In briefGPA-VGGT: A Training-Free Way to Speed Up VGGT While Keeping Geometric Accuracy
SummaryVGGT is a feed-forward model that recovers camera parameters, depth maps, and point maps from one to hundreds of views in a single pass. Its global attention cost grows quadratically with the number of frames, which makes long sequences slow. GPA-VGGT accelerates it without retraining by using frame attention with shared cross-frame keys and values in early blocks, merging only keys and values in later blocks, and randomly projecting the matching metric to 128 dimensions. The paper reports speedups comparable to prior work while retaining over 99% of VGGT's pose accuracy under AUC@3.
In briefOnline gas classification at the edge: a low-power, insect-inspired neural network
SummaryThis paper presents a neural network inspired by the insect olfactory system that classifies raw gas sensor array readings using a sparse, high-dimensional binary representation. It uses only binary weights and Hebbian local learning, making it suited to ultra-low-power neuromorphic processors. In a hazardous gas detection experiment on the mixed-signal neuromorphic chip TEXEL, the on-chip classifier reached 74.6% accuracy at 482 nJ per inference and 27 μW static power.
In briefCaLLM: Continual Adaptation of LLMs via Non-Parametric Adapter Routing
SummaryThis paper presents CaLLM, a framework for continually adapting large language models to changing task distributions while reducing catastrophic forgetting. It combines metric-based meta-learning with modular parameter isolation using lightweight adapters, routing each input batch to the adapter matching its nearest prototype in a task-agnostic embedding space. The authors report that CaLLM outperforms traditional isolation and replay methods across diverse fine-grained continual learning benchmarks.
In briefSpectRAG: Graph-Based Context Selection for Multi-Hop Retrieval-Augmented Generation
SummaryThis paper proposes SpectRAG, a retrieval framework for retrieval-augmented generation that builds a weighted graph over document chunks and uses spectral relevance propagation. It expands candidate passages over this graph and assembles a context within a fixed token budget. Experiments on several datasets report improved evidence recovery and generation quality competitive with baselines, along with trade-offs in early precision and computational cost.
Zafer Sahinoglu, Abdulazeez Mohammed, Buse Yaren Kazangirler et al.
In briefThe State of Multilingual LLM Evaluation: An Audit of 51 Benchmarks
SummaryThis paper audits 51 recent multilingual benchmarks covering 242 datasets and 219 languages, examining coverage, representativeness, and rigor. It finds that many languages appear in only one benchmark, low-resource languages are evaluated on far fewer task categories, and 56% of dataset-language instances are translated from English. The authors also report benchmark contamination. They conclude with recommendations for natively constructed, culturally grounded, contamination-aware evaluation frameworks.
WhyA data-driven audit of 51 multilingual benchmarks across coverage, representativeness, and rigor offers a fresh, well-evidenced critique.
Manan Uppadhyay, Himanshu Beniwal, Prashant Kodali et al.
In briefA testbed game for studying self-governing groups of humans and AI agents
SummaryThis paper introduces Megabike, a multi-agent simulator and serious game for studying whether mixed human-machine groups can govern themselves. Riders sharing a megabike must agree on steering, pedalling and resource distribution, and choose their political regime. In forty sessions with language-model riders, the agents chose regimes, elected leaders and expelled one another. The results show that different combinations of agent types can produce unexpected and counterproductive outcomes.
In briefLearnPrunerV2: A Learnable Method for Gradually Removing Visual Tokens in Vision-Language Models
SummaryVision-language models process many visual tokens, which makes them computationally expensive. This paper proposes LearnPrunerV2, which adds lightweight learnable modules to every layer to score token importance by combining visual saliency with query relevance. It also introduces Budget-Deterministic Pruning, a fixed per-layer schedule based on a staircase redundancy pattern. The authors report that on LLaVA-v1.5 it keeps 99.8% of the original performance while pruning 77.8% of visual tokens.
In briefAdvancement or Acceleration? How Data Curation Improves LLM Training
SummaryThis paper distinguishes two effects of data curation for LLM training: advancement, which raises the performance a model can ultimately reach, and acceleration, which reaches the same performance with less data. The authors propose a Δ-test, the gap between the uncurated baseline's training and target losses, which can be computed before curation. In experiments, a clearly nonzero Δ predicts persistent gains, while gains near Δ=0 tend to shrink as the training budget grows.
In briefNeural network-guided road network reduction for city-wide centralized traffic routing
SummaryThis paper studies centralized traffic routing, where a central planner assigns routes to origin-destination vehicle flows to minimize congestion-dependent segment costs. The authors propose a learning-based network reduction method: a graph neural network scores road segments for each OD flow, highly ranked segments are assembled into complete routes, and the original segment-flow MILP is solved on the reduced network. Experiments on four real urban networks show the smallest relative objective difference to a 30-second full-graph reference in every city under both traffic regimes, and traffic simulations indicate more efficient routing.
Nikolai Antonov, Luk\'a\v{s} Chrpa, Andrii Nyporko et al.
In briefAutoVend: Benchmarking LLM Agents on Long-Term Autonomous Vending Machine Operations
SummaryThis paper introduces AutoVend, a benchmark where one LLM agent manages 20 vending machines across five site types through a Web interface over 180 simulated days. The simulator includes a shared warehouse, procurement lead times, paid replenishment, inventory expiration, and recurring operating costs. Four of the five evaluated LLM agents underperform a rule-based inventory-control baseline, and inventory and execution costs often erode gains in individual metrics.
In briefExplicit Regional Maps for Judging Video Quality
SummaryThis paper addresses the problem that end-to-end video quality assessment methods leave their regional evidence largely implicit. The authors introduce a Quality-Relevance Heatmap (QRH), a pixel-level explicit regional representation that combines localized degradation evidence with perceptual priors under vision-language model guidance, without requiring pixel-level annotations. Applied to the DOVER model through two deterministic fragment-sampling strategies, the quality-aware sampling variant improves most evaluated metrics on popular VQA benchmarks.
In briefVC dimension of autoregressive linear threshold functions
SummaryThis paper studies the VC dimension of autoregressive linear threshold functions, a class of Boolean predictors that repeatedly apply a linear threshold to the most recent d bits and return the bit produced after T steps. The authors prove that for deterministic autoregressive LTFs the VC dimension is Θ(d²) for every T ≥ d², resolving a known quadratic gap. They also adapt the technique to show that replacing the sign function with logistic sampling can raise the VC dimension of the final majority prediction class from quadratic to exponential in d.
In briefPaper on a memory method that helps event-driven AI agents act proactively
SummaryThis paper proposes ProactMem, a memory method that organizes history into recent events, temporal trajectories, and behavioral regularities. Trajectory facts are checked against source events, and regularities are validated by historical replay before being matched to current conditions. The authors also build ProactiveStream, a longitudinal benchmark across three application scopes, where ProactMem reaches an overall F1 of 52.6%, a 21.4% relative gain over the strongest of five baselines.
