With the growing popularity of general-purpose Large Language Models (LLMs), comes a need for more global explanations of model behaviors. Concept-based explanations arise as a promising avenue for explaining high-level patterns learned by LLMs. Yet their evaluation poses unique challenges, especially due to their non-local nature and high dimensional representation in a model’s hidden space. Current methods approach concepts from different perspectives, lacking a unified formalization. This makes evaluating the core measures of concepts, namely faithfulness or readability, challenging. To bridge the gap, we introduce a formal definition of concepts generalizing to diverse concept-based explanations’ settings. Based on this, we quantify the faithfulness of a concept explanation via perturbation. We ensure adequate perturbation in the high-dimensional space for different concepts via an optimization problem. Readability is approximated via an automatic and deterministic measure, quantifying the coherence of patterns that maximally activate a concept while aligning with human understanding. Finally, based on measurement theory, we apply a meta-evaluation method for evaluating these measures, generalizable to other types of explanations or tasks as well. Extensive experimental analysis has been conducted to inform the selection of explanation evaluation measures.

Hallucination detection is a challenging task for large language models (LLMs), and existing studies heavily rely on powerful closed-source LLMs such as GPT-4. In this paper, we propose an autonomous LLM-based agent framework, called HaluAgent, which enables relatively smaller LLMs (e.g. Baichuan2-Chat 7B) to actively select suitable tools for detecting multiple hallucination types such as text, code, and mathematical expression. In HaluAgent, we integrate the LLM, multi-functional toolbox, and design a fine-grained three-stage detection framework along with memory mechanism. To facilitate the effectiveness of HaluAgent, we leverage existing Chinese and English datasets to synthesize detection trajectories for fine-tuning, which endows HaluAgent with the capability for bilingual hallucination detection. Extensive experiments demonstrate that only using 2K samples for tuning LLMs, HaluAgent can perform hallucination detection on various types of tasks and datasets, achieving performance comparable to or even higher than GPT-4 without tool enhancements on both in-domain and out-of-domain datasets.

Aligning large language models (LLMs) with human expectations requires high-quality instructional dialogues, which can be achieved by raising diverse, in-depth, and insightful instructions that deepen interactions. Existing methods target instructions from real instruction dialogues as a learning goal and fine-tune a user simulator for posing instructions. However, the user simulator struggles to implicitly model complex dialogue flows and pose high-quality instructions. In this paper, we take inspiration from the cognitive abilities inherent in human learning and propose the explicit modeling of complex dialogue flows through instructional strategy reuse. Specifically, we first induce high-level strategies from various real instruction dialogues. These strategies are applied to new dialogue scenarios deductively, where the instructional strategies facilitate high-quality instructions. Experimental results show that our method can generate diverse, in-depth, and insightful instructions for a given dialogue history. The constructed multi-turn instructional dialogues can outperform competitive baselines on the downstream chat model.

Although chain-of-thought (CoT) prompting combined with language models has achieved encouraging results on complex reasoning tasks, the naive greedy decoding used in CoT prompting usually causes the repetitiveness and local optimality. To address this shortcoming, ensemble-optimization tries to obtain multiple reasoning paths to get the final answer assembly. However, current ensemble-optimization methods either simply employ rule-based post-processing such as self-consistency, or train an additional model based on several task-related human annotations to select the best one among multiple reasoning paths, yet fail to generalize to realistic settings where the type of input questions is unknown or the answer format of reasoning paths is unknown. To avoid their limitations, we propose Self-Agreement, a generalizable ensemble-optimization method applying in almost all scenarios where the type of input questions and the answer format of reasoning paths may be known or unknown. Self-agreement firstly samples from language model’s decoder to generate a diverse set of reasoning paths, and subsequently prompts the language model one more time to determine the optimal answer by selecting the most agreed answer among the sampled reasoning paths. Self-agreement simultaneously achieves remarkable performance on six public reasoning benchmarks and superior generalization capabilities.

