DT2IT-MRM: Debiased Preference Construction and Iterative Training for Multimodal Reward Modeling

Multimodal reward models (MRMs) play a crucial role in aligning Multimodal Large Language Models (MLLMs) with human preferences. Training a good MRM requires high-quality multimodal preference data. However, existing preference datasets face three key challenges: lack of granularity in preference strength, textual style bias, and unreliable preference signals. Besides, existing open-source multimodal preference datasets suffer from substantial noise, yet there is a lack of effective and scalable curation methods to enhance their quality. To address these limitations, we propose \textbf{DT2IT-MRM}, which integrates a \textbf{D}ebiased preference construction pipeline, a novel reformulation of text-to-image (\textbf{T2I}) preference data, and an \textbf{I}terative \textbf{T}raining framework that curates existing multimodal preference datasets for \textbf{M}ultimodal \textbf{R}eward \textbf{M}odeling. Our experimental results show that DT2IT-MRM achieves new \textbf{state-of-the-art} overall performance on three major benchmarks: VL-RewardBench, Multimodal RewardBench, and MM-RLHF-RewardBench.

FLASH: Fast Learning via GPU-Accelerated Simulation for High-Fidelity Deformable Manipulation in Minutes

Simulation frameworks such as Isaac Sim have enabled scalable robot learning for locomotion and rigid-body manipulation; however, contact-rich simulation remains a major bottleneck for deformable object manipulation. The continuously changing geometry of soft materials, together with large numbers of vertices and contact constraints, makes it difficult to achieve high accuracy, speed, and stability required for large-scale interactive learning. We present FLASH, a GPU-native simulation framework for contact-rich deformable manipulation, built on an accurate NCP-based solver that enforces strict contact and deformation constraints while being explicitly designed for fine-grained GPU parallelism. Rather than porting conventional single-instruction-multiple-data (SIMD) solvers to GPUs, FLASH redesigns the physics engine from the ground up to leverage modern GPU architectures, including optimized collision handling and memory layouts. As a result, FLASH scales to over 3 million degrees of freedom at 30 FPS on a single RTX 5090, while accurately simulating physical interactions. Policies trained solely on FLASH-generated synthetic data in minutes achieve robust zero-shot sim-to-real transfer, which we validate on physical robots performing challenging deformable manipulation tasks such as towel folding and garment folding, without any real-world demonstration, providing a practical alternative to labor-intensive real-world data collection.

CoDaS: AI Co-Data-Scientist for Biomarker Discovery via Wearable Sensors

Scientific discovery in digital health requires converting continuous physiological signals from wearable devices into clinically actionable biomarkers. We introduce CoDaS (AI Co-Data-Scientist), a multi-agent system that structures biomarker discovery as an iterative process combining hypothesis generation, statistical analysis, adversarial validation, and literature-grounded reasoning with human oversight using large-scale wearable datasets. Across three cohorts totaling 9,279 participant-observations, CoDaS identified 41 candidate digital biomarkers for mental health and 25 for metabolic outcomes, each subjected to an internal validation battery spanning replication, stability, robustness, and discriminative power. Across two independent depression cohorts, CoDaS surfaced circadian instability-related features in both datasets, reflected in sleep duration variability (DWB, ρ= 0.252, p < 0.001) and sleep onset variability (GLOBEM, ρ= 0.126, p < 0.001). In a metabolic cohort, CoDaS derived a cardiovascular fitness index (steps/resting heart rate; ρ= -0.374, p < 0.001), and recovered established clinical associations, including the hepatic function ratio (AST/ALT; ρ= -0.375, p < 0.001), a known correlate of insulin resistance. Incorporating CoDaS-derived features alongside demographic variables led to modest but consistent improvements in predictive performance, with cross-validated ΔR^2 increases of 0.040 for depression and 0.021 for insulin resistance. These findings suggest that CoDaS enables systematic and traceable hypothesis generation and prioritization for biomarker discovery from large-scale wearable data.