In briefDo bots feel bad? Reviewing evidence for functional valence in LLMs
SummaryThis paper examines whether large language models have valenced states, meaning states that feel pleasant or unpleasant. It assesses recent claims of functional pleasure, pain, and emotions in LLMs against the philosophical and cognitive science understanding of valence. The author argues that key aspects of valence remain untested and that current experiments fail to separate functional valence from preferences, value judgments, and role-play. The paper closes by proposing new experimental approaches aimed at valence itself.
In briefLive-OPSD: On-Policy Self-Distillation for Proactive Streaming Video Assistants
SummaryThis paper introduces Live-OPSD, a training method for streaming multimodal assistants that must decide when and how to respond while video events are still unfolding. It uses hints available only during training to guide the model on its own generated outputs. On Qwen3-Omni-30B-A3B, Live-OPSD scores 59.64 on LiveProBench, 5.26 points above the strongest continued-SFT baseline, and 65.84 on OVO-Bench.
In briefPailitaoMMSearch2.0 Technical Report: Scaling Generative Retrieval to Unified End-to-End Multimodal Search
SummaryThis paper presents PailitaoMMSearch, a system that generates product Semantic IDs directly from a raw query image and free-form instruction, aiming for unified end-to-end multimodal search. It scales along data, intent, and scenario dimensions, using continual pre-training, supervised fine-tuning, and reinforcement learning. The system is deployed in Taobao Pailitao, and a controlled online A/B test reports statistically significant gains in item page views, transactions, and gross merchandise value.
WhyProposes a unified end-to-end multimodal generative retrieval system with CPT and RL; a controlled Taobao A/B test shows significant gains, indicating broad impact.
In briefUsing Nineteenth-Century Editions to Aid Recognition of a Thirteenth-Century Papal Register
SummaryThis paper examines how nineteenth-century regesta and partial editions can support handwritten text recognition of the registers of Pope Honorius III. Using the regnal years with existing modern transcriptions as a benchmark, the authors test what this legacy material contributes at the orthographic, lexical, and topological levels. They conclude that such editions cannot replace modern transcription but can guide, check, and segment recognition across untranscribed registers.
In briefDense learning and differentiable shape modeling for detecting spine landmarks in X-rays
SummaryThis paper treats spine landmark detection in spinal X-rays as a pixel-level contrastive representation learning problem, matching dense pixel embeddings to landmark prototypes. Even with a generic U-Net backbone, it produces more discriminative heatmaps than traditional heatmap regression, improving discrimination between vertebral corners on different vertebrae. The heatmaps are combined with learnable shape priors over the global configuration of spinal landmarks to resolve location ambiguities. The approach is competitive in landmark accuracy and clinical parameter estimation on the public AASCE dataset and two private spinal X-ray datasets.
Augustin Daridon, Sofiane Sifaoui, Ayman Assi et al.
In briefDCKV: A trained codec that compresses prompt KV caches in hybrid language models.
SummaryThis paper presents DCKV, a differentiable codec for the key-value cache of hybrid language models and described as the first such codec trained end to end. Its transform, quantizer and entropy models are conditioned on prompt tokens and already decoded neighbouring positions, and are optimized jointly against the frozen language model. On Qwen3.5-2B, DCKV stores a compressed token in more than 55 times fewer bytes than bf16 with at most a 1.8-point loss on three long-context benchmarks, and in more than 100 times fewer bytes with at most a 3.6-point loss.
Egor Maximov, Aleksei Merkelov, Vasiliy Alekseev et al.
In briefRemembering dialogue inside a frozen language model's parameters, without external memory
SummaryThis paper asks whether a language model can store a dialogue inside its own parameters and later answer questions about it, with no external memory. The proposed method, SKIM, has the model generate its own varied questions for each dialogue turn and write them into its parameters as keys for retrieving that turn's content. Evaluated on five language models, SKIM recalls most stored dialogue turns from later user questions, answering from the parameters alone.
In briefBeyond QASecClaw: Agent Scaling and Evidence Reliability in Software Assurance
SummaryThis paper asks whether adding specialist language-model contexts improves software assurance when deterministic tool evidence is already available. The authors run five experiments covering Java warning controls, OWASP Benchmark, Defects4J test generation, SV-COMP verdicts, and misleading evidence packages. Increasing context count sharply raises token use without improving unsafe clearance, while verifier feedback and verdict aggregation show mixed results. The findings support deterministic evidence admission in the tested settings and expose limits in feedback interpretation and evaluation coverage.
S M Abdulla Al Mamun, Akif Islam, Mohd. Ruhul Ameen
In briefAlgorithmic monoculture risks in machine learning fertilizer recommendations
SummaryThe paper argues that machine learning fertilizer recommendations are usually evaluated one model at a time, yet correlated errors can accumulate across interconnected farms and create systemic costs. It models social loss from shared and farm-specific errors, using the Herfindahl-Hirschman Index and a covariance-generalized formulation to analyze model concentration and error dependence. Calibrated on nine model families across five Moroccan cereal systems, it finds that accuracy-constrained diversification lowers modeled social loss by 10.89-41.53% under a 10% regret cap, though these counterfactual results are not agronomic or monetary welfare estimates.
In briefTDLM: Adaptive Temporal Integration for Diffusion Language Models
SummaryThis paper introduces TDLM, a temporal self-conditioning diffusion language model that lets unresolved token positions attend to their own history of hidden representations inside a frozen denoiser. Training uses self-generated trajectories and a cross-entropy objective, with causal temporal attention enabling parallel evaluation. On LLaDA and Dream, the main comparison reports up to 7.15 times faster generation than the original models, with competitive answer quality on the evaluated tasks.
In briefJevRec: Variable-length item codes for LLM-based recommendation
SummaryJevRec is a generative recommendation method using LLMs that encodes items as variable-length semantic identifiers, giving popular items shorter codes and long-tail items richer ones. It uses separate output heads to predict the code length and the code content, sharing the interaction-history representation. The authors report that it outperforms all compared baselines on three public datasets, with lower training and inference time than TIGER.
In briefGaussian process bandit method for optimizing constrained black-box allocation levers
SummaryThis paper addresses how to optimize discrete budget allocation weights across segments when feedback is delayed, noisy, aggregated, and weight changes are rate-limited. It proposes a GP-UCB approach that places a Gaussian process prior on the reward function so neighboring configurations share information. A synthetic simulator identifies a composite Matérn plus Linear kernel as the most robust across stationary, drifting, and reactive settings. A large randomized controlled trial in production reports a statistically significant efficiency gain over an ε-greedy baseline.
Bram van den Akker, Hidde Fokkema, Anastasiia Kornilova
In briefDetecting Provenance of LoRA-Adapted Image Diffusion Models
SummaryThis paper proposes a protocol for determining whether an online service is using a Low Rank Adaptation (LoRA) version of a standard image diffusion model. It considers an open-world setting in which the possible adaptations and alternative models are unknown. Using hypothesis testing, the method builds a black-box signature based only on the intrinsic dimension of sets of generated images, distinguishing the base model, LoRA adaptations, and entirely different model families.
