In briefAutonomous vehicle routing that accounts for effects on human drivers
SummaryThis paper studies how autonomous vehicles (AVs) choosing routes selfishly can slow down human-driven vehicles in mixed traffic. The authors add a marginal-cost term to the AV reward, measuring each AV's impact by comparing other road users' travel times with a simulation where that AV is removed, and train routing policies with multi-agent reinforcement learning. Experiments on the Ingolstadt and Provins road networks show that including this social component reduces the travel-time increase for human drivers and shortens AV travel times, improving overall system performance.
Anastasia Psarou, Mikolaj Rams, Blazej Torbus et al.
In briefRollout-Time Intervention: Pushing Agent RL Past Its Policy Limits
SummaryThis paper proposes Rollout-Time Intervention (RTI), which uses reward-side evaluation rubrics during rollouts to turn erroneous trajectories into successful ones, supplying learning signals for tasks the policy rarely solves alone. The authors find that these off-policy intervention trajectories can still guide policy optimization once those with excessive distributional shift are filtered out. Across five benchmarks, RTI consistently outperforms RL baselines, bringing the post-trained policy close to substantially stronger contemporary LLMs and beyond its original capability boundary.
In briefLocal diagnostics of learned event representations for anomaly detection at the HL-LHC
SummaryThe paper studies anomaly detection for collision events at the High-Luminosity LHC using low-dimensional representations learned by neural networks. Supervised and self-supervised embeddings are evaluated on COLLIDE-2V, a 750-million-event simulated dataset, with autoencoder reconstruction error in the latent space used as the anomaly score. The authors also introduce a local diagnostic pipeline that localizes flagged events and ranks high-level observables by their local discriminating power.
Donatella Genovese, Philip Ploner, Simone Scardapane et al.
In briefPPG-Bench: A Unified Benchmark for Testing Photoplethysmography Foundation Models
SummaryThis paper presents PPG-Bench, an open-source benchmark that evaluates five public photoplethysmography (PPG) foundation models across nine open-access datasets and 26 tasks. The models are compared against a random/constant baseline, a deterministic pulse-rate-variability feature set, and the generic time-series model MOMENT. The authors find that 35.3% of the 68 task-regime combinations show no meaningful advantage over the random/constant baseline, and in 83.8% no PPG-specific model outperforms both PRV and MOMENT. They conclude that current PPG foundation models cannot yet be claimed to provide generalised, transferable representations, and that future evaluations should include simple non-learned controls.
Arvasu Kulkarni, Mahir Jain, Aditya Ray Mishra et al.
In briefCNN-Transformer with Sigmoid Labels for Sub-Degree Direction-of-Arrival Estimation
SummaryThis paper tackles the grid quantization error in deep-learning-based direction-of-arrival (DOA) estimation, where accuracy is limited by grid step and label peak width. It proposes a sigmoid label that encodes source DOAs as sums of angle-shifted monotonic sigmoid functions. A CNN-Transformer trained with this label, using a 0.25-degree grid, reportedly achieves sub-degree accuracy and 1-degree resolution. Compared with Gaussian, 0-1 hard, and binary cross-entropy labels under the same backbone, the sigmoid label gives the highest accuracy and best resolution.
In briefDoes Financial Speech Need Continual Learning Regularization? EWC and mini-KFAC on Earnings Call ASR
SummaryThis paper studies continual learning for earnings call speech recognition by adapting Whisper-small from a general speech corpus to the Earnings22 benchmark. On held-out financial speech from the same domain, all evaluated methods improved rather than forgot, so regularization was counterproductive in this setting. When regularization is needed, the authors report that layer-wise mini-KFAC EWC achieves lower word error rates and more stable optimization than diagonal EWC.
In briefWhen AI Advisors Agree: Forecast Risk vs. Forecast Reliability
SummaryThis paper examines whether agreement among seven AI investment advisors, built from ten Gemini configurations, signals a reliable forecast. Across 232 assets and first-quarter outcomes, the council's error rose with disagreement, but a no-change baseline rose almost identically, suggesting agreement mostly reflects asset risk. The authors also show that directional accuracy depends on the aggregation rule, so agreement, asset risk, and aggregation skill should be evaluated separately.
In briefA Tutorial Guide to Time Series Causal Discovery Across Fields
SummaryThis tutorial paper gives a unified overview of time series causal discovery, which detects causal links between time series, across statistics, dynamical systems, and machine learning. It introduces the key principles behind modern discovery methods and discusses their limitations and assumptions. The paper also covers how these methods are evaluated, benchmarked, and fine-tuned, and notes their relevance to agriculture, Earth science, healthcare, and AI systems. A companion repository links to code and resources.
Kurt Butler, Damian Machlanski, Michael P.J. Camilleri et al.
In briefSADL: A Size-Adaptive Decision Layer for Chronic Stroke Lesion Segmentation
SummaryThis paper tackles segmenting chronic ischemic stroke lesions on T1-weighted MRI, where lesions resemble healthy tissue and vary widely in volume. The authors combine an eleven-model nnU-Net ensemble with SADL, a decision layer that sets per-lesion-size thresholds, prunes components, and rescues empty predictions. On a site-disjoint 91-case holdout, SADL reports Dice 0.7322, lesion-F1 0.7461 and PR-AUC 0.8430, with the lesion-F1 gain significant at p=0.024, largest for small lesions.
Marek Wodzinski, Gniewosz Drwiega, Wojciech Szymanski
In briefWeb$hell: A Unified Interface for Agents Accessing the Web
SummaryThis paper presents Web$hell, an interface that unifies web access for large language model agents under a single access model. WebCompiler converts live webpages into addressable information and action objects, which are stored in WebCache and handled through shared operations. The authors report strong task performance at lower cost across web search, web navigation, and deep research benchmarks.
In briefTextVBench: Benchmark for Text Rendering and Preservation in Video Generation
SummaryTextVBench is a benchmark and evaluation framework for checking whether video generators draw text correctly and keep it consistent across frames. It has two tracks: T2V, which tests text specified in prompts, and Ref2V, which tests preserving text from reference images. The framework scores spatial accuracy per frame and temporal consistency across frames, with attribute-based slices for diagnosis. The paper reports a T2V TextVScore of 68.32 for MiniMax-H3 and a best Ref2V score of 56.51, with performance dropping as more reference text regions appear.
In briefIn-context learning for system identification: a meta-model that predicts system outputs without fitting a model first
SummaryTraditional system identification estimates a model from training data and then predicts outputs, while this paper uses in-context learning, where an offline-trained meta model maps training data and query inputs directly to outputs. The authors propose a unified probabilistic meta-identification architecture combining a Transformer encoder-decoder with mechanisms for long training sequences, and show a pre-trained meta model can generate uncertainty-aware synthetic data for classical methods. Evaluations on standard nonlinear system identification benchmarks and an out-of-distribution scenario report accuracy competitive with state-of-the-art methods and strong generalization.
