In briefVitaBench: A Benchmark for Memory Lifecycles Beyond Retrieval
SummaryThis paper argues that existing long-term memory benchmarks for language agents mainly test whether past information can be retrieved, while not directly evaluating whether a memory should stay exact, be compressed, consolidated, or deleted. The authors formalize memory lifecycle evaluation and introduce VitaBench, which has a controlled Diagnostic tier and a Naturalistic tier of generated customer-service-style interaction streams. Results show adaptive lifecycle policies outperform the best fixed strategy, and many states change their preferred action between short and full horizons. The authors conclude that retrieval and decisions about how memory persists are distinct evaluation problems.
In briefSelf-improving synthetic data for scalable embedding model adaptation
SummaryThis paper introduces SYNTRA, a self-improving agentic framework that turns a few retrieval examples into reusable instructions for generating training data for embedding models. Across 24 datasets in 15 domains, the authors report that the synthesis workflow improves retrieval even for strong embedding models, and that self-improvement outperforms human calibration on four of six evaluated tasks. The reported gains persist across a 32-fold range of synthetic training-data sizes on MS MARCO.
In briefMHWA: A Multi-timescale Hierarchical World-Action Model for Decision Making
SummaryThis paper presents MHWA, a world-action model for sample-efficient decision making that reduces compounding drift in long autoregressive rollouts. It uses a strided high-level context channel to update global context less often, and a gating network that routes a low-level mixture of experts based on global context and current observations. On a 14-game Atari subset with 10% offline data, MHWA (77M parameters) reaches an IQM human normalized score of 0.884, above the matched JOWA-70M baseline (0.529) and numerically comparable to JOWA-150M with about half the parameters.
In briefLightVAE: A Compression Framework for Faster, Compact Video Autoencoders
SummaryThis paper presents LightVAE, a method for compressing video variational autoencoders used in latent diffusion video generation, which are costly to run. It splits causal Conv3D layers into a spatial response and a temporal correction, applies low-rank compression only to the temporal part, and reduces depth through width-preserving block pruning with feature alignment. Reported results show decoding throughput gains of up to 2.24x on H100 and 1.11x on RTX 5090 for the Wan VAE family, and up to 2.97x on MiniMax-H3 while keeping 99.51% of the original VBench-2.0 score.
In briefAdding AIRL-based reward estimation to a fuzzy inference and policy gradient control method
SummaryThis paper proposes a fuzzy control method that adds a reward estimation mechanism based on adversarial inverse reinforcement learning (AIRL) to a base model that integrates policy gradient with fuzzy inference. To counter the unstable learning of the GAN framework used in AIRL, the policy is treated as the generator and the membership functions serve as the discriminator's cost function, so reward and policy updates run simultaneously on shared parameters without estimating a value function. In a vehicle speed control simulation, the method acquired policies with high generalization performance through stable and efficient learning.
In briefSemantic grounding method for discovering new classes at test time
SummaryThe paper proposes SEGA, a framework for Test-Time Discovery that recognizes known classes and detects novel ones in an unlabeled test stream after deployment. Existing methods give discovered classes only anonymous labels, whereas SEGA uses large language models to generate attribute-rich descriptions and, at test time, uses Relational Semantic Grounding to produce and rerank candidate concepts for unknown samples. This turns anonymous clusters into interpretable semantic concepts. Experiments show improvements over existing TTD baselines.
In briefMulti-token random-walk learning with meeting-triggered information exchange
SummaryThis paper studies decentralized learning in which multiple tokens random-walk over a network and exchange information when they meet. It compares two exchange policies, averaging stochastic gradients and averaging models, and derives iteration complexity lower bounds on complete graphs for non-convex, convex, and strongly convex objectives. Experiments show that Average-Model achieves lower mean-token loss, while Average-Gradient achieves lower best-token loss.
In briefPixARG: Pixel-Asynchronous Retrieval-Augmented Diffusion for Faster, More Realistic Image Generation
SummaryThis paper proposes PixARG, a framework that adds real-world image retrieval to diffusion models to improve efficiency and visual realism. It builds a pixel-level affinity map between the text prompt and a retrieved image, using it to set different noise levels across regions and to guide a pixel-wise asynchronous denoising scheduler. The authors report that experiments on LAION and COYO datasets in offline, online, and user-specific settings show better realism and semantic alignment with far fewer diffusion steps.
In briefDeKoR: An efficient method that fuses GNNs and LLMs by routing decomposed knowledge
SummaryThis paper addresses the mismatch between the knowledge spaces of Graph Neural Networks (GNNs) and Large Language Models (LLMs), along with the high computational cost of combining them. The proposed framework, DeKoR, decomposes LLM knowledge into complementary local and global views and distills it from the LLM, used as an offline teacher, into lightweight experts. An adaptive router learns how much each node should rely on each expert. Experiments on multiple benchmarks report that DeKoR outperforms existing baselines on node and subgraph classification.
In briefVela: Using Real User Behavior as Training Supervision for LLMs
SummaryThe paper introduces Vela, a framework that turns real user behavior records into trainable tasks and trajectories for large language models. From anonymized records of 100,000 users, it generates 100,000 tasks across 91 templates and retains about 75,000 after filtering. Fine-tuning Qwen3.6-35B-A3B on these trajectories improves user-related tasks by up to 11.5 points and general benchmarks by an average of 4 points.
In briefHow to Build a Benchmark That Measures Real Taste Judgment
SummaryThis paper introduces TasteBench, a pairwise good-versus-bad benchmark for taste across visual, brand, book, and tech items, built on expert-only gold labels and a control protocol separating craft judgment from test-gaming. In a pilot with three human experts and 250 pairs, the authors identify three validity threats: position bias (always choosing slot A scores 0.956), a filename lexical cue scoring 0.912, and a difficulty ceiling in craft domains. They release an eval harness and a side-randomization tool, which removes the position skew, while noting the craft-domain ceiling remains unresolved.
In briefLingdan-V2: Large language models for reasoning about traditional Chinese medicine prescriptions
SummaryThis paper presents Lingdan-V2, a family of traditional Chinese medicine language models with 4B, 8B, and 14B parameters, trained with continued pretraining, reasoning supervision, and prescription-centered reinforcement learning. The authors evaluate them on a public benchmark of 25,620 items and a private dataset of 4,348 spleen and stomach disease cases. The 14B model reaches a prescription generation F1 of 37.0 on the benchmark, versus 32.3 for DeepSeek-R1, and a herb-level F1 of 52.8 on the clinical dataset.
