Yali Du

Xidong Feng, Yicheng Luo, Ziyan Wang et al. · cmu

When solving decision-making tasks, humans typically depend on information from two key sources: (1) Historical policy data, which provides interaction replay from the environment, and (2) Analytical insights in natural language form, exposing the invaluable thought process or strategic considerations. Despite this, the majority of preceding research focuses on only one source: they either use historical replay exclusively to directly learn policy or value functions, or engaged in language model training utilizing mere language corpus. In this paper, we argue that a powerful autonomous agent should cover both sources. Thus, we propose ChessGPT, a GPT model bridging policy learning and language modeling by integrating data from these two sources in Chess games. Specifically, we build a large-scale game and language dataset related to chess. Leveraging the dataset, we showcase two model examples ChessCLIP and ChessGPT, integrating policy learning and language modeling. Finally, we propose a full evaluation framework for evaluating language model's chess ability. Experimental results validate our model and dataset's effectiveness. We open source our code, model, and dataset at https://github.com/waterhorse1/ChessGPT.

Shangding Gu, Long Yang, Yali Du et al.

Reinforcement Learning (RL) has achieved tremendous success in many complex decision-making tasks. However, safety concerns are raised during deploying RL in real-world applications, leading to a growing demand for safe RL algorithms, such as in autonomous driving and robotics scenarios. While safe control has a long history, the study of safe RL algorithms is still in the early stages. To establish a good foundation for future safe RL research, in this paper, we provide a review of safe RL from the perspectives of methods, theories, and applications. Firstly, we review the progress of safe RL from five dimensions and come up with five crucial problems for safe RL being deployed in real-world applications, coined as "2H3W". Secondly, we analyze the algorithm and theory progress from the perspectives of answering the "2H3W" problems. Particularly, the sample complexity of safe RL algorithms is reviewed and discussed, followed by an introduction to the applications and benchmarks of safe RL algorithms. Finally, we open the discussion of the challenging problems in safe RL, hoping to inspire future research on this thread. To advance the study of safe RL algorithms, we release an open-sourced repository containing the implementations of major safe RL algorithms at the link: https://github.com/chauncygu/Safe-Reinforcement-Learning-Baselines.git.

Yali Du, Yinwei Wei, Wei Ji et al.

The booming development and huge market of micro-videos bring new e-commerce channels for merchants. Currently, more micro-video publishers prefer to embed relevant ads into their micro-videos, which not only provides them with business income but helps the audiences to discover their interesting products. However, due to the micro-video recording by unprofessional equipment, involving various topics and including multiple modalities, it is challenging to locate the products related to micro-videos efficiently, appropriately, and accurately. We formulate the microvideo-product retrieval task, which is the first attempt to explore the retrieval between the multi-modal and multi-modal instances. A novel approach named Multi-Queue Momentum Contrast (MQMC) network is proposed for bidirectional retrieval, consisting of the uni-modal feature and multi-modal instance representation learning. Moreover, a discriminative selection strategy with a multi-queue is used to distinguish the importance of different negatives based on their categories. We collect two large-scale microvideo-product datasets (MVS and MVS-large) for evaluation and manually construct the hierarchical category ontology, which covers sundry products in daily life. Extensive experiments show that MQMC outperforms the state-of-the-art baselines. Our replication package (including code, dataset, etc.) is publicly available at https://github.com/duyali2000/MQMC.

Xingzhou Lou, Jiaxian Guo, Junge Zhang et al.

Zero-shot human-AI coordination holds the promise of collaborating with humans without human data. Prevailing methods try to train the ego agent with a population of partners via self-play. However, these methods suffer from two problems: 1) The diversity of a population with finite partners is limited, thereby limiting the capacity of the trained ego agent to collaborate with a novel human; 2) Current methods only provide a common best response for every partner in the population, which may result in poor zero-shot coordination performance with a novel partner or humans. To address these issues, we first propose the policy ensemble method to increase the diversity of partners in the population, and then develop a context-aware method enabling the ego agent to analyze and identify the partner's potential policy primitives so that it can take different actions accordingly. In this way, the ego agent is able to learn more universal cooperative behaviors for collaborating with diverse partners. We conduct experiments on the Overcooked environment, and evaluate the zero-shot human-AI coordination performance of our method with both behavior-cloned human proxies and real humans. The results demonstrate that our method significantly increases the diversity of partners and enables ego agents to learn more diverse behaviors than baselines, thus achieving state-of-the-art performance in all scenarios. We also open-source a human-AI coordination study framework on the Overcooked for the convenience of future studies.

Ruiqi Zhang, Jing Hou, Florian Walter et al. · berkeley

Reinforcement Learning (RL) is a potent tool for sequential decision-making and has achieved performance surpassing human capabilities across many challenging real-world tasks. As the extension of RL in the multi-agent system domain, multi-agent RL (MARL) not only need to learn the control policy but also requires consideration regarding interactions with all other agents in the environment, mutual influences among different system components, and the distribution of computational resources. This augments the complexity of algorithmic design and poses higher requirements on computational resources. Simultaneously, simulators are crucial to obtain realistic data, which is the fundamentals of RL. In this paper, we first propose a series of metrics of simulators and summarize the features of existing benchmarks. Second, to ease comprehension, we recall the foundational knowledge and then synthesize the recently advanced studies of MARL-related autonomous driving and intelligent transportation systems. Specifically, we examine their environmental modeling, state representation, perception units, and algorithm design. Conclusively, we discuss open challenges as well as prospects and opportunities. We hope this paper can help the researchers integrate MARL technologies and trigger more insightful ideas toward the intelligent and autonomous driving.

Mengyue Yang, Zhen Fang, Yonggang Zhang et al.

Out-of-distribution (OOD) generalization is indispensable for learning models in the wild, where testing distribution typically unknown and different from the training. Recent methods derived from causality have shown great potential in achieving OOD generalization. However, existing methods mainly focus on the invariance property of causes, while largely overlooking the property of \textit{sufficiency} and \textit{necessity} conditions. Namely, a necessary but insufficient cause (feature) is invariant to distribution shift, yet it may not have required accuracy. By contrast, a sufficient yet unnecessary cause (feature) tends to fit specific data well but may have a risk of adapting to a new domain. To capture the information of sufficient and necessary causes, we employ a classical concept, the probability of sufficiency and necessary causes (PNS), which indicates the probability of whether one is the necessary and sufficient cause. To associate PNS with OOD generalization, we propose PNS risk and formulate an algorithm to learn representation with a high PNS value. We theoretically analyze and prove the generalizability of the PNS risk. Experiments on both synthetic and real-world benchmarks demonstrate the effectiveness of the proposed method. The details of the implementation can be found at the GitHub repository: https://github.com/ymy4323460/CaSN.