In briefFaster Linear Bandits via Persistent Candidate Sets
SummaryLinear bandits with many actions require scoring every action each round, which is costly. The paper proposes PC-LinUCB, which keeps k active actions and scores only these plus q challengers sampled from the rest. Experiments show it substantially speeds up LinUCB with comparable regret, and the candidate mechanism can extend beyond LinUCB.
Bingkui Tong, Daria Korovaitceva, Marina Sheshukova et al.
In briefForecasting desertification in semi-arid Brazil with deep learning
SummaryThe study uses a convolutional LSTM trained on satellite Earth Observation data from 2000 to 2023 to forecast year-to-year changes in a land degradation index in Ceará, Brazil. The model reproduces part of the regional drift but explains little of the variation at individual grid cells. The authors conclude that satellite-based desertification forecasts are reliable for the regional trend but not for individual cells.
Filipe Moura Da Silva, Cl\'aubio Landney Lima Bandeira
In briefWhy simple linear models can still predict well in changing environments
SummaryThe paper argues that linear models can predict well in complex domains even when the true mechanisms are nonlinear, because the world's complexity exceeds what can be reliably learned from finite, partially observed data. Flexible models may capture curvature, thresholds, and interactions that hold in one environment but fail in others. When this instability lies mainly in the nonlinear component, omitting it can lower a linear model's prediction error in new settings. Proposition 1 gives a squared-error case supporting this claim.
In briefRegulaTutor: A Theory-Informed Framework for Long-Term Personalized Learning
SummaryRegulaTutor is a framework for long-term personalized tutoring that links a learner's history to their interpretation and teaching strategies, guided by self-regulated learning research. Its policies keep links to supporting evidence and are updated as new observations arrive. The authors also propose PairRoute, an offline method that separates harness, model, and mixed failures to guide targeted improvements. They report average gains in historical-evidence accuracy, learner diagnosis, and judged teaching quality across evaluated model configurations.
In briefToward the Roof of the World: Progress in Tibetan NLP Research
SummaryThis paper presents a systematic bibliometric survey of Tibetan natural language processing, covering 140 publications from 2004 to 2026 in the ACL Anthology and DBLP. Each publication is annotated by venue, research area, first-author country, and resource release. The authors find rapid growth since 2018, expansion across the full NLP pipeline up to large language models, and that only 29% of the works release data, code, or models. The paper identifies gaps in dialect coverage, evaluation, and visibility, and recommends open science and community participation.
In briefModel-free Bayesian Clustering with Multimodal Posterior Summaries
SummaryThis paper proposes a model-free, noise-aware Bayesian clustering framework that builds a martingale posterior over clusterings from a predictive Bayesian viewpoint, without explicit likelihood and prior specification or costly MCMC sampling. To interpret the high-dimensional posterior, the authors introduce a Wasserstein-type summarisation method that compresses many posterior draws into a few weighted representatives using a novel noise-aware loss. They also marginalise uncertainty in the density level and establish theoretical properties of the loss and guarantees for the approximation. Results show the framework captures clustering and noise uncertainty.
In briefProgramGym: Scaling Training Environments for Long-Horizon Program Reproduction Agents
SummaryThis paper introduces ProgramGym, a framework with Program-Env for building verifiable training environments from repositories and Program-Agent for long-horizon program reproduction. The authors construct 1,648 training environments spanning C, C++, Go, and Rust. Experiments on ProgramBench report that fine-tuning raises Qwen3.8-27B from 34.87% to 42.84%, and Program-Agent reaches 55.71% (base) and 56.25% (SFT).
In briefPlacePulse: Learning Reusable Place Embeddings from Structural Geographic Data
SummaryThis paper introduces PlacePulse, a geospatial foundation model that compresses 572 structural attributes about U.S. places into 256-dimensional embeddings for about 1.06 million H3 resolution-7 cells. The authors also propose a two-axis benchmark that separately measures intrinsic representation quality and downstream predictive utility. Across 74 held-out external targets, PlacePulse outperforms a dimensionality-matched feature-selection baseline on unrelated targets and is competitive with Google's Population Dynamics Foundation Model, with strengths varying by domain.
Luc\'ia Garc\'ia-Duarte, Miguel \'Alvarez, Diego Vicente et al.
In briefRadiomics-based thyroid nodule classifier that flags uncertain cases for clinician review
SummaryThis paper proposes a radiomics framework that extracts features from thyroid ultrasound images with PyRadiomics and classifies nodules using a Generalized Additive Model. Label-conditional cross-conformal prediction is added to quantify uncertainty, so ambiguous cases are referred to clinicians instead of receiving a binary label. On the public TN5000 dataset, the model reached an AUC of 0.89 on the validation set. At the chosen significance level, 63% of test nodules received a certain prediction, and 37% were referred to the clinician.
Laura Testa, Pierpaolo Brutti, Umberto Ferraro Petrillo
In briefMoPrO-SD: Modular Prompt Optimization for Detecting Stance in Text
SummaryThis paper presents MoPrO-SD, a framework that improves prompts for stance detection, the task of identifying whether text supports or opposes a position, which is hard for language models because of sarcasm and implicit opinions. It splits prompts into semantic modules and refines them step by step using feedback from an LLM, testing three refinement strategies. The authors also introduce VaccineStance, a human-annotated vaccine stance benchmark, and report that MoPrO-SD improves over a human-written baseline prompt in many settings. Code and dataset are available on GitHub.
Meysam Sabbaghan, Nikesh Gyawali, Arman Zareian Jahromi et al.
In briefBLOKS: Scaling Kernel Solvers via Sketched Block Preconditioning
SummaryThis paper presents BLOKS, a conjugate gradient-based solver for Nyström-sketched kernel ridge regression. It uses a new preconditioner that combines a block Jacobi approximation with sketching, and its memory cost does not depend on dataset size or the number of parameters. The authors report a GPU-accelerated implementation and show that BLOKS matches or outperforms state-of-the-art methods on standard large-scale benchmark datasets while keeping favorable time and memory complexity.
Giordano Vitale, Giacomo Meanti, Marco Rando et al.
In briefFrom Structure to Validity: Mechanism-Guided Generation and Refinement of LLM Rubrics
SummaryThis paper systematically analyzes how rubric design works across criterion organization, criterion operationalization, and score aggregation. It finds that structural organization enables more comprehensive criteria, semantic specification alone does not ensure effective criterion-level judgments, and sophisticated score aggregation gives limited extra gains. Based on these findings, the authors propose ARTS-Hybrid, which combines adaptive tree construction with semantic-first, response-calibrated refinement. Experiments show it improves end-to-end evaluation performance and intrinsic rubric quality, and that its rubrics also provide useful signals for model optimization.
In briefUsing expert selection biases to correct data imbalance in MoE models
SummaryThe paper addresses how data imbalance across categories can leave minority categories concentrated in few experts or spread inefficiently in Mixture-of-Experts models. The authors propose applying category-conditioned expert-selection biases to route tokens from each category toward specialized expert subsets while still allowing sharing. They report performance gains across modalities, low-resource languages, and code in models up to 10.6B parameters.
