In briefExplainable machine learning for estimating multilayer planar winding inductance
SummaryThis paper presents an explainable machine learning framework that combines feature attribution (SHAP and permutation importance) with Kolmogorov-Arnold Network-guided symbolic regression (SR-KAN) to derive closed-form equations for estimating self-inductance of multilayer planar windings. On over 10,000 finite element simulations, tree-based models showed extrapolation errors above 36% out of distribution, while the SR-KAN expression achieved an 8.22% relative error. Tests on 55 printed circuit board prototypes (up to 8 layers) gave a mean absolute relative error of 6.26%. The dataset and measurements are released open-source.
WhyCombining SHAP with KAN-guided symbolic regression for planar winding inductance is novel; validated on over 10,000 FEA simulations and 55 physical prototypes, offering practical value for power conve
Spyros Rigas, Theofilos Papadopoulos, Georgios Alexandridis et al.
In briefPointwise or Pairwise: When Pairwise Losses Help Reward Learning
SummaryThis paper asks when pairwise losses outperform pointwise losses for reward learning, even when pointwise rewards are observed. It compares Value Regression, which regresses rewards directly, with Value Difference Regression, which regresses reward differences between action pairs under the same context, in a grouped offline contextual-bandit setting. The authors prove finite-sample regression guarantees and offline-regret bounds, showing VDR removes the misspecification term for finite function classes while for linear classes a feature-geometry-dependent bias-variance tradeoff arises, supported by numerical experiments.
In briefLearning to Prove, Not Just Answer: Formal-Verification-Based Reinforcement Learning for Natural-Language Logical Reasoning
SummaryThis paper addresses the problem that language models can give correct logical answers without verifiable proofs. It proposes Proof-R1, a reinforcement learning framework that accepts intermediate conclusions only after they pass UNSAT-based machine-checkable formal verification. The method also traces the proof dependencies that support the final answer and aligns outcome rewards with them. Experiments report improved answer accuracy on three benchmarks and four backbone models, along with better reasoning-process verifiability than training-free and other training-based methods.
In briefProbabilistic symbolic model of droplet collision for high-pressure spray simulation
SummaryThis paper presents a probabilistic symbolic-distillation model of droplet collision outcomes. A machine-learning model is trained on nearly forty thousand experimental events to learn the joint outcome-probability landscape, and symbolic regression then distils it into eight class-specific analytical expressions. The model replaces abrupt regime switching with finite-width fuzzy boundaries and outperforms the evaluated conventional analytical boundary models. The authors also present a sampling scheme intended for Eulerian-Lagrangian spray simulation.
In briefAdaptive Reward Routing: Dynamic Reward Optimization for Joint Audio-Video Diffusion
SummaryThis paper proposes Adaptive Reward Routing, a method for training joint audio-video diffusion models with multiple rewards via forward-process reinforcement learning. It dynamically chooses where reward-driven updates are applied, using bidirectional cross-attention as a proxy for cross-modal influence, and adjusts how competing rewards are weighted, keeping predefined priors and adding gradient-based corrections after warm-up. The authors report consistent gains over strong RL baselines in modality quality, semantic consistency, and audio-video synchronization, with ablations supporting the complementary benefit of both components.
Songlin Yang, Xiaotong Zhao, Jiacheng Zhang et al.
In briefV-Engram: Trigger-Indexed External Memory for Modular Text-to-Image Personalization
SummaryThis paper proposes V-Engram, an external memory mechanism for Stable Diffusion 3.5 in which each concept has an explicit trigger that retrieves its memory. The retrieved memory is added as relative residuals to frozen text-encoder and MMDiT context states, keeping it separate from backbone adaptation. Experiments report that V-Engram broadly matches DreamBooth-LoRA in overall subject fidelity and shows advantages in some settings, such as contextual subject preservation.
In briefUsing in-context learning to read handwritten text without retraining models
SummaryThis paper studies whether general-purpose vision-language models can transcribe handwritten text by learning from example images in their prompt, without updating model parameters. The authors test this across several collections and models, including a cross-domain setting where examples come from a different collection than the query. They report that as the number of context examples grows, error ranges narrow toward average performance, and that some ways of choosing examples work better than others.
In briefJoyAI-Voice 2.0: A Speech Generation Model Using Joint Semantic and Acoustic Representations
SummaryThe paper presents JoyAI-Voice 2.0, an end-to-end speech generation model built on a fully continuous dual-encoder design that fuses semantic and acoustic representations. A causal autoregressive Transformer plans each speech patch, and a local diffusion Transformer renders it at 48 kHz. The authors report an average word error rate of 2.51% on Seed-TTS, a 14.9% relative reduction over the strongest baseline.
In briefHaPRL: Human-Anchored Process Reinforcement Learning for Visual Search Agents
SummaryThis paper studies multi-turn visual search agents that decide where to look in high-resolution images. Standard reinforcement learning rewards only the final answer, so flawed search routes with correct answers often go unchecked. The authors propose HaPRL, which collects over 1,000 human-annotated search traces with fine-grained behavioral signals and uses them to score agent rollouts. Experiments report that HaPRL outperforms outcome-based RL, and that early process supervision yields 6.7x more improvement in later outcome-based scaling.
In briefSparse linear bandits with budget limits, robust to corrupted data
SummaryThis paper studies sparse linear contextual bandits with knapsack constraints where both rewards and resource consumption can be corrupted. The authors propose Robust Optimistic Primal–Dual (ROPD), which combines corruption-aware confidence widths, online resource prices, and a budget-safety rule. They give regret bounds under different exploration and design coverage conditions, and introduce Shared-Grid for when the corruption level is unknown. Both methods keep observed budgets within limits on every realization.
In briefInsightMap: Structured spatial modeling for robots that follow language instructions while navigating.
SummaryThis paper presents InsightMap, which uses top-down maps as both spatial memory and targets for predicting the map after an action, jointly learning navigation actions and map generation. On the validation-unseen splits of R2R-CE and RxR-CE, it reaches success rates of 56.9% and 54.9%. Adding map-prediction supervision improves R2R-CE SR by 4.3 and SPL by 3.2 percentage points.
In briefFaster simulation of random systems using learned surrogate models
SummaryThis paper addresses the high computational cost of simulating dynamical and stochastic systems. It proposes a framework that uses a Variational Autoencoder with convolutional or graph structure to reduce system dimensionality, then propagates the latent vector over time with a Temporal Fusion Transformer that encodes long- and short-range effects and static covariates. Tested on three distinct cases, the framework reportedly matched simulation results closely while requiring only a fraction of the time, and it provides built-in uncertainty quantification.
Marco Jochum, Ioannis Kouroudis, Gohar Ali Siddiqui et al.
In briefEgoHumanoid-V2: Transferring human whole-body movement skills to humanoid robots for loco-manipulation
SummaryThe paper presents EgoHumanoid-V2, a framework that transfers coordinated whole-body skills from egocentric human demonstrations to humanoid robots for loco-manipulation. Its core is coarse-to-fine action alignment, combining kinematic reference correction with dynamics-aware refinement to improve end-effector pose accuracy while keeping whole-body coordination. Robot-arm rendering and training-time image augmentation are used to reduce the visual embodiment gap. On four real-world tasks, vision-language-action policies trained on aligned human data transfer zero-shot without target-task robot demonstrations, with scores comparable to policies trained on teleoperation data at lower collection cost.
