Recent advances in language model (LM) agents and function calling have enabled autonomous, feedback-driven systems to solve problems across various digital domains. To better understand the unique limitations of LM agents, we introduce RefactorBench, a benchmark consisting of 100 large handcrafted multi-file refactoring tasks in popular open-source repositories. Solving tasks within RefactorBench requires thorough exploration of dependencies across multiple files and strong adherence to relevant instructions. Every task is defined by 3 natural language instructions of varying specificity and is mutually exclusive, allowing for the creation of longer combined tasks on the same repository. Baselines on RefactorBench reveal that current LM agents struggle with simple compositional tasks, solving only 22% of tasks with base instructions, in contrast to a human developer with short time constraints solving 87%. Through trajectory analysis, we identify various unique failure modes of LM agents, and further explore the failure mode of tracking past actions. By adapting a baseline agent to condition on representations of state, we achieve a 43.9% improvement in solving RefactorBench tasks. We further extend our state-aware approach to encompass entire digital environments and outline potential directions for future research. RefactorBench aims to support the study of LM agents by providing a set of real-world, multi-hop tasks within the realm of code. "Repetition is the root of all software evil" -- Martin Fowler Large language models (LLMs) have been quickly acquiring new capabilities (Bubeck et al., 2023), leading towards adoption of AI-powered systems in various formats and domains. The increasing usage of LLM powered tools like Github Copilot have greatly improved the capability of developers in software development tasks (Peng et al., 2023). More recently, an emphasis on multi-step execution through LLM feedback loops has unlocked the ability to solve harder problems within a variety of fields (Reed et al., 2022; Sumers et al., 2024; Yao & Narasimhan, 2023), including parts of software engineering. This new paradigm of solving larger software tasks has led to the construction of a variety of new automated software engineering (ASE) systems, most being structured as LM agents (Wang et al., 2024c; Cognition.ai, Evaluations for such systems are currently largely comprised from real world data on Github (Jimenez et al., 2024; LaBash et al., 2024). While being the strongest open-source signal for software engineering tasks at scale, Github is inherently noisy through its snapshot nature, also requiring strong filtration and validation testing for reliable evaluations (Chowdhury et al., 2024; Bowman & Dahl, 2021).

Recently, with the development of Multi-agent reinforcement learning (MARL), adaptive traffic signal control (ATSC) has achieved satisfactory results. In traffic scenarios with multiple intersections, MARL treats each intersection as an agent and optimizes traffic signal control strategies through learning and real-time decision-making. Considering that observation distributions of intersections might be different in real-world scenarios, shared parameter methods might lack diversity and thus lead to high generalization requirements in the shared-policy network. A typical solution is to increase the size of network parameters. However, simply increasing the scale of the network does not necessarily improve policy generalization, which is validated in our experiments. Accordingly, an approach that considers both the personalization of intersections and the efficiency of parameter sharing is required. To this end, we propose Hyper-Action Multi-Head Proximal Policy Optimization (HAMH-PPO), a Centralized Training with Decentralized Execution (CTDE) MARL method that utilizes a shared PPO policy network to deliver personalized policies for intersections with non-iid observation distributions. The centralized critic in HAMH-PPO uses graph attention units to calculate the graph representations of all intersections and outputs a set of value estimates with multiple output heads for each intersection. The decentralized execution actor takes the local observation history as input and output distributions of action as well as a so-called hyper-action to balance the multiple values estimated from the centralized critic to further guide the updating of TSC policies. The combination of hyper-action and multi-head values enables multiple agents to share a single actor-critic while achieving personalized policies.

The key to robot-assisted rehabilitation lies in the design of the human-machine interface, which must accommodate the needs of both patients and machines. Current interface designs primarily focus on machine control algorithms, often requiring patients to spend considerable time adapting. In this paper, we introduce a novel approach based on the Cooperative Adaptive Markov Decision Process (CAMDPs) model to address the fundamental aspects of the interactive learning process, offering theoretical insights and practical guidance. We establish sufficient conditions for the convergence of CAMDPs and ensure the uniqueness of Nash equilibrium points. Leveraging these conditions, we guarantee the system's convergence to a unique Nash equilibrium point. Furthermore, we explore scenarios with multiple Nash equilibrium points, devising strategies to adjust both Value Evaluation and Policy Improvement algorithms to enhance the likelihood of converging to the global minimal Nash equilibrium point. Through numerical experiments, we illustrate the effectiveness of the proposed conditions and algorithms, demonstrating their applicability and robustness in practical settings. The proposed conditions for convergence and the identification of a unique optimal Nash equilibrium contribute to the development of more effective adaptive systems for human users in robot-assisted rehabilitation.