In briefDocXEval: A Systematic Evaluation of Explainability in Multimodal Document Understanding
SummaryThis paper introduces DocXEval, a comprehensive explainability framework for multimodal document understanding that combines text, layout, and image. It adapts existing explainable AI methods and metrics to this setting and adds multi-target variants for tasks such as semantic entity recognition. The authors evaluate 13 XAI methods across 3 tasks, 8 datasets, 4 models, and 17 metrics, finding that Occlusion is the most faithful and interpretable overall. The work also includes a meta-evaluation of metrics and dataset-wide analyses of model behavior.
WhyFirst XAI framework for multimodal document understanding, systematically evaluating 13 methods and 17 metrics across 8 datasets and 4 models, with solid evidence.
Saifullah Saifullah, David Dembinsky, Stefan Agne et al.
In briefVision-language agents that learn inside their own latent world
SummaryThis paper introduces Dwell, a model-based reinforcement learning framework that treats a pretrained vision-language model as a recurrent latent transition operator. The agent predicts, adjusts, and acts in latent space, learning from both real interactions and latent imagination. The authors report that across five interactive visual environments, Dwell achieves the strongest overall performance among trainable methods while substantially reducing imagination compute.
In briefTrawMem: Task-Oriented Role-Adaptive Memory Workspaces for AI Agents
SummaryThis paper argues that current memory systems for large language model agents retrieve memories by relevance to the current query, which does not capture how each memory helps solve a task. The authors propose TrawMem, which organizes interaction histories into local threads and builds task-oriented workspaces by assigning roles to memories and jointly selecting complementary ones. Experiments on LoCoMo and MemGallery with four backbone models show consistent gains over representative memory systems in macro-F1.
In briefCan Lightweight Decision Models Classify Why Retrieved Sources Conflict in RAG?
SummaryWhen retrieved sources disagree, the appropriate response depends on the type of conflict. This paper tests whether decision models, which output probabilities over fixed options in a single forward pass, can classify conflict types quickly and cheaply. On the CONFLICTS benchmark, 10 of 11 decision models match or exceed Gemini 2.5 Flash (61.3%) in accuracy, at 4 to 39 times lower billed cost.
WhyApplies decision models to RAG conflict typing, comparing against LLMs and cascades on a benchmark with extensive experiments; useful for the subfield.
In briefBuilding a Basque Speech Language Model for Spoken Conversation
SummaryThis paper presents the first open speech language model for Basque, designed for low-latency, end-to-end spoken interaction. It builds on the LLaMA-Omni architecture with a Latxa backbone, fine-tuning several modules to carry Basque language ability into speech. To address data scarcity, the English VoiceAssistant-400K corpus was translated into Basque and synthesized with a Basque TTS system. Reported results are 3.86/5 for content quality, 9.94% WER for intelligibility, and 3.07 UTMOS for audio quality.
Andoni Sudupe, Inma Hernaez, Ander Arriandiaga et al.
In briefNothing Left Behind: Finding Evidence in Large Trial Collections
SummaryThe paper compares two ways to find relevant evidence in a large collection: tagging items with labels and searching on them, versus using text embeddings to find candidates by meaning and then having a language model keep only the relevant ones. In a corpus of 7,354 randomized trials, label search found 45% of the evidence a systematic review would include, while the new approach found 92%. The authors propose a two-stage architecture that combines broad candidate retrieval with query-specific evidence filtering.
Giuliano Martinelli, Abelardo Carlos Martinez Lorenzo, Satvik Garg et al.
In briefHiLo-RD: Cross-Budget Self-Distillation for Masked Diffusion Language Models
SummaryHiLo-RD is a method that improves masked diffusion language models under a limited generation budget. It pairs states from a large-window teacher with cropped small-window views and matches their hidden representations at committed positions, guiding both token prediction and commitment order. On mathematics benchmarks it raises low-budget accuracy and shortens responses, and on GSM8K accuracy rises by 9.4 percentage points while mean readable length falls by about 25%.
In briefMining natural hard negatives from positives for instruction-following retrieval
SummaryThis paper addresses the scarcity of instruction-conditioned contrastive supervision for instruction-following information retrieval. It proposes Natural Hard-Negative Mining (NHNM), which turns original positive passages into instruction hard negatives and builds high-contrast training data through automated candidate cleaning, reversal construction, and geometry filtering. Applied to MSMARCO, it yields NHNM-MARCO, a training set of nearly 7K instances, and models of 568M to 4B parameters fine-tuned on it perform strongly on InstructIR and FollowIR, matching or surpassing much larger retrievers in several settings.
In briefPREEN: Color style transfer by predicting smartphone editor presets from a reference
SummaryPREEN treats color style transfer as predicting the 29 slider values of a deterministic image editor, producing an editor preset instead of a raster image. It trains with pseudo-presets obtained by inverting a style-transfer teacher's outputs into the editor's parameter space, then refines a CLIP-conditioned regressor with LoRA. On the PST50 benchmark it leads baselines on pixel and perceptual fidelity metrics, and a single inference-time strength knob moves continuously from the default look to a vivid one.
In briefNoise Steering: a reinforcement learning method that uses turbine yaw to limit wind farm noise while keeping power output high
SummaryThis paper proposes Noise Steering, a model-free reinforcement learning strategy that jointly controls turbine yaw angles and power setpoints to maximise farm power under a local noise constraint. Training uses a low-fidelity acoustic model coupled with the steady-state FLORIS simulator. On a six-turbine test farm, the policy satisfies a 40 dB noise limit across turbulent winds with an average power deficit of about 2%, compared with 17% for the industry-standard noise-reduced operation baseline.
Mart\'in de Frutos, Baptiste Corban, Claire Bizon Monroc et al.
In briefJoint Neural Architecture and Hyperparameter Search via Evolutionary Optimization for Reinforcement Learning (HPNAS-RL)
SummaryThis paper proposes a dual-branch evolutionary framework that simultaneously performs neural architecture search and hyperparameter optimization for reinforcement learning in autonomous driving. It builds on the authors' earlier EMNAS method, evolving discrete network architectures and continuous hyperparameters in parallel under a unified multi-objective strategy, with parallel evaluation and teacher-student policy transfer. Experiments in a high-fidelity driving simulator show that the joint approach outperforms standalone NAS, yielding policies with higher rewards, fewer parameters, and improved stability.
In briefRAG-CSI: Global-Local Retrieval for CSI Prediction via Joint State and Dynamics Matching
SummaryThis paper proposes RAG-CSI, a retrieval-enhanced Transformer that predicts future channel state information (CSI) from historical observations. It builds a global-local CSI knowledge store and uses joint state-dynamics retrieval (JSDR) to find relevant CSI sequences, which a cross-attention Transformer fuses with the current CSI history. On a multi-scene CSI dataset, it improves all five evaluation metrics by over 10% relative to the best baseline for each, including a 15.42% reduction in normalized mean squared error.
In briefRubricAtlas: Type-level scoring criteria mined from expert reports to evaluate deep research agents
SummaryThe paper introduces RubricAtlas, which derives type-level analytical requirements from expert reports on different objects within the same research type and links binary checkpoints to report passages. Built from 3,079 Chinese financial research reports covering 99 research types, it finds that individual reports cover only about half of their type's analytical questions. Across ten agents, two judge pipelines show checkpoint agreement of κ = 0.71 and agent-ranking agreement of Kendall τ = 0.87.