In briefFaster topology-preserving image segmentation via sparse cubical filtrations
SummaryThis paper proposes sparse cubical filtrations as a new basis for persistent homology computation, aiming to preserve topological structure in image segmentation at lower cost. The authors report reductions in downstream computational cost of up to 100 times on real datasets. They evaluate it as an optimization objective in training settings, including 3D data with large patches, and report topological accuracy gains of up to 80% across six datasets while keeping pixel- and region-based accuracy.
Alexander H. Berger, Marco Fontana, Daniel Rueckert et al.
In briefAbsorbed in Inertia: Activation Analysis for Computer-Use Agents
SummaryThis paper studies computer-use agents that carry out tasks on live desktops from natural-language instructions. The authors find these agents can repeat actions that are ineffective even when they recognize the actions fail, a behavior they call inertia. By analyzing the models' internal activation values, they link inertia to an absorbing region of activation space that is hard to leave. They propose R³ (Reset, Reroute, Restore), which temporarily resets the context to escape that region and then restores history to finish the task, reducing measured inertia by 17-55% across models.
Giulio Segalini, Zhi Wen Soi, J\'er\'emie Decouchant et al.
In briefVision-language model fine-tuning for end-to-end combinatorial optimization
SummaryThis paper proposes a general-purpose vision-language solver that adds visual representations derived from the input to textual problem descriptions. One model is applied across several combinatorial optimization tasks and is trained with supervised fine-tuning followed by verifier-guided reinforcement learning. The experiments report that it generally improves solution quality over a text-only version, with clear gains on complex problems such as CVRP and JSSP. The benefit of visual information is larger at bigger problem scales.
In briefSimpleEvol: A Lightweight Agent-Loop Framework for Automated Heuristic Design with Few Human Priors
SummaryThe paper argues that current LLM-driven automated heuristic design embeds models as narrow components inside heavily hand-engineered evolutionary frameworks. The authors introduce two metrics, handcraftedness (AHI) and intelligence conversion efficiency (ICE), and evaluate ten LLMs on three combinatorial optimization problems. They find that frameworks with fewer human priors consistently achieve higher ICE. Based on this, they propose SimpleEvol, an agent-loop framework that removes nearly all human priors and reports the highest ICE overall.
In briefClosing Semantic Gaps in RAG by Fusing Generated Context Knowledge
SummaryThis paper addresses a mismatch between queries and retrieved passages in retrieval-augmented generation (RAG). It proposes Knowledge-Aware Semantic Bridging (KASB), which fuses generative and retrieval-based knowledge in a multistage process to align these semantic spaces. The authors evaluate KASB on three open-domain question answering datasets.
In briefInterpolated Policy Distillation: a controllable middle ground between off-policy and on-policy distillation
SummaryThe paper proposes Interpolated Policy Distillation (IPD), which linearly mixes the student and teacher next-token distributions at each decoding step, with a coefficient that controls the trade-off between trajectory quality and student learnability. A new speculative-decoding rule makes sampling practical while exactly preserving the interpolated distribution. Experiments on text-only and multimodal reasoning benchmarks show IPD outperforms SFT, OPD, their two-stage combination, and recent heuristic segment-interleaving methods.
In briefTrajectory Soup: spreading LLM mid-training compute across diverse runs and merging them
SummaryThe paper observes that serial compute in LLM mid-training gives diminishing downstream gains and can even degrade some capabilities. It proposes Trajectory Soup, which splits the mid-training budget across several independent branches forked from a shared checkpoint, then merges the best validation-selected checkpoints via intra- and inter-trajectory averaging. Across model scales, learning-rate schedules, token budgets and trajectory counts, the method outperforms the strongest single-trajectory average under matched budgets and keeps improving as budgets grow.
In briefRemoving spurious correlations in vision-language-action robot models via domain-invariant latent lookahead prediction
SummaryVision-language-action (VLA) models are fragile under visual distribution shifts and often rely on spurious correlations tied to domain-specific factors. The paper proposes DILL, which supervises policies with domain-invariant future latents learned from domain-transformed trajectories, using a Task-Domain Encoder trained with contrastive and Gaussian disentanglement objectives. On LIBERO-Plus, DILL reports 69.1% average success, 11.4 points above the strongest baseline, and real-world manipulation experiments are also presented.
In briefEnd-to-end self-supervised RGB-T tracking that avoids pseudo-labels and dense annotation
SummaryThis paper proposes ESMTrack, a self-supervised framework for tracking objects in paired visible (RGB) and thermal infrared video, trained end to end without offline pseudo-labels or dense per-frame boxes. It uses only the standard first-frame annotation, learning from a grounding triplet loss on annotated frames and a cross-frame temporal triplet loss on unlabeled frames. A three-branch design with unimodal branches is used to reduce dominance of one modality, and experiments on five benchmarks report competitive results and real-time speed.
In briefVoting Hides Failure: How Aggregation Choice Shapes Noise Robustness in Heart Murmur Detection
SummaryThis paper compares two automated heart murmur detection pipelines, HMS-Net-CNN and BiLSTM-LSTM, on phonocardiogram recordings under controlled noise of several severities. Using noise-augmented fine-tuning and held-out generalization tests, it finds that the BiLSTM pipeline outperforms HMS-Net across conditions when aggregation is matched. The authors show that the aggregation method can change noise-robustness conclusions substantially, so evaluation choices matter as much as the models.
Nicholaus Dismas Ladislaus, Olatunji Damilare Emmanuel, Samuel Chol Buol
In briefGLASS: a slot-based set decoder that generates all atoms of a crystal in parallel, aiming at larger all-atom structures
SummaryThis paper proposes GLASS, which encodes crystal structures into a permutation-invariant global latent space and learns their distribution with flow matching. A learned-slot decoder builds all atoms in parallel, avoiding atom-wise correspondence during generation. On MP20 it is competitive with particle-space models, and on a QMOF subset it generates metal-organic frameworks with up to 150 atoms per unit cell, approaching the validity of the training data. The authors also report a validity-novelty tradeoff in flow training.
Hendrik Kra{\ss}, Seyed Mohamad Moosavi, Mathias Niepert
In briefFrom learner behavior to reusable skills for more effective and efficient learner simulation
SummaryThe paper proposes Learner2Skill, which turns the learner-specific simulation ability gained from past interactions into a persistent, reusable Simulation Skill. The Skill captures the learner's current state and recurring response patterns, and it evolves as new real interactions arrive. It can be transferred to a new LLM through lightweight executor calibration without rebuilding the learner from scratch. Experiments report more faithful reproduction of fine-grained learner behavior, lower overall token cost, and effective reuse of the same Skills across different LLM executors.
In briefLucid Dreaming for World Models: teaching AI agents to doubt their imagined predictions and act on trust.