Lukas Schäfer, Oliver Slumbers, Stephen McAleer et al. · microsoft-research

Multi-agent reinforcement learning (MARL) requires agents to explore within a vast joint action space to find joint actions that lead to coordination. Existing value-based MARL algorithms commonly rely on random exploration, such as $ε$-greedy, to explore the environment which is not systematic and inefficient at identifying effective actions in multi-agent problems. Additionally, the concurrent training of the policies of multiple agents during training can render the optimisation non-stationary. This can lead to unstable value estimates, highly variant gradients, and ultimately hinder coordination between agents. To address these challenges, we propose ensemble value functions for multi-agent exploration (EMAX). EMAX is a framework to seamlessly extend value-based MARL algorithms. EMAX leverages an ensemble of value functions for each agent to guide their exploration, reduce the variance of their optimisation, and makes their policies more robust to miscoordination. EMAX achieves these benefits by (1) systematically guiding the exploration of agents with a UCB policy towards parts of the environment that require multiple agents to coordinate. (2) EMAX computes average value estimates across the ensemble as target values to reduce the variance of gradients and make optimisation more stable. (3) During evaluation, EMAX selects actions following a majority vote across the ensemble to reduce the likelihood of miscoordination. We first instantiate independent DQN with EMAX and evaluate it in 11 general-sum tasks with sparse rewards. We show that EMAX improves final evaluation returns by 185% across all tasks. We then evaluate EMAX on top of IDQN, VDN and QMIX in 21 common-reward tasks, and show that EMAX improves sample efficiency and final evaluation returns across all tasks over all three vanilla algorithms by 60%, 47%, and 538%, respectively.

Shangding Gu, Alap Kshirsagar, Yali Du et al.

Deployment of Reinforcement Learning (RL) algorithms for robotics applications in the real world requires ensuring the safety of the robot and its environment. Safe Robot RL (SRRL) is a crucial step towards achieving human-robot coexistence. In this paper, we envision a human-centered SRRL framework consisting of three stages: safe exploration, safety value alignment, and safe collaboration. We examine the research gaps in these areas and propose to leverage interactive behaviors for SRRL. Interactive behaviors enable bi-directional information transfer between humans and robots, such as conversational robot ChatGPT. We argue that interactive behaviors need further attention from the SRRL community. We discuss four open challenges related to the robustness, efficiency, transparency, and adaptability of SRRL with interactive behaviors.

Yunqiu Xu, Meng Fang, Ling Chen et al.

Text-based games provide an interactive way to study natural language processing. While deep reinforcement learning has shown effectiveness in developing the game playing agent, the low sample efficiency and the large action space remain to be the two major challenges that hinder the DRL from being applied in the real world. In this paper, we address the challenges by introducing world-perceiving modules, which automatically decompose tasks and prune actions by answering questions about the environment. We then propose a two-phase training framework to decouple language learning from reinforcement learning, which further improves the sample efficiency. The experimental results show that the proposed method significantly improves the performance and sample efficiency. Besides, it shows robustness against compound error and limited pre-training data.

Yali Du, Chengdong Ma, Yuchen Liu et al.

Reinforcement learning algorithms require a large amount of samples; this often limits their real-world applications on even simple tasks. Such a challenge is more outstanding in multi-agent tasks, as each step of operation is more costly requiring communications or shifting or resources. This work aims to improve data efficiency of multi-agent control by model-based learning. We consider networked systems where agents are cooperative and communicate only locally with their neighbors, and propose the decentralized model-based policy optimization framework (DMPO). In our method, each agent learns a dynamic model to predict future states and broadcast their predictions by communication, and then the policies are trained under the model rollouts. To alleviate the bias of model-generated data, we restrain the model usage for generating myopic rollouts, thus reducing the compounding error of model generation. To pertain the independence of policy update, we introduce extended value function and theoretically prove that the resulting policy gradient is a close approximation to true policy gradients. We evaluate our algorithm on several benchmarks for intelligent transportation systems, which are connected autonomous vehicle control tasks (Flow and CACC) and adaptive traffic signal control (ATSC). Empirically results show that our method achieves superior data efficiency and matches the performance of model-free methods using true models.

Wei Liu, Siya Qi, Yali Du et al.

Large language models (LLMs) make it plausible to build systems that improve through self-evolving loops, but many existing proposals are better understood as self-play and often plateau quickly. A central failure mode is that the loop synthesises more data without increasing learnable information for the next iteration. Through experiments on a self-play coding task, we reveal that sustainable self-evolution requires a self-synthesised data pipeline with learnable information that increases across iterations. We identify triadic roles that self-evolving LLMs play: the Proposer, which generates tasks; the Solver, which attempts solutions; and the Verifier, which provides training signals, and we identify three system designs that jointly target learnable information gain from this triadic roles perspective. Asymmetric co-evolution closes a weak-to-strong-to-weak loop across roles. Capacity growth expands parameter and inference-time budgets to match rising learnable information. Proactive information seeking introduces external context and new task sources that prevent saturation. Together, these modules provide a measurable, system-level path from brittle self-play dynamics to sustained self-evolution.

Chandler Smith, Marwa Abdulhai, Manfred Diaz et al.

Large Language Model (LLM) agents have demonstrated impressive capabilities for social interaction and are increasingly being deployed in situations where they might engage with both human and artificial agents. These interactions represent a critical frontier for LLM-based agents, yet existing evaluation methods fail to measure how well these capabilities generalize to novel social situations. In this paper, we introduce a method for evaluating the ability of LLM-based agents to cooperate in zero-shot, mixed-motive environments using Concordia, a natural language multi-agent simulation environment. Our method measures general cooperative intelligence by testing an agent's ability to identify and exploit opportunities for mutual gain across diverse partners and contexts. We present empirical results from the NeurIPS 2024 Concordia Contest, where agents were evaluated on their ability to achieve mutual gains across a suite of diverse scenarios ranging from negotiation to collective action problems. Our findings reveal significant gaps between current agent capabilities and the robust generalization required for reliable cooperation, particularly in scenarios demanding persuasion and norm enforcement.

Runji Lin, Ye Li, Xidong Feng et al.

The pretrain-finetuning paradigm in large-scale sequence models has made significant progress in natural language processing and computer vision tasks. However, such a paradigm is still hindered by several challenges in Reinforcement Learning (RL), including the lack of self-supervised pretraining algorithms based on offline data and efficient fine-tuning/prompt-tuning over unseen downstream tasks. In this work, we explore how prompts can improve sequence modeling-based offline reinforcement learning (offline-RL) algorithms. Firstly, we propose prompt tuning for offline RL, where a context vector sequence is concatenated with the input to guide the conditional policy generation. As such, we can pretrain a model on the offline dataset with self-supervised loss and learn a prompt to guide the policy towards desired actions. Secondly, we extend our framework to Meta-RL settings and propose Contextual Meta Transformer (CMT); CMT leverages the context among different tasks as the prompt to improve generalization on unseen tasks. We conduct extensive experiments across three different offline-RL settings: offline single-agent RL on the D4RL dataset, offline Meta-RL on the MuJoCo benchmark, and offline MARL on the SMAC benchmark. Superior results validate the strong performance, and generality of our methods.

