Large Language Models (LLMs), trained on a large amount of corpus, have demonstrated remarkable abilities. However, it may not be sufficient to directly apply open-source LLMs like Llama to certain real-world scenarios, since most of them are trained for \emph{general} purposes. Thus, the demands for customizing publicly available LLMs emerge, but are currently under-studied. In this work, we consider customizing pre-trained LLMs with new human preferences. Specifically, the LLM should not only meet the new preference but also preserve its original capabilities after customization. Drawing inspiration from the observation that human preference can be expressed as a reward model, we propose to cast LLM customization as optimizing the sum of two reward functions, one of which (denoted as $r_1$) was used to pre-train the LLM while the other (denoted as $r_2$) characterizes the new human preference. The obstacle here is that both reward functions are unknown, making the application of modern reinforcement learning methods infeasible. Thanks to the residual Q-learning framework, we can restore the customized LLM with the pre-trained LLM and the \emph{residual Q-function} without the reward function $r_1$. Moreover, we find that for a fixed pre-trained LLM, the reward function $r_2$ can be derived from the residual Q-function, enabling us to directly learn the residual Q-function from the new human preference data upon the Bradley-Terry model. We name our method Q-Adapter as it introduces an adapter module to approximate the residual Q-function for customizing the pre-trained LLM towards the new preference. Experiments based on the Llama-3.1 model on the DSP dataset and HH-RLHF dataset illustrate the superior effectiveness of Q-Adapter on both retaining existing knowledge and learning new preferences. Our code is available at \url{https://github.com/LAMDA-RL/Q-Adapter}.

Learning from multi-agent expert demonstrations, known as Multi-Agent Imitation Learning (MAIL), provides a promising approach to sequential decision-making. However, existing MAIL methods including Behavior Cloning (BC) and Adversarial Imitation Learning (AIL) face significant challenges: BC suffers from the compounding error issue, while the very nature of adversarial optimization makes AIL prone to instability. In this work, we propose \textbf{M}ulti-\textbf{A}gent imitation by learning and sampling from \textbf{F}actor\textbf{I}zed \textbf{S}oft Q-function (MAFIS), a novel method that addresses these limitations for both online and offline MAIL settings. Built upon the single-agent IQ-Learn framework, MAFIS introduces the value decomposition network to factorize the imitation objective at agent level, thus enabling scalable training for multi-agent systems. Moreover, we observe that the soft Q-function implicitly defines the optimal policy as an energy-based model, from which we can sample actions via stochastic gradient Langevin dynamics. This allows us to estimate the gradient of the factorized optimization objective for continuous control tasks, avoiding the adversarial optimization between the soft Q-function and the policy required by prior work. By doing so, we obtain a tractable and \emph{non-adversarial} objective for both discrete and continuous multi-agent control. Experiments on common benchmarks including the discrete control tasks StarCraft Multi-Agent Challenge v2 (SMACv2), Gold Miner, and Multi Particle Environments (MPE), as well as the continuous control task Multi-Agent MuJoCo (MaMuJoCo), demonstrate that MAFIS achieves superior performance compared with baselines. Our code is available at https://github.com/LAMDA-RL/MAFIS.

Large Language Models (LLMs) have demonstrated remarkable progress in complex reasoning tasks through both post-training and test-time scaling laws. While prevalent test-time scaling approaches are often realized by using external reward models to guide the model generation process, we find that only marginal gains can be acquired when scaling a model post-trained on specific reasoning tasks. We identify that the limited improvement stems from distribution discrepancies between the specific post-trained generator and the general reward model. To address this, we propose a framework that incentivizes LLMs to self-verify their own answers. By unifying answer generation and verification within a single reinforcement learning (RL) process, we train models that can effectively assess the correctness of their own solutions. The trained model can further scale its performance at inference time by verifying its generations, without the need for external verifiers. We train our self-verification models based on Qwen2.5-Math-7B and DeepSeek-R1-Distill-Qwen-1.5B, demonstrating their capabilities across varying reasoning context lengths. Experiments on multiple mathematical reasoning benchmarks show that our models can not only improve post-training performance but also enable effective test-time scaling. Our code is available at https://github.com/mansicer/self-verification.

Imitation Learning (IL) learns from experts, on which most existing studies assume that the imitator will be deployed in stationary environments. However, real-world scenarios commonly involve perturbations, necessitating robust imitators for non-stationary scenarios. To fulfill this, we leverage a multi-modal expert dataset encompassing diverse dynamics, while still adhering to the shared goal between the experts and imitator. Different from conventional multi-modal IL work that considers reproducing the demonstrated different behaviors, we aim to imitate a policy that rapidly adapts to sudden dynamic changes, even when encountering dynamics unseen during training. We propose a method called Generalizable Multi-modal Adversarial Imitation Learning (GMAIL) for non-stationary dynamics, which adversarially trains a discriminator and a generator. Due to dynamic mismatch between the experts and the imitator, the optimal next state for the imitator may require several steps for the experts to reach, inspiring us to propose to take the state-next-state pairs within multiple steps in the demonstrated trajectories to facilitate imitation under dynamic mismatch. For quick identification of the changed dynamic, GMAIL learns a dynamics-sensitive generator by introducing a history-based context encoder. On a wide range of navigation, locomotion and autonomous driving tasks, empirical results illustrate the effectiveness of GMAIL.

