Haotian Lin

Zelin He, Haotian Lin, Boran Han et al.

Agentic reinforcement learning (RL) enables LLM agents to improve continuously from environment rewards, yet the resulting policies do not systematically accumulate reusable strategies that generalize across tasks. Modular skills can provide such reusable strategies, yet existing skill-augmented RL methods decouple skill creation from policy optimization, risking adopting skills that conflict with the evolving policy. Inspired by Anthropic's Skill Creator, we introduce ReSkill, an RL-in-the-loop skill creation framework that reconciles skill evolution with policy learning. ReSkill exploits the group-wise structure of GRPO to naturally embed three mechanisms with only marginal additional overhead: (1) an assertion-driven skill creator that diagnoses failures from past experience and proposes conditional, trigger-based skill revisions; (2) within-group rollout sampling that enables controlled comparison of skill versions, capturing which version best supports the policy's ongoing learning; and (3) Thompson Sampling with adaptive discounting to balance exploration and exploitation in skill version selection as the policy evolves. Across several domains, ReSkill consistently outperforms existing memory and skill-based RL methods, with the largest gains on unseen tasks. Analysis of the skill lifecycle shows skills being automatically created, tested, refined, and pruned as the policy improves, demonstrating reconciled skill-policy co-evolution.

Haotian Lin, Matthew Reimherr

We study the transfer learning (TL) for the functional linear regression (FLR) under the Reproducing Kernel Hilbert Space (RKHS) framework, observing that the TL techniques in existing high-dimensional linear regression are not compatible with the truncation-based FLR methods, as functional data are intrinsically infinite-dimensional and generated by smooth underlying processes. We measure the similarity across tasks using RKHS distance, allowing the type of information being transferred to be tied to the properties of the imposed RKHS. Building on the hypothesis offset transfer learning paradigm, two algorithms are proposed: one conducts the transfer when positive sources are known, while the other leverages aggregation techniques to achieve robust transfer without prior information about the sources. We establish asymptotic lower bounds for this learning problem and show that the proposed algorithms enjoy a matching upper bound. These analyses provide statistical insights into factors that contribute to the dynamics of the transfer. We also extend the results to functional generalized linear models. The effectiveness of the proposed algorithms is demonstrated via extensive synthetic data as well as real-world data applications.

Haotian Lin, Matthew Reimherr

Many existing mechanisms to achieve differential privacy (DP) on infinite-dimensional functional summaries often involve embedding these summaries into finite-dimensional subspaces and applying traditional DP techniques. Such mechanisms generally treat each dimension uniformly and struggle with complex, structured summaries. This work introduces a novel mechanism for DP functional summary release: the Independent Component Laplace Process (ICLP) mechanism. This mechanism treats the summaries of interest as truly infinite-dimensional objects, thereby addressing several limitations of existing mechanisms. We establish the feasibility of the proposed mechanism in multiple function spaces. Several statistical estimation problems are considered, and we demonstrate one can enhance the utility of sanitized summaries by oversmoothing their non-private counterpart. Numerical experiments on synthetic and real datasets demonstrate the efficacy of the proposed mechanism.

Wenli Xiao, Haotian Lin, Andy Peng et al.

Supervised fine-tuning (SFT) has become the de facto post-training strategy for large vision-language-action (VLA) models, but its reliance on costly human demonstrations limits scalability and generalization. We propose Probe, Learn, Distill (PLD), a three-stage plug-and-play framework that improves VLAs through residual reinforcement learning (RL) and distribution-aware data collection. In Stage 1, we train lightweight residual actors to probe failure regions of the VLA generalist. In Stage 2, we use a hybrid rollout scheme that aligns collected trajectories with the generalist's deployment distribution while capturing recovery behaviors. In Stage 3, we distill the curated trajectories back into the generalist with standard SFT. PLD achieves near-saturated 99% task success on LIBERO, over 50% gains in SimplerEnv, and 100% success on real-world Franka and YAM arm manipulation tasks. Ablations show that residual probing and distribution-aware replay are key to collecting deployment-aligned data that improves both seen and unseen tasks, offering a scalable path toward self-improving VLA models.