Yang Li, Shao Zhang, Jichen Sun et al.

Zero-shot coordination in cooperative artificial intelligence (AI) remains a significant challenge, which means effectively coordinating with a wide range of unseen partners. Previous algorithms have attempted to address this challenge by optimizing fixed objectives within a population to improve strategy or behaviour diversity. However, these approaches can result in a loss of learning and an inability to cooperate with certain strategies within the population, known as cooperative incompatibility. To address this issue, we propose the Cooperative Open-ended LEarning (COLE) framework, which constructs open-ended objectives in cooperative games with two players from the perspective of graph theory to assess and identify the cooperative ability of each strategy. We further specify the framework and propose a practical algorithm that leverages knowledge from game theory and graph theory. Furthermore, an analysis of the learning process of the algorithm shows that it can efficiently overcome cooperative incompatibility. The experimental results in the Overcooked game environment demonstrate that our method outperforms current state-of-the-art methods when coordinating with different-level partners. Our demo is available at https://sites.google.com/view/cole-2023.

Sirui Chen, Zhaowei Zhang, Yaodong Yang et al.

Centralized Training with Decentralized Execution (CTDE) has been proven to be an effective paradigm in cooperative multi-agent reinforcement learning (MARL). One of the major challenges is credit assignment, which aims to credit agents by their contributions. While prior studies have shown great success, their methods typically fail to work in episodic reinforcement learning scenarios where global rewards are revealed only at the end of the episode. They lack the functionality to model complicated relations of the delayed global reward in the temporal dimension and suffer from inefficiencies. To tackle this, we introduce Spatial-Temporal Attention with Shapley (STAS), a novel method that learns credit assignment in both temporal and spatial dimensions. It first decomposes the global return back to each time step, then utilizes the Shapley Value to redistribute the individual payoff from the decomposed global reward. To mitigate the computational complexity of the Shapley Value, we introduce an approximation of marginal contribution and utilize Monte Carlo sampling to estimate it. We evaluate our method on an Alice & Bob example and MPE environments across different scenarios. Our results demonstrate that our method effectively assigns spatial-temporal credit, outperforming all state-of-the-art baselines.

Yang Li, Shao Zhang, Jichen Sun et al.

Securing coordination between AI agent and teammates (human players or AI agents) in contexts involving unfamiliar humans continues to pose a significant challenge in Zero-Shot Coordination. The issue of cooperative incompatibility becomes particularly prominent when an AI agent is unsuccessful in synchronizing with certain previously unknown partners. Traditional algorithms have aimed to collaborate with partners by optimizing fixed objectives within a population, fostering diversity in strategies and behaviors. However, these techniques may lead to learning loss and an inability to cooperate with specific strategies within the population, a phenomenon named cooperative incompatibility in learning. In order to solve cooperative incompatibility in learning and effectively address the problem in the context of ZSC, we introduce the Cooperative Open-ended LEarning (COLE) framework, which formulates open-ended objectives in cooperative games with two players using perspectives of graph theory to evaluate and pinpoint the cooperative capacity of each strategy. We present two practical algorithms, specifically \algo and \algoR, which incorporate insights from game theory and graph theory. We also show that COLE could effectively overcome the cooperative incompatibility from theoretical and empirical analysis. Subsequently, we created an online Overcooked human-AI experiment platform, the COLE platform, which enables easy customization of questionnaires, model weights, and other aspects. Utilizing the COLE platform, we enlist 130 participants for human experiments. Our findings reveal a preference for our approach over state-of-the-art methods using a variety of subjective metrics. Moreover, objective experimental outcomes in the Overcooked game environment indicate that our method surpasses existing ones when coordinating with previously unencountered AI agents and the human proxy model.

Cody Kommers, Ruth Ahnert, Maria Antoniak et al.

Generative AI systems are increasingly recognized as cultural technologies, yet current evaluation frameworks often treat culture as a variable to be measured rather than fundamental to the system's operation. Drawing on hermeneutic theory from the humanities, we argue that GenAI systems function as "context machines" that must inherently address three interpretive challenges: situatedness (meaning only emerges in context), plurality (multiple valid interpretations coexist), and ambiguity (interpretations naturally conflict). We present computational hermeneutics as an emerging framework offering an interpretive account of what GenAI systems do, and how they might do it better. We offer three principles for hermeneutic evaluation -- that benchmarks should be iterative, not one-off; include people, not just machines; and measure cultural context, not just model output. This perspective offers a nascent paradigm for designing and evaluating contemporary AI systems: shifting from standardized questions about accuracy to contextual ones about meaning.

Jiarui Jin, Xianyu Chen, Weinan Zhang et al.

Learning-to-rank is a core technique in the top-N recommendation task, where an ideal ranker would be a mapping from an item set to an arrangement (a.k.a. permutation). Most existing solutions fall in the paradigm of probabilistic ranking principle (PRP), i.e., first score each item in the candidate set and then perform a sort operation to generate the top ranking list. However, these approaches neglect the contextual dependence among candidate items during individual scoring, and the sort operation is non-differentiable. To bypass the above issues, we propose Set-To-Arrangement Ranking (STARank), a new framework directly generates the permutations of the candidate items without the need for individually scoring and sort operations; and is end-to-end differentiable. As a result, STARank can operate when only the ground-truth permutations are accessible without requiring access to the ground-truth relevance scores for items. For this purpose, STARank first reads the candidate items in the context of the user browsing history, whose representations are fed into a Plackett-Luce module to arrange the given items into a list. To effectively utilize the given ground-truth permutations for supervising STARank, we leverage the internal consistency property of Plackett-Luce models to derive a computationally efficient list-wise loss. Experimental comparisons against 9 the state-of-the-art methods on 2 learning-to-rank benchmark datasets and 3 top-N real-world recommendation datasets demonstrate the superiority of STARank in terms of conventional ranking metrics. Notice that these ranking metrics do not consider the effects of the contextual dependence among the items in the list, we design a new family of simulation-based ranking metrics, where existing metrics can be regarded as special cases. STARank can consistently achieve better performance in terms of PBM and UBM simulation-based metrics.

Runze Liu, Yali Du, Fengshuo Bai et al.

