Yang Zhao, Yangou Ouyang, Xiao Ding, Hepeng Wang, Bibo Cai, Kai Xiong, Jinglong Gao, Zhouhao Sun, Li Du, Bing Qin, Ting Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Yang Zhao, Hepeng Wang, Xiao Ding, Yangou Ouyang, Bibo Cai, Kai Xiong, Jinglong Gao, Zhouhao Sun, Li Du, Bing Qin, Ting Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Yang Zhao, Li Du, Xiao Ding, Yangou Ouyang, Hepeng Wang, Kai Xiong, Jinglong Gao, Zhouhao Sun, Dongliang Xu, Qing Yang, Dongchen Li, Bing Qin, Ting Liu. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

A primary impediment to scaling reinforcement learning (RL) for large language model (LLM) training is the substantial computational cost, predominantly arising from the necessity of multi-sampling for policy optimization and evaluation. This underscores the critical yet challenging nature of efficient training data selection. Drawing inspiration from the Zone of Proximal Development (ZPD) theory, which posits that learners acquire knowledge more effectively from tasks of intermediate difficulty, we hypothesize that LLMs exhibit optimal learning from data they have not yet mastered but demonstrate the potential to comprehend. Conventional methodologies for assessing data difficulty or informativeness typically rely on computationally intensive multi-sampling or iterative procedures. To address this limitation, we introduce UFO-RL (**U**ncertainty-**F**ocused **O**ptimization for **R**einforcement **L**earning), a novel framework that employs a computationally efficient single-pass uncertainty estimation technique to identify informative training instances. This method, requiring only a single forward pass and obviating the need for iterative next-token computation, achieves a significant acceleration (up to 185$\times$) in data evaluation compared to multi-sampling approaches. UFO-RL leverages this efficient metric to select data within the model's estimated ZPD for training. Extensive experimentation across diverse LLMs and mathematical benchmarks demonstrates that training with a mere 10\% of the data, carefully selected by UFO-RL, yields performance comparable to or even surpassing that of full-data training. Furthermore, this targeted data selection results in up to a 16$\times$ reduction in overall training time, concurrently enhancing training stability and improving generalization capabilities. Thus, UFO-RL presents a practical and highly efficient strategy for scaling RL fine-tuning of LLMs by focusing learning efforts on the most informative and valuable data, thereby mitigating the computational bottlenecks associated with traditional RL training.

Zhouhao Sun, Li Du, Xiao Ding, Yixuan Ma, Yang Zhao, Kaitao Qiu, Ting Liu, Bing Qin. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Large language models (LLMs) have achieved significant performance in various natural language reasoning tasks. However, they still struggle with performing first-order logic reasoning over formal logical theories expressed in natural language. This is because the previous LLMs-based reasoning systems have the theoretical incompleteness issue. As a result, it can only address a limited set of simple reasoning problems, which significantly decreases their generalization ability. To address this issue, we propose a novel framework, named Generalizable and Faithful Reasoner (GFaiR), which introduces the paradigm of resolution refutation. Resolution refutation has the capability to solve all first-order logic reasoning problems by extending reasoning rules and employing the principle of proof by contradiction, so our system’s completeness can be improved by introducing resolution refutation. Experimental results demonstrate that our system outperforms previous works by achieving state-of-the-art performances in complex scenarios while maintaining performances in simple scenarios. Besides, we observe that GFaiR is faithful to its reasoning process.

Understanding temporal commonsense concepts, such as times of occurrence and durations is crucial for event-centric language understanding. Reasoning about such temporal concepts in a complex context requires reasoning over both the stated context and the world knowledge that underlines it. A recent study shows massive pre-trained LM still struggle with such temporal reasoning under complex contexts (e.g., dialog) because they only implicitly encode the relevant contexts and fail to explicitly uncover the underlying logical compositions for complex inference, thus may not be robust enough. In this work, we propose to augment LMs with the temporal logic induction ability, which frames the temporal reasoning by defining three modular components: temporal dependency inducer and temporal concept defuzzifier and logic validator. The former two components disentangle the explicit/implicit dependency between temporal concepts across context (before, after, ...) and the specific meaning of fuzzy temporal concepts, respectively, while the validator combines the intermediate reasoning clues for robust contextual reasoning about the temporal concepts. Extensive experimental results on TIMEDIAL, a challenging dataset for temporal reasoning over dialog, show that our method, Logic Induction Enhanced Contextualized TEmporal Reasoning (LECTER), can yield great improvements over the traditional language model for temporal reasoning.

Although achieving promising performance, current Natural Language Understanding models tend to utilize dataset biases instead of learning the intended task, which always leads to performance degradation on out-of-distribution (OOD) samples.To increase the performance stability, previous debiasing methods empirically capture bias features from data to prevent the model from corresponding biases.However, our analyses show that the empirical debiasing methods may fail to capture part of the dataset biases and mistake semantic information of input text as biases, which limits the effectiveness of debiasing.To address these issues, we propose a debiasing framework IEGDB that comprehensively detects the dataset biases to induce a set of biased features, and purify the biased features with the guidance of information entropy.Experimental results show that IEGDB can consistently improve the stability of performance on OOD datasets for a set of widely adopted NLU models.