In briefLinear-Time Non-Local Codec Placement for Compressing Machine Learning Models
SummaryThis paper presents an algorithm that places encoding and decoding operators into a computational graph so that compressed models work with target hardware. It colors graph edges so each subgraph has a special node, called the absolute dominator, where traffic can be observed and parametric codec operators learned. The algorithm runs in time linear in graph size. Compared with naive local placement, it reduces modeled operand traffic by 24% to 44% across four models, including 24% on SigLIP and 34% on SmolLM2.
In briefDSABench: A Benchmark for Multi-Database Agents' Source Discovery, Coordination, and Execution
SummaryDSABench is a benchmark that tests whether AI agents can find the right databases in a large catalog, coordinate their results, and execute multi-database workflows. It contains 815 human-reviewed tasks, 618 of which require several databases and 105 of which allow more than one interpretation of the needed sources. In routing-only tests the best RA-Set score is 60.55%, but the best RA-Graph score drops to 34.66% when coordination is required, and execution success is lower still. Clarification questions substantially improve exact source recovery on ambiguous tasks.
In briefA joint diffusion model for designing protein binding interfaces
SummaryThis paper presents PIDiff, a multimodal diffusion framework that jointly generates target-conditioned protein sequences and 3D structures for protein-protein complex design. Its core PIFormer architecture combines geometry-aware invariant point attention with global sequence self-attention. The model was pretrained on the authors' PIBench dataset and fine-tuned for minibinder and antibody design. The authors report that it outperforms baselines on sequence design success rate, binding energy, and sequence-structure co-designability across all tasks.
In briefPURE: Improving Feed-forward 3D Reconstruction Without Labels
SummaryThis paper proposes PURE, a post-training refinement framework that improves depth and camera pose estimation for feed-forward 3D reconstruction models without ground-truth annotations. It distills structural knowledge from state-of-the-art 2D priors using relative depth constraints, affine-invariant geometry distillation, and feature-correspondence-based pose refinement. Experiments on synthetic and real-world dynamic benchmarks show consistent gains across multiple state-of-the-art architectures.
In briefAlignAudio: A Unified Dual-Alignment Framework for Noise-Robust Speech-Conditioned Audio Generation
SummaryThis paper proposes AlignAudio, a framework that generates target audio from speech input while staying robust to noise. It uses dual alignment: matching representations of clean and noisy speech to preserve semantic cues, and enforcing generation-level consistency within a flow matching approach. Experiments on speech-to-audio and voice conversion show performance comparable to baselines on clean speech and clearly better results across various noisy conditions.
In briefHidden Spatial Biases in Bayesian Optimization
SummaryBayesian optimization chooses where to explore using posterior uncertainty, but this uncertainty is not spatially neutral. On bounded domains, Gaussian-process posterior variance is typically larger near boundaries, giving variance-based acquisition functions a systematic edge bias. The paper keeps the predictive model fixed while modifying the spatial preference of the exploration term, comparing three alternative exploration principles on benchmark functions and stochastic robot-pushing tasks. These alternatives often lower regret, particularly early in optimization.
In briefVoTSpeech: Separating Voice Design from Speech Generation in Instruction-Following TTS
SummaryInstruction-following text-to-speech usually leaves voice design implicit, since standard speaker embeddings encode reference audio rather than text-based targets. This paper proposes VoTSpeech, which uses an explicit intermediate voice latent predicted from text instructions by a flow-based module, and shares it as a condition for the language model and the Acoustic DiT. On InstructTTSEval-ZH, it performs competitively with strong open-source models while maintaining high naturalness.
In briefAcrossVAM2.0: Conflict-Aware Point Supervision for Efficient Video Diffusion Tuning
SummaryThis paper adapts a 5.01B-parameter text-to-video diffusion transformer using point correspondence supervision, proposing AcrossVAM2.0 with a confidence-weighted tracking head, zero-initialized residual feedback, and asymmetric PCGrad to limit conflicts with the main generation objective. Across three seeds and a 24-prompt evaluation, it lowers tracking cycle error relative to a matched control and stays near control on six VBench dimensions. However, the tracking gain varies by seed and I3D-FVD worsens, so the authors conclude it improves correspondence consistency without uniformly improving video quality.
In briefDexCast: Learning Real-World Robot Hand Control from Human Videos
SummaryThis paper presents DexCast, a framework that learns dexterous robotic hand control directly from monocular RGB-D videos using joint hand-object reconstructions, without teleoperated demonstrations. It also proposes a DexHand Motion Perception Module that captures inter-joint coordination and temporal dependencies in compact latent representations to condition a diffusion policy. Real-world experiments report strong performance across diverse dexterous manipulation tasks.
In briefSpeeding up cross-model LLM inference by reusing attention maps
SummaryThis paper presents MapFlow, a system for serving heterogeneous LLMs that reuses attention maps instead of model-specific KV states. Attention maps are more transferable across models, and the approach adds partial recomputation, lightweight linear translation, block-level sparse compression, and a fused retrieval kernel. Experiments report lower prefilling latency and better end-to-end serving efficiency while preserving generation quality across model pairs and multi-turn workloads.
In briefPatchwise adaptive normalization for one-way coupled transformer PDE surrogates
SummaryThe paper introduces Patchwise Adaptive Normalization (PAdaNorm), which partitions exogenous fields such as medium density or obstacle geometry into patches aligned with state tokenization for transformer-based PDE surrogates. On 60-step autoregressive rollouts across three PDE families, it matches or improves on parameter-matched baselines by up to 16% in mean long-horizon variance-scaled RMSE. For static conditioning, it caches embeddings and reduces inference FLOPs per step by up to 36%.
Emmanuel Gottlob, Hugo Thimonier, Ludovico Buizza et al.
In briefDemiurge: Generating Navigable 4D Worlds by Autoregressive Video Re-rendering
SummaryThis paper presents Demiurge, a pipeline that re-renders video from new camera viewpoints using a dynamic 4D point-cloud memory cache. It improves consistency across different camera trajectories through four components, including an appearance video diffusion model and a memory-consistent depth completion model. The authors report state-of-the-art results, including handling large camera motions such as 180-degree rotations, and outputs that can be optimized into 3D Gaussian Splatting representations.
In briefTriage-PiEvo: Evidence Admission for Principle-Evolving Scientific AI Agents
SummaryThis paper addresses inefficiency in LLM-based AI Scientists caused by noisy observations, which existing principle-evolving agents handle by revising on every anomaly, even when the anomaly is an evaluation artifact. The authors propose treating revision as evidence admission: anomalies guide search immediately, but only independently verified evidence may rewrite persistent principles, and contradicted principles are retracted. They implement this as TRIAGE-PIEVO, which uses a staged verification gate to control entry into principle state. Across four scientific tasks and three backbones, it reports higher average solution quality than the strongest baseline and reduced false evidence admissions when plugged into external research harnesses.