In briefGUI-JEPA: A Fast, Decoder-Free Action Policy for Mobile GUI Agents Using JEPA Latents
SummaryThis paper introduces GUI-JEPA, a decoder-free action policy for mobile GUI agents built on a Joint-Embedding Predictive Architecture. It selects actions in the latent space of an encoder pretrained on unlabelled GUI screenshots, using an action- and region-conditioned predictor to score all candidate actions in one batched pass. On AndroidControl, it is reported to run 2.8-4.0x faster than a fine-tuned Qwen2.5-VL-7B agent, using 6% of its parameters while reaching 86-92% of its in-distribution accuracy.
WhyA novel decoder-free JEPA latent policy for GUI agents; AndroidControl results show speed and parameter gains over a 7B baseline, but accuracy remains below it.
In briefHoloRecon: Globally Consistent Long-Range Panorama Reconstruction
SummaryHoloRecon is a framework that combines an ORB-based sparse panoramic mapping frontend with a fine-tuned feed-forward model to reconstruct long panoramic sequences. The predicted local geometry is used to verify feature matches across distant revisits, providing long-range camera constraints that optimize the global pose graph. The authors also introduce PanoRecon, a seven-scene benchmark with repetitive architecture and reference poses and LiDAR geometry. The method achieves the lowest mean absolute trajectory error and a high F-1 score among the evaluated methods.
In briefAdaGround: A training-free method that adaptively zooms and verifies clicks to locate GUI elements on screens
SummaryAdaGround is a training-free framework for locating interface elements on high-resolution screens, especially small targets. It adjusts zoom settings based on the model's own estimate of target size, and uses a Yes/No verifier with marked clicks to keep or reject zooms and trigger re-grounding. It also composes spreadsheet cells from their headers. On Qwen3-VL models, it outperforms existing training-free zooming methods and many trained GUI models, and its gains extend to the held-out UI-Vision benchmark.
Bingchao Wang, Yongcheng Jing, Chengyu Wang et al.
In briefRangeRectQuant: Compressing Scientific Maps So Any Future Rectangle Sum Stays Accurate
SummaryThis paper presents RangeRectQuant, a method for compressing scientific raster data so that the sum over any axis-aligned rectangle remains within a prescribed error budget. It controls errors in prefix sums during encoding, so users can later measure arbitrary rectangles without the source data or re-encoding. The method runs in linear time and needs workspace of about one row. Experiments on geographically held-out built-surface and water-occurrence maps show substantial storage savings over alternatives meeting the same requirements.
In briefTail Cues, Principal Corrections: A Lightweight Plug-and-Play Fix for Open-Set Test-Time Adaptation
SummaryThis paper addresses open-set test-time adaptation, where target data streams contain both covariate and semantic shifts. It introduces TALOR, a lightweight, backward-free module that uses the source linear head as a geometric anchor to split features into principal and tail coordinates. TALOR estimates an affine correction in the principal subspace via weighted ridge regression and uses it to rectify features before the source head. On ImageNet-C with six csOOD datasets, it outperforms the next-best UniEnt by 2.2% H-score while running 2.5x faster and using 12% of its GPU memory.
In briefSample-and-Score: Surrogate-Guided Coordinated Spatial Reuse in IEEE 802.11
SummaryThis paper addresses coordinated spatial reuse (Co-SR) in IEEE 802.11bn, where multiple access points transmit simultaneously to raise area throughput, a task hampered by a large combinatorial configuration space. The authors propose a sample-and-score method that randomly draws Co-SR configurations, scores each with a surrogate model, and selects the best. The surrogate is trained offline and deployed without online fine-tuning or sequential exploration. On unseen topologies with 2 to 16 access points, it recovers 52% to 77% of a perfect surrogate's throughput, and a perfect sample-and-score approach reaches 95% of the upper bound with 8192 candidates.
In briefIntrinsic Information-Theoretic Analysis of ReLU Nets
SummaryThis paper studies how mutual information in deep networks can be analysed without the bias of external estimators, by using the piecewise-linear structure of ReLU-type networks that partitions inputs into linear regions. It decomposes mutual information into geometric terms, with the dominant routing information having a plug-in estimator. The estimator agrees strongly with three established baselines, and a functional-equivalence quotient restores informative estimates when the partition saturates.
Johan Mylius-Kroken, Elisabeth Wetzer, Ali Ramezani-Kebrya et al.
In briefHorizON: A multi-horizon neural network for predicting how organic battery stacks perform over time.
SummaryOrganic electrode materials can reduce reliance on critical metals like lithium, cobalt, and nickel, but organic batteries still have limited long-term cycling stability. The paper introduces HorizON, a multi-fidelity, multi-horizon Deep Operator Network that learns DFT redox potentials and cycle-dependent specific-capacity trajectories within one model. It is trained on complete battery-stack representations and usage conditions so it captures interactions among stack components, not just the cathode. The framework is presented for long-term property prediction, inverse design, and process optimization of organic battery systems.
Subhash V.S. Ganti, Lukas Woelfel, Arla Diengdoh et al.
In briefActiveAlign: Aligning Robot Manipulation Through Active Perception
SummaryThis paper proposes ActiveAlign, a framework that uses active perception, such as moving the hand and touching objects, to close the gap between real-world and simulated environments. ActiveAlign-Waving performs hand-eye calibration using active hand motions and differentiable rendering, while ActiveAlign-Touching refines object alignment with tactile feedback to reach millimeter-level precision. The authors show that the method can curate simulation environments for uncalibrated robot datasets and enables a Unitree G1 robot to manipulate millimeter-scale objects such as pressing buttons and moving sliders.
In briefSwitchable unlearning that removes training data while keeping retrieval ability.
SummaryThe paper presents MoME, a framework that lets one model switch between Full mode for standard inference and Core mode for unlearned inference, supporting privacy requests to delete individuals' training data. Core parameters receive no gradients from the forget set, making Core independent of that data, and the authors introduce the Forgetting-Rank statistic to audit this independence. On Market-1501, VeRi-776, and RegDB, Core performs near a retrained reference under three membership attacks while retrieving above it.
WhyGradient-isolated switchable branches enable unlearning, audited by a rank statistic; validated across three datasets and several backbones.
In briefTesting AlphaEarth embeddings for crop classification across separability, time, and regions
SummaryThis paper proposes a representation-agnostic framework to compare 64-dimensional AlphaEarth embeddings from a geospatial foundation model against Sentinel-2 reflectance, NDVI phenology, and harmonic features for crop classification. The evaluation uses field-scale data from the conterminous US (2017–2023) and Northern India (2021). Embeddings give better distribution-aware separation and capture multi-year crop rotation, while harmonic and NDVI features transfer better from the US to India, and the combined feature space performs in a balanced way.
