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.

Recent studies have demonstrated the effectiveness of LLM test-time scaling. However, existing approaches to incentivize LLMs’ deep thinking abilities generally require large-scale data or significant training efforts. Meanwhile, it remains unclear how to improve the thinking abilities of less powerful base models. In this work, we introduce S^2R, an efficient framework that enhances LLM reasoning by teaching models to self-verify and self-correct during inference. Specifically, we first initialize LLMs with iterative self-verification and self-correction behaviors through supervised fine-tuning on carefully curated data. The self-verification and self-correction skills are then further strengthened by outcome-level and process-level reinforcement learning with minimized resource requirements. Our results demonstrate that, with only 3.1k behavior initialization samples, Qwen2.5-math-7B achieves an accuracy improvement from 51.0% to 81.6%, outperforming models trained on an equivalent amount of long-CoT distilled data. We also discuss the effect of different RL strategies on enhancing LLMs’ deep reasoning. Extensive experiments and analysis based on three base models across both in-domain and out-of-domain benchmarks validate the effectiveness of S^2R.

Evaluating the step-by-step reliability of large language model (LLM) reasoning, such as Chain-of-Thought, remains challenging due to the difficulty and cost of obtaining high-quality step-level supervision. In this paper, we introduce Self-Play Critic (SPC), a novel approach where a critic model evolves its ability to assess reasoning steps through adversarial self-play games, eliminating the need for manual step-level annotation. SPC involves fine-tuning two copies of a base model to play two roles, namely a "sneaky generator" that deliberately produces erroneous steps designed to be difficult to detect, and a "critic" that analyzes the correctness of reasoning steps. These two models engage in an adversarial game in which the generator aims to fool the critic, while the critic model seeks to identify the generator's errors. Using reinforcement learning based on the game outcomes, the models iteratively improve; the winner of each confrontation receives a positive reward and the loser receives a negative reward, driving continuous self-evolution. Experiments on three reasoning process benchmarks (ProcessBench, PRM800K, DeltaBench) demonstrate that our SPC progressively enhances its error detection capabilities (e.g., accuracy increases from 70.8% to 77.7% on ProcessBench) and surpasses strong baselines, including distilled R1 model. Furthermore, SPC can guide the test-time search of diverse LLMs and significantly improve their mathematical reasoning performance on MATH500 and AIME2024, surpassing those guided by state-of-the-art process reward models.

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.

Xin Zhou, Ping Nie, Yiwen Guo, Haojie Wei, Zhanqiu Zhang, Pasquale Minervini, Ruotian Ma, Tao Gui, Qi Zhang, Xuanjing Huang. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Ruotian Ma, Xuanting Chen, Zhang Lin, Xin Zhou, Junzhe Wang, Tao Gui, Qi Zhang, Xiang Gao, Yun Wen Chen. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Recently, many studies have illustrated the robustness problem of Named Entity Recognition (NER) systems: the NER models often rely on superficial entity patterns for predictions, without considering evidence from the context.Consequently, even state-of-theart NER models generalize poorly to outof-domain scenarios when out-of-distribution (OOD) entity patterns are introduced.Previous research attributes the robustness problem to the existence of NER dataset bias, where simpler and regular entity patterns induce shortcut learning.In this work, we bring new insights into this problem by comprehensively investigating the NER dataset bias from a dataset difficulty view.We quantify the entity-context difficulty distribution in existing datasets and explain their relationship with model robustness.Based on our findings, we explore three potential ways to de-bias the NER datasets by altering entity-context distribution, and we validate the feasibility with intensive experiments.Finally, we show that the de-biased datasets can transfer to different models and even benefit existing model-based robustness-improving methods, indicating that building more robust datasets is fundamental for building more robust NER systems. * Equal contribution. † Corresponding authors.Ms. Hall expects Cathay 's profit to grow around 13 % annually this year.

Large pre-trained language models (PLMs) have demonstrated superior performance in industrial applications. Recent studies have explored parameter-efficient PLM tuning, which only updates a small amount of task-specific parameters while achieving both high efficiency and comparable performance against standard fine-tuning. However, all these methods ignore the inefficiency problem caused by the task-specific output layers, which is inflexible for us to re-use PLMs and introduces non-negligible parameters. In this work, we focus on the text classification task and propose plugin-tuning, a framework that further improves the efficiency of existing parameter-efficient methods with a unified classifier. Specifically, we re-formulate both token and sentence classification tasks into a unified language modeling task, and map label spaces of different tasks into the same vocabulary space. In this way, we can directly re-use the language modeling heads of PLMs, avoiding introducing extra parameters for different tasks. We conduct experiments on six classification benchmarks. The experimental results show that plugin-tuning can achieve comparable performance against fine-tuned PLMs, while further saving around 50% parameters on top of other parameter-efficient methods.