In preference-based Reinforcement Learning (RL), obtaining a large number of preference labels are both time-consuming and costly. Furthermore, the queried human preferences cannot be utilized for the new tasks. In this paper, we propose Zero-shot Cross-task Preference Alignment and Robust Reward Learning (PEARL), which learns policies from cross-task preference transfer without any human labels of the target task. Our contributions include two novel components that facilitate the transfer and learning process. The first is Cross-task Preference Alignment (CPA), which transfers the preferences between tasks via optimal transport. The key idea of CPA is to use Gromov-Wasserstein distance to align the trajectories between tasks, and the solved optimal transport matrix serves as the correspondence between trajectories. The target task preferences are computed as the weighted sum of source task preference labels with the correspondence as weights. Moreover, to ensure robust learning from these transferred labels, we introduce Robust Reward Learning (RRL), which considers both reward mean and uncertainty by modeling rewards as Gaussian distributions. Empirical results on robotic manipulation tasks from Meta-World and Robomimic demonstrate that our method is capable of transferring preference labels across tasks accurately and then learns well-behaved policies. Notably, our approach significantly exceeds existing methods when there are few human preferences. The code and videos of our method are available at: https://sites.google.com/view/pearl-preference.

Shutong Ding, Jingya Wang, Yali Du et al.

Recent advances in constrained reinforcement learning (RL) have endowed reinforcement learning with certain safety guarantees. However, deploying existing constrained RL algorithms in continuous control tasks with general hard constraints remains challenging, particularly in those situations with non-convex hard constraints. Inspired by the generalized reduced gradient (GRG) algorithm, a classical constrained optimization technique, we propose a reduced policy optimization (RPO) algorithm that combines RL with GRG to address general hard constraints. RPO partitions actions into basic actions and nonbasic actions following the GRG method and outputs the basic actions via a policy network. Subsequently, RPO calculates the nonbasic actions by solving equations based on equality constraints using the obtained basic actions. The policy network is then updated by implicitly differentiating nonbasic actions with respect to basic actions. Additionally, we introduce an action projection procedure based on the reduced gradient and apply a modified Lagrangian relaxation technique to ensure inequality constraints are satisfied. To the best of our knowledge, RPO is the first attempt that introduces GRG to RL as a way of efficiently handling both equality and inequality hard constraints. It is worth noting that there is currently a lack of RL environments with complex hard constraints, which motivates us to develop three new benchmarks: two robotics manipulation tasks and a smart grid operation control task. With these benchmarks, RPO achieves better performance than previous constrained RL algorithms in terms of both cumulative reward and constraint violation. We believe RPO, along with the new benchmarks, will open up new opportunities for applying RL to real-world problems with complex constraints.

Zhenyi Shen, Hanqi Yan, Linhai Zhang et al.

Chain-of-Thought (CoT) reasoning enhances Large Language Models (LLMs) by encouraging step-by-step reasoning in natural language. However, leveraging a latent continuous space for reasoning may offer benefits in terms of both efficiency and robustness. Prior implicit CoT methods attempt to bypass language completely by reasoning in continuous space but have consistently underperformed compared to the standard explicit CoT approach. We introduce CODI (Continuous Chain-of-Thought via Self-Distillation), a novel training framework that effectively compresses natural language CoT into continuous space. CODI jointly trains a teacher task (Explicit CoT) and a student task (Implicit CoT), distilling the reasoning ability from language into continuous space by aligning the hidden states of a designated token. Our experiments show that CODI is the first implicit CoT approach to match the performance of explicit CoT on GSM8k at the GPT-2 scale, achieving a 3.1x compression rate and outperforming the previous state-of-the-art by 28.2% in accuracy. CODI also demonstrates robustness, generalizable to complex datasets, and interpretability. These results validate that LLMs can reason effectively not only in natural language, but also in a latent continuous space. Code is available at https://github.com/zhenyi4/codi.

Zhixun Chen, Ming Li, Yuxuan Huang et al.

Large Language Model (LLM) agents have demonstrated remarkable generalization capabilities across multi-domain tasks. Existing agent tuning approaches typically employ supervised finetuning on entire expert trajectories. However, behavior-cloning of full trajectories can introduce expert bias and weaken generalization to states not covered by the expert data. Additionally, critical steps, such as planning, complex reasoning for intermediate subtasks, and strategic decision-making, are essential to success in agent tasks, so learning these steps is the key to improving LLM agents. For more effective and efficient agent tuning, we propose ATLaS that identifies the critical steps in expert trajectories and finetunes LLMs solely on these steps with reduced costs. By steering the training's focus to a few critical steps, our method mitigates the risk of overfitting entire trajectories and promotes generalization across different environments and tasks. In extensive experiments, an LLM finetuned on only 30% critical steps selected by ATLaS outperforms the LLM finetuned on all steps and recent open-source LLM agents. ATLaS maintains and improves base LLM skills as generalist agents interacting with diverse environments.

Edoardo Pona, Milad Kazemi, Mehran Hosseini et al.

Monitoring LLM safety at scale requires balancing cost and accuracy: a cheap latent-space probe can screen every input, but hard cases should be escalated to a more expensive expert. Existing cascades delegate based on probe uncertainty, but uncertainty is a poor proxy for delegation benefit, as it ignores whether the expert would actually correct the error. To address this problem, we introduce Calibrate-Then-Delegate (CTD), a model-cascade approach that provides probabilistic guarantees on the computation cost while enabling instance-level (streaming) decisions. CTD builds on a novel delegation value (DV) probe, a lightweight model operating on the same internal representations as the safety probe that directly predicts the benefit of escalation. To enforce budget constraints, CTD calibrates a threshold on the DV signal using held-out data via multiple hypothesis testing, yielding finite-sample guarantees on the delegation rate. Evaluated on four safety datasets, CTD consistently outperforms uncertainty-based delegation at every budget level, avoids harmful over-delegation, and adapts budget allocation to input difficulty without requiring group labels.

Yali Du, Hui Sun, Ming Li

Programs implemented in various programming languages form the foundation of software applications. To alleviate the burden of program migration and facilitate the development of software systems, automated program translation across languages has garnered significant attention. Previous approaches primarily focus on pairwise translation paradigms, learning translation between pairs of languages using bilingual parallel data. However, parallel data is difficult to collect for some language pairs, and the distribution of program semantics across languages can shift, posing challenges for pairwise program translation. In this paper, we argue that jointly learning a unified model to translate code across multiple programming languages is superior to separately learning from bilingual parallel data. We propose Variational Interaction for Multilingual Program Translation~(VIM-PT), a disentanglement-based generative approach that jointly trains a unified model for multilingual program translation across multiple languages. VIM-PT disentangles code into language-shared and language-specific features, using variational inference and interaction information with a novel lower bound, then achieves program translation through conditional generation. VIM-PT demonstrates four advantages: 1) captures language-shared information more accurately from various implementations and improves the quality of multilingual program translation, 2) mines and leverages the capability of non-parallel data, 3) addresses the distribution shift of program semantics across languages, 4) and serves as a unified model, reducing deployment complexity.

Mark Towers, Yali Du, Christopher Freeman et al.

