Quyu Kong, Xu Zhang, Zhenyu Yang, Nolan Gao, Chen Liu, Panrong Tong, Chenglin Cai, Hanzhang Zhou, Jianan Zhang, Liangyu Chen, Zhidan Liu, Steven Hoi, Yue Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Yue Wang, Ruotian Ma, Xingyu Chen, Zhengliang Shi, Morunliu Yang, Wanshun Chen, Huang Liu, Jiadi Yao, Xin He, Qu Yang, Qingxuan Jiang, Fanghua Ye, Juntao Li, Zhaopeng Tu, Xiaolong Li, Liefeng Bo, Min Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Attributed text generation aims to enhance the reliability of content generated from large language models by providing citations for each claim, which thereby enables users to easily verify the correctness of the responses.However, the scarcity of high-quality training samples presents a significant challenge in aligning large language models to generate texts with citations, revealing considerable room for improvement in existing attribution systems.Besides, existing approaches of aligning large language models to follow user instructions can lead to an undue emphasis on irrelevant documents, which in turn reduces the quality of responses.To address the above problems, we propose Automatic Alignment Framework for Attributed Text Generation (A 3 ), a novel framework designed to automatically generate highquality attributed query-response pairs for both supervised fine-tuning and preference optimization stages without human annotation.With the help of A 3 , Mistral-7B can achieve a citation recall of 84.4 and a precision of 87.0 precision on ASQA, which notably surpasses GPT-4's citation recall of 73.0 and precision of 76.5. 1

Mixture-of-Experts (MoE) architectures within Large Reasoning Models (LRMs) have achieved impressive reasoning capabilities by selectively activating experts to facilitate structured cognitive processes. Despite notable advances, existing reasoning models often suffer from cognitive inefficiencies like overthinking and underthinking. To address these limitations, we introduce a novel inference-time steering methodology called Reinforcing Cognitive Experts (RICE), designed to improve reasoning depth and efficiency without additional training or complex heuristics. Leveraging normalized Pointwise Mutual Information (nPMI), we systematically identify specialized experts, termed cognitive experts that orchestrate meta-level reasoning operations characterized by tokens like <think>. Empirical evaluations with leading MoE-based LRMs (DeepSeek-R1 and Qwen3-235B) on rigorous quantitative and scientific reasoning benchmarks (AIME and GPQA Diamond) demonstrate noticeable and consistent improvements in reasoning accuracy, cognitive efficiency, and cross-domain generalization. Crucially, our lightweight approach substantially outperforms prevalent reasoning-steering techniques, such as prompt design and decoding constraints, while preserving the model's general instruction-following skills. These results highlight reinforcing cognitive experts as a promising, practical, and interpretable direction to enhance cognitive efficiency within advanced reasoning models.

Long reasoning models (LRMs) such as OpenAI's o1 and DeepSeek's R1 have demonstrated remarkable abilities in complex reasoning tasks by scaling test-time compute and exhibiting human-like deep thinking. However, we identify a phenomenon we term underthinking, where LRMs frequently switch between different reasoning thoughts without sufficiently exploring promising paths to reach a correct solution. This behavior leads to inadequate depth of reasoning and decreased performance, particularly on challenging mathematical problems. To systematically analyze this issue, we conduct experiments on three challenging test sets and two representative open-source LRMs, revealing that frequent thought switching correlates with incorrect responses. We introduce a novel metric to quantify underthinking by measuring token efficiency in incorrect answers. To address underthinking, we propose a decoding strategy with thought switching penalty (Tip) that discourages premature transitions between thoughts, encouraging deeper exploration of each reasoning path. Experimental results demonstrate that our approach improves accuracy across challenging datasets without requiring model fine-tuning. Our findings contribute to understanding reasoning inefficiencies in LRMs and offer a practical solution to enhance their problem-solving capabilities. Our code is open-source and available at https://github.com/wangyuenlp/underthinking.

Improving Multi-modal Large Language Models (MLLMs) in the post-training stage typically relies on supervised fine-tuning (SFT) or reinforcement learning (RL), which require expensive and manually annotated multi-modal data--an ultimately unsustainable resource. This limitation has motivated a growing interest in unsupervised paradigms as a third stage of post-training after SFT and RL. While recent efforts have explored this direction, their methods are complex and difficult to iterate. To address this, we propose MM-UPT, a simple yet effective framework for unsupervised post-training of MLLMs, enabling continual self-improvement without any external supervision. The training method of MM-UPT builds upon GRPO, replacing traditional reward signals with a self-rewarding mechanism based on majority voting over multiple sampled responses. Our experiments demonstrate that such training method effectively improves the reasoning ability of Qwen2.5-VL-7B (e.g., 66.3\%$\rightarrow$72.9\% on MathVista, 62.9\%$\rightarrow$68.7\% on We-Math), using standard dataset without ground truth labels. To further explore scalability, we extend our framework to a data self-generation setting, designing two strategies that prompt the MLLM to synthesize new training samples on its own. Additional experiments show that combining these synthetic data with the unsupervised training method can also boost performance, highlighting a promising approach for scalable self-improvement. Overall, MM-UPT offers a new paradigm for autonomous enhancement of MLLMs, serving as a critical third step after initial SFT and RL in the absence of external supervision. Our code is available at \url{https://github.com/waltonfuture/MM-UPT}.