In briefA hybrid kinematic GCN and elastic Transformer for efficient Indian Sign Language recognition
SummaryThis paper proposes an efficient hybrid framework for recognizing Indian Sign Language (ISL) glosses. It uses an entropy-driven SVD sampler to extract informative keyframes and an Elastic Transformer with continuous dual positional encoding to handle irregular time gaps. A kinematically-conditioned adaptive GCN models bimanual geometry, and the authors built a 120-class, 9,453-video ISL dataset, reporting 89.8% accuracy.
In briefHumanPhys: A Human-Centric Physical Perception Benchmark for Visual Reasoning
SummaryThis paper introduces HumanPhys, a benchmark for perceiving physical quantities related to people in real-world scenes, along with a training corpus built by the same pipeline. Questions cover depth, inter-person distance, gaze, displacement, velocity, blink counts, and contact deformation, with labels derived programmatically from calibration, metadata, or renderer state. The authors evaluate open- and closed-source multimodal LLMs across six dimensions and report that all remain challenging for current models.
In briefCAST-ASD: Causal Attention Across Speakers and Time for Live Speaker Detection
SummaryThis paper presents CAST-ASD, a causal model for active speaker detection that jointly reasons over all faces in a scene while labelling each frame without looking ahead. It applies attention across faces within the current frame and causal temporal attention over each face's history. On MSDWild, it reduces competing-speaker false alarms by 32% relative to the per-face causal detector it extends and outperforms all causal baselines on both validation splits. Code and models are publicly available.
In briefCompetition Domains Matter: Selective Head Globalization for Training-Free Sparse Attention in Video Diffusion Transformers
SummaryThis paper introduces EG-Sparse, a training-free sparse attention method that speeds up video diffusion transformer inference. It routes low-entropy attention heads to grouped cross-head selection while keeping per-row local selection for high-entropy heads. A controlled diagnostic shows grouped cross-head selection cuts reconstruction error for low-entropy heads by 9.8%, while forcing high-entropy heads into a global pool raises it by 7.9%. Across four configurations, the method improves video quality while reducing attention density and GPU denoising latency.
In briefSPARE: Sharing a visual encoder and retrieving past video clips to answer questions about live video
SummarySPARE is a framework for answering questions about streaming video. A lightweight gate first checks whether recent frames provide enough evidence, and if not, a multimodal embedding model retrieves relevant historical clips. LoRA instruction tuning teaches the model to use both sources together and to abstain when evidence remains insufficient, while the understanding and retrieval branches share a visual encoder. The paper reports 77.78% on the evaluated Real-time and Backward task groups of OVO-Bench and 73.31% on StreamingBench.
In briefLength-normalized policy updates can reduce task success
SummaryThe paper studies whether dividing each trajectory's score contribution by its realized length can reverse the direction of a reward-centered policy update. In a finite trajectory model, it derives an exact ascent threshold for a three-outcome softmax policy and gives a binary-reward example where the expected Euclidean update decreases task success. It also analyzes natural-gradient ascent and a third-order U-statistic estimator of the return change.
In briefTETRIS-4DGS: A Post-Training Method for Compressing 4D Gaussian Splatting
Summary4D Gaussian Splatting (4DGS) is a dynamic scene representation that renders in real time with high quality but requires very large storage. This paper proposes a general post-training compression framework that uses existing image and video codecs, applying importance-based adaptive quantization to Gaussians. It also reorganizes attributes into codec-friendly 2D maps using motion- and importance-aware packing. The authors report over 40x compression within seconds while maintaining high quality.
In briefDoes Failure Feedback Direct Agent Self-Modification?
SummaryThe paper asks whether self-improving agents use failure feedback to specifically repair targeted failures. Across 145 modifications in public Darwin Gödel Machine logs, no detectable target-specific advantage appeared. Controlled experiments show relevant feedback can improve target repair, but target repair did not reliably imply held-out improvement.
In briefRASSG: A Method That Decides When Voice-Triggered Tools Should Start or Be Cancelled
SummaryStreaming speech assistants can start tools before a user finishes speaking, but later self-corrections can make early arguments invalid, leading to irreversible stale executions. The paper proposes RASSG, which launches a tool when three consecutive speech recognition results share identical nonempty arguments and cancels it at the last feasible moment, using calibrated invalidity risk and execution costs. Tested on ReCAST, FDB-v3, and Audio2Tool with a controlled execution clock, RASSG cuts the stale execution rate on ReCAST's prospective-repair subset from 50.00% to 3.15%, with a mean delay of 3.676 seconds on successful trajectories.
In briefTGuide: Faster Autoregressive Video Diffusion via Historical Trajectory Guidance
SummaryThis paper presents TGuide, a framework that speeds up autoregressive video diffusion when only four denoising steps remain. Its core is a lightweight Anchor-Transport-Correct predictor guided by historical sampling trajectories, which combines features from the current chunk's preceding step with the previous chunk's same-timestep features. Experiments on Self Forcing, Causal Forcing, and HY-WorldPlay show preserved generation quality and denoising speedups of up to 1.89 times over the original 4-step generators.
In briefProspective Forcing: Teaching Video Generators to Look Further Ahead
SummaryThis paper addresses a limitation of autoregressive long video generators, which are usually trained to predict only the next chunk and so struggle to keep quality over long rollouts. The authors propose Prospective Forcing, a multi-head framework with lightweight heads that each predict chunks at different future temporal offsets from a shared history representation. At inference, accepted future-chunk predictions are used to skip full backbone executions to speed up sampling.
In briefTAG: Calibrating SAM3 for General Image Segmentation Without Dense Labels
SummaryThe paper finds that SAM3's main weakness is recognizing which concepts are present in an image, not the quality of its masks. The authors propose TAG, which uses only image-level category labels, adding image-conditioned low-rank adapters and Soft Boundary Ranking objectives to calibrate concept presence. It also couples a multimodal large language model with SAM3 through Confidence-Weighted Distillation. With only 3.1M trainable parameters, TAG performs well across diverse scenarios and surpasses some mask-supervised methods on several metrics.
In briefCan synthetic drone test data be trusted to choose AI models? A check against real data
SummaryThe paper examines whether synthetic drone (UAV) imagery can reliably replace scarce real data for selecting perception models. On a road-debris benchmark, synthetic data lowered absolute scores but preserved the model ranking and top model. On a separate business corpus with four detectors, synthetic testing changed the top-ranked detector and reversed three of six pairwise preferences. An audit of five UAV scenarios also found notable real-synthetic mismatches in class mix, object density, and target scale, especially under-representing small targets.