In briefUsing an information bottleneck to explain when classifier-free guidance should activate
SummaryThis paper studies when classifier-free guidance (CFG) should be activated in diffusion models, using an information-theoretic view that models the diffusion trajectory as a time-dependent information bottleneck. The authors find that conditional information becomes significantly stronger after a critical stage characterized by a Fisher-information balance condition. Based on this, they propose a training-free dynamic CFG method that activates guidance when the estimated guidance gain crosses a set threshold, and note that interval-based guidance is a limiting special case.
In briefPhysics-aware self-supervised pretraining for Raman spectra across multiple domains
SummaryThis paper proposes a multi-domain self-supervised framework based on a masked autoencoder to learn transferable representations from heterogeneous Raman spectra in biology, chemistry, forensics, and materials. The pretraining combines masked intensity reconstruction, a sparse Voigt-profile peak decoder, and a denoising objective. Experiments show the largest benefit on challenging tasks with limited labeled data, and attention-based probing extracts more useful information than linear probing.
Charalampos Davalas, Georgios Levantinos, Anastasios Dimitriou et al.
In briefAgentSWE: Can coding agents build the agent users actually need?
SummaryThis paper introduces AgentSWE, a benchmark testing whether coding agents can build agents that truly meet a user's requirements. It covers 25 tasks across three stages: creating an agent from a natural-language requirement, editing production codebases, and optimizing toward a measurable target. Acceptance uses held-out cases and requirement-derived rubrics, and the evaluated coding agents fall well short of delivering what was requested. The authors also identify recurring failure modes, including overfitting to visible cases, hollow runtimes, and unsupported claims of completion.
In briefSWE-Critix: Training General Verifiers to Judge Coding-Agent Task Trajectories
SummaryThis paper introduces SWE-Critix, a post-training pipeline for building generalist verifiers that assess whether coding agents completed a task based on their execution trajectories. The verifier predicts a completion label and produces a natural-language explanation of its judgment. Combining supervised fine-tuning with reinforcement learning, classification accuracy on a held-out test set rose from 57% for the base model to 66% after fine-tuning and 78% after reinforcement learning.
In briefAlphaNIST: Robot Assembly Skills Learned from Simple Binary Success Rewards
SummaryThis paper tests whether a robot can learn to assemble parts on the NIST Assembly Task Board #1 using only a sparse binary task reward and task diversity, without demonstrations or progress rewards. In simulation, a single policy trained from scratch across all 20 assembly tasks reaches 98.2% overall training success. The learned policies also show planning-like behaviors, such as choosing grasps for later assembly and using surrounding geometry to reorient parts.
Octi Zhang, Iretiayo Akinola, Rosario Scalise et al.
In briefEmoGavel: Reinforcement learning with verifiable rewards for multimodal emotion preference learning
SummaryThis paper presents EmoGavel, a framework for multimodal emotion preference learning that determines which open-ended emotion description better matches human judgment. It first applies grounded supervised fine-tuning to jointly learn preference prediction, error localization, repair, and missing-evidence recovery. It then optimizes these verifiable outcomes with grounding-aware group relative policy optimization. The authors report improved preference prediction and controlled-error discrimination on EmoPrefer and controlled-error evaluation sets.
In briefThe AI's Poker Face: Persuasion and Sycophancy in Voice Dialogues
SummaryA randomized controlled experiment (N=616) examined multi-turn spoken conversations with an LLM under non-directive and two opposing directive conditions on feminism. Both directive conditions shifted opinions toward their assigned stances, with larger shifts for pro-feminist positions, and the non-directive condition also shifted opinions pro-feminism. Social sycophancy was near-universal but was not associated with opinion change, instead shaping users' subjective experience.
In briefA hierarchical multi-agent AI system for finding replacement materials grounded in known data
SummaryThe paper presents MARS, a multi-agent AI system that proposes replacement materials while respecting scientific and manufacturing constraints. It is applied to substituting PFAS, working through requirement analysis, evidence-based candidate generation, and manufacturability assessment with process design. In a multi-seed ablation evaluation, MARS performs best across all three subsystems, and the gains persist when the backbone model is held fixed.
In briefHidden data problems in clinical machine learning: temporal, semantic, and population misalignment in MIMIC-IV preprocessing
SummaryThe paper argues that widely reused electronic health record preprocessing pipelines can embed assumptions about when information became available, what clinical episode a label represents, and what missingness means, assumptions rarely audited. Using MIMIC-IV as a case study, the authors quantify violations of temporal, semantic, and population alignment, such as temporal leakage and biased imputation. They offer recommendations for temporally valid, clinically aligned, reproducible preprocessing, with code publicly released.
Carolin Ciss\'ee, Dennis Fast, Raneen Younis et al.
In briefTest-Time Memory Construction: Building LLM memory on demand during reasoning
SummaryLarge language models struggle to answer questions over long interaction histories whose relevant evidence is spread across distant sessions. The paper proposes T²Mem, which extracts only lightweight episode-level metadata offline, keeps links to raw interactions, and lets the LLM build query-specific memory progressively at test time. Experiments on LoCoMo and LongMemEval show it consistently outperforms strong baselines across different backbone LLMs while substantially reducing memory construction overhead.
WhyDeferring memory construction to inference with query-adaptive organization meaningfully improves offline memory systems, consistently beating strong baselines on LoCoMo and LongMemEval across backbon
In briefWhy DiT models learn weaker features without long skip connections
SummaryThis paper studies diffusion features, intermediate activations from diffusion backbones used for visual understanding. The authors show that Diffusion Transformers with long skip connections (LSCs) produce better features when trained from scratch. They argue that diffusion noise, which encodes trajectory information, pollutes the image signal in the main backbone, and LSCs reduce this by routing noise around it.
In briefUnderstanding Mamba-3 Through Oscillatory Dynamics
SummaryMamba-3 adds input-dependent rotations to its recurrent state updates, and this paper interprets them as underdamped second-order dynamics to study the role of oscillation in sequence modeling. Controlled comparisons show underdamped dynamics outperform non-oscillatory damping on state-tracking tasks. Language-model pretraining shows benefits for long-context prediction and length generalization.
In briefVideoAWR: Improving video flow models with reward-weighted correction using decoupled timesteps
SummaryThis paper addresses reinforcement learning for video flow models, where existing regression-based methods struggle with classifier-free guidance. The proposed VideoAWR method separates sampling from the optimization target: an EMA reference policy generates guided rollouts, while the trainable policy is corrected against an unguided anchor. Experiments on Wan2.1-1.3B and 14B show stable optimization and better visual and motion quality, both with and without CFG.
