Automated red teaming is an effective method for identifying misaligned behaviors in large language models (LLMs). Existing approaches, however, often focus primarily on improving attack success rates while overlooking the need for comprehensive test case coverage. Additionally, most of these methods are limited to single-turn red teaming, failing to capture the multi-turn dynamics of real-world human-machine interactions. To overcome these limitations, we propose **HARM** (**H**olistic **A**utomated **R**ed tea**M**ing), which scales up the diversity of test cases using a top-down approach based on an extensible, fine-grained risk taxonomy. Our method also leverages a novel fine-tuning strategy and reinforcement learning techniques to facilitate multi-turn adversarial probing in a human-like manner. Experimental results demonstrate that our framework enables a more systematic understanding of model vulnerabilities and offers more targeted guidance for the alignment process.

Large language models (LLMs), despite their remarkable progress across various general domains, encounter significant barriers in medicine and healthcare. This field faces unique challenges such as domain-specific terminologies and reasoning over specialized knowledge. To address these issues, we propose MedAgents, a novel multi-disciplinary collaboration framework for the medical domain. MedAgents leverages LLM-based agents in a role-playing setting that participate in a collaborative multi-round discussion, thereby enhancing LLM proficiency and reasoning capabilities. This training-free framework encompasses five critical steps: gathering domain experts, proposing individual analyses, summarising these analyses into a report, iterating over discussions until a consensus is reached, and ultimately making a decision. Our work focuses on the zero-shot setting, which is applicable in real-world scenarios. Experimental results on nine datasets (MedQA, MedMCQA, PubMedQA, and six subtasks from MMLU) establish that our proposed MedAgents framework excels at mining and harnessing the medical expertise within LLMs, as well as extending its reasoning abilities. Our code can be found at https://github.com/gersteinlab/MedAgents.

Multimodal emotion recognition for video has gained considerable attention in recent years, in which three modalities (i.e., textual, visual and acoustic) are involved. Due to the diverse levels of informational content related to emotion, three modalities typically possess varying degrees of contribution to emotion recognition. More seriously, there might be inconsistencies between the emotion of individual modality and the video. The challenges mentioned above are caused by the inherent uncertainty of emotion. Inspired by the recent advances of quantum theory in modeling uncertainty, we make an initial attempt to design a quantum-inspired adaptive-priority-learning model (QAP) to address the challenges. Specifically, the quantum state is introduced to model modal features, which allows each modality to retain all emotional tendencies until the final classification. Additionally, we design Q-attention to orderly integrate three modalities, and then QAP learns modal priority adaptively so that modalities can provide different amounts of information based on priority. Experimental results on the IEMOCAP and MOSEI datasets show that QAP establishes new state-of-the-art results.

The technology of text-to-SQL has significantly enhanced the efficiency of accessing and manipulating databases. However, limited research has been conducted to study its vulnerabilities emerging from malicious user interaction. By proposing TrojanSQL, a backdoor-based SQL injection framework for text-to-SQL systems, we show how state-of-the-art text-to-SQL parsers can be easily misled to produce harmful SQL statements that can invalidate user queries or compromise sensitive information about the database. The study explores two specific injection attacks, namely boolean-based injection and union-based injection, which use different types of triggers to achieve distinct goals in compromising the parser. Experimental results demonstrate that both medium-sized models based on fine-tuning and LLM-based parsers using prompting techniques are vulnerable to this type of attack, with attack success rates as high as 99% and 89%, respectively. We hope that this study will raise more concerns about the potential security risks of building natural language interfaces to databases.

In recent years, multimodal sentiment analysis (MSA) has attracted more and more interest, which aims to predict the sentiment polarity expressed in a video. Existing methods typically 1) treat three modal features (textual, acoustic, visual) equally, without distinguishing the importance of different modalities; and 2) split the video into frames, leading to missing the global acoustic information. In this paper, we propose a global Acoustic feature enhanced Modal-Order-Aware network (AMOA) to address these problems. Firstly, a modal-order-aware network is designed to obtain the multimodal fusion feature. This network integrates the three modalities in a certain order, which makes the modality at the core position matter more. Then, we introduce the global acoustic feature of the whole video into our model. Since the global acoustic feature and multimodal fusion feature originally reside in their own spaces, contrastive learning is further employed to align them before concatenation. Experiments on two public datasets show that our model outperforms the state-of-the-art models. In addition, we also generalize our model to the sentiment with more complex semantics, such as sarcasm detection. Our model also achieves state-of-the-art performance on a widely used sarcasm dataset.

Reinforcement learning methods have emerged as a popular choice for training an efficient and effective dialogue policy. However, these methods suffer from sparse and unstable reward signals returned by a user simulator only when a dialogue finishes. Besides, the reward signal is manually designed by human experts, which requires domain knowledge. Recently, a number of adversarial learning methods have been proposed to learn the reward function together with the dialogue policy. However, to alternatively update the dialogue policy and the reward model on the fly, we are limited to policy-gradient-based algorithms, such as REINFORCE and PPO. Moreover, the alternating training of a dialogue agent and the reward model can easily get stuck in local optima or result in mode collapse. To overcome the listed issues, we propose to decompose the adversarial training into two steps. First, we train the discriminator with an auxiliary dialogue generator and then incorporate a derived reward model into a common reinforcement learning method to guide the dialogue policy learning. This approach is applicable to both on-policy and off-policy reinforcement learning methods. Based on our extensive experimentation, we can conclude the proposed method: (1) achieves a remarkable task success rate using both on-policy and off-policy reinforcement learning methods; and (2) has potential to transfer knowledge from existing domains to a new domain.

Dialogue policy learning for task-oriented dialogue systems has enjoyed great progress recently mostly through employing reinforcement learning methods. However, these approaches have become very sophisticated. It is time to re-evaluate it. Are we really making progress developing dialogue agents only based on reinforcement learning? We demonstrate how (1) traditional supervised learning together with (2) a simulator-free adversarial learning method can be used to achieve performance comparable to state-of-the-art reinforcement learning-based methods. First, we introduce a simple dialogue action decoder to predict the appropriate actions. Then, the traditional multi-label classification solution for dialogue policy learning is extended by adding dense layers to improve the dialogue agent performance. Finally, we employ the Gumbel-Softmax estimator to alternatively train the dialogue agent and the dialogue reward model without using reinforcement learning. Based on our extensive experimentation, we can conclude the proposed methods can achieve more stable and higher performance with fewer efforts, such as the domain knowledge required to design a user simulator and the intractable parameter tuning in reinforcement learning. Our main goal is not to beat RL with supervised learning, but to demonstrate the value of rethinking the role of reinforcement learning and supervised learning in optimizing task-oriented dialogue systems.