In briefBeyond Independent Values: Optimal Auctions for Correlated Advertiser Valuations in LLM Ads
SummaryThe paper proposes AffMyerson, an auction for sponsored slots in conversational LLM assistants, where advertisers read the same conversation and their willingness to pay is correlated. It models values with a multivariate normal prior under an affiliation sign condition and proves that every prior in this family yields a truthful, ex-post individually rational auction. In experiments with LLM-agent bidders, it earned 6.2% more than a tuned second-price auction, 6.1% more than Ronen's lookahead auction, and 9.2% more than Myerson's independent-values auction, with a 4.5% gain on real keyword search queries.
In briefShaping Latent Mixture-of-Experts: Expert Shape and Layerwise Capacity Allocation
SummaryThis paper studies design choices for mixture-of-experts (MoE) models. LatentMoE runs expert computation in a lower-dimensional space, making each expert's input dimension and intermediate width explicit design parameters. The authors show that balancing these two dimensions maximizes the rank ceiling at fixed per-expert parameters, and controlled sweeps favor near-square experts. They also use layerwise intrinsic dimension to allocate capacity across depth, reporting lower validation loss and stronger downstream performance than layer-homogeneous baselines.
In briefAegisVox: Privacy-Protecting Long Speech While Keeping Clinical Information
SummaryThis paper presents AegisVox, a framework that assigns consistent surrogate voices and replaces identifier-bearing speech turns with document-consistent substitutes to protect privacy in clinical audio. Tested on 390 public simulated or non-patient conversations totaling 63.6 hours, text-to-speech rendering realized 535 of 601 planned replacements, versus 453 for a codec-based speech editor. The authors find that better replacement realization can coexist with worse clinical or speaker-linkage measurements, and they state that the framework supports use-specific evaluation rather than guaranteeing identifier removal or clinical safety.
In briefPEARL: A Reinforcement Learning Method That Credits Intermediate Decisions in Multi-Turn Agents
SummaryWith only terminal rewards, it is hard to tell which intermediate decisions in a multi-turn agent should be adjusted. PEARL compares two continuations from the same history with different outcomes, locates their first divergence, and uses recurring local decision patterns from recent success and failure memories to selectively strengthen negative advantages at lower-reward decisions, without an extra critic. The paper reports gains over the ARPO baseline on AppWorld and ALFWorld.
In briefPVFH-Bench: A Benchmark for Premise-Induced Visual Fixation Hallucinations in Vision-Language Models
SummaryThis paper introduces Premise-Induced Visual Fixation Hallucination (PVFH), a failure mode where vision-language models anchor on visually confusable evidence, accept a false premise, and describe objects that are absent from the image. The authors present PVFH-Bench, which automatically builds visually confusable but absent targets and includes diagnostic subsets. Experiments on 10 state-of-the-art models show all are highly vulnerable, and explicit chain-of-thought reasoning can make the problem worse.
In briefAutomatic assembly modeling via attention-based assembly feature mate prediction
SummaryThis paper proposes a data-driven framework that automatically builds assembly models from unassembled B-Rep parts, using assembly features to group mating-interface geometry and narrow candidate constraints. The authors built a large dataset with an automatic annotation pipeline and trained an attention-based network to predict assembly feature mates, guided by assembly feature patterns. Reported assembly constraint accuracy is 74.79% on a six-assembly comparison set and 64.09% on an extended 16-assembly test set. With ground-truth face labels supplied, accuracies rise to 81.78% and 80.93%.
In briefAuditing Whether Open Jev-Style Retrieval Models Judge Each Question Independently
SummaryThis paper checks whether several open Jev-style models truly evaluate each typed question independently against shared evidence. Using 276 constructed three-passage states from 92 TREC queries, the authors apply paired interventions that change questions, displayed evidence, and candidate eligibility. Expanding the question set flips some binary judgments and changes selection-or-abstention actions, and field masking traces these flips to schema fields serialized into each scoring prompt.
In briefDuSRFlow: Dual-lens super-resolution using parallax-aware flow and complementary reference warping
SummaryThis paper presents DuSRFlow, a method that enhances a wide-angle low-resolution image using a telephoto reference image. It introduces DuFlowNet, a flow estimator built for dual-lens alignment, and a Complementary Reference Warping strategy that adapts to local reference reliability. The authors also release DuSR-RealV2, a real-world benchmark with large parallax and occlusion, and report state-of-the-art results across several benchmarks.
In briefTagBO: LLM-guided task-aware Bayesian optimization for scientific discovery
SummaryThe paper addresses mixed-variable optimization in scientific discovery, where standard Bayesian optimization relies on structural descriptors that may not reflect task performance similarity. It proposes TagBO, which combines LLM-derived task-aware relational priors with descriptor-based structural information. TagBO runs constrained Gaussian process Bayesian optimization over a joint discrete-continuous space. Experiments on chemistry and materials formulation tasks report improved sample efficiency under fixed evaluation budgets across LLMs of varying capabilities.
In briefFrom Existing Methods to New Research Ideas: Two Perspectives on Recursive Self-Improvement
SummaryThis paper examines recursive self-improvement (RSI), in which AI systems take part in their own continued development. The authors propose a taxonomy with nine design dimensions in four groups to place existing methods in a shared design space. They also draw on traditional deep learning mechanisms such as continual learning and evolutionary computation to suggest new research ideas, and review benchmarks for evaluating self-improvement.
In briefEarthAlign: Aligning Earth System Data with Language Representations
SummaryThis paper introduces EarthAlign, a large dataset from 2000 to 2024 combining 10 expert-curated natural disaster sources with normal-weather samples, each paired with time and location information and four caption styles aligned to 63-variable ERA5 data. The authors propose MeteoLIP, a dual-encoder model that learns meteorology-language representations using a geolocation- and time-conditioned tokenizer. MeteoLIP is reported to outperform CLIP and SigLIP baselines on cross-modal retrieval, weather-condition classification, and physical variable estimation.
Fadillah Adamsyah Maani, Numan Saeed, Jiaqi Ma et al.
In briefForecasting road-network functionality loss from rainfall using transferable graph learning
SummaryThe paper introduces G-R2R, a graph-learning framework that forecasts rainfall-induced declines in road-network functionality for unseen transportation networks without retraining on the target region. Tested on simulation-derived data from ten flood-prone regions in Wenzhou, China, it outperforms representative baselines, with a stabilized MAPE of 2.33% and R² of 0.855 on the fixed unseen-region benchmark. A replay of Typhoon Fitow also reproduces the simulation reference's temporal and spatial patterns.