MemoSight: Unifying Context Compression and Multi Token Prediction for Reasoning Acceleration

While Chain-of-thought (CoT) reasoning enables LLMs to solve challenging reasoning problems, as KV cache grows linearly with the number of generated tokens, CoT reasoning faces scaling issues in terms of speed and memory usage. In this work, we propose MemoSight (Memory-Foresight-based reasoning), a unified framework that integrates both context compression and multi-token prediction to mitigate the efficiency issues while maintaining CoT reasoning performance. Our framework adopts the same minimalist design for both context compression and multi-token prediction via special tokens and their corresponding position layout tailored to each token type. Comprehensive experiments on four reasoning benchmarks demonstrate that MemoSight reduces the KV cache footprint by up to 66% and accelerates inference by 1.56x, while outperforming existing CoT compression methods.

Scalable Trajectory Generation for Whole-Body Mobile Manipulation

Robots deployed in unstructured environments must coordinate whole-body motion -- simultaneously moving a mobile base and arm -- to interact with the physical world. This coupled mobility and dexterity yields a state space that grows combinatorially with scene and object diversity, demanding datasets far larger than those sufficient for fixed-base manipulation. Yet existing acquisition methods, including teleoperation and planning, are either labor-intensive or computationally prohibitive at scale. The core bottleneck is the lack of a scalable pipeline for generating large-scale, physically valid, coordinated trajectory data across diverse embodiments and environments. Here we introduce AutoMoMa, a GPU-accelerated framework that unifies AKR modeling, which consolidates base, arm, and object kinematics into a single chain, with parallelized trajectory optimization. AutoMoMa achieves 5,000 episodes per GPU-hour (over $80\times$ faster than CPU-based baselines), producing a dataset of over 500k physically valid trajectories spanning 330 scenes, diverse articulated objects, and multiple robot embodiments. Prior datasets were forced to compromise on scale, diversity, or kinematic fidelity; AutoMoMa addresses all three simultaneously. Training downstream IL policies further reveals that even a single articulated-object task requires tens of thousands of demonstrations for SOTA methods to reach $\approx 80\%$ success, confirming that data scarcity -- not algorithmic limitations -- has been the binding constraint. AutoMoMa thus bridges high-performance planning and reliable IL-based control, providing the infrastructure previously missing for coordinated mobile manipulation research. By making large-scale, kinematically valid training data practical, AutoMoMa showcases generalizable whole-body robot policies capable of operating in the diverse, unstructured settings of the real world.

Think Through Uncertainty: Improving Long-Form Generation Factuality via Reasoning Calibration

Large language models (LLMs) often hallucinate in long-form generation. Existing approaches mainly improve factuality through post-hoc revision or reinforcement learning (RL) with correctness-based rewards, but they do not teach the model to estimate which parts of its generation are reliable. As a result, models may still state incorrect claims confidently in their responses. Recent advances in reasoning have significantly improved LLM performance, and have been leveraged to estimate confidence by incorporating calibration into RL objectives. However, existing approaches remain limited to a single scalar confidence for the entire response, which is insufficient for long-form generation where uncertainty varies across individual claims. To mitigate this problem, we propose CURE, a framework that improves long-form factuality by teaching LLMs to reason about uncertainty at the claim level. We first introduce a Claim-Aware Reasoning Protocol, which structures outputs into atomic claims paired with explicit confidence estimates. We then develop a multi-stage training pipeline that aligns model confidence with claims' correctness and then optimizes on factuality. The resulting calibrated confidence further enables selective prediction, allowing the model to abstain from uncertain claims at inference time. Experiments on four long-form factuality benchmarks show that CURE consistently improves factual accuracy over competitive supervised and RL baselines, while maintaining factual recall. In particular, it improves claim-level accuracy by up to 39.9% on Biography generation. These gains are accompanied by improved calibration, as reflected by a 16.0% increase in AUROC on FactBench.