Future reward estimation is a core component of reinforcement learning agents; i.e., Q-value and state-value functions, predicting an agent's sum of future rewards. Their scalar output, however, obfuscates when or what individual future rewards an agent may expect to receive. We address this by modifying an agent's future reward estimator to predict their next N expected rewards, referred to as Temporal Reward Decomposition (TRD). This unlocks novel explanations of agent behaviour. Through TRD we can: estimate when an agent may expect to receive a reward, the value of the reward and the agent's confidence in receiving it; measure an input feature's temporal importance to the agent's action decisions; and predict the influence of different actions on future rewards. Furthermore, we show that DQN agents trained on Atari environments can be efficiently retrained to incorporate TRD with minimal impact on performance.

Olivier Schipper, Yudi Zhang, Yali Du et al.

LLM-based agents have shown promise in various cooperative and strategic reasoning tasks, but their effectiveness in competitive multi-agent environments remains underexplored. To address this gap, we introduce PillagerBench, a novel framework for evaluating multi-agent systems in real-time competitive team-vs-team scenarios in Minecraft. It provides an extensible API, multi-round testing, and rule-based built-in opponents for fair, reproducible comparisons. We also propose TactiCrafter, an LLM-based multi-agent system that facilitates teamwork through human-readable tactics, learns causal dependencies, and adapts to opponent strategies. Our evaluation demonstrates that TactiCrafter outperforms baseline approaches and showcases adaptive learning through self-play. Additionally, we analyze its learning process and strategic evolution over multiple game episodes. To encourage further research, we have open-sourced PillagerBench, fostering advancements in multi-agent AI for competitive environments.

Gurusha Juneja, Dylan Lu, Saaket Agashe et al.

Theory of Mind (ToM), the ability to track others epistemic state, makes humans efficient collaborators. AI agents need the same capacity in multi agent settings, yet existing benchmarks mostly test literal ToM by asking direct belief questions. The ability act optimally on implicit beliefs in embodied environments, called functional ToM, remains largely untested. We introduce EnactToM, an evolving benchmark of 300 embodied multi-agent tasks set in a 3D household with partial observability, private information, and constrained communication. Each task is formally verified for solvability and required epistemic depth, and new tasks are generated increase difficulty as models improve. On the hard split, all seven evaluated frontier models score 0.0% Pass^3 on functional task completion, while averaging 45.0% on literal belief probes. Manual analysis traces 93% of sampled failures to epistemic coordination breakdowns such as withheld information, ignored partner constraints, and misallocated messages, providing a concrete target for future work.

Mingze Zhong, Meng Fang, Zijing Shi et al.

The Spiral of Silence (SoS) theory holds that individuals with minority views often refrain from speaking out for fear of social isolation, enabling majority positions to dominate public discourse. When the 'agents' are large language models (LLMs), however, the classical psychological explanation is not directly applicable, since SoS was developed for human societies. This raises a central question: can SoS-like dynamics nevertheless emerge from purely statistical language generation in LLM collectives? We propose an evaluation framework for examining SoS in LLM agents. Specifically, we consider four controlled conditions that systematically vary the availability of 'History' and 'Persona' signals. Opinion dynamics are assessed using trend tests such as Mann-Kendall and Spearman's rank, along with concentration measures including kurtosis and interquartile range. Experiments across open-source and closed-source models show that history and persona together produce strong majority dominance and replicate SoS patterns; history signals alone induce strong anchoring; and persona signals alone foster diverse but uncorrelated opinions, indicating that without historical anchoring, SoS dynamics cannot emerge. The work bridges computational sociology and responsible AI design, highlighting the need to monitor and mitigate emergent conformity in LLM-agent systems.

Yali Du, Joel Z. Leibo, Usman Islam et al.

Cooperation in multi-agent learning (MAL) is a topic at the intersection of numerous disciplines, including game theory, economics, social sciences, and evolutionary biology. Research in this area aims to understand both how agents can coordinate effectively when goals are aligned and how they may cooperate in settings where gains from working together are possible but possibilities for conflict abound. In this paper we provide an overview of the fundamental concepts, problem settings and algorithms of multi-agent learning. This encompasses reinforcement learning, multi-agent sequential decision-making, challenges associated with multi-agent cooperation, and a comprehensive review of recent progress, along with an evaluation of relevant metrics. Finally we discuss open challenges in the field with the aim of inspiring new avenues for research.

Siyuan Guo, Yali Du, Hechang Chen et al.

Large language models (LLMs) have become a central foundation of modern artificial intelligence, yet their lifecycle remains constrained by a rigid separation between training and deployment, after which learning effectively ceases. This limitation contrasts with natural intelligence, which continually adapts through interaction with its environment. In this paper, we formalise deployment-time learning (DTL) as the third stage in the LLM lifecycle that enables LLM agents to improve from experience during deployment without modifying model parameters. We present CASCADE (CASe-based Continual Adaptation during DEployment), a general and principled framework that equips LLM agents with an explicit, evolving episodic memory. CASCADE formulates experience reuse as a contextual bandit problem, enabling principled exploration-exploitation trade-offs and establishing no-regret guarantees over long-term interactions. This design allows agents to accumulate, select, and refine task-relevant cases, transforming past experience into actionable knowledge. Across 16 diverse tasks spanning medical diagnosis, legal analysis, code generation, web search, tool use, and embodied interaction, CASCADE improves macro-averaged success rate by 20.9% over zero-shot prompting while consistently outperforming gradient-based and memory-based baselines. By reframing deployment as an adaptive learning process, this work establishes a foundation for continually improving AI systems.

Zijing Shi, Meng Fang, Shunfeng Zheng et al.

Multi-agent collaboration with Large Language Models (LLMs) demonstrates proficiency in basic tasks, yet its efficiency in more complex scenarios remains unexplored. In gaming environments, these agents often face situations without established coordination protocols, requiring them to make intelligent inferences about teammates from limited data. This problem motivates the area of ad hoc teamwork, in which an agent may potentially cooperate with a variety of teammates to achieve a shared goal. Our study focuses on the ad hoc teamwork problem where the agent operates in an environment driven by natural language. Our findings reveal the potential of LLM agents in team collaboration, highlighting issues related to hallucinations in communication. To address this issue, we develop CodeAct, a general agent that equips LLM with enhanced memory and code-driven reasoning, enabling the repurposing of partial information for rapid adaptation to new teammates.

Zhicheng Zhang, Ziyan Wang, Yali Du et al.