In briefLLM-guided Bayesian optimization for exploring chiplet design spaces more efficiently
SummaryThis paper proposes a framework that combines a large language model with Bayesian optimization to simulate and automatically optimize 2.5D and 3D CPU-GPU heterogeneous chiplet designs. The authors extend an open-source chiplet simulator to evaluate 3D designs with graph topologies. An LLM generates policies to guide the optimization, using adaptive directional policy reasoning and exploration gain-aware fine-tuning. Experiments on 2.5D and 3D chiplets show improved design space exploration efficiency.
In briefEviScene: Agentic Image Factuality Assessment Beyond “Is It AI?”
SummaryThe paper introduces Image Factuality Assessment, where a system judges whether the situation depicted in an image is factually supported or refuted, with evidence-based explanations. The authors release EviLens, a benchmark of 10,174 expert-reviewed images covering twelve factual error types, and EviScene, a multimodal ReAct agent that inspects images, retrieves external evidence, and refines its questions. A trained Qwen3.5-9B version reaches 79.05% balanced accuracy and 52.26% Strict Evidence Sufficiency Rate on 1,527 test images, outperforming its base agent. Code, models, and data are slated for release.
In briefSeparating Knowledge Forgetting from Refusal Learning in Multimodal LLM Unlearning
SummaryThis paper argues that current multimodal unlearning evaluations only check whether a model stops outputting target information, which can hide hallucinated facts or refusals that leave the knowledge intact. The authors propose Forget-Aware Self-Distillation (FASD), which splits the goal into two channels: a behavior channel that distills refusals from an ignorance-prompted copy of the model with target tokens masked, and a knowledge channel that directly suppresses ground-truth answer spans. According to the authors, experiments show FASD unlearns target knowledge, prevents leakage in open-ended tasks, and preserves performance on retained data.
In briefCompLit: A Literature Search Benchmark for Implicit, Cumulative, and Unmet Needs
SummaryThis paper introduces CompLit, a benchmark for testing AI agents on finding scientific papers across all eight arXiv domains. It adds three diagnostic settings: implicit requirements, requirements that build up over a search, and needs that no paper satisfies. The strongest agents score 86-87% on standard queries but drop to 66-68% with implicit requirements and abstain on at most 53% of unmet queries.
In briefBidirectional Adaptive Transformers for multimodal anomaly detection: aligning RGB and 3D point cloud features
SummaryThis paper addresses the difficulty of fusing RGB image and 3D point cloud features for multimodal anomaly detection. It proposes the Bidirectional Adaptive Transformers (BAT) framework, whose Bidirectional Feature Alignment module uses mutual prediction to link 2D texture and 3D geometry. Lightweight image and point cloud adapters enable parameter-efficient fine-tuning of pre-trained encoders. The authors report state-of-the-art results on MVTec 3D-AD and Eyecandies.
In briefToF-Wheel Odometry: Robot Localization That Adapts When Sensing Is Limited
SummaryThis paper proposes ToF-Wheel Odometry (TWO), a filter that tightly combines wheel motion data with geometric constraints from Time-of-Flight cameras for ground robot localization. Its core is a degeneracy-aware method that adjusts process noise by direction according to how observable the geometry is. The approach suppresses geometric corrections when geometry is degenerate and accepts them as real wheel violations when geometry is distinct. Experiments reportedly show better trajectory accuracy and robustness than static and adaptive Gaussian filters.
In briefAttribute-Centric Vision-Language Matching for Detailed Gait Analysis
SummaryThis paper introduces RA-GAR, a large multimodal benchmark for gait attribute recognition with 1,001 subjects, 242,729 walking sequences, and 15 static and motion attributes. It proposes Gait-VL, which matches attribute-specific visual features from multiple frames with textual class prototypes built from LLM-expanded descriptions. The authors report state-of-the-art attribute recognition on RA-GAR and consistent gains when the attribute-pretrained encoder is used in identity recognition pipelines.
In briefFlow-support-weighted candidate proposals for test-time verified VLA policies
SummaryThe paper observes that fine-tuned vision-language-action models produce nearly identical action samples, limiting the benefit of test-time verification. It proposes Action Dynamic Sampling (ADS), which during training weights additional trajectories by flow support and corrects the policy's action distribution toward them. Deployment cost is unchanged. Reported results include higher proposal diversity and improved success on simulation and real-robot tasks.
In briefReCAP: Keeping Only Key Memory for Long-Horizon Robot World Models
SummaryThis paper addresses the instability of autoregressive robot video world models built on discrete visual tokens over long rollouts. It proposes ReCAP, a training-free context-scheduling method that keeps the initial scene anchor and a window of recent visual-action blocks while evicting older generated history. Evaluated on RT-1, BridgeV2, LIBERO-90, and CALVIN, it reports better visual fidelity and action consistency than Full Context decoding, with LPIPS reduced by 63.3% on BridgeV2.
In briefFine-tuning strategy, not just data: a rigorous audit of gender and stereotype bias in embedding models for HR recommendations
SummaryThis paper audits six multilingual sentence-embedding models used in job-candidate matching for gender and stereotype bias, using the Bias-in-Bios and StereoSet benchmarks. It corrects a non-independence error in a naive significance test and applies profession-level paired tests, bootstrap intervals, and permutation tests. Knowledge-distillation fine-tuning of E5-Base significantly increased measured bias, while binary-classification fine-tuning showed no statistically distinguishable change, suggesting fine-tuning strategy is an actionable design lever for HR recommenders.
In briefUIDPO: Aligning Language Models with Human Preferences via Uniform Information Density
SummaryThis paper applies the psycholinguistic Uniform Information Density theory to preference optimization for large language models. The authors assume humans prefer language with evenly distributed information, since this reduces cognitive effort. They build UIDPO, which combines global standard deviation and local second-order differences into an information-uniformity metric within the optimization objective. UIDPO performs comparably to, and sometimes better than, SimPO on AlpacaEval2, ArenaHard, and MT-Bench, and better reflects uniform information tendencies in complement clause analysis.
In briefIterate on a frontier LLM, then distill live behavior into a small model to cut inference cost
SummaryThe paper proposes first refining desired behavior via prompts on a frontier LLM using live traffic feedback, then distilling recorded production decisions into a small language model through supervised fine-tuning. In a home-care scheduling agent's yes/no safety decisions, a LoRA-tuned Qwen3-8B reached 0.913 agreement with recorded decisions, while the teacher model agreed with itself only 72.7% on re-runs. When they disagreed, a blinded professional scheduler sided with the student 48 to 29. The SLM also showed roughly 2.5 times lower median latency at a fraction of the cost.
In briefCB-Agent: Concept-Based In-Context Learning for Choosing Sequential Actions
SummaryThe paper proposes CB-Agent, which learns a reusable concept vector from expert trajectories by fitting the conditional distribution over candidate actions with maximum likelihood. During ReAct interaction, the agent selects actions from this distribution without placing full demonstrations in the action-scoring context. Across TextWorld, ALFWorld, and ScienceWorld, it achieves the highest cross-benchmark average at all three tested model sizes.
