Distilling reasoning paths from teacher to student models via supervised fine-tuning (SFT) provides a shortcut for improving the reasoning ability of the smaller Large Language Models (LLMs). However, the reasoning paths generated by teacher models often reflect only surface-level traces of their underlying authentic reasoning. Insights from cognitive neuroscience suggest that authentic reasoning involves a complex interweaving between meta-reasoning that selects the appropriate sub-problem from multiple candidates, and solving, which addresses the sub-problem. It means that authentic reasoning has implicit multi-branch structure. Supervised fine-tuning collapses this rich structure into a flat sequence of token prediction in teacher's reasoning path, which cannot distill this structure to student. To address this limitation, we propose RLKD, a reinforcement learning (RL)-based distillation framework guided by a novel Generative Structure Reward Model (GSRM). Our GSRM converts the reasoning path into multiple meta-reasoning-solving steps and gives the reward to measure the alignment between the reasoning structures of student and teacher. Our RLKD combines this reward with RL, enables the student LLM to internalize the teacher’s implicit multi-branch structure in authentic reasoning, rather than merely mimicking fixed teacher's output paths. Experiments show that RLKD, even when trained on only 0.1% of the data under an RL-only regime, surpasses the performance of standard SFT-RL pipelines and further unleashes the potential reasoning ability of the student LLM than SFT-based distillation.

Zenghao Duan, Zhiyi Yin, Zhichao Shi, Liang Pang, Shaoling Jing, Zihe Huang, Jiayi Wu, Yu Yan, Jingcheng Deng, Huawei Shen, Xueqi Cheng. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Zihao Wei, Liang Pang, Jiahao Liu, Wenjie Shi, Jingcheng Deng, Shicheng Xu, Zenghao Duan, Jingang Wang, Fei Sun, Huawei Shen, Xueqi Cheng. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The extensive utilization of large language models (LLMs) underscores the crucial necessity for precise and contemporary knowledge embedded within their intrinsic parameters. Existing research on knowledge editing primarily concentrates on monolingual scenarios, neglecting the complexities presented by multilingual contexts and multi-hop reasoning. To address these challenges, our study introduces MLaKE (Multilingual Language Knowledge Editing), a novel benchmark comprising 4072 multi-hop and 5360 single-hop questions designed to evaluate the adaptability of knowledge editing methods across five languages: English, Chinese, Japanese, French, and German. MLaKE aggregates fact chains from Wikipedia across languages and utilizes LLMs to generate questions and answer. We assessed the effectiveness of current multilingual knowledge editing methods using the MLaKE dataset. Our results show that due to considerable inconsistencies in both multilingual performance and encoding efficiency, these methods struggle to generalize effectively across languages. The accuracy of these methods when editing English is notably higher than for other languages. The experimental results further demonstrate that models encode knowledge and generation capabilities for different languages using distinct parameters, leading to poor cross-lingual transfer performance in current methods. Transfer performance is notably better within the same language family compared to across different families. These findings emphasize the urgent need to improve multilingual knowledge editing methods.

Hengxing Cai, Jinhan Dong, Jingjun Tan, Jingcheng Deng, Sihang Li, Zhifeng Gao, Haidong Wang, Zicheng Su, Agachai Sumalee, Renxin Zhong. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Jingcheng Deng, Zhongtao Jiang, Liang Pang, Zihao Wei, Liwei Chen, Kun Xu, Yang Song, Huawei Shen, Xueqi Cheng. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Recent knowledge editing methods have primarily focused on modifying structured knowledge in large language models. However, this task setting overlooks the fact that a significant portion of real-world knowledge is stored in an unstructured format, characterized by long-form content, noise, and a complex yet comprehensive nature. Techniques like "local layer key-value storage" and "term-driven optimization", as used in previous methods like MEMIT, are not effective for handling unstructured knowledge. To address these challenges, we propose a novel Unstructured Knowledge Editing method, namely UnKE, which extends previous assumptions in the layer dimension and token dimension. Firstly, in the layer dimension, we propose non-local block key-value storage to replace local layer key-value storage, increasing the representation ability of key-value pairs and incorporating attention layer knowledge. Secondly, in the token dimension, we replace "term-driven optimization" with "cause-driven optimization", which edits the last token directly while preserving context, avoiding the need to locate terms and preventing the loss of context information. Results on newly proposed unstructured knowledge editing dataset (UnKEBench) and traditional structured datasets demonstrate that UnKE achieves remarkable performance, surpassing strong baselines. In addition, UnKE has robust batch editing and sequential editing capabilities.

With the application of generation models, internet is increasingly inundated with AI-generated content (AIGC), causing both real and AI-generated content indexed in corpus for search. This paper explores the impact of AI-generated images on text-image search in this scenario. Firstly, we construct a benchmark consisting of both real and AI-generated images for this study. In this benchmark, AI-generated images possess visual semantics sufficiently similar to real images. Experiments on this benchmark reveal that text-image retrieval models tend to rank the AI-generated images higher than the real images, even though the AI-generated images do not exhibit more visually relevant semantics to the queries than real images. We call this bias as invisible relevance bias. This bias is detected across retrieval models with different training data and architectures. Further exploration reveals that mixing AI-generated images into the training data of retrieval models exacerbates the invisible relevance bias. These problems cause a vicious cycle in which AI-generated images have a higher chance of exposing from massive data, which makes them more likely to be mixed into the training of retrieval models and such training makes the invisible relevance bias more and more serious. To mitigate this bias and elucidate the potential causes of the bias, firstly, we propose an effective method to alleviate this bias. Subsequently, we apply our proposed debiasing method to retroactively identify the causes of this bias, revealing that the AI-generated images induce the image encoder to embed additional information into their representation. This information makes the retriever estimate a higher relevance score. We conduct experiments to support this assertion.

The task of response selection in multi-turn dialogue is to find the best option from all candidates.In order to improve the reasoning ability of the model, previous studies pay more attention to using explicit algorithms to model the dependencies between utterances, which are deterministic, limited and inflexible.In addition, few studies consider differences between the options before and after reasoning.In this paper, we propose an Implicit Relational Reasoning Graph Network to address these issues, which consists of the Utterance Relational Reasoner (URR) and the Option Dual Comparator (ODC).URR aims to implicitly extract dependencies between utterances, as well as utterances and options, and make reasoning with relational graph convolutional networks.ODC focuses on perceiving the difference between the options through dual comparison, which can eliminate the interference of the noise options.Experimental results on two multi-turn dialogue reasoning benchmark datasets Mu-Tual and MuTual plus show that our method significantly improves the baseline of four pretrained language models and achieves state-ofthe-art performance.The model surpasses human performance for the first time on the Mu-Tual dataset.Our code is released in the link. 1