Recent advancements in robot learning have used imitation learning with large models and extensive demonstrations to develop effective policies. However, these models are often limited by the quantity, quality, and diversity of demonstrations. This paper explores improving offline-trained imitation learning models through online interactions with the environment. We introduce Policy Decorator, which uses a model-agnostic residual policy to refine large imitation learning models during online interactions. By implementing controlled exploration strategies, Policy Decorator enables stable, sample-efficient online learning. Our evaluation spans eight tasks across two benchmarks-ManiSkill and Adroit-and involves two state-of-the-art imitation learning models (Behavior Transformer and Diffusion Policy). The results show Policy Decorator effectively improves the offline-trained policies and preserves the smooth motion of imitation learning models, avoiding the erratic behaviors of pure RL policies. See our project page (https://policydecorator.github.io) for videos.

Scientific idea generation has been extensively studied in creativity theory and computational creativity research, providing valuable frameworks for understanding and implementing creative processes. However, recent work using Large Language Models (LLMs) for research idea generation often overlooks these theoretical foundations. We present a framework that explicitly implements combinatorial creativity theory using LLMs, featuring a generalization-level retrieval system for cross-domain knowledge discovery and a structured combinatorial process for idea generation. The retrieval system maps concepts across different abstraction levels to enable meaningful connections between disparate domains, while the combinatorial process systematically analyzes and recombines components to generate novel solutions. Experiments on the OAG-Bench dataset demonstrate our framework's effectiveness, consistently outperforming baseline approaches in generating ideas that align with real research developments (improving similarity scores by 7\%-10\% across multiple metrics). Our results provide strong evidence that LLMs can effectively realize combinatorial creativity when guided by appropriate theoretical frameworks, contributing both to practical advancement of AI-assisted research and theoretical understanding of machine creativity.

The analysis of dynamical systems is a fundamental tool in the natural sciences and engineering. It is used to understand the evolution of systems as large as entire galaxies and as small as individual molecules. With predefined conditions on the evolution of dy-namical systems, the underlying differential equations have to fulfill specific constraints in time and space. This class of problems is known as boundary value problems. This thesis presents novel approaches to learn time-reversible deterministic and stochastic dynamics constrained by initial and final conditions. The dynamics are inferred by machine learning algorithms from observed data, which is in contrast to the traditional approach of solving differential equations by numerical integration. The work in this thesis examines a set of problems of increasing difficulty each of which is concerned with learning a different aspect of the dynamics. Initially, we consider learning deterministic dynamics from ground truth solutions which are constrained by deterministic boundary conditions. Secondly, we study a boundary value problem in discrete state spaces, where the forward dynamics follow a stochastic jump process and the boundary conditions are discrete probability distributions. In particular, the stochastic dynamics of a specific jump process, the Ehrenfest process, is considered and the reverse time dynamics are inferred with machine learning. Finally, we investigate the problem of inferring the dynamics of a continuous-time stochastic process between two probability distributions without any reference information. Here, we propose a novel criterion to learn time-reversible dynamics of two stochastic processes to solve the Schr\"odinger Bridge Problem.

With the growth of social media and large language models, content moderation has become crucial. Many existing datasets lack adequate representation of different groups, resulting in unreliable assessments. To tackle this, we propose a socio-culturally aware evaluation framework for LLM-driven content moderation and introduce a scalable method for creating diverse datasets using persona-based generation. Our analysis reveals that these datasets provide broader perspectives and pose greater challenges for LLMs than diversity-focused generation methods without personas. This challenge is especially pronounced in smaller LLMs, emphasizing the difficulties they encounter in moderating such diverse content.

Ordering has been extensively studied in many visualization applications, such as axis and matrix reordering, for the simple reason that the order will greatly impact the perceived pattern of data. Many quality metrics concerning data pattern, perception, and aesthetics are proposed, and respective optimization algorithms are developed. However, the optimization problems related to ordering are often difficult to solve (e.g., TSP is NP-complete), and developing specialized optimization algorithms is costly. In this paper, we propose Versatile Ordering Network (VON), which automatically learns the strategy to order given a quality metric. VON uses the quality metric to evaluate its solutions, and leverages reinforcement learning with a greedy rollout baseline to improve itself. This keeps the metric transparent and allows VON to optimize over different metrics. Additionally, VON uses the attention mechanism to collect information across scales and reposition the data points with respect to the current context. This allows VONs to deal with data points following different distributions. We examine the effectiveness of VON under different usage scenarios and metrics. The results demonstrate that VON can produce comparable results to specialized solvers. The code is available at https://github.com/sysuvis/VON.