SummaryThe paper proposes the Lucid World Model (LucidWM), which estimates uncertainty in world models that agents use to plan in imagination. It integrates Subjective Logic into categorical latent transitions to assign each transition a doubt level, with its complement defining trust that is multiplied along imagined trajectories to reweight returns and guide actions. Evaluated on four base world models against seventeen uncertainty readouts, it detects environmental changes and signals uncertainty under action-corrupted rollouts. In a controlled navigation case study, acting on trust reduced steps to reach the goal from 362 to 190.
In briefDShuttleNet: A Badminton Doubles Dataset for Measuring Teamwork
SummaryThis paper presents DShuttleNet, a badminton doubles dataset built from elite BWF World Tour matches. Besides trajectory data, it includes expert-labeled loss reasons for diagnosing tactics. The authors define geometric features and two new indices that measure how well players cover the court and share defensive duties. Experiments show that conventional models underfit when these teamwork-focused features are left out.
Guan-Yi Jhang, Chiencheng Chiang, Wei-Kai Chen et al.
In briefFlowMAS: Learning Multi-Agent Workflow Topology via Information-guided Generative Flow Network
SummaryThis paper proposes FlowMAS, a method that uses Generative Flow Networks (GFlowNets) to learn the topology of multi-agent workflows, modeling workflow generation as reward-guided flow over the topology space. It combines a GFlowNet-based generation backbone, a curiosity-driven module that encourages exploring structurally novel workflows, and an information-guided module that evaluates information contribution and communication efficiency of operators. Experiments on six benchmark datasets with three LLM backbones show that FlowMAS consistently outperforms multiple baselines.
In briefDetecting Self-Compassion Across Video, Audio and Text in Reflective Dialogue
SummaryThis paper studies how hard-to-observe psychological traits can be inferred from video, audio and text, using self-compassion as the example: responding to setbacks with patience rather than harsh self-criticism. The authors collected and annotated 51 reflective dialogue sessions from a technology-mediated training setting and built a window-based pipeline aligning the three modalities. A simple probability-level fusion gives modest but consistent gains over the best single modality.
In briefImproved Distributional Diffusion Models: A Stochastic Few-Step Image Generator
SummaryDistributional Diffusion Models replace the standard mean-prediction denoiser with one trained by a scoring rule to approximate the full conditional distribution. This paper defers particle expansion to late transformer layers and adds time-dependent scoring rule schedules, making training practical on class-conditional ImageNet-256². A single DiT-XL/2 model trained in one stage reaches FID 4.48 at 4 steps and 2.38 at 50 steps, and the same recipe transfers to text-to-image generation.
WhyAdds time-dependent scoring rule schedules and late-layer particle expansion to DDMs, with strong ImageNet-256 FID from single-stage training.
Tommaso Martorella, Alexandre Galashov, Felix Krause et al.
In briefRethinking LLM-as-a-Judge: From Judgment Quality to Downstream Utility
SummaryThis paper studies using large language models as judges to evaluate responses on open-ended tasks without ground-truth answers. The authors vary how judgments are elicited (verdict granularity, critique usage, batching) and how they are used, extending them beyond policy training to test-time methods such as Best-of-N selection, judge-guided revision, and beam search. They find that judgment quality and downstream utility do not always align, that judge protocol design strongly affects both, and that judge guidance can turn test-time compute into performance gains.
In briefLoLBench: Testing coding agents on turning long proposals into code for large software
SummaryLoLBench is a multilingual benchmark of 100 tasks across 29 large software systems, where each task pairs a human-written enhancement proposal with the implementation it requires. It tests whether coding agents can both understand high-level intent and produce the correct code changes. Across 28 agents, the best resolved only 14% of tasks, with incomplete code localization identified as a major bottleneck. Providing reference file trees and API specifications raised resolved rates by 16 to 22 percentage points.
WhyIntroduces proposal-to-implementation evaluation of coding agents; large experiments across 28 agents and 100 tasks offer useful insights for the subfield.
In briefMSTypography: Semantic typography for multi-character words that keeps text readable while showing object meaning
SummaryThis paper presents a framework for semantic typography on multi-character words, where letter shapes visually convey the word's meaning while staying legible. It first approximates the overall silhouette using masks, then refines local details under semantic guidance, with structural and OCR-based constraints preserving readability and diffusion priors pushing glyphs toward the target concept. The authors claim it is the first method to balance word legibility and object recognizability for multiple characters, outperforming prior methods across English, Chinese, Japanese, Korean, and Arabic. Code will be open-sourced.
In briefBandPC: Learned Residual-Band Preconditioner Combinations for Flexible Conjugate Gradient Solvers
SummaryThis paper presents BandPC, a graph neural network framework that predicts which sequence of preconditioners a flexible conjugate gradient solver should use across its iterations, based on matrix structure. The iteration is split into three residual-norm bands, with 125 candidate combinations drawn from five classical preconditioners. On a benchmark of synthetic and SuiteSparse matrices, BandPC achieves the best iteration count on 41.3% and the best solution time on 33.1% of test matrices.
In briefTaoFlowForge: Progressive Native Mesh Generation via Cascaded Flow Matching
SummaryTaoFlowForge is an artistic mesh foundation model that generates production-ready 3D meshes. It splits mesh generation into vertex generation, done coarse-to-fine in two stages, and connectivity prediction, which estimates vertex connections and per-vertex normals. Trained on a combined dataset of hand-crafted assets and public topology data, it outperforms autoregressive methods in image-conditioned generation and achieves state-of-the-art results among open-source mesh topology generators.
In briefMulti-granularity language guidance for imitation learning through instruction decomposition
SummaryThis paper proposes a method that splits an overall task description into finer, subtask-level language instructions to guide robot policy learning. The goal is to help the robot tell apart the behaviors needed at different stages of multi-stage manipulation tasks. The authors evaluate the method in a multi-task imitation learning setting and report that it improves learning efficiency and performance.
In briefTrain Ahead, Distill Back: A Bootstrapped Self-Distillation Method for Language Models
SummaryThe paper proposes Bootstrapped On-Policy Self-Distillation (B-OPSD), which first trains the policy forward to create a future teacher, then resets the student to its original state and supervises it with that teacher. The future teacher provides more reliable privileged trajectories and more informative token-level targets. On mathematical reasoning tasks with Qwen3-4B and Qwen3-8B, the method consistently outperforms standard OPSD, improving scores from 27.50 to 41.30 and from 48.80 to 64.44 in the rollout-privileged setting.
In briefSkillCome: Group-contrast skill optimization for language models using dual memory
SummaryThis paper proposes SkillCome, a method for improving large language model skills by analyzing trajectories generated under a given skill. Instead of relying on a single trajectory per question, it generates multiple trajectories and contrasts successful and failed ones to pinpoint key behavioral differences. A dual memory system accumulates evidence from historical steps to capture patterns shared across groups. Experiments on six benchmarks and five models show consistent gains over baselines, up to +5.69 points.