Controllable and Verifiable Tool-Use Data Synthesis for Agentic Reinforcement Learning

Existing synthetic tool-use corpora are primarily designed for offline supervised fine-tuning, yet reinforcement learning (RL) requires executable environments that support reward-checkable online rollouts. We propose COVERT, a two-stage pipeline that first generates reliable base tool-use trajectories through self-evolving synthesis with multi-level validation, and then applies oracle-preserving augmentations that systematically increase environmental complexity. These augmentations introduce distractor tools, indirect or ambiguous user queries, and noisy, multi-format, or erroneous tool outputs, while strictly preserving oracle tool calls and final answers as ground truth. This design enables automatic reward computation via reference matching for standard cases and lightweight judge-assisted verification for special behaviors such as error detection, supporting RL optimization of tool-calling policies. On Qwen2.5-Instruct-14B, COVERT-RL improves overall accuracy on BFCL v3 from 56.5 to 59.9 and on ACEBench from 53.0 to 59.3, with minimal regressions on general-ability benchmarks; when stacked on SFT, it further reaches 62.1 and 61.8, confirming additive gains. These results suggest that oracle-preserving synthetic environments offer a practical RL refinement stage, complementary to SFT, for improving tool-use robustness under ambiguity and unreliable tool feedback.

Insights from Visual Cognition: Understanding Human Action Dynamics with Overall Glance and Refined Gaze Transformer

Recently, Transformer has made significant progress in various vision tasks. To balance computation and efficiency in video tasks, recent works heavily rely on factorized or window-based self-attention. However, these approaches split spatiotemporal correlations between regions of interest in videos, limiting the models' ability to capture motion and long-range dependencies. In this paper, we argue that, similar to the human visual system, the importance of temporal and spatial information varies across different time scales, and attention is allocated sparsely over time through glance and gaze behavior. Is equal consideration of time and space crucial for success in video tasks? Motivated by this understanding, we propose a dual-path network called the Overall Glance and Refined Gaze (OG-ReG) Transformer. The Glance path extracts coarse-grained overall spatiotemporal information, while the Gaze path supplements the Glance path by providing local details. Our model achieves state-of-the-art results on the Kinetics-400, Something-Something v2, and Diving-48, demonstrating its competitive performance. The code will be available at https://github.com/linuxsino/OG-ReG.

Physical Sensitivity Kernels Can Emerge in Data-Driven Forward Models: Evidence From Surface-Wave Dispersion

Data-driven neural networks are increasingly used as surrogate forward models in geophysics, but it remains unclear whether they recover only the data mapping or also the underlying physical sensitivity structure. Here we test this question using surface-wave dispersion. By comparing automatically differentiated gradients from a neural-network surrogate with theoretical sensitivity kernels, we show that the learned gradients can recover the main depth-dependent structure of physical kernels across a broad range of periods. This indicates that neural surrogate models can learn physically meaningful differential information, rather than acting as purely black-box predictors. At the same time, strong structural priors in the training distribution can introduce systematic artifacts into the inferred sensitivities. Our results show that neural forward surrogates can recover useful physical information for inversion and uncertainty analysis, while clarifying the conditions under which this differential structure remains physically consistent.

Training LLMs for Multi-Step Tool Orchestration with Constrained Data Synthesis and Graduated Rewards

Multi-step tool orchestration, where LLMs must invoke multiple dependent APIs in the correct order while propagating intermediate outputs, remains challenging. State-of-the-art models frequently fail on full sequence execution, with parameter value errors accounting for a significant portion of failures. Training models to handle such workflows faces two obstacles: existing environments focus on simple per-turn function calls with simulated data, and binary rewards provide no signal for partial correctness. We present a framework addressing both challenges. First, we construct a reinforcement learning environment backed by a large-scale cache of real API responses, enabling a data synthesis pipeline that samples valid multi-step orchestration traces with controllable complexity and significantly higher generation efficiency than unconstrained methods. Second, we propose a graduated reward design that decomposes correctness into atomic validity (individual function call correctness at increasing granularity) and orchestration (correct tool sequencing with dependency respect). On ComplexFuncBench, our approach demonstrates substantial improvements in turn accuracy. Ablation studies confirm both reward components are essential: using either alone significantly degrades performance.