In briefAttention Sinks in Text-to-Image Transformers: Diffusion Versus Autoregressive Models
SummaryThis paper studies "attention sinks," tokens or positions that receive disproportionately high attention, in text-to-image Transformers, an area largely unexplored compared with large language models. The authors compare bidirectional diffusion models with causal autoregressive models and find that sinks appear in both but differ in location: the autoregressive model focuses on the first and global tokens, while diffusion models show sinks near modality boundaries, within text regions, and at spatial corners. Perturbation experiments and the introduction of register tokens are used to assess how these sinks functionally affect generation.
In briefAgentTune: A unified Python library for training and optimizing tool-using LLM agents
SummaryThis paper presents AgentTune, a source-available Python library for building, training, and optimizing agents that use tools. Its two-tier EventLog keeps agent trajectories and supports consistent use in training, evaluation, and auditing. The library includes tool-calling reinforcement learning, agentic distillation, retrieval-augmented generation, and evaluation utilities, demonstrated on tasks such as legal contract search.
R E Zera Marveen Lyngkhoi, Abhivansh Gupta, Vidushee Vats et al.
In briefPDE-PFN: A Prior-Data Fitted Neural Solver for Partial Differential Equations
SummaryThis paper explores predicting partial differential equation (PDE) solution fields in context using approximate pre-training data, inspired by large language models. It proposes PDE-PFN, a Transformer-based prior-data fitted network pre-trained on noisy solutions generated by physics-informed neural networks. Experiments on two-dimensional PDEs report generalization across different equations, robustness to noisy priors, and inference without gradient updates. The authors state that predictive accuracy is competitive with or better than the evaluated baselines.
In briefMasked diffusion language agent for tool-using chemical reasoning
SummaryThis paper introduces ChemDiffAgent, the first masked diffusion language agent for chemical reasoning that integrates external tools. It treats interleaved reasoning and tool-use trajectories as iterative denoising, so tool decisions are made with full-trajectory context. The authors use a two-stage post-training method, combining agentic supervised fine-tuning and variance-reduced preference optimization, and build a benchmark covering six chemistry tasks. Under matched budgets, the approach reportedly outperforms autoregressive agents by about 20%, and its 8B model is said to match or sometimes exceed much larger frontier models.
WhyApplies masked diffusion language models to tool-integrated chemical reasoning, offering a new full-trajectory denoising paradigm; experimental breadth and comparison details need further verification
In briefDuoRate: Two-Rate Per-Expert Precision Allocation for Efficient Mixture-of-Experts Serving
SummaryThis paper observes that serving a Mixture-of-Experts model pays for expert precision in two ways: resident storage covers all experts, while each decode step reads only the selected ones. The authors propose DuoRate, which stores one uniform-precision weight stream offline and, online, profiles a workload's prefill to allocate per-expert precision for decoding via a Lagrangian rate-distortion solve. On Qwen3-30B-A3B, DuoRate at 2.3 bits per weight outperforms all 2-bit-class baselines, and at 2.9 bits it matches 4-bit formats that read 1.4 to 1.6 bits more.
In briefMoving tokenization from the CPU host to the GPU for agentic LLM serving
SummaryAgentic workloads resend long, nearly unchanged contexts, and prefix caching cannot avoid the full tokenization each request still requires, which dominates time-to-first-token. The paper presents GraphTok, a GPU tokenizer that runs the entire encoding pipeline as one replayable CUDA graph and reproduces reference token IDs bit-for-bit. It integrates with unmodified vLLM and NVIDIA Dynamo, and the authors report large speedups and higher sustainable request load over CPU tokenizers.
Mohammad Siavashi, Pascal Schwinte, Ali Banaei et al.
In briefCoVA: Composing VAE features to improve diffusion model training without extra encoders
SummaryCoVA is a lightweight framework that composes features from a generative model's own VAE into stage-wise guidance for flow-matching training. It aligns intermediate SiT features with these composed representations early in training, then removes that alignment and uses detached composer posteriors for soft label conditioning. On ImageNet 256×256, it reduces SiT-XL/2 FID from 17.2 to 7.4 in 400K steps while adding only 8.5M training-only parameters and no inference overhead.
In briefHost-Aware Web Crawling for Better LLM Pretraining Data
SummaryThis paper studies how web crawlers should spread their budget across and within hosts to collect more high-quality data for LLM pretraining. It finds that about half of the variation in quality classifier scores lies between hosts. Sampling a few pages per host and then crawling deeper into higher-scoring hosts collects several times more high-quality pages than uniform or centrality-based allocation.
In briefSBNO: Schrödinger Bridge Neural Operator for Forward and Inverse PDE Problems with Missing Data
SummaryThis paper introduces SBNO, an equation-free framework that learns to solve forward and inverse partial differential equation (PDE) problems from paired data when observations are sparse. It uses diffusion Schrödinger bridge matching with a mask-aware coupling to learn drifts for both problem directions. At inference, it produces deterministic reconstructions by integrating a probability-flow ODE, without tuning guidance weights. On five PDE benchmarks with 5% observations, SBNO reports accuracy comparable or superior to baselines, with at least 9.7 times faster inference and at least 4.3 times lower GPU memory use than guided-diffusion methods.
In briefThinkDrag: Drag-based image editing that reasons about content meaning
SummaryThis paper presents ThinkDrag, a multimodal framework for drag-based image editing that interprets drags as semantic object and scene changes rather than only geometric moves. It can optionally produce a reasoning trace explaining the intended edit, and uses Drag Endpoint Markers to ground each drag in image space. The authors also build a supervised dataset of semantic drag transformations and introduce the DragBench++ benchmark. Experiments report state-of-the-art generation quality and plausibility while keeping competitive point-following precision.
Armando Fortes, Tianyi Wei, Chindanai Trakantannarong et al.
In briefJEPA-kMMVAE: Predicting Wireless Channel State for Time-Varying Links
SummaryThis paper proposes JEPA-kMMVAE, a neural network method that predicts future channel state information (CSI) for time-varying multi-antenna wireless channels from noisy past observations. It combines a joint-embedding predictive architecture with a k-memory Markov variational autoencoder to capture temporal structure in latent representations. The framework supports both direct prediction and residual corrections based on the latest latent state and noisy CSI. Simulations show the residual configuration beats the evaluated baselines at moderate and high input CSI SNR, while sequence baselines are more accurate under stronger noise.
Efstathios Chatziloizos, Konstantinos Vandikas, Aneta Vulgarakis Feljan et al.
In briefNovaSplat: Model-free 6D object pose estimation using 3D Gaussian Splatting
SummaryThis paper presents NovaSplat, a method for estimating the 6D pose of objects without needing CAD models. It trains a 3D Gaussian Splatting model from a short smartphone video and uses it for detection, coarse pose estimation, and pose refinement. NovaSplat outperforms model-based methods on a household-object benchmark and outperforms other model-free methods on a textureless industrial benchmark. A faster variant, NovaSplat-Lite, runs 3.1 times faster with negligible accuracy loss.
WhyUses 3DGS as both template and differentiable refiner, beating model-based methods on household and industrial benchmarks with substantial gains.