Multilingual BERT (mBERT) has demonstrated considerable cross-lingual syntactic ability, whereby it enables effective zero-shot cross-lingual transfer of syntactic knowledge.The transfer is more successful between some languages, but it is not well understood what leads to this variation and whether it fairly reflects difference between languages.In this work, we investigate the distributions of grammatical relations induced from mBERT in the context of 24 typologically different languages.We demonstrate that the distance between the distributions of different languages is highly consistent with the syntactic difference in terms of linguistic formalisms.Such difference learnt via self-supervision plays a crucial role in the zero-shot transfer performance and can be predicted by variation in morphosyntactic properties between languages.These results suggest that mBERT properly encodes languages in a way consistent with linguistic diversity and provide insights into the mechanism of crosslingual transfer.

Xin Zhou, Jinzhu Lu, Tao Gui, Ruotian Ma, Zichu Fei, Yuran Wang, Yong Ding, Yibo Cheung, Qi Zhang, Xuanjing Huang. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022.

Input distribution shift is one of the vital problems in unsupervised domain adaptation (UDA). The most popular UDA approaches focus on domain-invariant representation learning, trying to align the features from different domains into a similar feature distribution. However, these approaches ignore the direct alignment of input word distributions between domains, which is a vital factor in word-level classification tasks such as cross-domain NER. In this work, we shed new light on cross-domain NER by introducing a subword-level solution, X-Piece, for input word-level distribution shift in NER. Specifically, we re-tokenize the input words of the source domain to approach the target subword distribution, which is formulated and solved as an optimal transport problem. As this approach focuses on the input level, it can also be combined with previous DIRL methods for further improvement. Experimental results show the effectiveness of the proposed method based on BERT-tagger on four benchmark NER datasets. Also, the proposed method is proved to benefit DIRL methods such as DANN.

Ruotian Ma, Xin Zhou, Tao Gui, Yiding Tan, Linyang Li, Qi Zhang, Xuanjing Huang. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Ruotian Ma, Tao Gui, Linyang Li, Qi Zhang, Xuanjing Huang, Yaqian Zhou. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Pre-Trained Models have been widely applied and recently proved vulnerable under backdoor attacks: the released pre-trained weights can be maliciously poisoned with certain triggers.When the triggers are activated, even the fine-tuned model will predict pre-defined labels, causing a security threat.These backdoors generated by the poisoning methods can be erased by changing hyper-parameters during fine-tuning or detected by finding the triggers.In this paper, we propose a stronger weight-poisoning attack method that introduces a layerwise weight poisoning strategy to plant deeper backdoors; we also introduce a combinatorial trigger that cannot be easily detected.The experiments on text classification tasks show that previous defense methods cannot resist our weight-poisoning method, which indicates that our method can be widely applied and may provide hints for future model robustness studies.

Recently, many works have tried to augment the performance of Chinese named entity recognition (NER) using word lexicons.As a representative, Lattice-LSTM (Zhang and Yang, 2018) has achieved new benchmark results on several public Chinese NER datasets.However, Lattice-LSTM has a complex model architecture.This limits its application in many industrial areas where real-time NER responses are needed.In this work, we propose a simple but effective method for incorporating the word lexicon into the character representations.This method avoids designing a complicated sequence modeling architecture, and for any neural NER model, it requires only subtle adjustment of the character representation layer to introduce the lexicon information.Experimental studies on four benchmark Chinese NER datasets show that our method achieves an inference speed up to 6.15 times faster than those of state-ofthe-art methods, along with a better performance.The experimental results also show that the proposed method can be easily incorporated with pre-trained models like BERT. 1 * Equal contribution.

Adversarial attacks for discrete data (such as texts) have been proved significantly more challenging than continuous data (such as images) since it is difficult to generate adversarial samples with gradient-based methods.Current successful attack methods for texts usually adopt heuristic replacement strategies on the character or word level, which remains challenging to find the optimal solution in the massive space of possible combinations of replacements while preserving semantic consistency and language fluency.In this paper, we propose BERT-Attack, a high-quality and effective method to generate adversarial samples using pre-trained masked language models exemplified by BERT.We turn BERT against its fine-tuned models and other deep neural models in downstream tasks so that we can successfully mislead the target models to predict incorrectly.Our method outperforms state-of-theart attack strategies in both success rate and perturb percentage, while the generated adversarial samples are fluent and semantically preserved.Also, the cost of calculation is low, thus possible for large-scale generations.The code is available at https://github.com/ LinyangLee/BERT-Attack.

Character-level Chinese named entity recognition (NER) that applies long short-term memory (LSTM) to incorporate lexicons has achieved great success. However, this method fails to fully exploit GPU parallelism and candidate lexicons can conflict. In this work, we propose a faster alternative to Chinese NER: a convolutional neural network (CNN)-based method that incorporates lexicons using a rethinking mechanism. The proposed method can model all the characters and potential words that match the sentence in parallel. In addition, the rethinking mechanism can address the word conflict by feeding back the high-level features to refine the networks. Experimental results on four datasets show that the proposed method can achieve better performance than both word-level and character-level baseline methods. In addition, the proposed method performs up to 3.21 times faster than state-of-the-art methods, while realizing better performance.