Exploration remains a key bottleneck for reinforcement learning (RL) post-training of large language models (LLMs), where sparse feedback and large action spaces can lead to premature collapse into repetitive behaviors. We propose Verbalized Action Masking (VAM), which verbalizes an action mask in the prompt and enforces that the model outputs an action from the masked set. Building on this interface, we introduce iterative action-space pruning: if the target action is not sampled, we remove valid sampled actions from the mask and resample under the reduced candidate set, repeating until the target is sampled or a fixed budget is exhausted. We study VAM in chess and evaluate it under two training regimes: an engine-play regime that generates states via play against an engine opponent and a fixed-dataset regime that trains from a fixed dataset of positions with verifier scores. Across held-out chess puzzles and full-game play measured by average centipawn loss (ACPL), VAM improves learning efficiency and final performance over strong baselines, highlighting verbalized masking as a practical mechanism for controllable exploration in LLM RL post-training.

Hao Liang, Jiayu Cheng, Sean R. Sinclair et al.

Exogenous MDPs (Exo-MDPs) capture sequential decision-making where uncertainty comes solely from exogenous inputs that evolve independently of the learner's actions. This structure is especially common in operations research applications such as inventory control, energy storage, and resource allocation, where exogenous randomness (e.g., demand, arrivals, or prices) drives system behavior. Despite decades of empirical evidence that greedy, exploitation-only methods work remarkably well in these settings, theory has lagged behind: all existing regret guarantees for Exo-MDPs rely on explicit exploration or tabular assumptions. We show that exploration is unnecessary. We propose Pure Exploitation Learning (PEL) and prove the first general finite-sample regret bounds for exploitation-only algorithms in Exo-MDPs. In the tabular case, PEL achieves $\widetilde{O}(H^2|Ξ|\sqrt{K})$. For large, continuous endogenous state spaces, we introduce LSVI-PE, a simple linear-approximation method whose regret is polynomial in the feature dimension, exogenous state space, and horizon, independent of the endogenous state and action spaces. Our analysis introduces two new tools: counterfactual trajectories and Bellman-closed feature transport, which together allow greedy policies to have accurate value estimates without optimism. Experiments on synthetic and resource-management tasks show that PEL consistently outperforming baselines. Overall, our results overturn the conventional wisdom that exploration is required, demonstrating that in Exo-MDPs, pure exploitation is enough.

Wei Liu, Peijie Yu, Michele Orini et al.

The agency expected of Agentic Large Language Models goes beyond answering correctly, requiring autonomy to set goals and decide what to explore. We term this investigatory intelligence, distinguishing it from executional intelligence, which merely completes assigned tasks. Data Science provides a natural testbed, as real-world analysis starts from raw data rather than explicit queries, yet few benchmarks focus on it. To address this, we introduce Deep Data Research (DDR), an open-ended task where LLMs autonomously extract key insights from databases, and DDR-Bench, a large-scale, checklist-based benchmark that enables verifiable evaluation. Results show that while frontier models display emerging agency, long-horizon exploration remains challenging. Our analysis highlights that effective investigatory intelligence depends not only on agent scaffolding or merely scaling, but also on intrinsic strategies of agentic models.

Fengshuo Bai, Runze Liu, Yali Du et al.

Evaluating deep reinforcement learning (DRL) agents against targeted behavior attacks is critical for assessing their robustness. These attacks aim to manipulate the victim into specific behaviors that align with the attacker's objectives, often bypassing traditional reward-based defenses. Prior methods have primarily focused on reducing cumulative rewards; however, rewards are typically too generic to capture complex safety requirements effectively. As a result, focusing solely on reward reduction can lead to suboptimal attack strategies, particularly in safety-critical scenarios where more precise behavior manipulation is needed. To address these challenges, we propose RAT, a method designed for universal, targeted behavior attacks. RAT trains an intention policy that is explicitly aligned with human preferences, serving as a precise behavioral target for the adversary. Concurrently, an adversary manipulates the victim's policy to follow this target behavior. To enhance the effectiveness of these attacks, RAT dynamically adjusts the state occupancy measure within the replay buffer, allowing for more controlled and effective behavior manipulation. Our empirical results on robotic simulation tasks demonstrate that RAT outperforms existing adversarial attack algorithms in inducing specific behaviors. Additionally, RAT shows promise in improving agent robustness, leading to more resilient policies. We further validate RAT by guiding Decision Transformer agents to adopt behaviors aligned with human preferences in various MuJoCo tasks, demonstrating its effectiveness across diverse tasks.

Xidong Feng, Ziyu Wan, Mengyue Yang et al. · cmu

Reinforcement Learning (RL) has shown remarkable abilities in learning policies for decision-making tasks. However, RL is often hindered by issues such as low sample efficiency, lack of interpretability, and sparse supervision signals. To tackle these limitations, we take inspiration from the human learning process and introduce Natural Language Reinforcement Learning (NLRL), which innovatively combines RL principles with natural language representation. Specifically, NLRL redefines RL concepts like task objectives, policy, value function, Bellman equation, and policy iteration in natural language space. We present how NLRL can be practically implemented with the latest advancements in large language models (LLMs) like GPT-4. Initial experiments over tabular MDPs demonstrate the effectiveness, efficiency, and also interpretability of the NLRL framework.

Yudi Zhang, Lu Wang, Meng Fang et al.

Distilling large language models (LLMs) typically involves transferring the teacher model's responses through supervised fine-tuning (SFT). However, this approach neglects the potential to distill both data (output content) and reward signals (quality evaluations). Extracting reliable reward signals directly from teacher models is challenging, as LLMs are optimized for generation rather than evaluation, often resulting in biased or inconsistent assessments. To address this limitation, we propose a novel distillation pipeline that transfers both responses and rewards. Our method generates pseudo-rewards through a self-supervised mechanism that leverages the inherent structure of both teacher and student responses, enabling reward learning without explicit external evaluation. The reward model subsequently guides reinforcement learning (RL), allowing iterative refinement of the student model after an SFT warm-up phase. Experiments on GSM8K and MMLU-PRO demonstrate that our method consistently outperforms traditional SFT-based approaches, enabling student models to surpass the performance of their teachers. This work highlights the potential for scalable, efficient distillation through structured self-supervised reward learning, reducing dependence on external reward supervision.

Yang Li, Wenhao Zhang, Jianhong Wang et al.

Among the research topics in multi-agent learning, mixed-motive cooperation is one of the most prominent challenges, primarily due to the mismatch between individual and collective goals. The cutting-edge research is focused on incorporating domain knowledge into rewards and introducing additional mechanisms to incentivize cooperation. However, these approaches often face shortcomings such as the effort on manual design and the absence of theoretical groundings. To close this gap, we model the mixed-motive game as a differentiable game for the ease of illuminating the learning dynamics towards cooperation. More detailed, we introduce a novel optimization method named \textbf{\textit{A}}ltruistic \textbf{\textit{G}}radient \textbf{\textit{A}}djustment (\textbf{\textit{AgA}}) that employs gradient adjustments to progressively align individual and collective objectives. Furthermore, we theoretically prove that AgA effectively attracts gradients to stable fixed points of the collective objective while considering individual interests, and we validate these claims with empirical evidence. We evaluate the effectiveness of our algorithm AgA through benchmark environments for testing mixed-motive collaboration with small-scale agents such as the two-player public good game and the sequential social dilemma games, Cleanup and Harvest, as well as our self-developed large-scale environment in the game StarCraft II.

