Outcome-based reinforcement learning has made notable advances in training language models (LMs) for reasoning. However, without explicit incentives and controls, this paradigm has limitations and instability in eliciting high-quality reasoning trajectories with diverse actions—particularly for models whose pretraining lacked extensive reasoning-related data. To this end, we introduce MetaAct-RL, a new RL framework that frames LMs’ thinking as sequential decision making over meta-actions. In this framework, the model chooses and executes a high-level action at each step—such as forward reasoning, critique, or refinement—to gradually reach the correct answer. To encourage deeper exploration, richer action diversity, and to improve sampling efficiency in the RL optimization process, MetaAct-RL incorporates appropriate length-based reward and regularization, and a key-state restart mechanism. Extensive experiments across six benchmarks show that MetaAct-RL improves reasoning performance by 7.99 on Llama3.2-1B and 7.17 on Llama3.1-8B relative to vanilla RL method. Moreover, on the challenging AIME-2024, our method outperforms the vanilla RL by 7.5 with Qwen2.5-1.5B.

Deming Ding, Shichun Liu, Enhui Yang, Jiahang Lin, Ziying Chen, Shihan Dou, Honglin Guo, Weiyu Cheng, Pengyu Zhao, Chengjun Xiao, Qunhong Zeng, Qi Zhang, Xuanjing Huang, Qidi Xu, Tao Gui. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Shaofan Liu, Guoqiang Zhang, Shihan Dou, Huiyuan Zheng, Yiming Zhou, Junjie Ye, Shaowen Wang, Shichun Liu, Jiazheng Zhang, Tao Gui, Qi Zhang, Xuanjing Huang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Ming Zhang, Yujiong Shen, Jingyi Deng, Yuhui Wang, Huayu Sha, Kexin Tan, Qiyuan Peng, Yue Zhang, Junzhe Wang, Shichun Liu, Yueyuan Huang, Jingqi Tong, Changhao Jiang, Yilong Wu, Zhihao Zhang, Mingqi Wu, Mingxu Chai, Zhiheng Xi, Shihan Dou, Tao Gui, Qi Zhang, Xuanjing Huang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Shihan Dou, Jiayi Chen, Chenhao Huang, Feng Chen, Wei Chengzhi, Huiyuan Zheng, Shichun Liu, Yan Liu, Chenxiao Liu, Chao Xin, Lin Yan, Zongzhang Zhang, Tao Gui, Qi Zhang, Xuanjing Huang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

We offer a novel perspective on reward modeling by formulating it as a policy discriminator, which quantifies the difference between two policies to generate a reward signal, guiding the training policy towards a target policy with desired behaviors. Based on this conceptual insight, we propose a scalable pre-training method named POLicy DiscriminAtive LeaRning (POLAR), which trains a reward model (RM) to discern identical policies and discriminate different ones. Unlike traditional reward modeling methods relying on absolute preferences, POLAR captures the relative difference between one policy and an arbitrary target policy, which is a scalable, high-level optimization objective suitable for modeling generic ranking relationships. Leveraging the POLAR pre-training paradigm, we present a series of RMs with parameter scales from 1.8B to 7B. Empirical results show that POLAR substantially outperforms traditional non-pre-trained methods, significantly enhancing RM performance. For instance, POLAR-7B could improve preference accuracy from 54.8% to 81.0% on STEM tasks and from 57.9% to 85.5% on creative writing tasks compared to SOTA baselines. POLAR also shows robust generalization capabilities in RLHF using Reinforcement Fine-tuning (RFT), providing reliable reward signals and markedly enhancing policy performance—improving LLaMa3.1-8B from an average of 47.36% to 56.33% and Qwen2.5-32B from 64.49% to 70.47% on 20 benchmarks. Moreover, scaling experiments reveal a clear power-law relationship between computation and performance, supported by linear correlation coefficients approaching 0.99. The impressive performance, strong generalization, and scaling properties suggest that POLAR is a promising direction for developing general and strong reward models.

We introduce EvaLearn, a pioneering benchmark designed to evaluate large language models (LLMs) on their learning capability and efficiency in challenging tasks, a critical, yet underexplored aspect of model potential. EvaLearn contains 648 challenging problems across six task types, grouped into 182 sequences, each sequence dedicated to one task type. Diverging from most existing benchmarks that evaluate models in parallel, EvaLearn requires models to solve problems sequentially, allowing them to leverage the experience gained from previous solutions. EvaLearn provides five comprehensive automated metrics to evaluate models and quantify their learning capability and efficiency. We extensively benchmark nine frontier models and observe varied performance profiles: some models, such as Claude-3.7-sonnet, start with moderate initial performance but exhibit strong learning ability, while some models struggle to benefit from experience and may even show negative transfer. Moreover, we investigate model performance under two learning settings and find that instance-level rubrics and teacher-model feedback further facilitate model learning. Importantly, we observe that current LLMs with stronger static abilities do not show a clear advantage in learning capability across all tasks, highlighting that EvaLearn evaluates a new dimension of model performance. We hope EvaLearn provides a novel evaluation perspective for assessing LLM potential and understanding the gap between models and human capabilities, promoting the development of deeper and more dynamic evaluation approaches. All datasets, the automatic evaluation framework, and the results studied in this paper are available in the supplementary materials.

Recently, the evaluation of Large Language Models has emerged as a popular area of research. The three crucial questions for LLM evaluation are ``what, where, and how to evaluate''. However, the existing research mainly focuses on the first two questions, which are basically what tasks to give the LLM during testing and what kind of knowledge it should deal with. As for the third question, which is about what standards to use, the types of evaluators, how to score, and how to rank, there hasn't been much discussion. In this paper, we analyze evaluation methods by comparing various criteria with both manual and automatic evaluation, utilizing onsite, crowd-sourcing, public annotators and GPT-4, with different scoring methods and ranking systems. We propose a new dataset, LLMEval and conduct evaluations on 20 LLMs. A total of 2,186 individuals participated, leading to the generation of 243,337 manual annotations and 57,511 automatic evaluation results. We perform comparisons and analyses of different settings and conduct 10 conclusions that can provide some insights for evaluating LLM in the future. The dataset and the results are publicly available at https://github.com/llmeval. The version with the appendix are publicly available at https://arxiv.org/abs/2312.07398.

Ming Zhang, Caishuang Huang, Yilong Wu, Shichun Liu, Huiyuan Zheng, Yurui Dong, Yujiong Shen, Shihan Dou, Jun Zhao, Junjie Ye, Qi Zhang, Tao Gui, Xuanjing Huang. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

In this paper, we propose R$^3$: Learning Reasoning through Reverse Curriculum Reinforcement Learning (RL), a novel method that employs only outcome supervision to achieve the benefits of process supervision for large language models. The core challenge in applying RL to complex reasoning is to identify a sequence of actions that result in positive rewards and provide appropriate supervision for optimization. Outcome supervision provides sparse rewards for final results without identifying error locations, whereas process supervision offers step-wise rewards but requires extensive manual annotation. R$^3$ overcomes these limitations by learning from correct demonstrations. Specifically, R$^3$ progressively slides the start state of reasoning from a demonstration’s end to its beginning, facilitating easier model exploration at all stages. Thus, R$^3$ establishes a step-wise curriculum, allowing outcome supervision to offer step-level signals and precisely pinpoint errors. Using Llama2-7B, our method surpasses RL baseline on eight reasoning tasks by $4.1$ points on average. Notably, in program-based reasoning, 7B-scale models perform comparably to larger models or closed-source models with our R$^3$.