In briefHow boundary shape sets the best shallow ReLU approximation rate under weight limits
SummaryThis paper studies how the boundary shape of a convex body affects approximation of functions by shallow ReLU neural networks in the L^q norm. The target classes are unit balls of weighted variation spaces, which extend classical spaces by building domain geometry into constraints on output coefficients. The authors give upper and lower bounds for arbitrary convex bodies and determine optimal exponents for planar convex bodies and their products with polytopes, including cubes and cylinders. They show that boundary geometry can change the optimal exponent even when dimension, error norm, and weighting rule are fixed.
In briefLingGS: Separating Feed-Forward Geometry Prediction from Camera Tracking for High-Fidelity Gaussian SLAM
SummaryThis paper presents an online monocular RGB Gaussian splatting SLAM system that uses a frozen streaming 3D foundation model as a virtual RGB-D sensor for depth estimation. An explicit geometric solver estimates camera motion via feature reprojection, and a streaming Gaussian mapper integrates observations into a dense scene representation. The authors report improved reconstruction quality on indoor and outdoor datasets, about 7 FPS on a single RTX 4090, and 2-3 times the throughput of evaluated prior SLAM systems.
In briefuTHCD-C: A Robustness Benchmark for Tamil Handwritten Character Recognition under Corruptions
SummaryThis paper introduces uTHCD-C, a benchmark for testing how well Tamil handwritten character recognition systems withstand image corruptions. Built on the unconstrained Tamil Handwritten Character Database (uTHCD), it defines ten corruption types, including a handwriting-specific stroke-thinning corruption, at five severity levels. Nine convolutional and transformer-based architectures are compared using mean Corruption Error normalized to VGG16-BN. ConvNeXt and Swin achieved the lowest mCE values (0.75 and 0.83), while ShuffleNet and MnasNet showed higher values (1.55 and 1.85).
In briefWhen All Rollouts Fail: Policy-Compatible Joint Decoding for Hard Problems
SummaryThis paper addresses a problem in reinforcement learning with verifiable rewards, where models get no learning signal on problems beyond their current ability because every attempt fails. It proposes PICO, which uses distribution-constrained joint decoding so expert data guides generation while staying compatible with the current policy. These trajectories are optimized together with on-policy rollouts using a mixed-behavior objective. Experiments on diverse reasoning datasets and agentic benchmarks show PICO outperforms standard RLVR and prior off-policy baselines.
In briefEvolving Agent Teams: A Multi-Agent Framework Escaping Local Optima
SummaryThe paper finds that existing coding agents converge to local optima when optimizing real-world expert-written GPU kernels such as FlashInfer. The authors propose Evolving Agent Teams, in which a central supervisor diagnoses local optima and restructures the agent team, while a persistent file system stores accumulated experience. With an extended token budget, the generated CUDA kernel reaches a 1.2x mean speedup over the production FlashInfer MLA paged decode kernel and wins on 36 of 47 FlashInfer-Bench workloads.
WhyA supervisor restructures agent teams using persistent memory, with large speedups on several real kernel and robotics tasks.
In briefA conservative policy iteration method for zero-sum partially observable games
SummaryThis paper presents an approximate policy-iteration algorithm for finite-horizon zero-sum partially observable stochastic games (zs-POSGs). It keeps one policy per player and updates both along occupancy trajectories produced by conservative policy updates, using a single supporting hyperplane from the current joint policy instead of a full central value function. On finite-horizon zero-sum POSG benchmarks, it is competitive at small horizons and often gives a better runtime-exploitability trade-off at larger horizons than PBVI, HSVI, and CFR+.
Erwan C. Escudie, Matthia Sabatelli, Jilles Steeve Dibangoye
In briefAxisQ: Depth-Adaptive Channel-Frequency Quantization for Camouflaged Object Detection
SummaryThis paper introduces AxisQ, a post-training quantization framework for camouflaged object detection models. It allocates bit precision in the channel domain for shallow features and in the local-frequency domain for deep features under a fixed bit budget, guided by COD-aware sensitivity and a boundary-aware Hessian approximation. Experiments report better accuracy than strong baselines at low bit widths, with relative weighted F-measure gains of 12.7% for CFRN and 33.3% for CamoFormer-P on CAMO under W2A2 quantization.
In briefMVRCap: Comprehensive Video Captioning with Hierarchical Multi-Dimensional Verifiable Rewards
SummaryThe paper introduces MVRCap, a framework that treats video captioning as a reinforcement learning task guided by structured verification. Its core is a fine-grained verifiable reward with three parts: a hierarchical concept coverage reward, a factuality verification reward, and a conciseness constraint. The authors also build the MVRCap-QA and MVRCap-SFT datasets. The paper reports state-of-the-art results on benchmarks including VDC, CAPability, and VC4VG.
In briefBoosting-based joint distributional copula model for insurance claim frequency and severity
SummaryThis paper proposes a model-based boosting method for fitting structured additive distributional copula regression models with bivariate responses, focusing on jointly modeling claim frequency and claim severity. The approach builds on generalized additive models for location, scale, and shape (GAMLSS), allowing multiple distribution parameters to depend on covariates and keeping the model interpretable at the additive predictor level. Evaluations on synthetic and real-world data show that it recovers covariate effects and dependence structure, identifies informative variables, and achieves competitive predictive performance against benchmark models.]
In brief"Drop the Details, Get the Gist": A paper on dynamic concept representations for video recognition
SummaryThe paper extends concept-based image recognition to video by representing a multi-step video as a temporal sequence of segment-level concept score vectors, which a Transformer encoder uses to predict the class label. It also explores estimating each concept's contribution to improve interpretability. Tests on cooking and household-activity videos with three concept sets of different granularity found that coarser-grained concepts are sufficient for classification.
In briefRight Click, Wrong Box: Judging GUI Clicks by Their Outcomes
SummaryThe paper argues that GUI grounding benchmarks count a click as correct only inside one annotated box, even though interfaces often expose the same action through several elements. The authors define a click as correct when it produces the same outcome as the annotated target, and build outcome maps and an execution judge by executing about a thousand clicks per page on live web applications. Testing three strong grounders on a code forge and an online shop, they find that many box-rejected clicks still reach the intended outcome, and an audit on ScreenSpot-v2 yields corrections of 0.2 to 1.8 points.
WhyProposes outcome-based grading of GUI clicks, validated via live execution probes and a blind audit on ScreenSpot-v2.
In briefWhen the Critic Falls Behind: Co-Evolving Rubrics and Harnesses for Continual Agent Self-Improvement
SummaryThis paper proposes RuHarness, a framework in which an AI agent's evaluation rubrics and its execution harness evolve together during continual self-improvement. It continually discovers and reassesses behavioral criteria from execution trajectories, prioritizes the most relevant unresolved weaknesses, and uses them to guide targeted harness updates while the task objective stays fixed. Experiments across diverse agent tasks and backbone models show it outperforms strong harness-evolution baselines, especially on unseen held-out tasks.