Zelin He, Boran Han, Xiyuan Zhang et al.

Time-series diagnostic reasoning is essential for many applications, yet existing solutions face a persistent gap: general reasoning large language models (GRLMs) possess strong reasoning skills but lack the domain-specific knowledge to understand complex time-series patterns. Conversely, fine-tuned time-series LLMs (TSLMs) understand these patterns but lack the capacity to generalize reasoning for more complicated questions. To bridge this gap, we propose a hybrid knowledge-injection framework that injects TSLM-generated insights directly into GRLM's reasoning trace, thereby achieving strong time-series reasoning with in-domain knowledge. As collecting data for knowledge injection fine-tuning is costly, we further leverage a reinforcement learning-based approach with verifiable rewards (RLVR) to elicit knowledge-rich traces without human supervision, then transfer such an in-domain thinking trace into GRLM for efficient knowledge injection. We further release SenTSR-Bench, a multivariate time-series-based diagnostic reasoning benchmark collected from real-world industrial operations. Across SenTSR-Bench and other public datasets, our method consistently surpasses TSLMs by 9.1%-26.1% and GRLMs by 7.9%-22.4%, delivering robust, context-aware time-series diagnostic insights.

Zhe Liu, Yuhao Huang, Lian Liu et al.

Color Doppler echocardiography is a crucial tool for diagnosing mitral regurgitation (MR). Recent studies have explored intelligent methods for MR diagnosis to minimize user dependence and improve accuracy. However, these approaches often fail to align with clinical workflow and may lead to suboptimal accuracy and interpretability. In this study, we introduce an automated MR diagnosis model (MReg) developed on the 4-chamber cardiac color Doppler echocardiography video (A4C-CDV). It follows comprehensive feature mining strategies to detect MR and assess its severity, considering clinical realities. Our contribution is threefold. First, we formulate the MR diagnosis as a regression task to capture the continuity and ordinal relationships between categories. Second, we design a feature selection and amplification mechanism to imitate the sonographer's diagnostic logic for accurate MR grading. Third, inspired by the Mixture-of-Experts concept, we introduce a feature summary module to extract the category-level features, enhancing the representational capacity for more accurate grading. We trained and evaluated our proposed MReg on a large in-house A4C-CDV dataset comprising 1868 cases with three graded regurgitation labels. Compared to other weakly supervised video anomaly detection and supervised classification methods, MReg demonstrated superior performance in MR diagnosis. Our code is available at: https://github.com/cskdstz/MReg.

Mingxiao Huo, Pengliang Ji, Haotian Lin et al.

We introduce a pioneering unified library that leverages depth anything, segment anything models to augment neural comprehension in language-vision model zero-shot understanding. This library synergizes the capabilities of the Depth Anything Model (DAM), Segment Anything Model (SAM), and GPT-4V, enhancing multimodal tasks such as vision-question-answering (VQA) and composition reasoning. Through the fusion of segmentation and depth analysis at the symbolic instance level, our library provides nuanced inputs for language models, significantly advancing image interpretation. Validated across a spectrum of in-the-wild real-world images, our findings showcase progress in vision-language models through neural-symbolic integration. This novel approach melds visual and language analysis in an unprecedented manner. Overall, our library opens new directions for future research aimed at decoding the complexities of the real world through advanced multimodal technologies and our code is available at \url{https://github.com/AnthonyHuo/SAM-DAM-for-Compositional-Reasoning}.

Xiaoning Liu, Zongwei Wu, Ao Li et al.

This paper reviews the NTIRE 2024 low light image enhancement challenge, highlighting the proposed solutions and results. The aim of this challenge is to discover an effective network design or solution capable of generating brighter, clearer, and visually appealing results when dealing with a variety of conditions, including ultra-high resolution (4K and beyond), non-uniform illumination, backlighting, extreme darkness, and night scenes. A notable total of 428 participants registered for the challenge, with 22 teams ultimately making valid submissions. This paper meticulously evaluates the state-of-the-art advancements in enhancing low-light images, reflecting the significant progress and creativity in this field.