In briefLLM unbranding: removing brand identity from AI text while keeping it useful
SummaryThis paper defines LLM Unbranding, the task of stopping language models from producing text that reveals brand identity, such as signature phrases, slogans, and stylistic markers. Because such outputs can cause trademark dilution, false attribution, and brand defamation, the authors built an evaluation dataset of prominent brands across several commercial domains. They found that existing state-of-the-art machine unlearning methods struggle with selective textual unbranding. They then propose MUTE, an inference-time method that iteratively refines system instructions to remove brand leakage while preserving general model utility, without fragile parameter updates.
WhyIntroduces a new LLM unbranding task and an inference-time iterative method, MUTE, with benchmark evaluation of existing unlearning baselines.
Kajetan O\.z\'og, Alicja Wojciechowska, Dawid Malarz et al.
In briefSafety certificates for robots that stay safe even when moving obstacles act adversarially
SummaryThe paper proposes an adversarially robust geometric safety certificate for nonholonomic robots facing obstacles that can maneuver within bounded capabilities. It contracts the certificate parameters in closed form to account for the worst-case effect of admissible obstacle maneuvers, and gives sufficient conditions for forward invariance under kinematic bicycle dynamics. Simulations report substantial reductions in barrier violations and collisions, and show that pointwise robustification of the barrier derivative cannot replace contraction of its geometry.
Chandan Kumar Sah, Bazeela Banday, and Jishnu Keshavan
In briefWhen agents should check outside conditions: budgeting observations for stored intentions
SummaryThe paper studies when an AI agent that holds a stored intention tied to a future condition should spend resources checking whether that condition holds. It introduces BudgetPM, a resource-allocation method under a shared observation budget, with static and sequential policy variants. Experiments on two benchmarks report competitive or better results than baselines using fewer observations in some settings.
In briefEviViT: Evidence-Adaptive Vision Transformers for Fine-Grained Perception
SummaryEviViT is a lightweight module that can be attached to a pretrained vision transformer to decide where the model should acquire more visual detail. It uses human visual-search traces to supervise a question-conditioned evidence density, which guides re-reading of regions from the original pixels and allocation of visual tokens. Experiments across nine host models report consistent gains in average fine-grained accuracy, and matched-budget comparisons show it outperforms global-only processing while using fewer visual tokens.
In briefLearning the structure of triangular transport maps, with a method that learns the map, variable ordering and sparsity together
SummaryTriangular transport maps are sampling-based tools for density estimation, generative modeling and Bayesian inference. This paper introduces Self-Structuring Transport Maps (SSTM), which jointly learn the map, the variable ordering and the sparsity pattern, and uses a monotone BatchEnsemble to keep the approach scalable. On synthetic and real data, joint learning of structure and map gives better density estimates than learning the structure first. The authors report that SSTM is competitive with autoregressive flows on large datasets.
Morten Bl{\o}rstad, Pekka Parviainen, Berent {\AA}nund Str{\o}mnes Lunde
In briefCross-linguistic effects in bilingual phoneme-based baby language models
SummaryThe paper trains bilingual small language models on phonemic speech representations to study how a first language affects learning English as a second language. English is fixed as L2 while German, Swedish, Persian, and Basque serve as L1s. The authors find that phonemic input yields stronger L1-related variation in grammatical learning than orthographic input, and early lexical differences align with phoneme-inventory similarity.
Nikitas Theodoropoulos, Maria Lymperaiou, Giorgos Filandrianos
In briefReusing the Critic: Reward-Free Policy Optimization for LLM Post-Training
SummaryThe paper introduces Reward-Free Policy Optimization (RFPO), which reuses a single calibrated, frozen critic as a rollout-level reward, a baseline for generalized advantage estimation, and a predictor of success for unfinished prefixes. The authors argue that instability in critic-based RL for long reasoning chains is largely an optimization artifact that small, low-variance policy updates can resolve. They report that, with binarized scores, RFPO matches supervised PPO without labels in the training loop while reducing compute and memory costs.
In briefMultichannel Audio Quality Assessment: Reusing Pretrained Perceptual Models for Spatial Audio
SummaryThis paper studies how perceptual quality models pretrained on limited channel configurations can be reused for 5.1-channel spatial audio. The authors compare four levels of multichannel integration (signal, prediction, latent, and feature) and propose two learned approaches, latent-level aggregation of spatial-group representations and the feature-level Feature-Band Group Attention (FGAtt). Across five 5.1-channel test sets, FGAtt achieves the strongest overall performance.
In briefNRF-GS: Neural Residual Fields for Compact, Expressive Gaussian Splatting
SummaryThis paper examines how 3D Gaussian Splatting (3DGS) models appearance, arguing that low-order spherical harmonics cannot capture high-frequency view-dependent reflections, so extra splats are often added to compensate. The authors propose NRF-GS, which replaces per-splat spherical harmonic bases with a shared neural residual field predicted by a lightweight global MLP. According to the paper, NRF-GS achieves comparable or better rendering quality while using up to 50% fewer Gaussians, with improved specular and high-frequency details.
In briefMaking intrinsic dimension estimates interpretable by calibrating distance and angle separately
SummaryThe paper reformulates the DANCo intrinsic-dimension estimator so that its distance and angular components are kept separate, making the source of each estimate identifiable. For the distance component, it derives a closed-form KL divergence for the Gride estimator, reporting that mean percentage error on 24 manifolds falls from 27.7% to 17.6% under noise. Controlled and observational experiments, including a Gaussian scale mixture and CIFAR-10 and ImageNet, examine how sample amplitude heterogeneity affects the angular component.
In briefNeural constitutive learning for reaction-diffusion equations: learning the constitutive laws with neural networks and time-stepping them with a shared integrator
SummaryThe paper proposes the NCL-MCT Solver, which learns PDE-specific constitutive responses (mobility and driving force for transport, relative reaction rates for reactions) with neural networks, while time evolution is handled by a shared MCT integrator. These responses depend on the current density rather than the initial condition or elapsed time, so they can be reused across initial conditions and longer horizons. Across seven systems, separately trained modules reach relative rollout L² errors of 10^-4 to 10^-2. The authors also show that constitutive laws can be learned without generating solution trajectories when the laws are known.
Shang-Ke Chen, Yu-Peng Wang, Shih-Hsuan Hung et al.
In briefLearning from Viable Failure Prefixes: A Policy Optimization Method for Long-Horizon LLM Agents
SummaryThe paper addresses reinforcement learning for long-horizon LLM agents under sparse, delayed rewards, noting that existing methods such as GRPO and GiGPO fail when compared rollout returns show no variation, a problem it calls zero-credit failure. It proposes MVPO, which estimates prefix potential over Union-Find viability regions, repairs zero-credit groups with potential-difference advantages, and attenuates the potential branch by relative performance progress. Experiments with Qwen2.5-1.5B-Instruct report that MVPO outperforms eight baselines, improving over GiGPO by 4.4 success points on ALFWorld and 5.3 on WebShop at the same training length.
