LLM-as-a-Judge refers to the automatic modeling of preferences for responses generated by Large Language Models (LLMs), which is of significant importance for both LLM evaluation and reward modeling. Although generative LLMs have made substantial progress in various tasks, their performance as LLM-Judge still falls short of expectations. In this work, we propose Think-J, which improves generative LLM-as-a-Judge by learning how to think. We first utilized a small amount of curated data to develop the model with initial judgment thinking capabilities. Subsequently, we optimize the judgment thinking traces based on reinforcement learning (RL). We propose two methods for judgment thinking optimization, based on offline and online RL, respectively. The offline method requires training a critic model to construct positive and negative examples for learning. The online method defines rule-based reward as feedback for optimization. Experimental results showed that our approach can significantly enhance the evaluation capability of generative LLM-Judge, surpassing both generative and classifier-based LLM-Judge without requiring extra human annotations.

Pei Wang, Yanan Wu, Xiaoshuai Song, Weixun Wang, Gengru Chen, Zhongwen Li, Kezhong Yan, Qi Liu, Ken Deng, Shuaibing Zhao, Shaopan Xiong, Xuepeng Liu, Xuefeng Chen, Wanxi Deng, Wenbo Su, Bo Zheng. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Baolin Zheng, Guanlin Chen, Qingyang Teng, Hongqiong Zhong, Yingshui Tan, Zhendong Liu, Weixun Wang, Jiaheng Liu, Jian Yang, Huiyun Jing, Jincheng Wei, Wenbo Su, Xiaoyong Zhu, Bo Zheng, Kaifu Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Yancheng He, Shilong Li, Jiaheng Liu, Yingshui Tan, Weixun Wang, Hui Huang, Xingyuan Bu, Hangyu Guo, Chengwei Hu, Boren Zheng, Zhuoran Lin, Dekai Sun, Zhicheng Zheng, Wenbo Su, Bo Zheng. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Yancheng He, Shilong Li, Jiaheng Liu, Weixun Wang, Xingyuan Bu, Ge Zhang, Z.y. Peng, Zhaoxiang Zhang, Zhicheng Zheng, Wenbo Su, Bo Zheng. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Despite many breakthroughs in recent years, it is still hard for MultiAgent Reinforcement Learning (MARL) algorithms to directly solve complex tasks in MultiAgent Systems (MASs) from scratch. In this work, we study how to use Automatic Curriculum Learning (ACL) to reduce the number of environmental interactions required to learn a good policy. In order to solve a difficult task, ACL methods automatically select a sequence of tasks (i.e., curricula). The idea is to obtain maximum learning progress towards the final task by continuously learning on tasks that match the current capabilities of the learners. The key question is how to measure the learning progress of the learner for better curriculum selection. We propose a novel ACL framework, PrOgRessive mulTiagent Automatic curricuLum (PORTAL), for MASs. PORTAL selects curricula according to two critera: 1) How difficult is a task, relative to the learners’ current abilities? 2) How similar is a task, relative to the final task? By learning a shared feature space between tasks, PORTAL is able to characterize different tasks based on the distribution of features and select those that are similar to the final task. Also, the shared feature space can effectively facilitate the policy transfer between curricula. Experimental results show that PORTAL can train agents to master extremely hard cooperative tasks, which can not be achieved with previous state-of-the-art MARL algorithms.

Knowledge transfer in cooperative multi-agent reinforcement learning (MARL) has drawn increasing attention in recent years. Unlike generalizing policies in single-agent tasks, it is more important to consider coordination knowledge than individual knowledge in multi-agent transfer learning. However, most of the existing methods only focus on knowledge transfer of the individual agent policy, which leads to coordination bias and finally affects the final performance in cooperative MARL. In this paper, we propose a level-adaptive MARL framework called “LA-QTransformer”, to realize the knowledge transfer on the coordination level via efficiently decomposing the agent coordination into multi-level coalition patterns for different agents. Compatible with centralized training with decentralized execution (CTDE) regime, LA-QTransformer utilizes the Level- Adaptive Transformer to generate suitable coalition patterns and then realizes the credit assignment for each agent. Besides, to deal with unexpected changes in the number of agents in the coordination transfer phase, we design a policy network called “Population invariant agent with Transformer (PIT)” to adapt dynamic observation and action space. We evaluate the LAQTransformer and PIT in the StarCraft II micro-management benchmark by comparing them with several state-of-the-art MARL baselines. The experimental results demonstrate the superiority of LA-QTransformer and PIT and verify the feasibility of coordination knowledge transfer.