Ziyan Wang, Zhicheng Zhang, Fei Fang et al. · cmu

Designing effective reward functions in multi-agent reinforcement learning (MARL) is a significant challenge, often leading to suboptimal or misaligned behaviors in complex, coordinated environments. We introduce Multi-agent Reinforcement Learning from Multi-phase Human Feedback of Mixed Quality ($\text{M}^3\text{HF}$), a novel framework that integrates multi-phase human feedback of mixed quality into the MARL training process. By involving humans with diverse expertise levels to provide iterative guidance, $\text{M}^3\text{HF}$ leverages both expert and non-expert feedback to continuously refine agents' policies. During training, we strategically pause agent learning for human evaluation, parse feedback using large language models to assign it appropriately and update reward functions through predefined templates and adaptive weights by using weight decay and performance-based adjustments. Our approach enables the integration of nuanced human insights across various levels of quality, enhancing the interpretability and robustness of multi-agent cooperation. Empirical results in challenging environments demonstrate that $\text{M}^3\text{HF}$ significantly outperforms state-of-the-art methods, effectively addressing the complexities of reward design in MARL and enabling broader human participation in the training process.

Yue Wang, Qizhou Wang, Feng Liu et al.

Large language model (LLM) unlearning has demonstrated its essential role in removing privacy and copyright-related responses, crucial for their legal and safe applications. However, the pursuit of complete unlearning often comes with substantial costs due to its compromises in their general functionality, leading to a notorious trade-off between unlearning and retention. It motivates this paper to explore enhanced unlearning schemes that can mitigate this trade-off. Specifically, we propose Gradient Rectified Unlearning (GRU), an improved framework that regulates the directions of gradient updates during the unlearning procedure such that their side impacts on other, unrelated responses can be minimized. GRU is easy and general to implement, demonstrating practical effectiveness across a variety of well-established unlearning benchmarks.

Ziyan Wang, Yali Du, Yudi Zhang et al. · cmu

Offline Multi-agent Reinforcement Learning (MARL) is valuable in scenarios where online interaction is impractical or risky. While independent learning in MARL offers flexibility and scalability, accurately assigning credit to individual agents in offline settings poses challenges because interactions with an environment are prohibited. In this paper, we propose a new framework, namely Multi-Agent Causal Credit Assignment (MACCA), to address credit assignment in the offline MARL setting. Our approach, MACCA, characterizing the generative process as a Dynamic Bayesian Network, captures relationships between environmental variables, states, actions, and rewards. Estimating this model on offline data, MACCA can learn each agent's contribution by analyzing the causal relationship of their individual rewards, ensuring accurate and interpretable credit assignment. Additionally, the modularity of our approach allows it to seamlessly integrate with various offline MARL methods. Theoretically, we proved that under the setting of the offline dataset, the underlying causal structure and the function for generating the individual rewards of agents are identifiable, which laid the foundation for the correctness of our modeling. In our experiments, we demonstrate that MACCA not only outperforms state-of-the-art methods but also enhances performance when integrated with other backbones.

Runze Liu, Chenjia Bai, Jiafei Lyu et al.

Reward engineering is one of the key challenges in Reinforcement Learning (RL). Preference-based RL effectively addresses this issue by learning from human feedback. However, it is both time-consuming and expensive to collect human preference labels. In this paper, we propose a novel \textbf{V}ision-\textbf{L}anguage \textbf{P}reference learning framework, named \textbf{VLP}, which learns a vision-language preference model to provide preference feedback for embodied manipulation tasks. To achieve this, we define three types of language-conditioned preferences and construct a vision-language preference dataset, which contains versatile implicit preference orders without human annotations. The preference model learns to extract language-related features, and then serves as a preference annotator in various downstream tasks. The policy can be learned according to the annotated preferences via reward learning or direct policy optimization. Extensive empirical results on simulated embodied manipulation tasks demonstrate that our method provides accurate preferences and generalizes to unseen tasks and unseen language instructions, outperforming the baselines by a large margin.

Wei Liu, Siya Qi, Xinyu Wang et al. · tsinghua

Recent advances such as DeepSeek R1-Zero highlight the effectiveness of incentive training, a reinforcement learning paradigm that computes rewards solely based on the final answer part of a language model's output, thereby encouraging the generation of intermediate reasoning steps. However, these methods fundamentally rely on external verifiers, which limits their applicability to domains like mathematics and coding where such verifiers are readily available. Although reward models can serve as verifiers, they require high-quality annotated data and are costly to train. In this work, we propose NOVER, NO-VERifier Reinforcement Learning, a general reinforcement learning framework that requires only standard supervised fine-tuning data with no need for an external verifier. NOVER enables incentive training across a wide range of text-to-text tasks and outperforms the model of the same size distilled from large reasoning models such as DeepSeek R1 671B by 7.7 percent. Moreover, the flexibility of NOVER enables new possibilities for optimizing large language models, such as inverse incentive training.

Xingzhou Lou, Junge Zhang, Ziyan Wang et al. · cmu

Safe reinforcement learning (RL) agents accomplish given tasks while adhering to specific constraints. Employing constraints expressed via easily-understandable human language offers considerable potential for real-world applications due to its accessibility and non-reliance on domain expertise. Previous safe RL methods with natural language constraints typically adopt a recurrent neural network, which leads to limited capabilities when dealing with various forms of human language input. Furthermore, these methods often require a ground-truth cost function, necessitating domain expertise for the conversion of language constraints into a well-defined cost function that determines constraint violation. To address these issues, we proposes to use pre-trained language models (LM) to facilitate RL agents' comprehension of natural language constraints and allow them to infer costs for safe policy learning. Through the use of pre-trained LMs and the elimination of the need for a ground-truth cost, our method enhances safe policy learning under a diverse set of human-derived free-form natural language constraints. Experiments on grid-world navigation and robot control show that the proposed method can achieve strong performance while adhering to given constraints. The usage of pre-trained LMs allows our method to comprehend complicated constraints and learn safe policies without the need for ground-truth cost at any stage of training or evaluation. Extensive ablation studies are conducted to demonstrate the efficacy of each part of our method.

Zihao Guo, Shuqing Shi, Richard Willis et al.

