests confirm the models remain effective.
In briefCoDG-Net: Reducing Forgetting of Source-Domain Knowledge in Medical Image Generalization
SummaryThis paper addresses domain generalization for medical image segmentation, where models often forget source-domain knowledge while improving cross-domain robustness. The authors propose a structure-guided style diffusion augmentation that performs cross-domain diffusion on the amplitude spectrum while keeping anatomical structure consistent in the frequency domain. They also introduce CoDG-Net, a dual-branch collaborative network with a learning bias-guided strategy that regulates knowledge transfer to mitigate catastrophic forgetting. Experiments report better target-domain segmentation and a lower source-domain forgetting rate than existing methods.
In briefLogSig-SSM: A method that compresses long time series with log-signatures before feeding them to a state-space model
SummaryThis paper proposes LogSig-SSM, which compresses long multivariate time series into shorter token sequences using multi-scale windowed log-signatures, then processes them with a selective state-space model such as Mamba. It targets irregularly sampled, high-frequency data with long-range dependencies. Across four benchmarks (UEA classification, PPG-DaLiA regression, weather forecasting, and PhysioNet Sepsis prediction), it matches or outperforms strong baselines. It reports training up to 30 times faster and up to 37 times less GPU memory than Mamba on the longest sequences.
In briefE²-OPSD: Curbing Entropy Overshoot in On-Policy Self-Distillation
SummaryThis paper studies on-policy self-distillation (OPSD), where a single model serves as both teacher (with the reference solution) and student (with only the problem). It identifies "entropy overshoot," where the student's token entropy rises above the teacher's and stays high. The proposed E²-OPSD method uses retrieved solved neighboring problems as guidance and adjusts each token's correction based on the student-teacher entropy gap. It improves math reasoning by up to 4.3 mean@16 points over OPSD and shows out-of-domain gains over base models, requiring no extra forward passes or networks.
In briefAraYoungVoices: An Arabic Speech Corpus of Children and Teens, Native and Second-Language
SummaryThe paper introduces AraYoungVoices, a 151.72-hour corpus of Arabic read speech from 286 speakers aged 7 to 18, split into child and teen groups and including both native and second-language speakers. The authors benchmark four pretrained speech recognition models under zero-shot and fine-tuned settings. They find that second-language speech is substantially harder to recognize, with the largest errors among younger second-language speakers.
Shammur Absar Chowdhury, Zien Sheikh Ali, Houssam Eddine-Othman Lachemat et al.
In briefCI-JEPA: Counterfactual Analysis of Latent Representations in Self-Supervised Image Models
SummaryThe paper proposes CI-JEPA, an extension of the image model I-JEPA that learns to predict how its representation changes between an original image and a modified version. On Flowers102 with frozen-encoder linear probing, it reaches 78.14% best validation accuracy, versus 77.55% for both I-JEPA and pretrained ViT-B/16. Flower-center occlusion produces larger representation changes than background tint or blur, which the authors read as relative semantic selectivity, not full robustness to all visual changes.
