Imitation learning (IL) has proven effective for enabling robots to acquire visuomotor skills through expert demonstrations. However, traditional IL methods are limited by their reliance on high-quality, often scarce, expert data, and suffer from covariate shift. To address these challenges, recent advances in offline IL have incorporated suboptimal, unlabeled datasets into the training. In this paper, we propose a novel approach to enhance policy learning from mixed-quality offline datasets by leveraging task-relevant trajectory fragments and rich environmental dynamics. Specifically, we introduce a state-based search framework that stitches state-action pairs from imperfect demonstrations, generating more diverse and informative training trajectories. Experimental results on standard IL benchmarks and real-world robotic tasks showcase that our proposed method significantly improves both generalization and performance.

The success of OpenAI's ChatGPT in 2023 has spurred financial enterprises into exploring Generative AI applications to reduce costs or drive revenue within different lines of businesses in the Financial Industry. While these applications offer strong potential for efficiencies, they introduce new model risks, primarily hallucinations and toxicity. As highly regulated entities, financial enterprises (primarily large US banks) are obligated to enhance their model risk framework with additional testing and controls to ensure safe deployment of such applications. This paper outlines the key aspects for model risk management of generative AI model with a special emphasis on additional practices required in model validation.

Text-to-video (T2V) generation has made significant strides with diffusion models. However, existing methods still struggle with accurately binding attributes, determining spatial relationships, and capturing complex action interactions between multiple subjects. To address these limitations, we propose MagicComp, a training-free method that enhances compositional T2V generation through dual-phase refinement. Specifically, (1) During the Conditioning Stage: We introduce the Semantic Anchor Disambiguation to reinforces subject-specific semantics and resolve inter-subject ambiguity by progressively injecting the directional vectors of semantic anchors into original text embedding; (2) During the Denoising Stage: We propose Dynamic Layout Fusion Attention, which integrates grounding priors and model-adaptive spatial perception to flexibly bind subjects to their spatiotemporal regions through masked attention modulation. Furthermore, MagicComp is a model-agnostic and versatile approach, which can be seamlessly integrated into existing T2V architectures. Extensive experiments on T2V-CompBench and VBench demonstrate that MagicComp outperforms state-of-the-art methods, highlighting its potential for applications such as complex prompt-based and trajectory-controllable video generation. Project page: https://hong-yu-zhang.github.io/MagicComp-Page/.

Existing Large Reasoning Models (LRMs) have shown the potential of reinforcement learning (RL) to enhance the complex reasoning capabilities of Large Language Models~(LLMs). While they achieve remarkable performance on challenging tasks such as mathematics and coding, they often rely on their internal knowledge to solve problems, which can be inadequate for time-sensitive or knowledge-intensive questions, leading to inaccuracies and hallucinations. To address this, we propose \textbf{R1-Searcher}, a novel two-stage outcome-based RL approach designed to enhance the search capabilities of LLMs. This method allows LLMs to autonomously invoke external search systems to access additional knowledge during the reasoning process. Our framework relies exclusively on RL, without requiring process rewards or distillation for a cold start. % effectively generalizing to out-of-domain datasets and supporting both Base and Instruct models. Our experiments demonstrate that our method significantly outperforms previous strong RAG methods, even when compared to the closed-source GPT-4o-mini.

Malicious users attempt to replicate commercial models functionally at low cost by training a clone model with query responses. It is challenging to timely prevent such model-stealing attacks to achieve strong protection and maintain utility. In this paper, we propose a novel non-parametric detector called Account-aware Distribution Discrepancy (ADD) to recognize queries from malicious users by leveraging account-wise local dependency. We formulate each class as a Multivariate Normal distribution (MVN) in the feature space and measure the malicious score as the sum of weighted class-wise distribution discrepancy. The ADD detector is combined with random-based prediction poisoning to yield a plug-and-play defense module named D-ADD for image classification models. Results of extensive experimental studies show that D-ADD achieves strong defense against different types of attacks with little interference in serving benign users for both soft and hard-label settings.

In this report, we present the third technical report on the development of slow-thinking models as part of the STILL project. As the technical pathway becomes clearer, scaling RL training has become a central technique for implementing such reasoning models. We systematically experiment with and document the effects of various factors influencing RL training, conducting experiments on both base models and fine-tuned models. Specifically, we demonstrate that our RL training approach consistently improves the Qwen2.5-32B base models, enhancing both response length and test accuracy. Furthermore, we show that even when a model like DeepSeek-R1-Distill-Qwen-1.5B has already achieved a high performance level, it can be further refined through RL training, reaching an accuracy of 39.33% on AIME 2024. Beyond RL training, we also explore the use of tool manipulation, finding that it significantly boosts the reasoning performance of large reasoning models. This approach achieves a remarkable accuracy of 86.67% with greedy search on AIME 2024, underscoring its effectiveness in enhancing model capabilities. We release our resources at the STILL project website: https://github.com/RUCAIBox/Slow_Thinking_with_LLMs.