A significant challenge facing researchers in the area of multi-agent reinforcement learning (MARL) pertains to the identification of a library that can offer fast and compatible development for multi-agent tasks and algorithm combinations, while obviating the need to consider compatibility issues. In this paper, we present MARLlib, a library designed to address the aforementioned challenge by leveraging three key mechanisms: 1) a standardized multi-agent environment wrapper, 2) an agent-level algorithm implementation, and 3) a flexible policy mapping strategy. By utilizing these mechanisms, MARLlib can effectively disentangle the intertwined nature of the multi-agent task and the learning process of the algorithm, with the ability to automatically alter the training strategy based on the current task's attributes. The MARLlib library's source code is publicly accessible on GitHub: https://github.com/Replicable-MARL/MARLlib.

In many real-world multi-agent systems, the sparsity of team rewards often makes it difficult for an algorithm to successfully learn a cooperative team policy. At present, the common way for solving this problem is to design some dense individual rewards for the agents to guide the cooperation. However, most existing works utilize individual rewards in ways that do not always promote teamwork and sometimes are even counterproductive. In this paper, we propose Individual Reward Assisted Team Policy Learning (IRAT), which learns two policies for each agent from the dense individual reward and the sparse team reward with discrepancy constraints for updating the two policies mutually. Experimental results in different scenarios, such as the Multi-Agent Particle Environment and the Google Research Football Environment, show that IRAT significantly outperforms the baseline methods and can greatly promote team policy learning without deviating from the original team objective, even when the individual rewards are misleading or conflict with the team rewards.

Learning in real-world multiagent tasks is challenging due to the usual partial observability of each agent. Previous efforts alleviate the partial observability by historical hidden states with Recurrent Neural Networks, however, they do not consider the multiagent characters that either the multiagent observation consists of a number of object entities or the action space shows clear entity interactions. To tackle these issues, we propose the Agent Transformer Memory (ATM) network with a transformer-based memory. First, ATM utilizes the transformer to enable the unified processing of the factored environmental entities and memory. Inspired by the human’s working memory process where a limited capacity of information temporarily held in mind can effectively guide the decision-making, ATM updates its fixed-capacity memory with the working memory updating schema. Second, as agents' each action has its particular interaction entities in the environment, ATM parses the action space to introduce this action’s semantic inductive bias by binding each action with its specified involving entity to predict the state-action value or logit. Extensive experiments on the challenging SMAC and Level-Based Foraging environments validate that ATM could boost existing multiagent RL algorithms with impressive learning acceleration and performance improvement.

Multi-agent reinforcement learning is difficult to apply in practice, partially because of the gap between simulated and real-world scenarios. One reason for the gap is that simulated systems always assume that agents can work normally all the time, while in practice, one or more agents may unexpectedly “crash” during the coordination process due to inevitable hardware or software failures. Such crashes destroy the cooperation among agents and lead to performance degradation. In this work, we present a formal conceptualization of a cooperative multi-agent reinforcement learning system with unexpected crashes. To enhance the robustness of the system to crashes, we propose a coach-assisted multi-agent reinforcement learning framework that introduces a virtual coach agent to adjust the crash rate during training. We have designed three coaching strategies (fixed crash rate, curriculum learning, and adaptive crash rate) and a re-sampling strategy for our coach agent. To our knowledge, this work is the first to study unexpected crashes in a multi-agent system. Extensive experiments on grid-world and StarCraft II micromanagement tasks demonstrate the efficacy of the adaptive strategy compared with the fixed crash rate strategy and curriculum learning strategy. The ablation study further illustrates the effectiveness of our re-sampling strategy.