In briefA tool for stakeholders to negotiate shared meaning when assessing AI in regulatory sandboxes
SummaryThe paper notes that AI regulatory sandboxes (AIRSes) have been studied mainly for their legal and institutional design, leaving socio-technical interpretation to stakeholders and creating a gap in sense-making. Using Research-through-Design, the authors iteratively developed a Boundary Negotiating Artifact, an interface that lets different stakeholders collaborate on AI assessment. They position it as a technical backbone for established AIRS frameworks and report design insights, with qualitative evaluation left to future work.
Idoia Landa-Oregi, Tom Deckenbrunnen, Alessio Buscemi et al.
In briefWaypoint-1.5: A Real-Time Video World Model for Consumer Hardware
SummaryThis paper presents Waypoint 1.5, a real-time diffusion world model that generates interactive video on consumer-grade hardware. It is pre-trained on 100,000 hours of video from hundreds of games and produces playable video responding to full keyboard and mouse input. The authors distinguish rendered FPS, latent FPS, and control rate, and evaluate interactivity through latency and throughput. The paper also discusses safety and ethical considerations specific to interactive world models.
Rajit Rajpal, Shahbuland Matiana, Liew Wei Pyn et al.
In briefVACE: Validation-gated alternating co-evolution of agent models and their task harnesses
SummaryThis paper proposes VACE, a method that alternates reinforcement learning for language model agents with revisions to the harness that guides task execution. After each RL stage, a proposed harness revision is adopted only if it improves validation performance. The authors report gains over weight-only RL and over alternation without validation gating on OfficeQA and AutomationBench, and note that many proposed harness revisions were rejected.
In briefWhat Post-Training Changes in Multilingual Reasoning: An Audit of Open Reasoning Models
SummaryThis paper audits open-source reasoning models on competition mathematics problems in eleven languages. It finds that, in the ten non-English languages, only a small share of problems receive a correct, complete solution with visible reasoning in the requested language, compared with 92.9% in English. Comparing thirteen checkpoints from one model family, the authors report that multilingual supervised fine-tuning restores target-language reasoning but lowers accuracy, while reinforcement learning restores termination and keeps the target language only when the reward includes a language term.
In briefV2X-WAM: A Cooperative World-Action Model for Self-Driving Cars That Predicts Future Traffic
SummaryThis paper proposes V2X-WAM, a model that links vehicle-infrastructure cooperative perception, driving action generation, and future traffic prediction. It fuses onboard and roadside observations into a reliability-aware representation, compressing roadside data into compact quantized messages to save communication bandwidth. The planner's predicted trajectories are used to forecast future occupancy and dynamic flow, and these predictions are fed back to refine the trajectory. Experiments on a large real-world cooperative driving dataset report better planning accuracy and safety, stronger future prediction, and lower communication overhead than representative methods.
In briefRevealing hidden weaknesses in LLM peer reviewers via adaptive fuzzing
SummaryThis paper examines whether LLM-based reviewers reliably penalize manipulated content such as overclaiming, noting that prior conclusions rest on a narrow set of static perturbation templates. The authors build a three-level evaluation framework covering surface presentation, argumentative logic, and value judgment, finding that vulnerability depends on whether a paper's original score is high or low, and that a single template misses weaknesses. They propose SCOPE-Fuzzer, which uses feedback-driven strategy selection and adaptive content mutation, and report that it uncovers vulnerabilities overlooked by static evaluation and other baselines.
In briefWhy Multimodal LLMs Struggle to Count, and the ConvStack Fix
SummaryThis paper examines why multimodal large language models perform poorly at fine-grained visual counting. It identifies two bottlenecks in their global attention pipeline: an individuation problem caused by image patchification, and a collapse in counting aggregation as numerosity grows. The authors propose ConvStack, a lightweight architecture that aggregates and injects local spatial structure in visual token space. Fine-tuned only on counting tasks, it shows gains on dense object counting and spatial understanding benchmarks without losing general visual capabilities.
In briefLLM Multi-Agent Systems over Wireless Networks: A Joint Agent-Network Design View
SummaryThe paper examines how LLM-based agents are increasingly distributed across wireless edge nodes, so wireless networks must support the multi-agent reasoning process itself. It proposes a joint design approach that coordinates agent-side and network-side decisions, including interaction scheduling with resource allocation and co-designed message selection and transmission. In a vehicle-to-everything (V2X) case study, jointly adapting agent and network operations improves task completion under communication and edge-resource constraints compared with separate agent-only or wireless-only designs.
In briefResidual vector quantization for task-oriented visual feature compression in device-edge multimodal inference
SummaryThis paper addresses the delay caused by sending visual data over limited uplinks when large multimodal models run across a device and an edge server. It proposes Q-TOFC, which uses residual vector quantization to encode features as compact codebook index sequences, adds query relevance to feature aggregation, and applies a quantization error compensation adapter. On seven multimodal benchmarks, it reduces the visual payload by 53.6% relative to TOFC while keeping comparable average normalized task performance, and it lowers end-to-end latency under bandwidth-constrained uplinks.
In briefReal2Gym: Turning Videos into Simulation Gyms to Teach Robots Skills
SummaryThis paper presents Real2Gym, a framework that reconstructs interactive simulation environments from human or robot demonstration videos. Within these simulated gyms, an agent generates manipulation code, learns from successes and failures, and distills reusable skills that transfer to real robots. The authors report a 16.7% higher success rate than GPT-6 Astra Direct Mode in simulation with about 74.9% fewer policy-execution tokens, and a 33.3% higher success rate on four tasks with a real Franka robot.
In briefMeasuring Energy Use When Training Modern Vision Models
SummaryThis study measures the energy used to train seven modern computer vision models on a 10,000-image ImageNet subset with a uniform 40-epoch setup on two NVIDIA Tesla P100 GPUs. Energy was recorded via NVML and compared with each model's computational cost in GFLOPs, which correlated at 0.85 with energy use. Models with similar GFLOPs differed in energy consumption by up to 3.1 times, and EfficientNet variants gave the best balance of accuracy and efficiency, while Vision Transformers used the most energy and scored lower under these conditions.
In briefARGOS: Reinforcement Learning for Multidimensional Elastic Service Orchestration
SummaryThe paper presents ARGOS, an end-to-end controller that uses reinforcement learning to adjust both analytics quality (coverage, sample, freshness) and cluster pressure when resource scaling alone is insufficient. It combines this with capacity-aware admission control. In controlled and multi-tenant heterogeneous cluster tests, deep reinforcement learning policies outperform non-learning baselines and approach the best fixed reference. Live evaluation shows gains over a static midpoint with no recorded CPU or memory violations, but coverage violations remain.
Javier Mateos-Bravo, Sergio Laso, Juan Luis Herrera et al.
In briefTesting whether genetic ranking resists hidden bias better than learned rankers
SummaryThis paper compares label-free genetic multi-criteria sorting with learning-to-rank models trained on labels, such as LightGBM and XGBoost. Experiments on MSLR-WEB10K, Baidu-ULTR, WANDS, and Expedia test whether rankings align with undesignated proxy signals under proxy skew and label contamination. The genetic methods show much lower rank-proxy correlation, and their bias stays flat as labels are contaminated, while LightGBM and XGBoost show rising bias.