Georgios Floros, Evangelos Sartinas, Orestis Vangelis et al.
In briefSustainable cloud-fog-edge systems for smart buildings using memory-limited adaptive learning and dynamic orchestration
SummaryThis paper proposes an edge computing architecture and algorithms that let smart buildings run autonomous, resilient, privacy-preserving learning systems. It covers reliable data acquisition with low-cost Arduino hardware and secure storage, a lightweight orchestration framework for dynamically managing service level agreements, and a batch-based adaptation of the Generalized Hebbian Algorithm for principal component analysis on low-power microcontrollers. Experiments on Tour Perret sensor data show the method converges to results visually equivalent to full-memory processing and reliably identifies two operational states.
In briefStabilizing Half-Precision PPO by Protecting Key Numerical Quantities
SummaryThis paper studies how to run proximal policy optimization (PPO), a reinforcement learning method, in half precision (FP16) on commodity GPUs while keeping training stable and efficient. The authors propose PPO-FP16, which adds three modules that protect global scalar statistics, small differences between current and behavior policies, and optimizer state with parameter writeback. Across six continuous-control tasks and five seeds, it achieves the lowest mean return-AUC rank among four methods and speeds up PPO updates by 1.16 to 2.03 times in most configurations. It also reduces optimizer-state memory by about 72 to 75 percent relative to full FP32 Adam.
In briefMORAD: RNA inverse design via reinforcement learning with multiple rewards
SummaryMORAD is an online reinforcement learning framework that trains one shared policy for RNA diffusion models conditioned on many target backbone structures. It jointly optimizes six measurements covering tertiary geometry, base pairing, ensemble behavior, and sequence composition, combined through a weighted geometric mean. Applied to the pretrained RIDE model, the paper reports gains in GDT-TS and pairing MCC and a reduced normalized ensemble defect. The authors state it outperforms RiboDiffusion, gRNAde, and RDesign on the test set.
In briefFutureLens: Tracking Topic Trends in Papers and Their Future-Work Sections
SummaryThis paper studies how topics in 5,655 ACL papers from 2012 to 2022, including future-work text for 4,005 of them, change over time. It compares NMF and LDA topic models on prevalence, stability, and agreement between input text and references. NMF shows higher coherence and assignment stability, while LDA shows greater keyword diversity. The authors frame the work as an empirical measurement study, not a new algorithm or forecasting system.
In briefLearning DAG structure when each node has its own heavy-tailed noise
SummaryThis paper studies learning directed acyclic graph (DAG) structures from data under linear structural equation models, where each node's disturbance has heavy tails and its own scale. The authors propose a node-varying Student-t version of DAGMA that jointly estimates the graph, per-node disturbance scales, and the degrees-of-freedom parameter, and they establish identifiability of the graph and disturbance parameters. Experiments on synthetic data show better structure recovery under heterogeneous heavy-tailed disturbances, and on real finance data the method achieves lower test negative log-likelihood than the baselines.
In briefLabel-free rubric edits make LLM judges agree more with each other but not with humans
SummaryThe paper tests a label-free procedure that finds the most contested sub-criteria of an evaluation rubric, based on disagreement among a panel of LLM judges, and has an editor model rewrite them. Evaluated on SummEval, HealthBench, and USR Topical-Chat, the procedure raised inter-judge Krippendorff's α by 0.069 on SummEval over the original rubric. Agreement with expert ratings showed no detectable change.
In briefPAGE: Part-aware guidance for generating gesture portrait videos from speech
SummaryThis paper presents PAGE, a backbone-agnostic optimization framework for diffusion models that generates audio-driven portrait videos from a single reference image and speech. It emphasizes key regions such as lips, eyes, and hands during training and corrects audio-visual inconsistencies at inference using an audio-conditioned classifier, while refining facial and body motion separately. The authors also introduce the Chinese Announcer Speech (CAS) dataset, and experiments on PATS and CAS report state-of-the-art results in perceptual quality, gesture diversity, and audio-visual alignment.
In briefSonic-AMG: generating human motion from spatial sound as it streams in
SummaryThis paper introduces Sonic-AMG, a framework that generates human motion reactions to spatial audio in a streaming fashion, without needing the full sound sequence in advance. It encodes sound content and body-relative source position as separate tokens, which condition a causal transformer with a flow head. Rehearsal training and spatial feedback are proposed to reduce the mismatch between training and streaming inference. The authors report state-of-the-art FID and R-precision on the SAM dataset and support for real-time online generation.
In briefDecoding register tokens keeps whole-scene details in image reconstruction
SummaryThis paper examines whether non-spatial register tokens in pretrained vision transformers, which encode global properties like illumination, blur, and style, improve image decoding when retained alongside patch tokens. The authors use representation autoencoders with frozen encoders and a decoder that jointly attends to registers and all patches. On ImageNet-1K, reconstruction FID improves by 9.5% with DINOv2 and 5.2% with DINOv3 relative to patch-only baselines, and generation FID improves by 9.8% for DINOv2 models.
In briefMulti-Source, Multi-Dimensional Information Fusion for Resource-Limited Edge Networks
SummaryThis paper proposes a framework that fuses semantic information from multiple sources and dimensions in edge networks under limited communication and computation resources. It combines the Hyper-Slot fusion method, a contribution-based incentive mechanism, and a bi-level Stackelberg-Nash game for coordinating terminal uploads. On a tri-modal extension of COCO2014 and Flickr30K image-text retrieval, Hyper-Slot improves mAP by 3.66 points and RSum by 15.9 points over the strongest baselines while reducing parameters by over 80%.
In briefPRIME: A pre-retrieval indexing method that improves how documents are searched for AI systems
SummaryThis paper proposes PRIME, a framework that optimizes the index side of retrieval before search, using LLM-based semantic restructuring and deterministic metadata with hierarchical retrieval. The authors report that deterministic metadata matches structure-aware chunking in retrieval quality while cutting embedding computation by 45.8x and index vectors by 22.4x. Full deployment with semantic restructuring improves Answer-F1 by 13.7%. Experiments on QASPER, BRIGHT, and NFCorpus identify three regimes where index-side optimization helps or hurts, and a strong reranker neutralizes these effects.
In briefLearning a Three-Body Interaction from Motion Data with Lagrangian Neural Networks
SummaryThis paper examines whether Lagrangian Neural Networks (LNNs) can learn a many-body interaction that cannot be written as a sum of pairwise terms, using only motion trajectories. The authors focus on the Axilrod-Teller three-body dispersion interaction, learning only the potential as a function of interparticle distances and angles, with training data concentrated where the three-body signal is strongest. The learned potential reproduces the interaction's radial scaling, angular dependence, and sign reversal, and symbolic regression recovers its closed-form expression.
Sena Kalabalik, Sukru Caglar, Abdullah Umut Hamzaogullari et al.