In briefLearning After Execution with Action-Specific Credit
SummaryThis paper proposes learning after execution, which treats the result of running a generated program as an intermediate state for an extra learning step by the same policy, rather than only as a final score. Action-specific credit assigns credit separately to the initial action and the corrective action. On 945 held-out image-to-TikZ examples, single-call render success reaches 92.70%, compared with 88.25% for continued single-stage GRPO with the same number of additional iterations.
In briefEvolving to Match Capabilities: Improving LLM agent self-improvement by adapting skills to weaker models
SummaryThis paper studies how LLM agents improve by revising their harness, the guidance skills they follow, without changing model parameters. The authors find that skills helping strong models can hurt weaker ones, and propose skill adaptation, which rewrites skills using a fixed meta-skill that targets the executor's capabilities. Across seven models, the meta-skill raised final scores by 2.0 to 5.1 points over the same self-evolution loop without it.
In briefReward-Aligned Distillation: Weighting Teacher Supervision by Its Predicted Reward Effect
SummaryThis paper proposes Reward-Aligned Distillation (RAD), a method for improving on-policy distillation of large language models. It uses a bilevel framework to weight each distillation signal by its predicted effect on the student's reward after a temporary update, rather than distilling every token equally. At a matched training-response budget, RAD reports higher mean accuracy than the strongest baselines on six competition-math benchmarks and higher pass@1 on LiveCodeBench-v6.
In briefBalancing Carbon Emissions and Latency Targets in Geo-Distributed LLM Serving
SummaryThis paper studies how to reduce the carbon emissions of large language model inference across geo-distributed datacenters while meeting strict latency service-level objectives (SLOs). The authors build a latency-centric simulator calibrated with hardware measurements and compare schedulers ranging from greedy heuristics to a facility-level mixed-integer linear programming (MILP) approach. Across three LLMs and MLPerf SLOs, the MILP scheduler reduces carbon emissions by up to 60% relative to round-robin, with an average SLO violation of 3%.
In briefPsyMemBehav: A Psychology- and Memory-Based Two-Step Reasoning Method for Simulating Social Media Behavior
SummaryThe paper introduces PsyMemBehav, which uses psychology-informed user profiles and retrieved memories to reason from identity to platform actions, such as choosing a target and then commenting, liking, or reposting. The authors built a benchmark from 5.57M recorded behaviors of 5,862 real users across Reddit, Bluesky, and Twitter. Its 27B model ranks first in 11 of 12 individual-level metrics and reports the lowest group-level MAE and MSE among compared systems. The benchmark covers only recorded engagements and does not model no-action decisions.
In briefFrameFold: response-guided cache merging for long autoregressive video generation
SummaryAutoregressive video generators that keep all past key-value states become slower and use more memory as videos get longer. This paper proposes FrameFold, a training-free method that merges adjacent historical cache entries, guided by how recent queries respond before and after merging. Under matched cache budgets, it reports better consistency and visual quality on Self Forcing, Causal Forcing, and LongLive at 30 to 120 seconds. Eight-minute continuous Self Forcing rollouts also show less quality degradation than the baselines.
In briefStrong Gaussian approximation for constant-rate SGD, covering bias and whole-trajectory inference
SummaryThis paper develops strong Gaussian (KMT-type) approximations for stochastic gradient descent (SGD) with a constant learning rate, under heavy-tailed noise, where the number of iterations grows as the learning rate shrinks. It builds simple Gaussian processes on the same probability space and bounds the maximum difference between SGD errors and their limits. Unlike earlier central limit theorems that cover only single iterates, the results enable simultaneous inference over the whole trajectory and yield functional central limit theorems. The paper also addresses the trade-off between SGD bias and initialization effects, removing them with a burn-in scheme and Richardson-Romberg extrapolation, and includes applications to risk control and simultaneous inference with simulations.
In briefChunk-wise Model-based Policy Optimization: a reinforcement learning method that avoids recurrent model rollouts
SummaryThe paper proposes a model-based reinforcement learning method that uses a chunk-wise world model as a differentiable evaluator during training to optimize a chunk-level policy without repeated model calls. The world model and policy jointly predict H-step imagined rollouts in a single forward pass, allowing direct optimization of value expansion and avoiding costly, unstable backpropagation through time. The method is evaluated on 19 high-dimensional humanoid tasks from HumanoidBench and deployed on a real quadruped robot for bipedal walking.
In briefPGSB: A Pretrained-Guided Shared Basis for Merging LoRA Models
SummaryThis paper proposes PGSB, a two-stage framework for merging LoRA adaptations that were trained for different tasks into a single model. It first identifies update directions shared across the task-specific LoRA updates, then re-parameterizes those directions using their interaction with the pretrained weights. The authors report that PGSB consistently outperforms state-of-the-art baselines across multiple benchmarks.
In briefAdversarial Gradient Forcing: a method for robust few-step autoregressive video generation
SummaryAutoregressive video models generate videos block by block, but errors accumulate over long generations. This paper introduces Adversarial Gradient Forcing (AGF), which uses distribution matching gradients from future block predictions to select adversarial perturbations of self-generated histories, then trains the generator on both original and perturbed histories. Experiments report the highest VBench Total score for 60-second video generation and the best visual quality for four-minute videos. AGF requires only 5-second rollouts for training.
In briefLASER: A Search Method That Builds Revocable Certificates for Long-Range Protocol Dependencies
SummaryThe paper addresses protocol specifications whose supporting evidence is scattered and may later be corrected. It introduces LASER, which treats missing support and unresolved corrections as explicit obligations, searches typed evidence routes, and uses a symbolic checker to decide whether a behavior is warranted, refuted, or unresolved. On 91 BehaviorCheck families, LASER improved joint accuracy over the strongest local reader on two model families while reducing false discoveries by a factor of 2 to 3.
In briefLLM-Based Relevance Evaluation for Dynamic Product Advertising Retrieval
SummaryThis paper proposes a user-product relevance metric for offline evaluation of Dynamic Product Advertising retrieval systems. User engagement history is compressed into intent representatives, and large language models judge how relevant each recommended product is to each intent. The scores are aggregated into a user-level measure. The authors report that the metric correlates strongly with offline and online business metrics, supporting faster iteration before A/B testing.
In briefActive Neural Matter: A Framework for Predicting Responses in Physical and AI Systems
SummaryThis paper introduces Active Neural Matter (ANM), a finite-window framework for determining what past information a system's state must retain to predict future responses. Forward analysis predicts responses to perturbations from system dynamics, while inverse analysis tests whether a candidate state keeps enough information. The approach is checked with delayed-consensus simulations, flocking examples, and a controlled feedback task.