Vladyslav Shevchuk, Carlos Uziel P\'erez Malla, and Philipp Wissgott
In briefRecovering Governing Equations from Unstructured Data via Causal Representation Learning
SummaryThis paper studies recovering the governing ODE of a dynamical system from unstructured, high-dimensional observations such as images. The authors show that variables learned by causal representation learning generally cannot be used directly for equation discovery, and introduce SPEED-AE, which adds a component-wise autoencoder to learn transformations suited to sparse ODE discovery. Experiments on Lotka-Volterra, Lorenz, and a two-pendulum system show improved disentanglement, recovered ODEs closest to the ground truth, and state-of-the-art forecasting performance.
Alessandro Trenta, Riccardo Massidda, Davide Bacciu et al.
In briefTraverse: Teaching Web Search Agents When to Remember, Reset, and Redirect
SummaryThis paper addresses long-horizon information-seeking agents, which accumulate noisy context so that early mistakes persist and are hard to recover from. The authors propose Traverse, an autonomous search harness where the agent manages its search through Rubric, Answer, and Verify states and has a Seal Memory tool for active context management. To keep reinforcement learning stable, they train only the final segment after context management. Their 35B model scores 72.83 on BrowseComp and improves over the base model across several benchmarks.
In briefLDM-is-AE: A Latent Diffusion Model That Acts as Its Own Autoencoder for End-to-End Image Generation
SummaryThe paper argues that standard latent diffusion models use a two-stage pipeline in which an autoencoder is pre-trained separately, creating a mismatch between the latent space and the denoising process. It proposes LDM-is-AE, which splits the DiT backbone into reciprocal encoding and decoding parts and applies image-space supervision to intermediate features at all timesteps, enabling one-stage end-to-end training. The authors report FIDs of 1.80 and 1.90 for class-conditional image generation at 256x256 and 512x512 resolution.
Zhengqiang Zhang, Lingchen Sun, Rongyuan Wu et al.
In briefZeroDiff: Diffusion-Based Zero-Shot Time Series Reconstruction with Informed Priors
SummaryThe paper addresses reconstructing target time series at locations where no measurements exist, a task it calls zero-shot time series reconstruction. It proposes ZeroDiff, which builds an informed prior from exogenous variables alone and uses diffusion to calibrate reconstruction errors, trained on observed locations and generalized to unobserved ones. The authors report significant improvements over existing approaches across diverse real-world datasets.
In briefUnlearningSoup: Is Repeated Tuning Necessary for LLM Unlearning?
SummaryThis paper studies unlearning in large language models, which aims to delete harmful or unwanted content while keeping other knowledge. The authors observe that models from different unlearning runs lie in a shared performance basin in weight space. They propose UnlearningSoup, with EfficientSoup (binary-search interpolation for quickly finding a good model) and PerformanceSoup (reweighted souping for later-stage gains). Experiments across datasets and models report 2.4x to 3.3x efficiency gains in hyperparameter selection and consistent performance improvements.
In briefPhysics-informed state space models for next-day solar power forecasting
SummaryThis paper proposes a physics-informed state space framework that combines a differentiable photovoltaic conversion model with a selective state space backbone (PowerMamba) for day-ahead forecasting over a 24-hour horizon. The physics module provides a baseline based on solar geometry, irradiance, temperature effects and inverter limits, while the state space model learns residual corrections from numerical weather prediction inputs. On one year of utility-scale PV data, the method reduces RMSE by 38.5% and MAE by 45.8% relative to a physics-only baseline and outperforms a purely data-driven model.
In briefMemory Bandwidth, Not FLOPs: Speeding Up Segmentation Decoders on Edge GPUs
SummaryThis paper shows that the transformer decoders of query-based segmentation models are limited mainly by memory bandwidth on edge devices, a bottleneck that standard FLOP analysis does not reveal. The authors propose CompactInst, which uses parameter-free 4x spatial pooling to reduce the data moved inside the decoder and a two-convolution refinement head to recover spatial detail. On Jetson Orin Nano, CompactInst speeds up the decoder and the whole model with accuracy changes of only a few tenths of AP, and the paper concludes that decoder data volume, not FLOPs, determines latency.
Oliver Busemann (1, 2), Torben T\"oniges (2) et al.
In briefContext without Commitment: Robust Dense Correspondence under Non-Rigid Deformation
SummaryThis paper addresses non-rigid point-cloud registration, which finds matching target points for each point on a deforming surface. It proposes CoCo-Reg, which uses regional patches to enrich dense point features without letting patch predictions restrict the final point-level search. On 726 held-out ModelNet10 objects across nine deformation levels, it achieves a mean correspondence error of 0.0547, lower than two baselines at 0.1993 and 0.1921.
In briefPredictive Safety Curricula for Robust Legged Locomotion: a training method that focuses on rare safety failures
SummaryThe paper introduces Predictive Safety Curricula (PSC), which allocates legged robot locomotion training experience using learned predictions of future safety cost. A distributional safety critic prioritizes terrain contexts and randomized events, leaving task reward and policy loss unchanged. On ANYmal-D hardware, PSC reduced shank-collision incidence by 63% relative to a learning-progress curriculum across three matched seeds, and eliminated observed shank collisions in evaluated trials on a production stair-climbing platform.
Ivan Ovinnikov, Pascal Sutter, Christian Gehring et al.
In briefFACT: A Method for Building More Faithful Articulated Digital Twins
SummaryThis paper presents FACT, an agentic framework that progressively builds articulated digital twins to improve geometry, contact, and dynamic fidelity. It uses an evidence-diagnosis-revision loop on an editable representation to reconstruct articulated geometry from images, repair collision proxies, and fit response models from passive-response videos. Experiments reported in the paper show better geometric reconstruction than baselines and more accurate reproduction of held-out physical responses than direct parameter inference.
In briefBeyond Compression: How Post-Training Changes Math Reasoning
SummaryThis paper compares three post-training methods for large language models on math reasoning: off-policy distillation, Qwen3's combined distillation, and DeepSeek-Math's GRPO. It uses cross-surface pass@K tests over prompt variants and translations to diagnose what changed. On easier AMC problems, post-training mainly reduces sampling cost, while on harder AIME problems it raises the large-K ceiling, and GRPO does not beat sufficient off-policy distillation at large K. English-dominant distillation helps non-English reasoning but language gaps remain.
In briefRL-PaO: A Reinforcement Learning Method Treating Prediction as Action for Decisions Under Uncertainty
SummaryThis paper proposes RL-PaO, a reinforcement learning framework that integrates system formulation, optimization, and decision execution into one environment, treating prediction as an action. It aligns prediction error with realized cost without differentiating through the black-box solver. Evaluated on day-ahead energy scheduling with real historical data, it achieves the lowest annual cost among non-oracle baselines, about 10% on average.
In briefvSkipper: Turning Dynamic Layer Skipping into Faster LLM Serving
SummaryDynamic layer skipping lets each token run only part of a language model's layers to cut computation, but existing methods rely on specialized generation loops and do not fit modern serving engines, so skipping more layers does not always lower latency. The paper presents vSkipper, a virtualization layer that makes such skippers pluggable into serving engines while keeping batching and caching features, and implements it in SGLang. On GSM8K and BBH, it reduces mean end-to-end latency by 36.8% and 13.6% near the knee of the baseline's load curve, and raises request throughput by 11.3% and 7.4% under saturation.
