Yanrui Du, Fenglei Fan, Sendong Zhao, Jiawei Cao, Ming Ma, Danyang Zhao, Shuren Qi, Ting Liu, Bing Qin. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Large Language Models (LLMs) have gained significant attention for their exceptional performance across various domains. Despite their advancements, concerns persist regarding their implicit bias, which often leads to negative social impacts. Therefore, it is essential to identify the implicit bias in LLMs and investigate the potential threat posed by it. Our study focused on a specific type of implicit bias, termed the ''Yes-No'' implicit bias, which refers to LLMs' inherent tendency to favor ''Yes'' or ''No'' responses to a single instruction. By comparing the probability of LLMs generating a series of ''Yes'' versus ''No'' responses, we observed different inherent response tendencies exhibited by LLMs when faced with different instructions. To further investigate the impact of such bias, we developed an attack method called Implicit Bias In-Context Manipulation, attempting to manipulate LLMs' behavior. Specifically, we explored whether the ''Yes'' implicit bias could manipulate ''No'' responses into ''Yes'' in LLMs' responses to malicious instructions, leading to harmful outputs. Our findings revealed that the ''Yes'' implicit bias brings a significant security threat, comparable to that of carefully designed attack methods. Moreover, we offered a comprehensive analysis from multiple perspectives to deepen the understanding of this security threat, emphasizing the need for ongoing improvement in LLMs' security.

Artificial intelligence advances drug design by predicting drug properties through the encoding of drug molecules. Since different molecular representations contain complementary information, a large amount of research is currently dedicated to the fusion of these representations. However, current multimodal molecular studies rely more on single-granularity methods, resulting in the loss of atomic-level information between representations. Inspired by the success of multi-granularity approaches, we introduce MolFusion, an innovative method for multi-granularity molecular representation fusion. MolFusion comprises two core components: MolSim for molecular-level fusion and AtomAlign for atomic-level fusion. Comprehensive experiments show that our method outperforms all powerful baselines in terms of average performance across both classification and regression tasks, with a particular enhancement in regression tasks. Our code and dataset will be available on https://github.com/Mengqi97/MolFusion.

Large Language Models (LLMs) have demonstrated remarkable success in diverse natural language processing (NLP) tasks in general domains. However, LLMs sometimes generate responses with the hallucination about medical facts due to limited domain knowledge. Such shortcomings pose potential risks in the utilization of LLMs within medical contexts. To address this challenge, we propose knowledge-tuning, which leverages structured medical knowledge bases for the LLMs to grasp domain knowledge efficiently and facilitate trustworthy response generation. We also release cMedKnowQA, a Chinese medical knowledge question-answering dataset constructed from medical knowledge bases to assess the medical knowledge proficiency of LLMs. Experimental results show that the LLMs which are knowledge-tuned with cMedKnowQA can exhibit higher levels of accuracy in response generation compared with vanilla instruction-tuning and offer a new trustworthy way for the domain adaptation of LLMs. We release our code and data at https://github.com/SCIR-HI/Huatuo-Llama-Med-Chinese .

Molecule discovery serves as a cornerstone in numerous scientific domains, fueling the development of new materials and innovative drug designs. Recent developments of in-silico molecule discovery have highlighted the promising results of cross-modal techniques, which bridge molecular structures with their descriptive annotations. However, these cross-modal methods frequently encounter the issue of data scarcity, hampering their performance and application. In this paper, we address the low-resource challenge by utilizing artificially-real data generated by Large Language Models (LLMs). We first introduce a retrieval-based prompting strategy to construct high-quality pseudo data, then explore the optimal method to effectively leverage this pseudo data. Experiments show that using pseudo data for domain adaptation outperforms all existing methods, while also requiring a smaller model scale, reduced data size and lower training cost, highlighting its efficiency. Furthermore, our method shows a sustained improvement as the volume of pseudo data increases, revealing the great potential of pseudo data in advancing low-resource cross-modal molecule discovery.