In briefDo Prompt Optimizers Break Requirements They Are Not Scored On? A Preregistered Study
SummaryThis preregistered study compares three prompt optimizers, GEPA, MIPROv2, and TextGrad, on HotpotQA, SQuAD 2.0, and IFEval. It counts violations of requirements the target metric does not score, such as output format and citations, using deterministic checks. By the preregistered criterion all three regress, most clearly on citation validity for HotpotQA. A post-hoc analysis attributes much of the regression to responses cut off by the 200-token output cap, and among parseable outputs the net regression is near zero.
In briefBoundaryEvo: Data-model co-evolution at the capability boundary for distilling long reasoning chains
SummaryThis paper proposes BoundaryEvo, a two-stage framework for distilling long reasoning trajectories from large teacher models into small language models. It first directly distills a small set of teacher trajectories to build basic reasoning ability, then repeatedly selects questions the student has not mastered but the expert can reliably solve, updating training data as the student improves. The authors report that with only 1,000 training examples, it outperforms static and dynamic data selection baselines on 11 reasoning benchmarks, raising Qwen3.5-4B accuracy from 62.9 to 75.3 on IMO-AnswerBench and from 26.8 to 46.0 on AMO-Bench.
In briefQCLR: A plug-in module using quantization to make contrastive learning for skeleton-based action recognition more robust
SummaryThis paper addresses the vulnerability of contrastive self-supervised learning for skeleton-based human action recognition to adversarial attacks. The authors propose QCLR, a plug-in module that applies dual-level quantization across the encoding pipeline to break gradient propagation and absorb small adversarial perturbations. Experiments on several benchmark datasets report substantially improved robustness against both white-box and black-box attacks while keeping classification accuracy highly competitive.
In briefMHR-Router: Adaptive routing that picks the best vision-language model for each image-question query using retrieved past successes and failures
SummaryVision-language models have complementary strengths, but choosing the best one for a given image-question pair is difficult. The paper proposes MHR-Router, which first fuses frozen multimodal features into a query representation, then retrieves each expert's success and failure memories as local evidence. A shared scorer uses this evidence to adaptively select the optimal expert. Experiments reportedly show effective heterogeneous modality routing and consistent accuracy gains over traditional baselines.
In briefBeyond the Final Embedding: Retrieving Memories from Token-Level Evidence
SummaryThis paper studies how memory-augmented assistants retrieve relevant content from stored conversations and documents. It proposes a sparse token-matching method that needs no additional training: each query token is matched to a few tokens in candidate memories, and the scores are combined with whole-memory similarity for ranking. An optional low-rank query adapter can be trained while the encoder and stored memories stay fixed. The authors report improved retrieval across eight benchmarks and several model families, with gains carrying over to question-answering accuracy on long-term memory benchmarks.
In briefCopRoute: Dependence-Aware Conformal Routing for Language Models
SummaryThis paper proposes CopRoute, a method for routing queries among multiple language models. It orders models by how many new queries each answers correctly beyond those already selected, and decides per query how far down that order to go while meeting a target risk. On RouterBench and RouterEval, it uses the fewest certified calls on most datasets compared with the state-of-the-art risk-controlled policy, reducing calls by up to 79 percent.
In briefAgent4Spec: An LLM-based agent system that optimizes API specifications through automated validation and market analysis
SummaryThis paper presents Agent4Spec, a multi-agent framework with two pipelines: one that detects and self-corrects OpenAPI syntax errors, and another that researches competitors, finds functional gaps against industry leaders, and generates missing API-spec functionality. The authors evaluated it on two API specification datasets covering general-purpose and industrial scenarios, using 9 state-of-the-art LLMs. The top performer reportedly reached 100% syntax error detection, over 95% implementation success for proposed corrections, and over 50% adoption of suggested endpoint improvements.
In briefOWLET: Evidence-based training that helps video AI models learn from visual clues more precisely
SummaryThis paper proposes OWLET, a post-training method for Video Large Language Models. Existing reinforcement learning uses one trajectory-level reward for the whole response, which is too coarse for video reasoning, while on-policy distillation matches output distributions without directly supervising spatiotemporal representations. The authors observe that the same model reasons more accurately when its video view combines global temporal context with question-relevant details, and use this richer view as a training signal. OWLET adds evidence-aware credit redistribution and spatiotemporal state alignment to exploit this visual evidence.
In briefConClick: Element-Wise Contrastive Learning for Better GUI Element Location
SummaryThis paper studies GUI grounding, the task of locating the interface element a user's instruction refers to, and finds that most wrong clicks land on non-target elements rather than the background. The authors build ConClick-Data, with 4,292 target-distractor pairs, and train models with contrastive objectives that separate the target from plausible distractors. ConClick raises Qwen3-VL-8B's accuracy on three benchmarks, and the inference-time ReClick strategy adds a further 2.96 points on average.
WhyElement-wise contrastive learning targets distractor errors, with clear gains across three GUI grounding benchmarks; novelty is moderate.
In briefSelective Forgetting to Repair Poisoned LLM Agent Memory
SummaryThis paper proposes DREAM, a training-free memory repair framework that lets LLM agents selectively forget poisoned memories, inspired by neuroscience findings on memory decay. It works in three stages: perturbation-induced reactivation to expose fragile malicious memories, multi-dimensional activation-based anomaly diagnosis, and dynamic repair that suppresses harmful memories while reinforcing benign ones. The authors report over 95% reduction in backdoor attack success rates while preserving benign utility, plus a 2.13× speedup and 37.7% lower token use compared with A-MemGuard.
In briefCV-OPD: Contrastive vision on-policy distillation to make VLMs rely on visual evidence
SummaryThe paper finds that vision-language models trained with on-policy distillation still answer correctly when related image regions are masked or replaced with noise, suggesting they take a shortcut instead of using visual evidence. The authors propose CV-OPD, which contrasts teacher distributions from positive and negative crops and adds a margin-based contrastive loss with a discrepancy-aware gate to focus the student on visual tokens. Experiments on a range of fine-grained visual benchmarks show that CV-OPD reduces this shortcut behavior and improves the visual reasoning of existing VLMs.
In briefSheafPrune: Graph-based pruning of visual tokens for faster multimodal models
SummaryMultimodal large language models process every visual token, which raises prefill cost even when many tokens are irrelevant to the instruction. The paper proposes SheafPrune, which transfers relevance through a sparse text-visual graph, propagates scores on a visual graph with degree correction, and keeps a top-k token subset before a frozen language model. Across four backbones, ten benchmarks and seven pruning ratios, it achieves the best retention in 22 of 28 settings. On LLaVA-NeXT, keeping 10% of visual tokens gives a 5.15x net prefill speedup and an 84.0% end-to-end FLOPs reduction.
In briefDual-Stream Memory for Vision-Language-Action Robot Policies
SummaryThis paper proposes a dual-stream memory method for vision-language-action robot policies, targeting manipulation tasks that require remembering past execution state and unobservable visual evidence. One stream uses a Gated DeltaNet to compress history into a single temporal token, while the other uses a task-conditioned retriever to keep relevant frames encoded as spatial tokens. The complete model reaches 49.22% mean success across sixteen RoboMME tasks, versus 44.51% for the published reference, and 62.5% versus 37.5% on four real-robot tasks. Failures under cup swaps show that persistent identity tracking remains an open problem.