In briefWho knows what, and what does it support? Evidence calibration for multimodal theory of mind
SummaryThis paper addresses multimodal theory of mind, inferring an agent's beliefs, goals, and intentions from only the evidence available to that agent. The authors propose PIP-ToM, which builds agent-specific evidence stores, aligns them with structured mental-state claims, and calibrates hypothesis scores while keeping the Bayesian inference process. With GPT-4o, PIP-ToM reports average accuracies of 92.56% on MuMA-ToM, 95.00% on MMToM-QA, and 95.00% on Hi-ToM, outperforming existing ToM methods.
In briefTAPESTRY: One Backbone for Video Generation and Robot Control
SummaryThis paper presents TAPESTRY, which repurposes a pretrained video diffusion transformer into a robot policy by treating actions as single-patch tokens in the video latent space. The shared backbone is trained with three diffusion objectives, and at inference it produces a compact action token sequence in a single DDIM step without generating video. It runs at 84 ms per action chunk on consumer hardware. The authors report strong performance in simulation and real-robot experiments at substantially lower inference cost than comparable joint video-action policies.
Jeremy A. Collins, Seungjae Lee, Rachanon Wachakorn et al.
In briefIncremental Knowledge-Graph Digital Twin for Evidence-Based Early Warning
SummaryThis paper presents a lightweight, incremental learner digital twin built as a knowledge graph to support early warning in learning analytics. The twin is updated at temporal checkpoints by replaying learning traces and assessment records, and predictive features are extracted through graph queries. Evaluated on a public learning-analytics benchmark under a temporal protocol, graph-mediated features matched the predictive behavior of raw-table features, and generated alerts were linked to traceable evidence chains.
Ngoc Luyen Le, Marie-H\'el\`ene Abel, Bertrand Laforge
In briefVariable-length feedback coding with universal decoding and bandit channel selection
SummaryThis paper studies variable-length feedback coding over K additive white Gaussian noise channels whose noise variances are unknown to both ends but lie in a known interval. The receiver picks one channel at each time, reports the choice over noiseless feedback, and learns channel quality while decoding. The authors propose two joint coding and channel-selection schemes, VLSF-Bandit and VLF-Bandit, and prove second-order achievability bounds for both.
In briefWords or Deeds? How LLMs Predict User Decisions in Human-Agent Dialogue
SummaryThis study examines how large language models combine users' statements with their past decision records to predict whether users will cooperate in a repeated prisoner's dilemma-style game. Forty-two users played 1,260 rounds against five agents, each expressing a different Big Five trait, and could talk with the agent before each decision. Decision records alone separated cooperation from defection well, and adding dialogue gave no detectable gain. Predictions leaned too heavily on users' statements, so errors were largest when users broke their promises.
In briefWeakness Localization Is Not Enough: State-Aware Recursive Knowledge Improvement for LLMs
SummaryThis paper argues that localizing a large language model's knowledge weaknesses does not determine how to fix them, since the same weakness can stem from different knowledge states. The authors propose Recursive Knowledge Improvement (RKI), which organizes knowledge into graph-structured units and, each round, estimates the model's knowledge state to select an intervention before re-evaluating. On HotpotEval and SafeInternalEval, RKI outperforms existing knowledge incorporation methods by 18.41 points on average.
In briefLatent Visual Contrastive Reasoning with Multi-Branch Exploration: a method letting models explore several visual clues in parallel
SummaryThis paper addresses complex visual reasoning in multimodal large language models, which often must locate and compare several question-relevant image regions. It proposes Latent Visual Contrastive Reasoning, which generates multiple question-conditioned latent visual branches as soft hypotheses over image tokens, trained with bounding-box evidence annotations and a soft contrastive branch-evidence alignment objective. Instantiated on Qwen2.5-VL at 3B and 7B scales, the method reports overall gains over the backbone and single-path latent reasoning baselines on perception-intensive visual reasoning benchmarks.
In briefSetGRPO: Calibrating Grader-Induced Policy Optimization
SummaryThis paper studies how randomness in graders used for rubric-based reinforcement learning can alter GRPO training updates, beyond what score discrepancies alone reveal. It proposes Set-Valued Group Relative Policy Optimization (SetGRPO), built from three modules: joint reward set calibration, geometry-guided safe recovery, and selective risk-gated optimization to control direction-reversal risk. The authors report that on ten open-ended benchmarks across three domains, SetGRPO achieves the highest macro average among compared methods on Qwen3.5 and Llama3.2 model families.
In briefUsing LLM-elicited priors to improve out-of-distribution tabular prediction via cross-validated regularisation
SummaryThe paper addresses the problem that prediction models degrade when deployment populations differ from training populations. It uses coefficient priors elicited from large language models as the shrinkage target of a regularised regression instead of zero, testing out-of-distribution prediction on five tabular datasets, with providers and prior structures selected only by source-domain cross-validation. The priors improved results on four of five datasets, by up to 7.9% in out-of-distribution mean squared error, though variation across candidate priors exceeded the average benefit. On some datasets the gain in in-domain accuracy matched the out-of-domain gain, and elicited knowledge clearly improved transfer only where training data were large enough that the in-domain gain disappeared.
S{\o}lve Langseth, Kristoffer Herland Hellton, Hugo L. Hammer
In briefPersonality Latents: conditioning frozen language models on individuals' own writing
SummaryThe paper proposes Personality Latents (PL), which learn a latent representation for each person from their own writing, without personality labels or authored profiles, and inject it into a frozen language model as a soft prefix. An encoder predicts these latents so a new person needs one forward pass, with optional test-time refinement. On held-out people, PL outperforms persona embedding and prompt-based baselines on MBTI axes, and the ranking holds across three embedding spaces. A latent learned with Qwen2.5-3B-Instruct also transfers personality to Llama-3.1-8B-Instruct through a new adapter alone.
In briefAPI-TAIL: Detecting Hidden Model Swaps in LLM APIs Using Temperature-Aligned Likelihood Tails
SummaryCommercial language model APIs return only text, so a provider could secretly route some requests to a cheaper model while billing for the advertised one. The paper introduces API-TAIL, an audit that first aligns an effective temperature to remove batch-wide shifts caused by legitimate settings, then counts remaining anomalous responses against a calibrated threshold. Tested on real commercial APIs, it controls false alarms and detects low-rate partial substitution with few queries, outperforming existing rank- and MMD-based audits.
In briefLanguage Models Keep Abandoned Analyses: Residue and Adaptation after Garden-Path Sentences
SummaryThis paper examines whether language models keep using syntactic analyses they have abandoned after reading garden-path sentences. Across 24 open models, the authors find a residue of the abandoned analysis that raises later object attachment, plus an adaptation toward the resolved structure. Activation patching shows the residue is carried by noun-phrase states computed before disambiguation, and replacing them removes most of it in the larger models.
WhyActivation patching separates residue from adaptation across 24 models, a novel mechanistic framing with broad, controlled experimental support.
In briefAn algebra for reasoning about Transformers
SummaryPrior theoretical work on Transformers mainly studies computability, such as Turing-completeness and parallelizability for unbounded precision and limitations for bounded precision. This paper builds an algebra for reasoning about Transformers, relying on functional decomposition results of Transformers. It treats functional decompositions as atomic actions and includes sequential and parallel composition, recursion for feedback, and abstraction for hidden layers. The authors describe it as the first attempt to model Transformers logically, usable for verifying the correctness of Transformer-based applications.