Haotian Lin, Matthew Reimherr

Many existing two-phase kernel-based hypothesis transfer learning algorithms employ the same kernel regularization across phases and rely on the known smoothness of functions to obtain optimality. Therefore, they fail to adapt to the varying and unknown smoothness between the target/source and their offset in practice. In this paper, we address these problems by proposing Smoothness Adaptive Transfer Learning (SATL), a two-phase kernel ridge regression(KRR)-based algorithm. We first prove that employing the misspecified fixed bandwidth Gaussian kernel in target-only KRR learning can achieve minimax optimality and derive an adaptive procedure to the unknown Sobolev smoothness. Leveraging these results, SATL employs Gaussian kernels in both phases so that the estimators can adapt to the unknown smoothness of the target/source and their offset function. We derive the minimax lower bound of the learning problem in excess risk and show that SATL enjoys a matching upper bound up to a logarithmic factor. The minimax convergence rate sheds light on the factors influencing transfer dynamics and demonstrates the superiority of SATL compared to non-transfer learning settings. While our main objective is a theoretical analysis, we also conduct several experiments to confirm our results.

Zichun Guo, Yuling Shi, Wenhao Zeng et al.

Multimodal Large Language Models (MLLMs) have achieved remarkable performance in Visually Rich Document Understanding (VRDU) tasks, but their capabilities are mainly evaluated on pristine, well-structured document images. We consider content restoration from shredded fragments, a challenging VRDU setting that requires integrating visual pattern recognition with semantic reasoning under significant content discontinuities. To facilitate systematic evaluation of complex VRDU tasks, we introduce ShredBench, a benchmark supported by an automated generation pipeline that renders fragmented documents directly from Markdown. The proposed pipeline ensures evaluation validity by allowing the flexible integration of latest or unseen textual sources to prevent training data contamination. ShredBench assesses four scenarios (English, Chinese, Code, Table) with three fragmentation granularities (8, 12, 16 pieces). Empirical evaluations on state-of-the-art MLLMs reveal a significant performance gap: The method is effective on intact documents; however, once the document is shredded, restoration becomes a significant challenge, with NED dropping sharply as fragmentation increases. Our findings highlight that current MLLMs lack the fine-grained cross-modal reasoning required to bridge visual discontinuities, identifying a critical gap in robust VRDU research.

Haotian Lin, Pengcheng Wang, Jeff Schneider et al.

Model-based reinforcement learning algorithms that combine model-based planning and learned value/policy prior have gained significant recognition for their high data efficiency and superior performance in continuous control. However, we discover that existing methods that rely on standard SAC-style policy iteration for value learning, directly using data generated by the planner, often result in \emph{persistent value overestimation}. Through theoretical analysis and experiments, we argue that this issue is deeply rooted in the structural policy mismatch between the data generation policy that is always bootstrapped by the planner and the learned policy prior. To mitigate such a mismatch in a minimalist way, we propose a policy regularization term reducing out-of-distribution (OOD) queries, thereby improving value learning. Our method involves minimum changes on top of existing frameworks and requires no additional computation. Extensive experiments demonstrate that the proposed approach improves performance over baselines such as TD-MPC2 by large margins, particularly in 61-DoF humanoid tasks. View qualitative results at https://darthutopian.github.io/tdmpc_square/.

Yuning Wang, Zhiyuan Liu, Haotian Lin et al.

Scene understanding, defined as learning, extraction, and representation of interactions among traffic elements, is one of the critical challenges toward high-level autonomous driving (AD). Current scene understanding methods mainly focus on one concrete single task, such as trajectory prediction and risk level evaluation. Although they perform well on specific metrics, the generalization ability is insufficient to adapt to the real traffic complexity and downstream demand diversity. In this study, we propose PreGSU, a generalized pre-trained scene understanding model based on graph attention network to learn the universal interaction and reasoning of traffic scenes to support various downstream tasks. After the feature engineering and sub-graph module, all elements are embedded as nodes to form a dynamic weighted graph. Then, four graph attention layers are applied to learn the relationships among agents and lanes. In the pre-train phase, the understanding model is trained on two self-supervised tasks: Virtual Interaction Force (VIF) modeling and Masked Road Modeling (MRM). Based on the artificial potential field theory, VIF modeling enables PreGSU to capture the agent-to-agent interactions while MRM extracts agent-to-road connections. In the fine-tuning process, the pre-trained parameters are loaded to derive detailed understanding outputs. We conduct validation experiments on three datasets and two downstream tasks, i.e., trajectory prediction in urban scenario and intention recognition in highway scenario, to verify the model's generalization and understanding capabilities. Results show that compared with single-task-driven baselines, PreGSU achieves competitive performance on all datasets and downstream tasks, indicating its potential to be generalized to various scenes and targets. Ablation study shows the effectiveness of pre-train task design.