In briefAdaptive Mahalanobis metric for cross-head attention in Transformers
SummaryThis paper proposes DAMCHA, an attention mechanism that generalizes Multi-Head Attention by learning a full cross-head metric matrix that captures interactions between heads. The metric is made input-dependent so the attention geometry adapts to the data. The authors also add regularization such as stack-wise parameter sharing for efficiency and stable training. They report that DAMCHA can replace standard attention in Transformers and outperforms baselines on benchmark tasks.
In briefBeyond a Single Crop: Context-Aware Embeddings for Re-Identifying Players in Sports Broadcasts
SummaryThis paper proposes ContextReID, a framework that identifies players in broadcast sports videos by considering other players in the same frame and observations from other frames, not just a single player image. It uses a vision-language model that takes sets of player crops as input to produce context-aware embeddings. The model is trained with a combined contextual learning and retrieval objective. Experiments on three multi-object tracking benchmarks report improved Rank-1 and mAP over conventional single-crop methods.
In briefNeural Surrogate Models for Ultrasonic Waves in Mixed Materials
SummaryThis paper presents neural network surrogates that quickly approximate ultrasonic wave propagation, offering a faster alternative to full wave simulations. The authors find that Fourier Neural Operator methods miss key signal features such as reflections, while U-Net methods learn wave fields well but struggle across spatial scales. They propose a new surrogate that integrates FNOs and local-integral resampling into a U-Net, which trains on limited data, is robust to input noise, and works across spatial scales.
Burla Nur Korkmaz, Petro Moilanen, Ari Salmi et al.
In briefCAST: A credit-alignment method for cross-tokenizer on-policy distillation
SummaryThis paper addresses cross-tokenizer on-policy distillation, where decoded-text alignment identifies chunk-level correspondences but does not specify how a chunk's learning signal should be distributed across student tokens. The proposed CAST method redistributes each aligned chunk's advantage using predictive uncertainty and local teacher-signal smoothness, and uses a KL penalty to anchor the student to its initial policy where no teacher correspondence exists. The authors report average improvements over the strongest baseline across six mathematics, science, and code benchmarks.
In briefTabular Foundation Models Are Mostly Shallow: Removing Most Layers Keeps Performance
SummaryThis paper examines whether in-context tabular foundation models (TFMs) can be simplified by removing redundant transformer blocks. The authors introduce a closed-form, label-free affine translator that replaces up to 96% of a contiguous block window. Evaluating eight models on 171 classification datasets, replacing half the blocks costs at most 0.002 median AUROC, and the half-depth model still outperforms full-depth TFMs and tuned GBDTs. Removing depth saves up to 1.5B parameters and speeds up GPU inference by up to 6x.
In briefVision Transformers Learn Word Meaning from Image-Only Training
SummaryThe paper shows that the two largest DINOv3 vision transformers organize words rendered in images by meaning, even though they were pretrained on images alone without captions or labels. Brand names in the same industry and nouns from the same semantic domain cluster together, and country-capital pairs show consistent vector offsets. Contrastive alignment with text or speech encoders further strengthens this word-level semantic structure in one model.
Anders Gj{\o}lbye, Sebastian Ray Mason, Stella Frank et al.
In briefWhat Do Audio-Language Judges of Generated Music Listen For? Four Probes Against Expert Ratings
SummaryOpen audio-language models are used to rate generated music, but their correlation with expert ratings does not show whether they respond to the audio, the text prompt, or how the question is worded. The paper proposes four probes: removing the audio, replacing clips with other models' outputs for the same prompt, swapping clip order in pairwise judging, and rewording the instruction, with human ratings as controls. The probes are applied to five open judges on all 2,748 MusicEval clips, and the last two to SongEval songs.
In briefDeep learning method for real-time detection of labor contractions in dairy cows
SummaryThis paper presents a computer vision approach to automatically recognize labor contractions in cows. The authors collected videos of 100 cow births and extracted about 8,500 contraction and 25,000 non-contraction sequences as a dataset. Using CNNs and Vision Transformers to extract features from video frames, they report detection accuracy above 95%, suggesting potential for real-time monitoring during difficult calvings.
Sebastian Puchala, Dominik Gronkiewicz, Marcin Go{\l}\k{e}biewski et al.
In briefJag: A Method for Fast Direct Box Prediction in Visual Grounding
SummaryThis paper presents Jag, which adapts a compact pretrained multimodal model to predict continuous bounding box coordinates directly. A lightweight regression head reads the last valid input-token state and outputs all four box coordinates in a single forward pass, without generating coordinates sequentially or parsing text. According to the paper, Jag improves accuracy over its base model and achieves competitive accuracy against specialized grounding models, with lower single-request latency and GPU memory use.
In briefState-OPD: State-Grounded On-Policy Distillation for Multi-Turn Agents
SummaryThis paper studies how to transfer reasoning from frontier models to smaller models through on-policy distillation in multi-turn agent tasks. It proposes State-OPD, which grounds teacher supervision in interaction states and allocates sampling budget to states with high response entropy, generating multiple complete reasoning-action responses. On ALFWorld, ScienceWorld and WebShop, the authors report that for Qwen3-4B, macro-average success rises from 26.07% to 31.74%.
In briefPredictability Bias in Joint-Embedding Learning for CT Imaging
SummaryThis paper studies joint-embedding predictive architectures (JEPAs) applied to computed tomography (CT) medical imaging. The authors introduce TAP-JEPA, trained on 113,611 CT volumes, and audit how its prediction affects pathology versus anatomy representations. Prediction degrades pathology representations far more than anatomical ones, suggesting that contextual predictability is not a sufficient proxy for clinical relevance.
WhyProposes predictability bias and audits JEPA prediction on 113,611 CT volumes, showing pathology is suppressed while anatomy is preserved; novel and reasonably well-supported.
Fabian Drexel, Bernhard Kainz, Jakob N. Kather et al.
In briefHow Reliable Are AMR Parser Outputs? Estimating Uncertainty in Meaning Parsing
SummaryThis paper presents the first study of uncertainty estimation for Abstract Meaning Representation (AMR) parsers in a black-box setting. Uncertainty is estimated from pairwise similarities among top-k beam-search candidates and evaluated with the Prediction Rejection Ratio. The authors examine how parsing performance relates to uncertainty, how reliable beam candidate distributions are, and how evaluation metrics affect uncertainty assessment. They also identify parsers offering the best balance between performance and reliability.
In briefClawLoop: Self-improving AI agents through a built-in training loop
SummaryThis paper introduces ClawLoop, a framework that lets LLM-based agents working in sandbox environments improve themselves without an external teacher model. It cycles through task synthesis, training via SFT and/or RL, and diagnosis that feeds back into the next round. Using Qwen3.6-27B as the loop model, it reports consistent gains on SkillsBench, Claw-Eval-Live, and ClawGym-Bench. The refined tasks also improve smaller Qwen3.5-4B and 9B models across all benchmarks.