Transfer Learning has shown great potential to enhance single-agent Reinforcement Learning (RL) efficiency. Similarly, Multiagent RL (MARL) can also be accelerated if agents can share knowledge with each other. However, it remains a problem of how an agent should learn from other agents. In this paper, we propose a novel Multiagent Policy Transfer Framework (MAPTF) to improve MARL efficiency. MAPTF learns which agent's policy is the best to reuse for each agent and when to terminate it by modeling multiagent policy transfer as the option learning problem. Furthermore, in practice, the option module can only collect all agent's local experiences for update due to the partial observability of the environment. While in this setting, each agent's experience may be inconsistent with each other, which may cause the inaccuracy and oscillation of the option-value's estimation. Therefore, we propose a novel option learning algorithm, the successor representation option learning to solve it by decoupling the environment dynamics from rewards and learning the option-value under each agent's preference. MAPTF can be easily combined with existing deep RL and MARL approaches, and experimental results show it significantly boosts the performance of existing methods in both discrete and continuous state spaces.

In large-scale multi-agent systems, the large number of agents and complex game relationship cause great difficulty for policy learning. Therefore, simplifying the learning process is an important research issue. In many multi-agent systems, the interactions between agents often happen locally, which means that agents neither need to coordinate with all other agents nor need to coordinate with others all the time. Traditional methods attempt to use pre-defined rules to capture the interaction relationship between agents. However, the methods cannot be directly used in a large-scale environment due to the difficulty of transforming the complex interactions between agents into rules. In this paper, we model the relationship between agents by a complete graph and propose a novel game abstraction mechanism based on two-stage attention network (G2ANet), which can indicate whether there is an interaction between two agents and the importance of the interaction. We integrate this detection mechanism into graph neural network-based multi-agent reinforcement learning for conducting game abstraction and propose two novel learning algorithms GA-Comm and GA-AC. We conduct experiments in Traffic Junction and Predator-Prey. The results indicate that the proposed methods can simplify the learning process and meanwhile get better asymptotic performance compared with state-of-the-art algorithms.

A lot of efforts have been devoted to investigating how agents can learn effectively and achieve coordination in multiagent systems. However, it is still challenging in large-scale multiagent settings due to the complex dynamics between the environment and agents and the explosion of state-action space. In this paper, we design a novel Dynamic Multiagent Curriculum Learning (DyMA-CL) to solve large-scale problems by starting from learning on a multiagent scenario with a small size and progressively increasing the number of agents. We propose three transfer mechanisms across curricula to accelerate the learning process. Moreover, due to the fact that the state dimension varies across curricula, and existing network structures cannot be applied in such a transfer setting since their network input sizes are fixed. Therefore, we design a novel network structure called Dynamic Agent-number Network (DyAN) to handle the dynamic size of the network input. Experimental results show that DyMA-CL using DyAN greatly improves the performance of large-scale multiagent learning compared with state-of-the-art deep reinforcement learning approaches. We also investigate the influence of three transfer mechanisms across curricula through extensive simulations.

Bipartite b-matching is fundamental in algorithm design, and has been widely applied into diverse applications, such as economic markets, labor markets, etc. These practical problems usually exhibit two distinct features: large-scale and dynamic, which requires the matching algorithm to be repeatedly executed at regular intervals. However, existing exact and approximate algorithms usually fail in such settings due to either requiring intolerable running time or too much computation resource. To address this issue, based on a key observation that the matching instances vary not too much, we propose NeuSearcher which leverage the knowledge learned from previously instances to solve new problem instances. Specifically, we design a multichannel graph neural network to predict the threshold of the matched edges, by which the search region could be significantly reduced. We further propose a parallel heuristic search algorithm to iteratively improve the solution quality until convergence. Experiments on both open and industrial datasets demonstrate that NeuSearcher can speed up 2 to 3 times while achieving exactly the same matching solution compared with the state-of-the-art approximation approaches.

Transfer learning has shown great potential to accelerate Reinforcement Learning (RL) by leveraging prior knowledge from past learned policies of relevant tasks. Existing approaches either transfer previous knowledge by explicitly computing similarities between tasks or select appropriate source policies to provide guided explorations. However, how to directly optimize the target policy by alternatively utilizing knowledge from appropriate source policies without explicitly measuring the similarities is currently missing. In this paper, we propose a novel Policy Transfer Framework (PTF) by taking advantage of this idea. PTF learns when and which source policy is the best to reuse for the target policy and when to terminate it by modeling multi-policy transfer as an option learning problem. PTF can be easily combined with existing DRL methods and experimental results show it significantly accelerates RL and surpasses state-of-the-art policy transfer methods in terms of learning efficiency and final performance in both discrete and continuous action spaces.