Chat models (i.e. language models trained to follow instructions through conversation with humans) outperform base models (i.e. trained solely on unlabeled data) in both conversation and general task-solving abilities. These models are generally English-centric and require further adaptation for languages that are underrepresented in or absent from their training data. A common technique for adapting base models is to extend the model's vocabulary with target language tokens, i.e. vocabulary expansion (VE), and then continually pre-train it on language-specific data. Using chat data is ideal for chat model adaptation, but often, either this does not exist or is costly to construct. Alternatively, adapting chat models with unlabeled data is a possible solution, but it could result in catastrophic forgetting. In this paper, we investigate the impact of using unlabeled target language data for VE on chat models for the first time. We first show that off-the-shelf VE generally performs well across target language tasks and models in 71% of cases, though it underperforms in scenarios where source chat models are already strong. To further improve adapted models, we propose post-hoc techniques that inject information from the source model without requiring any further training. Experiments reveal the effectiveness of our methods, helping the adapted models to achieve performance improvements in 87% of cases.

We study the problem of aligning large language models (LLMs) with human preference data. Contrastive preference optimization has shown promising results in aligning LLMs with available preference data by optimizing the implicit reward associated with the policy. However, the contrastive objective focuses mainly on the relative values of implicit rewards associated with two responses while ignoring their actual values, resulting in suboptimal alignment with human preferences. To address this limitation, we propose calibrated direct preference optimization (Cal-DPO), a simple yet effective algorithm. We show that substantial improvement in alignment with the given preferences can be achieved simply by calibrating the implicit reward to ensure that the learned implicit rewards are comparable in scale to the ground-truth rewards. We demonstrate the theoretical advantages of Cal-DPO over existing approaches. The results of our experiments on a variety of standard benchmarks show that Cal-DPO remarkably improves off-the-shelf methods.

The emergence of Retrieval-augmented generation (RAG) has alleviated the issues of outdated and hallucinatory content in the generation of large language models (LLMs), yet it still reveals numerous limitations. When a general-purpose LLM serves as the RAG generator, it often suffers from inadequate response informativeness, response robustness, and citation quality. Past approaches to tackle these limitations, either by incorporating additional steps beyond generating responses or optimizing the generator through supervised fine-tuning (SFT), still failed to align with the RAG requirement thoroughly. Consequently, optimizing the RAG generator from multiple preference perspectives while maintaining its end-to-end LLM form remains a challenge. To bridge this gap, we propose Multiple Perspective Preference Alignment for Retrieval-Augmented Generation (PA-RAG), a method for optimizing the generator of RAG systems to align with RAG requirements comprehensively. Specifically, we construct high-quality instruction fine-tuning data and multi-perspective preference data by sampling varied quality responses from the generator across different prompt documents quality scenarios. Subsequently, we optimize the generator using SFT and Direct Preference Optimization (DPO). Extensive experiments conducted on four question-answer datasets across three LLMs demonstrate that PA-RAG can significantly enhance the performance of RAG generators. Our code and datasets are available at https://github.com/wujwyi/PA-RAG.

Learning policies from high-dimensional visual inputs, such as pixels and point clouds, is crucial in various applications. Visual reinforcement learning is a promising approach that directly trains policies from visual observations, although it faces challenges in sample efficiency and computational costs. This study conducts an empirical comparison of State-to-Visual DAgger, a two-stage framework that initially trains a state policy before adopting online imitation to learn a visual policy, and Visual RL across a diverse set of tasks. We evaluate both methods across 16 tasks from three benchmarks, focusing on their asymptotic performance, sample efficiency, and computational costs. Surprisingly, our findings reveal that State-to-Visual DAgger does not universally outperform Visual RL but shows significant advantages in challenging tasks, offering more consistent performance. In contrast, its benefits in sample efficiency are less pronounced, although it often reduces the overall wall-clock time required for training. Based on our findings, we provide recommendations for practitioners and hope that our results contribute valuable perspectives for future research in visual policy learning.

Continual learning in deep neural networks often suffers from catastrophic forgetting, where representations for previous tasks are overwritten during subsequent training. We propose a novel sample retrieval strategy from the memory buffer that leverages both gradient-conflicting and gradient-aligned samples to effectively retain knowledge about past tasks within a supervised contrastive learning framework. Gradient-conflicting samples are selected for their potential to reduce interference by re-aligning gradients, thereby preserving past task knowledge. Meanwhile, gradient-aligned samples are incorporated to reinforce stable, shared representations across tasks. By balancing gradient correction from conflicting samples with alignment reinforcement from aligned ones, our approach increases the diversity among retrieved instances and achieves superior alignment in parameter space, significantly enhancing knowledge retention and mitigating proxy drift. Empirical results demonstrate that using both sample types outperforms methods relying solely on one sample type or random retrieval. Experiments on popular continual learning benchmarks in computer vision validate our method's state-of-the-art performance in mitigating forgetting while maintaining competitive accuracy on new tasks.