Deep learning is now widely used in drug discovery, providing significant acceleration and cost reduction. As the most fundamental building block, molecular representation is essential for predicting molecular properties to enable various downstream applications. Most existing methods attempt to incorporate more information to learn better representations. However, not all features are equally important for a specific task. Ignoring this would potentially compromise the training efficiency and predictive accuracy. To address this issue, we propose a novel approach, which treats language models as an agent and molecular pretraining models as a knowledge base. The agent accentuates task-relevant features in the molecular representation by understanding the natural language description of the task, just as a tailor customizes clothes for clients. Thus, we call this approach MolTailor. Evaluations demonstrate MolTailor's superior performance over baselines, validating the efficacy of enhancing relevance for molecular representation learning. This illustrates the potential of language model guided optimization to better exploit and unleash the capabilities of existing powerful molecular representation methods. Our code and appendix are available at https://github.com/SCIR-HI/MolTailor.

Large Language Models (LLMs) are gaining widespread attention for their potential across various fields, particularly within the medical domain. Recent efforts have aimed to privatize general-domain LLMs into specialized medical-domain LLMs by feeding high-quality, medical-domain training data. However, an overlooked aspect in these privatization efforts is the Dual Logic Ability of LLMs. This ability enables LLMs to comprehend pairs of logically opposed questions, ensuring stance consistency in their responses. In our study, we investigate two primary questions: Q1. How does privatization affect the dual logic ability of LLMs? Q2. How can we maintain the robustness of LLMs’ dual logic ability after privatization? To explore these questions, we first constructed a medical-domain dual logic ability evaluation dataset comprising logically opposed question pairs, created manually by NLP experts. By examining the stance consistency in responses to logically opposed question pairs, our analysis demonstrates a significant decline in the dual logic ability of LLMs after privatization. Furthermore, we construct privatization data to investigate the effects of the pre-training and instruction fine-tuning stages on the dual logic ability of LLMs. Interestingly, our findings reveal that the instruction fine-tuning stage often inadvertently compromises the LLMs’ dual logic ability although it is not the trainers’ intention. To counteract this, we incorporated general-domain dual logic data derived from basic science during the instruction fine-tuning stage, which are automatically constructed by our designed pipeline. Experiment results show that privatized LLMs can generalize dual logic ability from general-domain dual logic data, leading to their enhanced performance in the medical domain. Our study underscores the importance of prioritizing LLMs’ dual logic ability during the privatization process and establishes a benchmark for future research. Our data and code are available at GitHub1.

Large Language Models (LLMs) are increasingly deployed in various applications. As their usage grows, concerns regarding their safety are rising, especially in maintaining harmless responses when faced with malicious instructions. Many defense strategies have been developed to enhance the safety of LLMs. However, our research finds that existing defense strategies lead LLMs to predominantly adopt a rejection-oriented stance, thereby diminishing the usability of their responses to benign instructions. To solve this problem, we introduce the MoGU framework, designed to enhance LLMs' safety while preserving their usability. Our MoGU framework transforms the base LLM into two variants: the usable LLM and the safe LLM, and further employs dynamic routing to balance their contribution. When encountering malicious instructions, the router will assign a higher weight to the safe LLM to ensure that responses are harmless. Conversely, for benign instructions, the router prioritizes the usable LLM, facilitating usable and helpful responses. On various open-sourced LLMs, we compare multiple defense strategies to verify the superiority of our MoGU framework. Besides, our analysis provides key insights into the effectiveness of MoGU and verifies that our designed routing mechanism can effectively balance the contribution of each variant by assigning weights. Our work released the safer Llama2, Vicuna, Falcon, Dolphin, and Baichuan2.

Recent research has revealed that deep neural networks often take dataset biases as a shortcut to make decisions rather than understand tasks, leading to failures in real-world applications. In this study, we focus on the spurious correlation between word features and labels that models learn from the biased data distribution of training data. In particular, we define the word highly co-occurring with a specific label as biased word, and the example containing biased word as biased example. Our analysis shows that biased examples are easier for models to learn, while at the time of prediction, biased words make a significantly higher contribution to the models' predictions, and models tend to assign predicted labels over-relying on the spurious correlation between words and labels. To mitigate models' over-reliance on the shortcut (i.e. spurious correlation), we propose a training strategy Less-Learn-Shortcut (LLS): our strategy quantifies the biased degree of the biased examples and down-weights them accordingly. Experimental results on Question Matching, Natural Language Inference and Sentiment Analysis tasks show that LLS is a task-agnostic strategy and can improve the model performance on adversarial data while maintaining good performance on in-domain data.