In briefA Self-Defining Operator That Learns From Past Incidents to Speed Up Fault Resolution
SummaryThis paper presents a self-defining operator (SDO) for distributed applications that learns from each resolved incident. It turns past experience into playbooks with reusable diagnostic instructions and into detectors, scripts that read telemetry and select relevant playbooks. The detectors run continuously, and LLMs are invoked only when one fires, so token cost scales with incident count. Across 11 open-source microservice applications, SDO reduced resolution time by 52% on recurring faults.
In briefInfraWorld: A Motion-Aware World Model for Roadside 4D Occupancy Forecasting
SummaryThis paper proposes InfraWorld, a model that predicts future semantic occupancy of roadside scenes in a fixed infrastructure coordinate system. It keeps a dense bird's-eye-view state for the whole scene and persistent motion queries for moving regions, updated through dynamic query initiation, long-short temporal fusion, and motion-aware query injection. Experiments on the InfraOcc and Occ3D-nuScenes datasets report state-of-the-art results, which the authors say also hold under ego-centric settings.
In briefDo Multimodal Benchmarks Really Measure Cross-Modal Reasoning? A Per-Item Diagnostic Method
SummaryThis paper argues that some multimodal benchmark items can be answered using only one modality or answer-option patterns, so they do not fully reflect models' cross-modal reasoning. The authors propose SRCoD, a framework that combines an item's multimodal information structure with model response behavior to diagnose how much each item depends on different modality evidence. Experiments under two validation protocols show that SRCoD outperforms existing benchmark diagnosis methods.
In briefGG-Nav: a goal-grounded reasoning chain for navigation models
SummaryThe paper finds that current navigation foundation models perform poorly on goal-oriented navigation, mainly because they under-use streaming observations and localize targets imprecisely. It proposes GG-Nav, which uses supervised fine-tuning to teach a structured chain-of-thought that gathers goal-related evidence from intermediate frames and then grounds the target to produce actions or clarification questions. It reports success-rate gains of 3.1%, 16.6%, and 5.3% on HM3D ObjectNav, VL-LN, and CoIN, and shows zero-shot generalization on a Unitree Go2 robot in unseen real environments.
In briefFDE-Bench: A Benchmark for End-to-End Delivery from Vague Business Requests
SummaryFDE-Bench is a benchmark built from real customer projects to test whether AI agents can deliver solutions when business requirements are incomplete. Its 49 cases cover combinatorial optimization and machine learning, and agents must clarify missing requirements with a simulated customer and iteratively improve their artifacts. Across 18 agent configurations, the best mean delivery score is 0.5510 against a forward-deployed engineer reference of 1.0, and the highest delivery success rate is 6.12%. The results indicate that reliable end-to-end delivery at this level remains beyond the evaluated agents.
In briefA generative latent model that learns beam dynamics while enforcing physical field constraints
SummaryParticle-in-cell simulations model particle accelerator beam dynamics but are too costly for many repeated analyses. This paper presents GLOBE, a physics-constrained generative latent operator that models 3D beam-field evolution across about 10^7 spatial cells. On two high-resolution 3D accelerator benchmarks, it reduces relative L2 error and spectral error compared with the strongest baseline, DPOT, while satisfying field constraints. The authors report that rollouts that once took hours of simulation now take about a minute.
In briefLongHisDoc: A Benchmark for Understanding Long Chinese Historical Documents
SummaryThis paper introduces LongHisDoc, the first benchmark for understanding long, full-book Chinese historical documents. It includes 1,012 expert-annotated question-answer pairs drawn from 101 PDF documents, covering single-page, cross-page, multi-hop, and unanswerable questions. The best large vision-language model and large language model reach only 62.9% and 69.9% accuracy, and the authors also evaluate various retrievers and end-to-end retrieval-augmented generation performance.
In briefDeciding which robots and object detections to trust in teams of robots, using only their observations of each other
SummaryThis paper presents NeST-Bayes, a framework that jointly estimates trust in robots and in their object hypotheses using only consistency among the robots' observations. It needs no pre-assigned trust, reputation data, or identity side-channels. Instantaneous evidence from ego-centric relational graphs is accumulated over time with Bayesian updates of Beta-distributed trust beliefs. Reported evaluations show high object precision and robot recall across benchmarks, including under optimized attacks and in an occlusion-rich Webots warehouse.
In briefEnergy Ring: Stabilizing feature energy for more reliable test-time class discovery
SummaryThis paper studies Test-Time Discovery, where a deployed model must recognize known classes and detect new ones from an unlabeled data stream. The authors find that the energy gap between known and unknown classes is unstable and tends to collapse. They propose Energy Ring, which anchors each class type on a separate energy ring using a lightweight adapter, a perturbation consistency module, and a stability-based correction step. Experiments reportedly show significant gains over existing TTD methods.
In briefScienceBuddy-PV: A Verifier-Governed Protocol for Measuring Support and Abstention in Vascular Intervention AI
SummaryThe paper adapts the ScienceBuddy recursive design to panvascular intervention research, using an evidence ledger, device-vessel ontology, abstention contract, and failure bank. It audits one fixed component on public ARCADE syntax and stenosis tasks, reporting Dice scores and how often a support gate accepts outputs versus abstaining. The authors state that acceptance does not establish correctness or clinical safety. They also state the work does not demonstrate recursive improvement, patient-level validity, or clinical device recommendations.
In briefIdentification and Estimation of Nonsymmetric Determinantal Point Processes
SummaryThis paper studies nonsymmetric determinantal point processes, which can model positive dependence among elements of random sets, unlike the usual symmetric specification. It proves equivalence of two sufficient conditions for a nonsymmetric kernel to define a valid process, develops identification theory, and proposes cardinality-truncated maximum likelihood and averaged stochastic gradient descent estimators. Under suitable conditions, both estimators are shown to be consistent and asymptotically normal, with simulations and two cyber applications illustrating the methods.
In briefQuery-adaptive neuron steering to improve LLM performance at inference time
SummaryThis paper asks whether internal steering of large language models can adapt to each query's activations. The authors propose Q-NeuS, which builds a compact subject-guided candidate neuron space and uses a lightweight Query-Adaptive Gate to produce neuron-wise modulation coefficients, trained with Neu-GRPO. The base model stays frozen, and no task or subject labels are needed at inference. Experiments across multiple LLMs and benchmarks report consistent performance gains, broad applicability across model scales, and cross-benchmark transferability.