WhyProposes a virtualization layer integrating dynamic layer skipping into SGLang, with real-workload latency and throughput gains and reasonably strong evidence.
Wei Da, Yavuz Ferhatosmanoglu, Evangelia Kalyvianaki
In briefA Unified View of Fairness via Distributional Stability
SummaryThis paper treats fairness as stability of a predictor under distributional shifts that change the composition of protected groups. Several classical fairness notions are shown to correspond to stability under specific perturbations. The authors propose a learning method based on convex combinations of reweighted predictors, formulated as a second-order cone program, and establish generalization bounds. Experiments on standard benchmarks illustrate the approach.
Gayane Taturyan, Charlotte Laclau, Stephan Cl\'emencon
In briefMessage passing achieves strong in-context graph learning with less computation
SummaryThis paper presents Ephris, a graph in-context learning model built on sparse message passing, so it avoids per-dataset training and its cost scales linearly with node-feature entries and graph edges. It is pretrained only on synthetic graphs generated from structural causal models. On 51 node-classification datasets, Ephris ranks first on Elo, improvability, average rank, and accuracy, and its inference cost is comparable to training one GNN once, over 10 times faster than prior graph ICL models.
WhyReplaces dense attention with sparse message passing for graph in-context learning, beating tuned GNNs on 51 datasets and running over 10x faster than prior graph ICL models.
In briefUsing LLM-guided search to find better binary linear codes with higher minimum distance
SummaryThis paper applies evolutionary program search guided by large language models to improve the best-known bounds for binary linear codes. The authors introduce LinCodeEvolve, which evolves code-construction programs against an exact minimum-distance evaluator and redirects the search with new strategies when progress stalls. The work reports seven record-breaking codes and, with standard modification techniques, improvements to 22 table entries, all verified by exhaustive enumeration.
WhyApplies LLM-guided evolutionary search to binary linear codes, yielding seven record codes and 22 improved table entries, verified exhaustively.
Amal Seddas, Vladyslav Shashkov, Maryna Viazovska et al.
In briefSpatial-OPSD: Label-free self-distillation to improve VLM spatial reasoning
SummaryThis paper proposes Spatial-OPSD, a self-improvement framework that enhances spatial reasoning in vision-language models without ground-truth answer labels. During training, a teacher with automatically obtained spatial cues such as depth, 3D relations, and camera geometry gives token-level supervision on the student's own sampled trajectories. A round-wise recursive scheme keeps the teacher frozen within each round and reinitializes from the improved student in the next. The authors report that one round improves the five-benchmark average across four VLM families, and three rounds bring a spatially specialized model to the highest average among open models.
In briefACTR: Aligning Reasoning and Responses for Safer Multilingual Reasoning AI
SummaryThis paper addresses a problem where reasoning large language models may give unsafe answers in non-high-resource languages under jailbreak attacks, even when their reasoning identifies the risk. The authors propose ACTR, which measures cross-lingual reasoning contributions, identifies safety-related neurons, and optimizes only those neurons using a frozen judge model that rewards consistency between reasoning and response safety categories. Across two reasoning models, ACTR reports lower average attack success rates than evaluated state-of-the-art methods on AdvBench-X and MultiJail, with gains extending to unseen languages, while maintaining multilingual knowledge and math performance and limiting false refusals.
In briefMatToolBench: A Real-World Test of AI Agents on Materials Science Software
SummaryThis paper introduces MatToolBench, the first real-environment benchmark for testing multimodal GUI agents on professional materials science software. It contains 204 tasks across 10 tools, covering GUI operation, OriginPro scripting, and code-based database queries. Experiments show that strong results on general benchmarks do not carry over to scientific workflows, with the best model reaching only 25% success on GUI tasks and 45% on code tasks.
In briefOmniRoute: Allocating Audio-Visual Token Budgets by Temporal Relevance for Faster Omni-LLMs
SummaryOmniRoute is a training-free, two-stage framework that compresses long audio-visual token sequences for omnimodal large language models to reduce prefill cost. The first stage, Temporal Evidence-Guided Budgeting, sets per-chunk leading-modality budgets from semantic relevance and local content variation. The second stage, Budget-Constrained Semantic Compression, compresses the leading modality and calibrates the follower's retention using the actual retained fraction. Experiments on four benchmarks reportedly show a better efficiency-performance trade-off than competitive baselines.
In briefNHO: A Neural Hamiltonian operator that locates a shape's region and matches points densely across it.
SummaryThis paper addresses matching a partial 3D shape to a full surface from sparse anchor points, which requires finding the matching region and then pairing points densely within it. The authors propose NHO, which learns a neural Hamiltonian operator from anchors and the partial shape's intrinsic geometry, so its eigenspace encodes both the region and point coordinates. A reciprocal refinement loop alternates between updating the operator and recovering correspondences to resolve spatial ambiguity. Experiments report competitive accuracy on both tasks and robustness to uniform scaling and rotation.
In briefMulti-Depth Temporal Fusion for Spiking Neural Networks Trained Locally
SummaryThis paper proposes a spiking neural network that processes static images and event streams using time-to-first-spike encoding. It combines an early-vision front end, a four-layer convolutional backbone trained with unsupervised STDP, a deterministic Multi-Depth Temporal Fusion module, and a classifier trained with reward-modulated STDP. The architecture is tested on MNIST, Fashion-MNIST, CIFAR-10, and N-MNIST, outperforming traditional STDP/R-STDP baselines by 18.2 points on Fashion-MNIST and 29.2 points on CIFAR-10.
In briefSpeed in the Blind Spot: Interpretability of VLM Dynamic Perception for Driving
SummaryThis paper uses speed understanding as a controlled diagnostic for vision-language models in autonomous driving, testing surrounding-agent speed, current ego speed, and short-horizon future ego speed on nuScenes. Current ego speed is often decodable from hidden representations with linear probes but is poorly verbalized, with probe errors of 4.7-5.8 km/h versus 10.2-16.8 km/h for verbal outputs. Future-speed gains mostly persist under frame shuffling, indicating limited temporal grounding. The driving-specialized Alpamayo-1.5 model shows stronger motion representations, but this does not guarantee reliable readout or temporal grounding.
Katharina Winter, Stefan Englmeier, Fabian B. Flohr
In briefLearning Reasoning Advantages via On-Policy Distillation
SummaryThis paper studies how to transfer the benefits of explicit intermediate reasoning (think mode) to student language models using on-policy distillation (OPD). The authors propose ThinkOPD, which routes supervision at the response level by combining group-relative reward gain with a trace-response divergence (TRD) compatibility proxy. Across mathematical reasoning and code generation tasks, ThinkOPD outperforms Uniform ThinkOPD in same-model and cross-model settings.