Cooperative multiagent reinforcement learning (MARL) has attracted significant attention and has the potential for many real-world applications. Previous arts mainly focus on facilitating the coordination ability from different aspects (e.g., nonstationarity and credit assignment) in single-task or multitask scenarios, ignoring the stream of tasks that appear in a continual manner. This ignorance makes the continual coordination an unexplored territory, neither in problem formulation nor efficient algorithms designed. Toward tackling the mentioned issue, this article proposes an approach, multiagent continual coordination via progressive task contextualization (MACPro). The key point lies in obtaining a factorized policy, using shared feature extraction layers but separated independent task heads, each specializing in a specific class of tasks. The task heads can be progressively expanded based on the learned task contextualization. Moreover, to cater to the popular centralized training with decentralized execution (CTDE) paradigm in MARL, each agent learns to predict and adopt the most relevant policy head based on local information in a decentralized manner. We show in multiple multiagent benchmarks that existing continual learning methods fail, while MACPro is able to achieve close-to-optimal performance. More results also disclose the effectiveness of MACPro from multiple aspects, such as high generalization ability.

Low sample efficiency is an enduring challenge of reinforcement learning (RL). With the advent of versatile large language models (LLMs), recent works impart common-sense knowledge to accelerate policy learning for RL processes. However, we note that such guidance is often tailored for one specific task but loses generalizability. In this article, we introduce a framework that harnesses LLMs to extract background knowledge of an environment, which contains general understandings of the entire environment, making various downstream RL tasks benefit from one-time knowledge representation. We ground LLMs by feeding a few precollected experiences and requesting them to delineate background knowledge of the environment. Afterward, we represent the output knowledge as potential functions for potential-based reward shaping, which has a good property for maintaining policy optimality from task rewards. We instantiate three variants to prompt LLMs for background knowledge, including writing code, annotating pReferences, and assigning goals. Our experiments show that these methods achieve significant sample efficiency improvements in a spectrum of downstream tasks from Minigrid and Crafter domains.

Human beings can make adaptive decisions in a preparatory manner, i.e., by making preparations in advance, which offers significant advantages in scenarios where both online and offline experiences are expensive and limited. Meanwhile, current reinforcement learning methods commonly rely on numerous environment interactions but hardly obtain generalizable policies. In this paper, we introduce the idea of \textit{rehearsal} into policy optimization, where the agent plans for all possible outcomes in mind and acts adaptively according to actual responses from the environment. To effectively rehearse, we propose ReDM, an algorithm that generates a diverse and eligible set of dynamics models and then rehearse the policy via adaptive training on the generated model set. Rehearsal enables the policy to make decision plans for various hypothetical dynamics and to naturally generalize to previously unseen environments. Our experimental results demonstrate that ReDM is capable of learning a valid policy solely through rehearsal, even with \emph{zero} interaction data. We further extend ReDM to scenarios where limited or mismatched interaction data is available, and our experimental results reveal that ReDM produces high-performing policies compared to other offline RL baselines.

Multi-agent pathfinding (MAPF) is essential to large-scale robotic coordination tasks. Planning-based algorithms show their advantages in collision avoidance while avoiding exponential growth in the number of agents. Reinforcement-learning (RL)-based algorithms can be deployed efficiently but cannot prevent collisions entirely due to the lack of hard constraints. This paper combines the merits of planning-based and RL-based MAPF methods to propose a deployment-efficient and collision-free MAPF algorithm. The experiments show the effectiveness of our approach.

We consider the problem of learning the best possible policy from a fixed dataset, known as offline Reinforcement Learning (RL). A common taxonomy of existing offline RL works is policy regularization, which typically constrains the learned policy by distribution or support of the behavior policy. However, distribution and support constraints are overly conservative since they both force the policy to choose similar actions as the behavior policy when considering particular states. It will limit the learned policy’s performance, especially when the behavior policy is sub-optimal. In this paper, we find that regularizing the policy towards the nearest state-action pair can be more effective and thus propose Policy Regularization with Dataset Constraint (PRDC). When updating the policy in a given state, PRDC searches the entire dataset for the nearest state-action sample and then restricts the policy with the action of this sample. Unlike previous works, PRDC can guide the policy with proper behaviors from the dataset, allowing it to choose actions that do not appear in the dataset along with the given state. It is a softer constraint but still keeps enough conservatism from out-of-distribution actions. Empirical evidence and theoretical analysis show that PRDC can alleviate offline RL’s fundamentally challenging value overestimation issue with a bounded performance gap. Moreover, on a set of locomotion and navigation tasks, PRDC achieves state-of-the-art performance compared with existing methods. Code is available at https://github.com/LAMDA-RL/PRDC