Humanoid robots, capable of assuming human roles in various workplaces, have become essential to the advancement of embodied intelligence. However, as robots with complex physical structures, learning a control model that can operate robustly across diverse environments remains inherently challenging, particularly under the discrepancies between training and deployment environments. In this study, we propose HWC-Loco, a robust whole-body control algorithm tailored for humanoid locomotion tasks. By reformulating policy learning as a robust optimization problem, HWC-Loco explicitly learns to recover from safety-critical scenarios. While prioritizing safety guarantees, overly conservative behavior can compromise the robot's ability to complete the given tasks. To tackle this challenge, HWC-Loco leverages a hierarchical policy for robust control. This policy can dynamically resolve the trade-off between goal-tracking and safety recovery, guided by human behavior norms and dynamic constraints. To evaluate the performance of HWC-Loco, we conduct extensive comparisons against state-of-the-art humanoid control models, demonstrating HWC-Loco's superior performance across diverse terrains, robot structures, and locomotion tasks under both simulated and real-world environments.

How can we build generalist robot systems? Scale may not be enough due to the significant multimodality of robotics tasks, lack of easily accessible data and the challenges of deploying on physical hardware. Meanwhile, most deployed robotic systems today are inherently modular and can leverage the independent generalization capabilities of each module to perform well. Therefore, this thesis seeks to tackle the task of building generalist robot agents by integrating these components into one: combining modularity with large-scale learning for general purpose robot control. The first question we consider is: how can we build modularity and hierarchy into learning systems? Our key insight is that rather than having the agent learn hierarchy and low-level control end-to-end, we can enforce modularity via planning to enable more efficient and capable robot learners. Next, we come to the role of scale in building generalist robot systems. To scale, neural networks require vast amounts of diverse data, expressive architectures to fit the data and a source of supervision to generate the data. We leverage a powerful supervision source: classical planning, which can generalize, but is expensive to run and requires access to privileged information to perform well in practice. We use these planners to supervise large-scale policy learning in simulation to produce generalist agents. Finally, we consider how to unify modularity with large-scale policy learning to build real-world robot systems capable of performing zero-shot manipulation. We do so by tightly integrating key ingredients of modular high and mid-level planning, learned local control, procedural scene generation and large-scale policy learning for sim2real transfer. We demonstrate that this recipe can produce a single, generalist agent that can solve challenging long-horizon manipulation tasks in the real world.

This paper addresses the challenge of solving Constrained Markov Decision Processes (CMDPs) with $d > 1$ constraints when the transition dynamics are unknown, but samples can be drawn from a generative model. We propose a model-based algorithm for infinite horizon CMDPs with multiple constraints in the tabular setting, aiming to derive and prove sample complexity bounds for learning near-optimal policies. Our approach tackles both the relaxed and strict feasibility settings, where relaxed feasibility allows some constraint violations, and strict feasibility requires adherence to all constraints. The main contributions include the development of the algorithm and the derivation of sample complexity bounds for both settings. For the relaxed feasibility setting we show that our algorithm requires $\tilde{\mathcal{O}} \left( \frac{d |\mathcal{S}| |\mathcal{A}| \log(1/\delta)}{(1-\gamma)^3\epsilon^2} \right)$ samples to return $\epsilon$-optimal policy, while in the strict feasibility setting it requires $\tilde{\mathcal{O}} \left( \frac{d^3 |\mathcal{S}| |\mathcal{A}| \log(1/\delta)}{(1-\gamma)^5\epsilon^2{\zeta_{\mathbf{c}}^*}^2} \right)$ samples.

User consumption behavior data, which records individuals' online spending history at various types of stores, has been widely used in various applications, such as store recommendation, site selection, and sale forecasting. However, its high worth is limited due to deficiencies in data comprehensiveness and changes of application scenarios. Thus, generating high-quality sequential consumption data by simulating complex user consumption behaviors is of great importance to real-world applications. Two branches of existing sequence generation methods are both limited in quality. Model-based methods with simplified assumptions fail to model the complex decision process of user consumption, while data-driven methods that emulate real-world data are prone to noises, unobserved behaviors, and dynamic decision space. In this work, we propose to enhance the fidelity and trustworthiness of the data-driven Generative Adversarial Imitation Learning (GAIL) method by blending it with the Exploration and Preferential Return EPR model . The core idea of our EPR-GAIL framework is to model user consumption behaviors as a complex EPR decision process, which consists of purchase, exploration, and preference decisions. Specifically, we design the hierarchical policy function in the generator as a realization of the EPR decision process and employ the probability distributions of the EPR model to guide the reward function in the discriminator. Extensive experiments on two real-world datasets of user consumption behaviors on an online platform demonstrate that the EPR-GAIL framework outperforms the best state-of-the-art baseline by over 19\% in terms of data fidelity. Furthermore, the generated consumption behavior data can improve the performance of sale prediction and location recommendation by up to 35.29% and 11.19%, respectively, validating its advantage for practical applications.