In briefSelect-TTT: Query-Guided Selective Weight Updates for Test-Time Training
SummaryThis paper studies test-time training, where a deployed language model stores incoming long context in a small set of fast weights. It finds that In-Place TTT, which writes every context chunk with equal strength, loses its benefit on realistic web long-context question answering because irrelevant content dilutes useful information. The proposed Select-TTT scales each chunk's write strength by its relevance to the user query, measured from the frozen output head without supervision or new parameters. The method consistently improves long-context question answering on two benchmarks and two backbone models.
In briefVisual Autoregressive Generation with Next Token Imagination: Letting Models Explore Plausible Images Beyond Training Data
SummaryVisual autoregressive models are normally trained only on the ground-truth next token, which limits how much of the visual distribution they learn. This paper proposes Next Token Imagination (NTI), which lets the model explore alternatives beyond the ground-truth next token under a criterion that keeps this exploration bounded by the visual distribution. Experiments on ImageNet show NTI consistently and substantially improves generated image fidelity, approaching state-of-the-art diffusion and autoregressive models with a smaller model and lower training cost.
In briefOmniGlimpse: An Asymmetric Identity Memory Method for Online Multi-View Tracking
SummaryThis paper studies online multi-view identity maintenance, which matches observations from different cameras to the correct targets without offline scene-specific training or camera calibration. The authors propose OmniGlimpse, an asymmetric propose-verify framework that separates adaptive short-term association from stable long-term identity consolidation. On the M3Track and MvMHAT benchmarks, it improves over the best evaluated online baseline on IDF1 and CVIDF1, and outperforms offline-trained references on IDF1.
In briefCVBL-CCTV-Face Dataset: Real-World Face Enhancement and Recognition for Surveillance
SummaryThis paper addresses how low resolution, motion blur, noise, and occlusion in CCTV face captures undermine reliable face recognition. The authors propose a framework for super-resolution and identity recognition for surveillance, and present the CVBL-CCTV-Face Dataset, which pairs low-quality face images from an 8MP IP dome CCTV camera with high-quality reference images taken in controlled settings in an unpaired manner. The modular system combines face detection, deep learning-based super-resolution, and identity recognition using L2-normalized embeddings. Enhancement models are trained and tested under three setups, and the system is deployed for identity verification in a real-time unconstrained environment.
Manish Das, Ajay Kumar Yadav, Shiv Ram Dubey et al.
In briefDoes LLM pre-training usually happen at the edge of stability?
SummaryThis paper revisits how well quadratic approximations model neural network training, especially in large language model (LLM) pretraining. It identifies an "Edge of Convexity" (EoC) failure mode, where persistent negative curvature makes the quadratic dynamics diverge regardless of learning rate, and proposes a generalized Edge of Stability (EoS) definition for stochastic adaptive optimizers. Across LLM pretraining runs with models up to 787M parameters, EoC is always observed, while EoS is common but not universal, disappearing at sufficiently small learning rates or batch sizes below a critical threshold.
WhyProposes Edge of Convexity and generalizes Edge of Stability, validated systematically across LLM pretraining runs up to 787M parameters.
In briefGraphs tracking how mathematical arguments depend on earlier steps, for review and measurement
SummaryThe paper proposes a framework that uses large language models to represent mathematical arguments as dependency graphs linking statements, equations, and assumptions to their sources. It defines structural measures such as depth, the longest prerequisite path, and bulk, the count of distinct ancestor edges, along with a standardized prompt protocol. The authors provide an evaluation design rather than empirical results, and note that suggested premises need mathematical validation and the measures describe the representation, not intrinsic difficulty or importance.
In briefRunning Several Candidate Edits at Once to Speed Up AI Prompt Tuning
SummaryThis paper extends the open-source GEPA optimizer so that each optimization step proposes and evaluates a batch of candidates rather than one at a time. In experiments on two tasks, most batched runs finished in half the wall-clock time or less, and the fastest in about a quarter to a third. Batched settings also scored higher on held-out tests. The authors offer a simple runtime model for the speedups under a finite worker pool and an initial look at splitting a fixed budget across parallel axes.
Jialin Zhang, Lakshya A Agrawal, Shangyin Tan et al.
In briefCan LLMs Reliably Grade Olympiad Proofs? A Study of Math Proof Verification
SummaryThis paper systematically studies whether large language models can verify natural-language mathematical proofs at the Olympiad level. The authors build an expert-graded dataset of USAMO 2025 candidate solutions and introduce TROJAN, an iterative self-critique pipeline for generating adversarial proofs at scale. They also propose MABGrader, a multi-armed bandit framework that outperforms the state-of-the-art ProofGrader by 26.4% in Quadratic Weighted Kappa and reduces grading error by 14%.
In briefEvaluation-Guided Demonstration Refinement for Vision-Language-Action Policies: Winning MetaFine Competition Entry
SummaryThis paper presents the winning entry in the MetaFine Open Competition. The authors fine-tuned a separate π0.5 policy for each of five tasks, revising demonstrated motions and adding expert continuations from states the policy reached to address grasping, pushing, and insertion failures. Checkpoints were compared on every instruction, and the submission, MAI, ranked first overall with a score of 96.66.
In briefBeyond the Evidence Coherence Trap: Agent-Driven Evidence Selection for Detecting Temporally Disordered Fake News
SummaryThis paper addresses fake news that rearranges real events into a false chronology, where individual facts may be accurate but their timing is misrepresented. The authors describe an "evidence coherence trap" in which retrieved evidence from a similar event in the wrong period leads to incorrect verification. They introduce ChronoFake, a benchmark of 30,706 articles, and propose DelSA, an agent that iteratively refines evidence selection and compares article and evidence chronologies to expose temporal conflicts. DelSA outperforms existing methods on ChronoFake and remains competitive on PolitiFact and GossipCop.
In briefThe Vocabulary Is Already Here: Adapting Pretrained Interpretable Features for Generative Recommendation
SummaryThis paper proposes API, a framework that reuses feature indices from pretrained sparse autoencoders as a semantic ID vocabulary for generative recommendation, without learning a new vocabulary. It introduces Pretrained Interpretable IDs and a Semantic Prediction Adapter that jointly predicts feature indices and activation strengths from interaction histories. Experiments report competitive performance across backbones and decoding schemes, along with explainable retrieval through feature contributions.
In briefMaking Self-Driving Labs Auditable: A Unified Framework for Safe, Traceable Automated Experiments
SummaryThis paper addresses the challenge of connecting flexible robot control with laboratory instrument interfaces in self-driving laboratories (SDLs). It proposes four components: LAM (Lab Action Model) for generating robot trajectories and instrument API calls, OpenSDL as an execution substrate, Safe-SDL as a runtime safety gate, and L3-Bench as an evaluation protocol. Together they provide an end-to-end design linking action generation, monitored execution, measurement feedback, and experimental evaluation.