Sequential social dilemmas pose a significant challenge in the field of multi-agent reinforcement learning (MARL), requiring environments that accurately reflect the tension between individual and collective interests. Previous benchmarks and environments, such as Melting Pot, provide an evaluation protocol that measures generalization to new social partners in various test scenarios. However, running reinforcement learning algorithms in traditional environments requires substantial computational resources. In this paper, we introduce SocialJax, a suite of sequential social dilemma environments and algorithms implemented in JAX. JAX is a high-performance numerical computing library for Python that enables significant improvements in operational efficiency. Our experiments demonstrate that the SocialJax training pipeline achieves at least 50\texttimes{} speed-up in real-time performance compared to Melting Pot RLlib baselines. Additionally, we validate the effectiveness of baseline algorithms within SocialJax environments. Finally, we use Schelling diagrams to verify the social dilemma properties of these environments, ensuring that they accurately capture the dynamics of social dilemmas.

Hui Sun, Yun-Ji Zhang, Zheng Xie et al.

Selecting LLM-generated code candidates using LLM-generated tests is challenging because the tests themselves may be incorrect. Existing methods either treat all tests equally or rely on ad-hoc heuristics to filter unreliable tests. Yet determining test correctness requires knowing which codes are correct, creating a \emph{circular dependency}. Our key insight is that we need not determine test correctness at all: \emph{test votes should rank, not merely count}. What matters is not how many codes pass a test, but whether the test can \emph{distinguish} correct from incorrect code. We break the circular dependency via leave-one-out evaluation: hold out one test, rank codes by their aggregate scores on all remaining tests, and measure whether the held-out test's pass/fail pattern agrees with this ranking. We formalize this agreement as the leave-one-out AUC~(LOO-AUC) and prove that the expected LOO-AUC is proportional to each test's ability to separate correct code from incorrect code. Building on this, we propose \textbf{ACES}~(\textbf{A}UC \textbf{C}onsist\textbf{E}ncy \textbf{S}coring) with two complementary variants: ACES-C provides closed-form weights that provably approximate the oracle in expectation under a mild assumption on average test quality; ACES-O drops this assumption and iteratively optimizes a differentiable LOO-AUC objective. Both operate solely on the binary pass matrix with negligible overhead, and achieve state-of-the-art Pass@$k$ on multiple code generation benchmarks.

Yali Du, Hui Sun, Ming Li

Code translation migrates codebases across programming languages. Recently, large language models (LLMs) have achieved significant advancements in software mining. However, handling the syntactic structure of source code remains a challenge. Classic syntax-aware methods depend on intricate model architectures and loss functions, rendering their integration into LLM training resource-intensive. This paper employs in-context learning (ICL), which directly integrates task exemplars into the input context, to post-incorporate code structural knowledge into pre-trained LLMs. We revisit exemplar selection in ICL from an information-theoretic perspective, proposing that list-wise selection based on information coverage is more precise and general objective than traditional methods based on combining similarity and diversity. To address the challenges of quantifying information coverage, we introduce a surrogate measure, Coverage of Abstract Syntax Tree (CAST). Furthermore, we formulate the NP-hard CAST maximization for exemplar selection and prove that it is a standard submodular maximization problem. Therefore, we propose a greedy algorithm for CAST submodular maximization, which theoretically guarantees a (1-1/e)-approximate solution in polynomial time complexity. Our method is the first training-free and model-agnostic approach to post-incorporate code structural knowledge into existing LLMs at test time. Experimental results show that our method significantly improves LLMs performance and reveals two meaningful insights: 1) Code structural knowledge can be effectively post-incorporated into pre-trained LLMs during inference, despite being overlooked during training; 2) Scaling up model size or training data does not lead to the emergence of code structural knowledge, underscoring the necessity of explicitly considering code syntactic structure.

Xingzhou Lou, Junge Zhang, Timothy J. Norman et al.

Multi-Agent Policy Gradient (MAPG) has made significant progress in recent years. However, centralized critics in state-of-the-art MAPG methods still face the centralized-decentralized mismatch (CDM) issue, which means sub-optimal actions by some agents will affect other agent's policy learning. While using individual critics for policy updates can avoid this issue, they severely limit cooperation among agents. To address this issue, we propose an agent topology framework, which decides whether other agents should be considered in policy gradient and achieves compromise between facilitating cooperation and alleviating the CDM issue. The agent topology allows agents to use coalition utility as learning objective instead of global utility by centralized critics or local utility by individual critics. To constitute the agent topology, various models are studied. We propose Topology-based multi-Agent Policy gradiEnt (TAPE) for both stochastic and deterministic MAPG methods. We prove the policy improvement theorem for stochastic TAPE and give a theoretical explanation for the improved cooperation among agents. Experiment results on several benchmarks show the agent topology is able to facilitate agent cooperation and alleviate CDM issue respectively to improve performance of TAPE. Finally, multiple ablation studies and a heuristic graph search algorithm are devised to show the efficacy of the agent topology.

Yudi Zhang, Pei Xiao, Lu Wang et al.

In-context learning (ICL) and Retrieval-Augmented Generation (RAG) have gained attention for their ability to enhance LLMs' reasoning by incorporating external knowledge but suffer from limited contextual window size, leading to insufficient information injection. To this end, we propose a novel framework, RuAG, to automatically distill large volumes of offline data into interpretable first-order logic rules, which are injected into LLMs to boost their reasoning capabilities. Our method begins by formulating the search process relying on LLMs' commonsense, where LLMs automatically define head and body predicates. Then, RuAG applies Monte Carlo Tree Search (MCTS) to address the combinational searching space and efficiently discover logic rules from data. The resulting logic rules are translated into natural language, allowing targeted knowledge injection and seamless integration into LLM prompts for LLM's downstream task reasoning. We evaluate our framework on public and private industrial tasks, including natural language processing, time-series, decision-making, and industrial tasks, demonstrating its effectiveness in enhancing LLM's capability over diverse tasks.

Zhixun Chen, Yali Du, David Mguni

Many leading language models (LMs) use high-intensity computational resources both during training and execution. This poses the challenge of lowering resource costs for deployment and faster execution of decision-making tasks among others. We introduce a novel plug-and-play LM framework named Language Optimising Network Distribution (LONDI) framework. LONDI learns to selectively employ large LMs only where complex decision-making and reasoning are required while using low-resource LMs (i.e. LMs require less GPU usage, but may not be able to solve the problem alone) everywhere else. LONDI consists of a system of two (off-)policy networks, an LM, a large LM (LLM), and a reinforcement learning module that uses switching controls to quickly learn which system states to call the LLM. We then introduce a variant of LONDI that maintains budget constraints on LLM calls and hence its resource usage. Theoretically, we prove LONDI learns the subset of system states to activate the LLM required to solve the task. We then prove that LONDI converges to optimal solutions while also preserving budgetary constraints on LLM calls almost surely enabling it to solve various tasks while significantly lowering computational costs. We test LONDI's performance in a range of tasks in ScienceWorld and BabyAI-Text and demonstrate that LONDI can solve tasks only solvable by resource-intensive LLMs while reducing GPU usage by up to 30%.