Reinforcement learning agents usually learn from scratch, which requires a large number of interactions with the environment. This is quite different from the learning process of human. When faced with a new task, human naturally have the common sense and use the prior knowledge to derive an initial policy and guide the learning process afterwards. Although the prior knowledge may be not fully applicable to the new task, the learning process is significantly sped up since the initial policy ensures a quick-start of learning and intermediate guidance allows to avoid unnecessary exploration. Taking this inspiration, we propose knowledge guided policy network (KoGuN), a novel framework that combines human prior suboptimal knowledge with reinforcement learning. Our framework consists of a fuzzy rule controller to represent human knowledge and a refine module to finetune suboptimal prior knowledge. The proposed framework is end-to-end and can be combined with existing policy-based reinforcement learning algorithm. We conduct experiments on several control tasks. The empirical results show that our approach, which combines suboptimal human knowledge and RL, achieves significant improvement on learning efficiency of flat RL algorithms, even with very low-performance human prior knowledge.

Reward shaping is an effective technique for incorporating domain knowledge into reinforcement learning (RL). Existing approaches such as potential-based reward shaping normally make full use of a given shaping reward function. However, since the transformation of human knowledge into numeric reward values is often imperfect due to reasons such as human cognitive bias, completely utilizing the shaping reward function may fail to improve the performance of RL algorithms. In this paper, we consider the problem of adaptively utilizing a given shaping reward function. We formulate the utilization of shaping rewards as a bi-level optimization problem, where the lower level is to optimize policy using the shaping rewards and the upper level is to optimize a parameterized shaping weight function for true reward maximization. We formally derive the gradient of the expected true reward with respect to the shaping weight function parameters and accordingly propose three learning algorithms based on different assumptions. Experiments in sparse-reward cartpole and MuJoCo environments show that our algorithms can fully exploit beneficial shaping rewards, and meanwhile ignore unbeneficial shaping rewards or even transform them into beneficial ones.

In E-commerce advertising, where product recommendations and product ads are presented to users simultaneously, the traditional setting is to display ads at fixed positions. However, under such a setting, the advertising system loses the flexibility to control the number and positions of ads, resulting in sub-optimal platform revenue and user experience. Consequently, major e-commerce platforms (e.g., Taobao.com) have begun to consider more flexible ways to display ads. In this paper, we investigate the problem of advertising with adaptive exposure: can we dynamically determine the number and positions of ads for each user visit under certain business constraints so that the platform revenue can be increased? More specifically, we consider two types of constraints: request-level constraint ensures user experience for each user visit, and platform-level constraint controls the overall platform monetization rate. We model this problem as a Constrained Markov Decision Process with per-state constraint (psCMDP) and propose a constrained two-level reinforcement learning approach to decompose the original problem into two relatively independent sub-problems. To accelerate policy learning, we also devise a constrained hindsight experience replay mechanism. Experimental evaluations on industry-scale real-world datasets demonstrate the merits of our approach in both obtaining higher revenue under the constraints and the effectiveness of the constrained hindsight experience replay mechanism.

The Iterated Prisoner's Dilemma has guided research on social dilemmas for decades. However, it distinguishes between only two atomic actions: cooperate and defect. In real-world prisoner's dilemmas, these choices are temporally extended and different strategies may correspond to sequences of actions, reflecting grades of cooperation. We introduce a Sequential Prisoner's Dilemma (SPD) game to better capture the aforementioned characteristics. In this work, we propose a deep multiagent reinforcement-learning approach that investigates the evolution of mutual cooperation in SPD games. Our approach consists of two phases. The first phase is offline: it synthesizes policies with different cooperation degrees and then trains a cooperation degree detection network. The second phase is online: an agent adaptively selects its policy based on the detected degree of opponent cooperation. The effectiveness of our approach is demonstrated in two representative SPD 2D games: the Apple-Pear game and the Fruit Gathering game. Experimental results show that our strategy can avoid being exploited by exploitative opponents and achieve cooperation with cooperative opponents.

The ongoing scaling of semiconductor technology is causing severe increase of on-chip power density and temperature in microprocessors. This urgently requires both power and thermal management during system design. In this paper, we propose a model checking-based technique using extended timed automata to solve the processor frequency assignment problem in a temperature and energy-constrained multitasking system. We also develop a polynomial time-approximation algorithm to address the state-space explosion problem caused by symbolic model checker. Our approximation scheme is guaranteed to not generate any false-positive answer, while it may return false-negative answer in rare cases. Our method is universally applicable since it is independent of any system and task characteristics. Experimental results demonstrate the usefulness of our approach.