Large Language Models (LLMs) hold considerable promise for artificial general intelligence, given their intrinsic abilities to accomplish a wide range of open-domain tasks either independently or in tandem with specialized expert models. However, despite these capabilities, the performance of LLMs has yet to be comprehensively evaluated in realistic scenarios. To this end, in this work, we introduce a novel task, the Realistic Chinese Spell Checking (RCSC), to evaluate the effectiveness of existing methods comprehensively. In contrast to existing works that solely address Chinese character misspellings or pinyin conversions, our task aims to convert the realistic Chinese text into the corresponding correct text. The realistic Chinese text may potentially contain both Chinese misspellings and pinyin conversions. We first present the Realistic Chinese Spell Checking Benchmark (RCSCB), which consists of two subsets and contains a total of 581,657 samples. Then, we benchmark the performance of various baselines and find that all the existing methods, including instruction-based LLMs, achieve unsatisfactory results on RCSCB. To further improve the performance on RCSCB, we propose Pinyin-Enhanced Spell Checker (PESC), which is specifically designed to address pinyin-related misspellings. Experimental results demonstrate that PESC can achieve state-of-the-art performance on RCSCB. Despite the progress made, the current state-of-the-art performance is still far from satisfactory. We expect further progress on this crucial and challenging task.

Misspellings in search engine queries may prevent search engines from returning accurate results. For Chinese mobile search engines, due to the different input methods (e.g., hand-written and T9 input methods), more types of misspellings exist, making this problem more challenging. As an essential module of search engines, Chinese Spelling Check~(CSC) models aim to detect and correct misspelled Chinese characters from user-issued queries. Despite the great value of CSC to the search engine, there is no CSC benchmark collected from real-world search engine queries. To fill this blank, we construct and release the Alipay Search Engine Query (AlipaySEQ) spelling check dataset. To the best of our knowledge, AlipaySEQ is the first Chinese Spelling Check dataset collected from the real-world scenario of Chinese mobile search engines. It consists of 15,522 high-quality human annotated and 1,175,151 automatically generated samples. To demonstrate the unique challenges of AlipaySEQ in the era of Large Language Models~(LLMs), we conduct a thorough study to analyze the difference between AlipaySEQ and existing SIGHAN benchmarks and compare the performance of various baselines, including existing task-specific methods and LLMs. We observe that all baselines fail to perform satisfactorily due to the over-correction problem. Especially, LLMs exhibit below-par performance on AlipaySEQ, which is rather surprising. Therefore, to alleviate the over-correction problem, we introduce a model-agnostic CSC Self-Refine Framework (SRF) to construct a strong baseline. Comprehensive experiments demonstrate that our proposed SRF, though more effective against existing models on both the AlipaySEQ and SIGHAN15, is still far from achieving satisfactory performance on our real-world dataset. With the newly collected real-world dataset and strong baseline, we hope more progress can be achieved on such a challenging and valuable task.

Randomness is widely introduced in neural network training to simplify model optimization or avoid the over-fitting problem. Among them, dropout and its variations in different aspects (e.g., data, model structure) are prevalent in regularizing the training of deep neural networks. Though effective and performing well, the randomness introduced by these dropout-based methods causes nonnegligible inconsistency between training and inference. In this paper, we introduce a simple consistency training strategy to regularize such randomness, namely R-Drop, which forces two output distributions sampled by each type of randomness to be consistent. Specifically, R-Drop minimizes the bidirectional KL-divergence between two output distributions produced by dropout-based randomness for each training sample. Theoretical analysis reveals that R-Drop can reduce the above inconsistency by reducing the inconsistency among the sampled sub structures and bridging the gap between the loss calculated by the full model and sub structures. Experiments on 7 widely-used deep learning tasks ( 23 datasets in total) demonstrate that R-Drop is universally effective for different types of neural networks (i.e., feed-forward, recurrent, and graph neural networks) and different learning paradigms (supervised, parameter-efficient, and semi-supervised). In particular, it achieves state-of-the-art performances with the vanilla Transformer model on WMT14 English → German translation ( 30.91 BLEU) and WMT14 English → French translation ( 43.95 BLEU), even surpassing models trained with extra large-scale data and expert-designed advanced variants of Transformer models.

Dropout is a powerful and widely used technique to regularize the training of deep neural networks. Though effective and performing well, the randomness introduced by dropout causes unnegligible inconsistency between training and inference. In this paper, we introduce a simple consistency training strategy to regularize dropout, namely R-Drop, which forces the output distributions of different sub models generated by dropout to be consistent with each other. Specifically, for each training sample, R-Drop minimizes the bidirectional KL-divergence between the output distributions of two sub models sampled by dropout. Theoretical analysis reveals that R-Drop reduces the above inconsistency. Experiments on $\bf{5}$ widely used deep learning tasks ($\bf{18}$ datasets in total), including neural machine translation, abstractive summarization, language understanding, language modeling, and image classification, show that R-Drop is universally effective. In particular, it yields substantial improvements when applied to fine-tune large-scale pre-trained models, e.g., ViT, RoBERTa-large, and BART, and achieves state-of-the-art (SOTA) performances with the vanilla Transformer model on WMT14 English$\to$German translation ($\bf{30.91}$ BLEU) and WMT14 English$\to$French translation ($\bf{43.95}$ BLEU), even surpassing models trained with extra large-scale data and expert-designed advanced variants of Transformer models. Our code is available at GitHub\footnote{\url{https://github.com/dropreg/R-Drop}}.