Yixiao Wang, Mingxiao Huo, Zhixuan Liang et al. · berkeley

Pretrained vision foundation models (VFMs) advance robotic learning via rich visual representations, yet individual VFMs typically excel only in specific domains, limiting generality across tasks. Distilling multiple VFMs into a unified representation for policy can mitigate this limitation but often yields inflexible task-specific feature selection and requires costly full re-training to incorporate robot-domain knowledge. We propose VER, a Vision Expert transformer for Robot learning. During pretraining, VER distills multiple VFMs into a vision expert library. It then fine-tunes only a lightweight routing network (fewer than 0.4% of parameters) to dynamically select task-relevant experts from the pretrained library for downstream robot tasks. We further introduce Patchwise Expert Routing with Curriculum Top-K Annealing to improve both flexibility and precision of dynamic expert selection. Moreover, VER supports parameter-efficient finetuning for scalable expert utilization and adaptive robot-domain knowledge integration. Across 17 diverse robotic tasks and multiple policy heads, VER achieves state-of-the-art performance. We find that VER reduces large-norm outliers in task-irrelevant regions (e.g., background) and concentrates on task-critical regions. Visualizations and codes can be found in https://yixiaowang7.github.io/ver_page/.

Pengcheng Wang, Qinghang Liu, Haotian Lin et al.

Learning domain adaptive policies that can generalize to unseen transition dynamics, remains a fundamental challenge in learning-based control. Substantial progress has been made through domain representation learning to capture domain-specific information, thus enabling domain-aware decision making. We analyze the process of learning domain representations through dynamical prediction and find that selecting contexts adjacent to the current step causes the learned representations to entangle static domain information with varying dynamical properties. Such mixture can confuse the conditioned policy, thereby constraining zero-shot adaptation. To tackle the challenge, we propose DADP (Domain Adaptive Diffusion Policy), which achieves robust adaptation through unsupervised disentanglement and domain-aware diffusion injection. First, we introduce Lagged Context Dynamical Prediction, a strategy that conditions future state estimation on a historical offset context; by increasing this temporal gap, we unsupervisedly disentangle static domain representations by filtering out transient properties. Second, we integrate the learned domain representations directly into the generative process by biasing the prior distribution and reformulating the diffusion target. Extensive experiments on challenging benchmarks across locomotion and manipulation demonstrate the superior performance, and the generalizability of DADP over prior methods. More visualization results are available on the https://outsider86.github.io/DomainAdaptiveDiffusionPolicy/.

Haotian Lin, Matthew Reimherr

When concept shifts and sample scarcity are present in the target domain of interest, nonparametric regression learners often struggle to generalize effectively. The technique of transfer learning remedies these issues by leveraging data or pre-trained models from similar source domains. While existing generalization analyses of kernel-based transfer learning typically rely on correctly specified models, we present a transfer learning procedure that is robust against model misspecification while adaptively attaining optimality. To facilitate our analysis and avoid the risk of saturation found in classical misspecified results, we establish a novel result in the misspecified single-task learning setting, showing that spectral algorithms with fixed bandwidth Gaussian kernels can attain minimax convergence rates given the true function is in a Sobolev space, which may be of independent interest. Building on this, we derive the adaptive convergence rates of the excess risk for specifying Gaussian kernels in a prevalent class of hypothesis transfer learning algorithms. Our results are minimax optimal up to logarithmic factors and elucidate the key determinants of transfer efficiency.