In briefIterative Exact Discrete Guidance (IEDG): a sampling method for discrete energy-based distributions
SummaryThis paper addresses sampling from unnormalized multimodal distributions over large discrete spaces. It proposes IEDG, which follows an annealing trajectory and learns stage-wise posterior corrections, with step sizes chosen by relative effective sample size. Experiments on Ising and Potts models and Max-Cut report improvements over neural baselines on several distribution-level and local-statistic metrics.
In briefGleanVID: A Training-Free Method for Efficient Video LLMs via Complementary Token Selection
SummaryVideo large language models understand video well but are costly to run because they process many visual tokens. The paper proposes GleanVID, a training-free framework that allocates token budget across frames by temporal novelty and selects tokens using local representativeness and subspace complementarity to reduce cross-frame redundancy. Keeping only 25% of visual tokens, it retains 98.6% of Qwen3-VL's original performance while cutting prefill latency by 44.7%.
In briefBeam Search as Test-Time Self-Distillation via Counterfactual Contexts
SummaryThis paper proposes a decoding-time method that brings the self-distillation idea to inference, without parameter updates, reward models, or training data. It replaces expert demonstrations with fixed counterfactual text templates and uses the log-odds ratio of a candidate answer under the two conditions as a reward signal. Beam search is used to approximate the resulting target distribution. On mathematical reasoning, code generation, and graduate-level science QA, the method outperforms standard sampling, low-temperature, beam search, and power sampling baselines on average.
In briefSelecting the Most Informative Tokens in Natural Language Autoencoders
SummaryThis paper studies which token positions an auditor should explain in natural language autoencoders to uncover threats such as prompt injection and concealment. Across 4.7 million explanations, a ranker trained only on chat structure usually selects more relevant explanations than model-computation signals, without needing a forward pass. On three of four datasets, explaining just 5% of positions retains nearly all the success rate of explaining every position. The paper also shows pretrained verbalizers can recover words models learned to conceal through fine-tuning, without extra training.
Federico Torrielli, Gianluca Barmina, Andrea Blasi N\'u\~nez et al.
In briefNowcastDiT: Standard Diffusion Transformers for Short-Term Precipitation Forecasting
SummaryThis paper examines whether a standard Diffusion Transformer (DiT) can serve as a basis for precipitation nowcasting, which means short-term rainfall forecasting. The authors propose NowcastDiT, which adds a dynamics-aware noise prior for temporally coherent forecasts and end-to-end reinforcement learning with timestep-aware rewards for meteorological skill. The paper reports state-of-the-art performance in perceptual quality and meteorological skill on the SEVIR and MRMS benchmarks.
In briefHigh-resolution brain connectivity generation via graph-variate flow matching
SummaryHigh-resolution dynamic functional connectivity captures fast brain-network interactions, but short time windows produce noisy, low-rank covariance estimates. The authors propose GVD-CFM, a class-conditional generative model that represents each trial as SPD matrices on a Riemannian manifold, maps them through a log-Euclidean diffeomorphism and an invertible DCT temporal basis, and generates full trajectories with a Transformer-based conditional flow. Across multiple EEG motor-imagery datasets, it achieves the strongest overall results in distributional fidelity, temporal-dynamics preservation, and synthetic-to-real classification, while remaining more computationally efficient than the raw-signal and GVD-space generative baselines.
In briefWatch-Think-Interact: Reinforcement learning for long streaming video question answering
SummaryThis paper addresses streaming video assistance, where models must answer questions arriving over time from a partially observed video under a fixed context budget. It proposes WTI, a closed-loop framework that keeps compact natural-language memory entries tagged with source-video time ranges and can recall past intervals when finer visual evidence is needed. The authors build the WTI-82K dataset of 82,335 timed questions and train with Stream-GDPO using trajectory-level feedback. Reported results show WTI reaching 83.3% on StreamingBench and 73.6% weighted overall accuracy on OVO-Bench, the best aggregate performance among the compared open-source streaming baselines.
In briefFedLAFP: Combining low-rank shared adaptation with full-rank personal adaptation in federated fine-tuning
SummaryThis paper proposes FedLAFP, a federated fine-tuning framework that pairs a globally aggregated LoRA branch with a client-private RandLoRA branch. Only the shared LoRA parameters are exchanged between clients. Across four visual recognition benchmarks, it reports an average personalized accuracy of 86.93%, 1.30 percentage points above the best baseline.
In briefLoRA direction extraction for controllable light on/off editing in FLUX.1 Kontext
SummaryThis paper proposes a fine-tuning method for flow-matching diffusion models that aims to model realistic artificial lighting without a large training dataset. It targets controllable interior image editing, turning artificial light sources on or off while keeping scene geometry, object placement, materials, and visual identity unchanged. The authors introduce a LoRA Direction Training Method that extracts the pure direction of the LoRA adapter's effect in the model's flow field, along with specialized losses for realistic inverse transformation. The resulting increment map is also used to adjust lighting color and temperature more precisely.
Petr Golenderov, Dmitry Mazyar, Natalia Sovpel et al.
In briefUniBuild: Unified building mapping from multi-source optical satellite imagery
SummaryThis paper presents UniBuild, a single framework for extracting buildings from multi-source RGB optical remote sensing images. It combines unified training across heterogeneous datasets, a detail-preserving decoder, and geometry-aware losses to sharpen boundaries and separate adjacent buildings. Experiments on 10 public high-resolution datasets and two self-collected low-resolution datasets report consistent gains and good generalization to unseen domains, up to 10 m resolution. The predicted masks can be converted into GIS-compatible building footprints through simple polygonization.
In briefMotionInsight: Diagnosing Object Motion Flaws in Generated Videos
SummaryVideo generation models often produce incorrect object motion, but existing evaluations mostly focus on aesthetics or text alignment. This paper assesses object-centric motion fidelity across object consistency, motion continuity, and physical plausibility. It introduces VidMotion, a dataset of 6,879 videos with fine-grained annotations, and MotionInsight, an evaluator that diagnoses motion in motion space. Experiments show its scores align with human judgments and come with grounded explanations.
In briefLongSpark: A Fixed-Cost Parallel Drafter for Faster Speculative Decoding
SummaryThis paper introduces LongSpark, a block-diffusion drafter that speeds up language model inference through speculative decoding. Existing drafters become more expensive as context grows, reducing their efficiency benefit, while LongSpark's decoding cost is made independent of prefix length. It extracts fixed-size, multiscale views from the target model's verification pass, avoiding a growing persistent state. The authors report state-of-the-art end-to-end efficiency across multiple model scales and lowest time-per-output-token on long-context tasks.
In briefBeyond Low-Rank Parameterization: Narrowing the LoRA–Full Fine-Tuning Gap via Gradient Decomposition
SummaryThis paper studies why LoRA, a popular efficient fine-tuning method, can still lag behind full fine-tuning. It shows that LoRA's usable gradient directions at each step match the tangent space of its current low-rank parameterization, which splits the full weight gradient into an orthogonal decomposition. The authors propose GDLoRA, which updates base weights with the orthogonal "normal" gradient component while keeping standard LoRA optimization. Experiments on language understanding, math and commonsense reasoning, and image classification show GDLoRA improves over LoRA and narrows the gap to full fine-tuning.