Reinforcement Learning (RL) has experienced rapid advancements in recent years. The widely studied RL algorithms mainly focus on a single input form, such as pixel-based image input or symbolic vector input. These two forms have different characteristics and, in many scenarios, will appear together, while few RL algorithms have studied the problems with mixed input types. Specifically, in the scenario where both pixel and symbolic inputs are available, symbolic input usually offers abstract features with specific semantics, which is more conducive to the agent's focus. Conversely, pixel input provides more comprehensive information, enabling the agent to make well-informed decisions. Tailoring the processing approach based on the properties of these two input types can contribute to solving the problem more effectively. To tackle the above issue, we propose an Internal Logical Induction (ILI) framework that integrates deep RL and rule learning into one system. ILI utilizes the deep RL algorithm to process the pixel input and the rule learning algorithm to induce propositional logic knowledge from symbolic input. To efficiently combine these two mechanisms, we further adopt a reward shaping technique by treating valuable knowledge as intrinsic rewards for the RL procedure. Experimental results demonstrate that the ILI framework outperforms baseline approaches in RL problems with pixel-symbolic input, and its inductive knowledge exhibits transferability advantages when pixel input semantics change.

Effective communication can improve coordination in cooperative multi-agent reinforcement learning (MARL). One popular communication scheme is exchanging agents' local observations or latent embeddings and using them to augment individual local policy input. Such a communication paradigm can reduce uncertainty for local decision-making and induce implicit coordination. However, it enlarges agents' local policy spaces and increases learning complexity, leading to poor coordination in complex settings. To handle this limitation, this paper proposes a novel framework named Multi-Agent Incentive Communication (MAIC) that allows each agent to learn to generate incentive messages and bias other agents' value functions directly, resulting in effective explicit coordination. Our method firstly learns targeted teammate models, with which each agent can anticipate the teammate's action selection and generate tailored messages to specific agents. We further introduce a novel regularization to leverage interaction sparsity and improve communication efficiency. MAIC is agnostic to specific MARL algorithms and can be flexibly integrated with different value function factorization methods. Empirical results demonstrate that our method significantly outperforms baselines and achieves excellent performance on multiple cooperative MARL tasks.

Value-based multi-agent reinforcement learning (MARL) methods hold the promise of promoting coordination in cooperative settings. Popular MARL methods mainly focus on the scalability or the representational capacity of value functions. Such a learning paradigm can reduce agents' uncertainties and promote coordination. However, they fail to leverage the task structure decomposability, which generally exists in real-world multi-agent systems (MASs), leading to a significant amount of time exploring the optimal policy in complex scenarios. To address this limitation, we propose a novel framework Multi-Agent Concentrative Coordination (MACC) based on task decomposition, with which an agent can implicitly form local groups to reduce the learning space to facilitate coordination. In MACC, agents first learn representations for subtasks from their local information and then implement an attention mechanism to concentrate on the most relevant ones. Thus, agents can pay targeted attention to specific subtasks and improve coordination. Extensive experiments on various complex multi-agent benchmarks demonstrate that MACC achieves remarkable performance compared to existing methods.

Representation learning of knowledge graphs aims to project both entities and relations as vectors in a continuous low-dimensional space. Relation Hierarchical Structure (RHS), which is constructed by a generalization relationship named subRelationOf between relations, can improve the overall performance of knowledge representation learning. However, most of the existing methods ignore this critical information, and a straightforward way of considering RHS may have a negative effect on the embeddings and thus reduce the model performance. In this paper, we propose a novel method named TransRHS, which is able to incorporate RHS seamlessly into the embeddings. More specifically, TransRHS encodes each relation as a vector together with a relation-specific sphere in the same space. Our TransRHS employs the relative positions among the vectors and spheres to model the subRelationOf, which embodies the inherent generalization relationships among relations. We evaluate our model on two typical tasks, i.e., link prediction and triple classification. The experimental results show that our TransRHS model significantly outperforms all baselines on both tasks, which verifies that the RHS information is significant to representation learning of knowledge graphs, and TransRHS can effectively and efficiently fuse RHS into knowledge graph embeddings.

In order to control the continuous passive motion (CPM) machine for injured fingers, we develop an embedded control platform. We first bring forward the philosophy of function modularization design for control platforms. Then we actually begin to develop an embedded control platform of the CPM machine by using the method of function modularization. The core of the control platform consists of two main parts: the data acquisition function module and the motor control function module, both are based on the serial peripheral interface (SPI) network. The whole control platform is open-ended for new functions and applications. It can be easily expanded if we add new modules to the SPI network. Primary experiments have proved that the control platform works well and the design method of function modularization provides a new method for the design of control platforms