Most LLM-driven conversational AI systems operate reactively, responding to user prompts without guiding the interaction. Most LLM-driven conversational AI systems operate reactively, responding to user prompts without guiding the interaction. However, many real-world applications-such as psychiatric diagnosis, consulting, and interviews-require AI to take a proactive role, asking the right questions and steering conversations toward specific objectives. Using mental health differential diagnosis as an application context, we introduce ProAI, a goal-oriented, proactive conversational AI framework. ProAI integrates structured knowledge-guided memory, multi-agent proactive reasoning, and a multi-faceted evaluation strategy, enabling LLMs to engage in clinician-style diagnostic reasoning rather than simple response generation. Through simulated patient interactions, user experience assessment, and professional clinical validation, we demonstrate that ProAI achieves up to 83.3% accuracy in mental disorder differential diagnosis while maintaining professional and empathetic interaction standards. These results highlight the potential for more reliable, adaptive, and goal-driven AI diagnostic assistants, advancing LLMs beyond reactive dialogue systems.

For partially observable cooperative tasks, multi-agent systems must develop effective communication and understand the interplay among agents in order to achieve cooperative goals. However, existing multi-agent reinforcement learning (MARL) with communication methods lack evaluation metrics for information weights and information-level communication modeling. This causes agents to neglect the aggregation of multiple messages, thereby significantly reducing policy learning efficiency. In this paper, we propose pluggable adaptive generative networks (PAGNet), a novel framework that integrates generative models into MARL to enhance communication and decision-making. PAGNet enables agents to synthesize global states representations from weighted local observations and use these representations alongside learned communication weights for coordinated decision-making. This pluggable approach reduces the computational demands typically associated with the joint training of communication and policy networks. Extensive experimental evaluations across diverse benchmarks and communication scenarios demonstrate the significant performance improvements achieved by PAGNet. Furthermore, we analyze the emergent communication patterns and the quality of generated global states, providing insights into operational mechanisms.

The novel neural networks show great potential in solving partial differential equations. For single-phase flow problems in subsurface porous media with high-contrast coefficients, the key is to develop neural operators with accurate reconstruction capability and strict adherence to physical laws. In this study, we proposed a hybrid two-stage framework that uses multiscale basis functions and physics-guided deep learning to solve the Darcy flow problem in high-contrast fractured porous media. In the first stage, a data-driven model is used to reconstruct the multiscale basis function based on the permeability field to achieve effective dimensionality reduction while preserving the necessary multiscale features. In the second stage, the physics-informed neural network, together with Transformer-based global information extractor is used to reconstruct the pressure field by integrating the physical constraints derived from the Darcy equation, ensuring consistency with the physical laws of the real world. The model was evaluated on datasets with different combinations of permeability and basis functions and performed well in terms of reconstruction accuracy. Specifically, the framework achieves R2 values above 0.9 in terms of basis function fitting and pressure reconstruction, and the residual indicator is on the order of $1\times 10^{-4}$. These results validate the ability of the proposed framework to achieve accurate reconstruction while maintaining physical consistency.

Recently, test-time scaling has garnered significant attention from the research community, largely due to the substantial advancements of the o1 model released by OpenAI. By allocating more computational resources during the inference phase, large language models (LLMs) can extensively explore the solution space by generating more thought tokens or diverse solutions, thereby producing more accurate responses. However, develop an o1-like reasoning approach is challenging, and researchers have been making various attempts to advance this open area of research. In this paper, we present a preliminary exploration into enhancing the reasoning abilities of LLMs through reward-guided tree search algorithms. This framework is implemented by integrating the policy model, reward model, and search algorithm. It is primarily constructed around a tree search algorithm, where the policy model navigates a dynamically expanding tree guided by a specially trained reward model. The implemented framework is denoted as STILL-1 (Slow Thinking with LLMs), marking the first model developed by our project, "Slow Thinking with LLMs". We thoroughly explore various design considerations necessary for implementing this framework and provide a detailed report of the technical aspects. To assess the effectiveness of our approach, we focus on mathematical reasoning tasks and conduct extensive evaluations on four challenging datasets, significantly enhancing the reasoning abilities of LLMs.