In recent years, efficient dynamic reconfiguration techniques have been widely employed for system optimization. Dynamic cache reconfiguration is a promising approach for reducing energy consumption as well as for improving overall system performance. It is a major challenge to introduce cache reconfiguration into real-time multitasking systems, since dynamic analysis may adversely affect tasks with timing constraints. This article presents a novel approach for implementing cache reconfiguration in soft real-time systems by efficiently leveraging static analysis during runtime to minimize energy while maintaining the same service level. To the best of our knowledge, this is the first attempt to integrate dynamic cache reconfiguration in real-time scheduling techniques. Our experimental results using a wide variety of applications have demonstrated that our approach can significantly reduce the cache energy consumption in soft real-time systems (up to 74%).

System optimization techniques are widely used to improve energy efficiency as well as overall performance. Dynamic voltage scaling (DVS) is well studied and known to be successful in reducing processor energy consumption. Due to the increasing significance of the memory subsystem's energy consumption, dynamic cache reconfiguration (DCR) techniques are recently proposed at the aim of improving cache subsystem's energy efficiency. As the manufacturing technology scales into the order of nanometers, leakage current, which leads to static power consumption, becomes a significant contributor in the overall power dissipation. In this paper, we consider various system components and study their impact on system-wide energy consumption under different processor voltage levels as well as cache configurations. Based on the observation, we efficiently integrate DVS and DCR techniques together to make decisions judiciously so that the total energy consumption is minimized. Our studies show that considering only DVS or DCR and ignoring the impact from other system components may lead to incorrect conclusions in overall energy savings. Experimental results demonstrate that our approach outperforms existing leakage-aware DVS techniques by 47.6% and leakage-oblivious DVS + DCR technique by up to 23.5%.

Multicore architectures, especially chip multi-processors, have been widely acknowledged as a successful design paradigm. Existing approaches primarily target application-driven partitioning of the shared cache to alleviate inter-core cache interference so that both performance and energy efficiency are improved. Dynamic cache reconfiguration is a promising technique in reducing energy consumption of the cache subsystem for uniprocessor systems. In this paper, we present a novel energy optimization technique which employs both dynamic reconfiguration of private caches and partitioning of the shared cache for multicore systems with real-time tasks. Our static profiling based algorithm is designed to judiciously find beneficial cache configurations (of private caches) for each task as well as partition factors (of the shared cache) for each core so that the energy consumption is minimized while task deadline is satisfied. Experimental results using real benchmarks demonstrate that our approach can achieve 29.29% energy saving on average compared to systems employing only cache partitioning.

System optimization techniques based on dynamic voltage scaling (DVS) are widely used with the aim of reducing processor energy consumption. Inter-task DVS assigns the same voltage level to all the instances of each task. Its intra-task counterpart exploits more energy savings by assigning multiple voltage levels within each task. In this paper, we propose a voltage scaling technique, named PreDVS, which assigns voltage levels based on the task set's preemptive scheduling for hard real-time systems. Our approach is based on an approximation scheme hence can guarantee to generate solutions within a specified quality bound (e.g., within 1% of the optimal) and is different from any existing inter- or intratask DVS techniques. PreDVS exploits static time slack at a finer granularity and achieves more energy saving than inter-task scaling without introducing any extra voltage switching overhead. Moreover, it can be efficiently employed together with existing intra-task scaling techniques. Experimental results demonstrate that PreDVS can significantly reduce energy consumption and outperform the optimal inter-task voltage scaling techniques by up to 24%.

The ongoing scaling of semiconductor technology is causing severe increase of on-chip power density and temperature in microprocessors. This has raised urgent requirement for both power and thermal management during each level of system design. In this paper, we propose a formal technique based on model checking using extended timed automata to solve the processor frequency assignment problem in a temperature- and energy- constrained multitasking system. The state space explosion problem is alleviated by transforming and solving a Pseudo-Boolean satisfiability problem. Our approach is capable of finding efficient solutions under various constraints and applicable to other problem variants as well. Our method is independent of any system and task characteristics. Experimental results demonstrate the usefulness of our approach.