In briefA new explanation for how diffusion transformer backbone layers behave
SummaryThis paper studies the depth profile of intermediate activations in diffusion transformer (DiT) backbones, which rises and then falls in an inverted-U shape across layers. The authors argue that the common explanation, that signal granularity shifts from coarse to fine, is inaccurate, and that the prediction target (such as ε, u, or x0) is the central goal, with semantic signal as a by-product. They also show that the choice of prediction target in training affects the curve's shape, and use this to predict near-optimal REPA placement.
In briefSEPM: Self-Evolving Procedural Memory for Multi-Agent Teamwork Based on Task Graphs
SummaryThe paper proposes SEPM, a framework that uses task dependency graphs to coordinate multiple AI agents and to evolve shared procedural memory. A shared blackboard records execution evidence and helps agents detect and correct divergences in goals, plans, or world state. After execution, verified actions are distilled into reusable procedures, which are created, refined, specialized, or retired based on success and cost. The authors report the highest task scores across six benchmark settings on ALFWorld, WebArena, and OfficeBench under AutoGen and DyLAN.
In briefCompositional Concept Erasing: Removing Object and Style Combinations from Diffusion Models via Domain Manifold Unlearning
SummaryDiffusion models can generate illicit content and infringe copyrights, prompting research into machine unlearning to remove harmful concepts or protected styles. The authors note that prior methods handle either objects or styles alone and struggle with object-style combinations, and that repeated unlearning requests cause catastrophic forgetting. They propose Domain Manifold Unlearning (DMU), which uses an EMA-updated feature proxy from intermediate diffusion features to represent the retained target-object direction, enabling selective removal of domain-specific components. Experiments reported in the paper show DMU substantially reduces catastrophic forgetting and outperforms state-of-the-art batch and sequential unlearning methods.
In briefTartanMatch: A Single Model for Dense Matching Across Sensor Types
SummaryRobots use many kinds of sensors, such as RGB, event, thermal, depth, and LiDAR, but most matching methods handle only a few sensor pairs. TartanMatch is one network that performs dense correspondence across all 25 ordered pairs of these five modalities. It adapts an RGB-pretrained model using lightweight modality adaptation and a mix of real and synthetic training data. The authors report reduced endpoint error on zero-shot real-world benchmarks compared with the strongest baselines, by 61.9% for cross-modal and 49.6% for same-modal matching.
In briefSpatiotemporal recognition of abdominal wall layers during optical trocar entry
SummaryThis retrospective study analyzed 71 recordings to test whether spatial context and temporal modeling improve recognition of abdominal wall tissues during optical trocar entry. A pipeline combining appearance, spatial regularization, and temporal constraints outperformed an emissions-only baseline in macro-F1 and onset timing. Recognition remained limited for preperitoneal fat and bowel, and external validation is needed before clinical use.
Shreeram Athreya, Andrew Melehy, Carter Prentice et al.
In briefMontageFlow: Physically valid, high-fidelity EEG signal generation
SummaryThe paper argues that existing EEG generative models ignore a physical constraint: each channel records the voltage difference between two scalp electrodes, so real recordings obey exact linear relations that current generators violate. The authors propose MontageFlow, a flow-matching model that restricts its output to signals satisfying these identities and conditions on a physiological state vector encoded from the whole EEG window. Across three large-scale benchmarks, the model improves TS-FID by 1.6 to 3.6 times over the closest rival while ensuring validity.
Sudarshan Regmi, Nabaraj Subedi, Arvind Pillai et al.
In briefRethinking Activation Functions: Reducing Bias Can Improve Learning
SummaryCommon activation functions such as ReLU, GELU, Swish, and Sigmoid mostly output positive values, giving a nonzero post-activation mean the authors call activation bias. In controlled models, activation bias speeds early loss reduction but comes with overshooting, non-smooth trajectories, and a steeper loss landscape. Reducing it slows initial learning yet yields a lower final loss, a pattern also seen in larger vision and language models.
In briefDSR-Net: Dynamic Skeletal Relation Learning for 3D Human Pose Estimation
SummaryThis paper proposes DSR-Net, a framework for estimating 3D human poses by learning how body joints relate to each other. Its Dynamic Skeletal Relation module learns adaptive joint dependencies beyond fixed skeletal connections, and a lightweight recalibration module stabilizes structure-aware features. On Human3.6M and MPI-INF-3DHP, it reports MPJPEs of 37.9 mm and 14.9 mm, improving over strong baselines.
In briefDEFT: A fine-tuning method that separates content from subject style in EEG models
SummaryEEG foundation models learn brain-signal representations from large multi-subject pretraining, but their embeddings often keep much subject-specific variation that burdens downstream classifiers, especially across subjects. The authors propose DEFT, a plug-and-play adapter that splits pretrained tokens into a task-oriented content branch and a subject-consistent style branch, predicting from the content branch. Experiments on multiple EEG datasets and four foundation model backbones show DEFT improves over matched fine-tuning baselines.
In briefReels-RSI: Program-Level Self-Improvement for Code-Rendered Explainer Videos
SummaryThe paper treats explainer videos as programs (scripts, storyboards and per-frame animation code) rendered deterministically into frames, rather than as pixel video generated directly. Reels-RSI, an open-source reel director, self-improves at the program level by logging revisions, pairing check findings with fixing diffs to derive lessons, and compiling statically checkable lessons into unit-tested rules. Across 111 reels, the authors report that a learned rule caught latent defects in earlier frames, while lessons shortened revision loops but did not compound or change judge preferences.
WhyTreats explainer videos as auditable programs and compiles learned fixes into unit-tested rules; 111-reel experiments with judge-bias analysis give solid evidence, though impact is confined to a niche
In briefA mass-conserving, interpretable model for simulating rainfall, evapotranspiration and runoff
SummaryThis paper presents GCHM, a gated hydrology model built on a mass-conserving perceptron that tracks five storage states such as snow, soil moisture and groundwater. Constrained gates learn how precipitation is partitioned and how water moves through evapotranspiration, infiltration, runoff and baseflow. The model was tested on 50 CAMELS catchments across several climate regimes, including arid ones, and compared against ablated variants and recurrent neural networks. The authors report competitive streamflow prediction with interpretable states and fluxes.
In briefMulti-modal Fusion for Long-term Microsurgical Anastomosis Trajectory Prediction
SummaryThis paper addresses microsurgical anastomosis by proposing a cross-modal feature fusion framework that predicts long-term 2D and 3D surgical instrument trajectories. The framework combines semantic recognition of surgical objects with consecutive motion estimation, and uses a coarse-to-fine progressive fusion strategy with an LSTM mechanism to capture temporal dependencies. The authors trained and evaluated it on their robotic platform and a dataset of phantom and ex-vivo scenes, reporting that it consistently outperforms existing methods and is more robust across surgical scenes.