Existing benchmarks do not test Large Multimodal Models (LMMs) on their interactive intelligence with human users, which is vital for developing generalpurpose AI assistants. We design InterFeedback, an interactive framework, which can be applied to any LMM and dataset to assess this ability autonomously. On top of this, we introduce InterFeedback-Bench that evaluates interactive intelligence using two representative datasets, MMMU-Pro and MathVerse, to test 10 different open-source LMMs. Additionally, we present InterFeedback-Human, a newly collected dataset of 120 cases designed for manually testing interactive performance in leading models such as OpenAI-o1 and Claude-3.5-Sonnet. Our evaluation results indicate that even the state-of-the-art LMM, OpenAI-o1, struggles to refine its responses based on human feedback, achieving an average score of less than 50%. Our findings point to the need for methods that can enhance LMMs' capabilities to interpret and benefit from feedback. In this paper, we are curious about the question "Can Large Multimodal Models evolve through Interactive Human Feedback?" It is central to developing general-purpose AI assistants with Large Multimodal Models (LMMs). While these models show exceptional performance on tackling multimodal tasks directly, their ability to interact with humans remains largely unknown. We argue that an LMM functioning as the general assistant should possess two capabilities: 1) exceptional problem-solving ability and 2) the ability to improve itself through feedback (e.g., human feedback, execution results).

Deep Reinforcement Learning (RL) has been explored and verified to be effective in solving decision-making tasks in various domains, such as robotics, transportation, recommender systems, etc. It learns from the interaction with environments and updates the policy using the collected experience. However, due to the limited real-world data and unbearable consequences of taking detrimental actions, the learning of RL policy is mainly restricted within the simulators. This practice guarantees safety in learning but introduces an inevitable sim-to-real gap in terms of deployment, thus causing degraded performance and risks in execution. There are attempts to solve the sim-to-real problems from different domains with various techniques, especially in the era with emerging techniques such as large foundations or language models that have cast light on the sim-to-real. This survey paper, to the best of our knowledge, is the first taxonomy that formally frames the sim-to-real techniques from key elements of the Markov Decision Process (State, Action, Transition, and Reward). Based on the framework, we cover comprehensive literature from the classic to the most advanced methods including the sim-to-real techniques empowered by foundation models, and we also discuss the specialties that are worth attention in different domains of sim-to-real problems. Then we summarize the formal evaluation process of sim-to-real performance with accessible code or benchmarks. The challenges and opportunities are also presented to encourage future exploration of this direction. We are actively maintaining a repository to include the most up-to-date sim-to-real research work to help domain researchers.

Large Language Model~(LLM) based agents have been increasingly popular in solving complex and dynamic tasks, which requires proper evaluation systems to assess their capabilities. Nevertheless, existing benchmarks usually either focus on single-objective tasks or use overly broad assessing metrics, failing to provide a comprehensive inspection of the actual capabilities of LLM-based agents in complicated decision-making tasks. To address these issues, we introduce DSGBench, a more rigorous evaluation platform for strategic decision-making. Firstly, it incorporates six complex strategic games which serve as ideal testbeds due to their long-term and multi-dimensional decision-making demands and flexibility in customizing tasks of various difficulty levels or multiple targets. Secondly, DSGBench employs a fine-grained evaluation scoring system which examines the decision-making capabilities by looking into the performance in five specific dimensions and offering a comprehensive assessment in a well-designed way. Furthermore, DSGBench also incorporates an automated decision-tracking mechanism which enables in-depth analysis of agent behaviour patterns and the changes in their strategies. We demonstrate the advances of DSGBench by applying it to multiple popular LLM-based agents and our results suggest that DSGBench provides valuable insights in choosing LLM-based agents as well as improving their future development. DSGBench is available at https://github.com/DeciBrain-Group/DSGBench.