In briefLearning Schwarzschild Orbits with Lagrangian Neural Networks
SummaryThis paper trains Lagrangian Neural Networks to learn Schwarzschild geodesic motion from trajectory data. It introduces a decoupled position-velocity normalization scheme and an energy-constrained, log-uniform sampling strategy informed by the innermost stable circular orbit. The model reproduces perihelion precession and time dilation and keeps energy and angular momentum conserved within a few percent over multiple orbits.
Sukru Caglar, Sena Kalabalik, Abdullah Umut Hamzaogullari et al.
In briefFPD-Drive: Distilling Future Knowledge for Safer End-to-End Driving
SummaryThis paper proposes FPD-Drive, an end-to-end driving planner that distills privileged future knowledge from a teacher with access to recorded future observations into a visual student policy. The method provides multiple trajectory targets beyond the single recorded human trajectory and corrects ranking errors among nearby safe and unsafe candidates. The student requires no future observations or teacher inference at deployment. The base model scores 94.9 PDMS on NAVSIM v1 navtest and 55.9 EPDMS on NAVSIM v2 navhard, while the scaled FPD-Drive-Pro reaches 95.5 and 58.1. Zero-shot closed-loop evaluation on HUGSIM yields an HD-Score of 45.1 and RC of 55.3.
In briefReSFlow: Rectifying categorical flows on statistical manifolds for one-step generation
SummaryThis paper proposes ReSFlow (Rectified Statistical Flow), a model for generating discrete categorical data. Inspired by Rectified Flow in Euclidean space, it learns trajectories along geodesics of the statistical manifold's Riemannian geometry, so noise can be mapped to discrete data in a single step. The authors report extensive experiments on image, text, and biological generation tasks, showing strong one-step generation quality.
In briefAutomatically detecting and encoding map symbols in historical maps
SummaryThis paper proposes a two-stage framework that first detects cartographic signs in historical maps and then generates interpretable embeddings capturing both semantics and stylistic variation. The authors also introduce CaSiDD, an open benchmark dataset covering multiple sign categories and diverse cartographic styles. Using YOLOv10-X, they show automated sign detection is practical, and their final method, based on an adapted RADIOv4 backbone, encodes both sign form and semantics.
In briefPredicting cancer risk levels by tensor fusion of multi-modal patient data
SummaryThis paper proposes an unsupervised machine learning framework that uses tensor decomposition to fuse genomic and clinical data, extracting latent features that are then clustered to stratify cancer patient risk. The framework is validated on LinkedOmics data from 11,158 patients across 32 cancer types, using tumor purity, pathological stage, and survival outcomes as reference indicators. The authors report that the tSVD+GMM configuration consistently outperformed other configurations on clinical benchmarks.
In briefUAV-based maritime vessel identification and tracking using multi-modal graph learning
SummaryThis paper proposes AMGT, an airborne multi-modal graph learning framework for identifying and tracking vessels in UAV-assisted maritime IoT. It fuses visible and infrared images through a perspective transformation and synchronizes AIS messages in time, then uses a mixed-modality detector and a tracking algorithm for small oriented vessels. A graph-based association network and a motion-aware coordinate unification strategy are introduced to handle temporal dynamics and UAV-induced misalignment. Experiments on the MAP-M2UAV50 dataset reportedly show high accuracy and efficiency.
In briefAdding Functional Interiors to AI-Generated 3D Shapes
SummaryAI 3D generators such as Hunyuan 3D and Meshy create appealing exteriors, but their interiors are often incomplete or implausible, limiting use in simulation and fabrication. The paper proposes an interior-aware part graph, built using VLM-based segmentation and attribute inference, to guide construction of each part's interior as a generalized cylinder. The cylinder parameters are optimized from the external geometry, preserving the original appearance. The method is demonstrated on vessels, containers, and storage furniture, with selected examples fabricated for functional validation.
In briefReference-free point cloud quality scoring that corrects for content complexity
SummaryThe paper proposes a no-reference point cloud quality assessment method that estimates the discrepancy caused by content complexity through successive surface reconstructions and uses it to calibrate the input-to-reconstruction discrepancy. Calibrated geometric and projected-view discrepancies are fused into a final degradation score, using frozen pretrained perceptual representations without PCQA-specific training. Experiments on three public databases show it consistently outperforms most generic baselines and performs comparably to several full-reference metrics.
In briefGolden Anchor for Video Generation: Tuning Text-to-Image Anchors Using Video Feedback
SummaryThis paper examines whether images produced directly by text-to-image models make good anchors for image-to-video generation, finding that common image quality scores correlate weakly or negatively with downstream video quality. The authors propose Golden Anchor Learning (GAL), which scores candidate anchors by the videos they produce under a frozen image-to-video model and fine-tunes the text-to-image generator with DPO using these video-derived preferences. Across four image-to-video models, GAL raises VBench averages by 2.97% to 8.46% and dynamic degree by 5.62% to 38.24%.
WhyProposes GAL, optimizing T2I anchors via downstream video feedback, with consistent VBench and motion gains across four I2V models.
In briefMonarch-MoE: Structured Experts for Continual Multimodal Instruction Tuning
SummaryThis paper proposes Monarch-MoE, which replaces the dense factors in LoRA experts with block-diagonal factors linked by a fixed permutation, aiming to reduce catastrophic forgetting when multimodal models learn new tasks sequentially. Evaluated on MLLM-DCL and UCIT with LLaVA-1.5-7B, it improves first-task retention and cuts adapter parameters by about 75% and 50%, respectively. The authors report that fine-tuning and inference fit on a single 24 GB RTX 4090.
In briefBeyond Actions: Learning Future Targets for Robot Vision-Language-Action Models
SummaryThis paper introduces AFCE, which makes a future physical Effect, rather than only the action commands, the main prediction target of robot policies. An offline tokenizer builds temporally ordered Effect tokens from demonstrated actions, robot responses, and visual transitions, and a flow-matching VLA predicts these Effects, which a state-conditioned decoder converts into action chunks. On the DexJoCo benchmark, AFCE reaches 50.97% overall success, outperforming π0.5 and UVT by 11.82 and 15.21 percentage points, and leads on all five bimanual tasks.
In briefHow Much Computation Does Each Forecast Need? Block Skipping in Time-Series Models
SummaryThis paper tests whether whole Transformer blocks can be skipped at inference in two pretrained time-series forecasting models, Chronos-Bolt and TimesFM 2.5. Masks are chosen on calibration windows and evaluated on 1,088 held-out windows across four datasets. Moderate skipping keeps forecasting quality nearly unchanged while reducing latency, but aggressive skipping harms Traffic and Weather. The authors conclude the results reveal structured redundancy, but do not yet establish a deployable dynamic inference method.