In briefGeoGSplat: Geometry-Grounded Non-Pixel-Aligned Gaussian Splatting for View Synthesis
SummaryThis paper presents GeoGSplat, a feed-forward Gaussian Splatting model that adds geometric priors to a non-pixel-aligned framework to fix misplaced Gaussian centers. A dual-branch encoder captures coarse 3D structure and fine scene features, and a hierarchical decoder progressively generates dense 3D points and Gaussian primitives. The model outperforms prior pixel-aligned and non-pixel-aligned methods on DTU and NRGBD for novel view synthesis and reconstruction, with the advantage growing as target views move farther from the input cameras.
In briefDocSkill: Agentic reasoning for multilingual multi-hop document question answering
SummaryThis paper presents DocSkill, an agentic framework that uses a learned policy to select, revisit, or skip skills while gathering evidence and reasoning over multilingual documents. Skill trajectories are generated offline with bounded Monte Carlo Tree Search and used for supervised fine-tuning and Direct Preference Optimization. The authors also introduce MMHop-DocBench, a dataset of 60K QA pairs across 10 languages and eight document domains. With Qwen3-VL-8B, DocSkill reaches 41.0 accuracy and 49.2 F1 on MMHop-DocBench, outperforming the strongest baseline, and performs competitively on MMLongBench-Doc and DocBench.
In briefReinforcement learning for diffusions with step-size-independent logarithmic regret
SummaryThe paper studies episodic value-based reinforcement learning for Euler discretizations of controlled diffusions, measuring performance by cumulative regret against the optimal policy of the Euler model. Under stability and regularity conditions, it proves an O(log K) regret bound whose constants do not depend on the timestep δ. For drift linear in unknown parameters with unknown constant diffusion covariance, it introduces a constrained model-based fitted-Q algorithm (CD-FQ) and obtains a high-probability O(log K log log K) regret bound.
In briefFaster remote sensing image-text retrieval using interaction-free recall and text-guided reranking.
SummaryThis paper proposes a fast-then-fine (FTF) framework for cross-modal retrieval between remote sensing images and text. It combines interaction-free multi-granular recall with parameter-free text-guided reranking of top-K candidates using a Balanced Text-guided Interaction Block (BTIB). An inter- and intra-modal loss is also designed to jointly optimize alignment across multi-granular representations. The authors report competitive accuracy and significantly improved efficiency on public benchmarks.
In briefRobust LLM Data Selection via Content-Invariant Gradient Attribution
SummaryThe paper shows that existing gradient-based data attribution methods for LLM instruction tuning are sensitive to surface-form changes such as wording or formatting, which can shift scores even when content is unchanged. It proposes CIGA, which selects layers whose gradients are both influential and stable under semantics-preserving perturbations, then fuses these signals with input semantic representations. Experiments across benchmarks report more stable rankings, improved average performance, and lower attribution cost.
In briefLearning to correct interpolation for reconstructing intermediate scientific fields
SummaryThe paper proposes iFLEX, a method for reconstructing unobserved intermediate scientific fields between two observation times. It first forms a time-weighted average of the two observations, then uses a conditional diffusion model to learn and generate the residual between this average and the true intermediate field. The authors evaluate it on Navier-Stokes Kraichnan turbulence, Shanghai radar precipitation, and sea surface temperature data, reporting lower reconstruction error than the compared baselines.
WhyLearns residuals over weighted averaging with a conditional diffusion model, benchmarked on three scientific datasets against optical-flow and video baselines; useful for the subfield.
In briefA multi-task Mamba-CNN model for MODIS spectral-temporal unmixing that accounts for endmember variability
SummaryThis paper proposes MDF-Mamba, a model for estimating land-cover abundances from MODIS time series via spectral-temporal unmixing. It separates spectral and temporal information using per-channel spectral experts and a bidirectional Mamba temporal branch, then merges their outputs with a mHCFusion module. Pixel-wise learnable weights model intra-class endmember variability. On MODIS data from Alberta and Saskatchewan, the model achieves the best result on 12 of 14 evaluation metrics.
Motasem Alkayid, Yimin Zhu, Naser El-Sheimy et al.
In briefUsing neurons' biological identity to analyze how neural populations behave across animals
SummaryThis paper proposes a framework that groups neurons into latent clusters based on their molecular, morphological, and electrophysiological profiles, then analyzes population dynamics through these group-level representations. Because the assignments rest on biological features shared across individuals, the learned organization transfers to unseen animals without retraining or neuron-to-neuron correspondence. The method is evaluated on whole-brain calcium imaging in larval zebrafish and large-scale electrophysiology in mouse visual cortex, where it reports stronger neural decoding and manifold analysis performance.
In briefSensitivity of Readout and Operating Points in Event-Stream Candidate Selection
SummaryThis paper examines the trade-off between capturing target events and triggering false alarms in event-triggered recording. Using real event accumulations from 520 DailyDVS-200 clips across 47 participants, it compares encoders, accumulation durations, and temporal view counts for candidate selection. The authors find that readout choices and threshold calibration can change which method appears better. They conclude the evidence covers only offline, action-defined selection and does not establish a JEPA-specific advantage.
In briefCAMR: A Method for Continual Learning in Large Language Models That Decides Which Examples to Learn and Replay
SummaryThis paper proposes CAMR, a continual learning framework for large language models that balances learning new information with retaining old knowledge. It uses compatibility checks and a learned memory gate to choose, for each incoming example, among learning with general replay, reinforcing compatible related examples, or skipping verified duplicates. Evaluated with Qwen3-8B-Base on factual updating and cross-task benchmarks, CAMR reports leading results over compared methods.
In briefHy-Nomas: Free-form Interactive World Generation
SummaryThis paper presents Hy-Nomas, an 8B diffusion-based video world model that follows free-form text instructions and continuous camera control across multiple interaction turns. It pairs structured video captions with action-aware causal attention, and uses a vision-language Prompt Enhancer to turn user requests and interaction histories into conditions for each segment. The model is trained with a multi-stage curriculum followed by diffusion distillation. The authors report superior human-evaluation results over existing world models and release weights, code and benchmarks.
Qiaoling Chen, Jiaxiang Cheng, Xueqing Deng et al.
In briefA method that progressively refines SAM to segment phytoplankton more accurately.
SummaryThis paper proposes PhySeg, a segmentation framework built on SAM with a lightweight refinement module that gradually recovers fine details through multi-scale feature fusion. The authors also manually annotated instance-level masks on an existing dataset to create a phytoplankton segmentation dataset. Experiments show that PhySeg improves segmentation accuracy.
In briefSAGE: A Saliency-Annealed Gated Encoder for General Vision Recognition
SummaryThis paper proposes SAGE, an enhanced ViT-Adapter backbone that adapts self-supervised Vision Transformers to dense prediction by exploiting saliency, global context, and high-frequency details. It combines a saliency-guided architecture, an annealing-based training scheme, and a two-stage training mechanism that shifts from focal loss to Varifocal Loss. The authors report state-of-the-art or competitive results on COCO, COCO keypoints, ADE20K, TinyImageNet, and DIOR-R object detection, using one shared backbone with task-specific decoders.