Reinforcement learning (RL) has been widely used in training large language models (LLMs) for preventing unexpected outputs, e.g., reducing harmfulness and errors. However, existing RL methods mainly adopt instance-level reward, which cannot provide fine-grained supervision for complex reasoning tasks. As a result, the RL training cannot be fully aware of the specific part or step that actually leads to the incorrectness in model response. To address it, we propose a new RL method named RLMEC that incorporates a generative model as the reward model, which is trained by the erroneous solution rewriting task under the minimum editing constraint, which can produce token-level supervision for RL training. Based 0on the generative reward model, we design the token-level RL objective for training and an imitation-based regularization for stabilizing RL process. And these two objectives focus on the revision of the key tokens for the erroneous solution, reducing the effect of other unimportant tokens. Experiment results on 8 tasks have demonstrated the effectiveness of our approach. Our code and data will be publicly released.

Large Language Models (LLMs) have gained increasing attention for their remarkable capacity, alongside concerns about safety arising from their potential to produce harmful content. Red teaming aims to find prompts that could elicit harmful responses from LLMs, and is essential to discover and mitigate safety risks before real-world deployment. However, manual red teaming is both time-consuming and expensive, rendering it unscalable. In this paper, we propose RTPE, a scalable evolution framework to evolve red teaming prompts across both breadth and depth dimensions, facilitating the automatic generation of numerous high-quality and diverse red teaming prompts. Specifically, in-breadth evolving employs a novel enhanced in-context learning method to create a multitude of quality prompts, whereas in-depth evolving applies customized transformation operations to enhance both content and form of prompts, thereby increasing diversity. Extensive experiments demonstrate that RTPE surpasses existing representative automatic red teaming methods on both attack success rate and diversity. In addition, based on 4,800 red teaming prompts created by RTPE, we further provide a systematic analysis of 8 representative LLMs across 8 sensitive topics.

The safety of Large Language Models (LLMs) has gained increasing attention in recent years, but there still lacks a comprehensive approach for detecting safety issues within LLMs’ responses in an aligned, customizable and explainable manner. In this paper, we propose ShieldLM, an LLM-based safety detector, which aligns with common safety standards, supports customizable detection rules, and provides explanations for its decisions. To train ShieldLM, we compile a large bilingual dataset comprising 14,387 query-response pairs, annotating the safety of responses based on various safety standards. Through extensive experiments, we demonstrate that ShieldLM surpasses strong baselines across four test sets, showcasing remarkable customizability and explainability. Besides performing well on standard detection datasets, ShieldLM has also been shown to be effective as a safety evaluator for advanced LLMs. ShieldLM is released at https://github.com/thu-coai/ShieldLM to support accurate and explainable safety detection under various safety standards.

Large language models (LLMs) have achieved remarkable breakthroughs in new dialogue capabilities by leveraging instruction tuning,which refreshes human impressions of dialogue systems. The long-standing goal of dialogue systems is to be human-like enough to establish long-term connections with users. Therefore, there has been an urgent need to evaluate LLMs as human-like dialogue systems. In this paper, we propose DialogBench, a dialogue evaluation benchmark that contains 12 dialogue tasks to probe the capabilities of LLMs as human-like dialogue systems should have. Specifically, we prompt GPT-4 to generate evaluation instances for each task. We first design the basic prompt based on widely used design principles and further mitigate the existing biases to generate higher-quality evaluation instances. Our extensive tests on English and Chinese DialogBench of 26 LLMs show that instruction tuning improves the human likeness of LLMs to a certain extent, but most LLMs still have much room for improvement as human-like dialogue systems. Interestingly, results also show that the positioning of assistant AI can make instruction tuning weaken the human emotional perception of LLMs and their mastery of information about human daily life.

Humans often interact with large language models (LLMs) in multi-turn interaction to obtain desired answers or more information. However, most existing studies overlook the multi-turn instruction following ability of LLMs, in terms of training dataset, training method, and evaluation benchmark. In this paper, we introduce Parrot, a solution aiming to enhance multi-turn instruction following for LLMs. First, we introduce an efficient but effective method for collecting multi-turn instructions that feature human-like queries, such as anaphora and ellipsis. Second, we propose a context-aware preference optimization strategy to further enhance LLMs for complex queries in multi-turn interaction. Moreover, to quantitatively evaluate LLMs in multi-turn instruction following, we manually build a multi-turn benchmark derived from existing ones. Extensive experiments show that Parrot improves current LLMs by up to 7.2% in multi-turn instruction following. Our dataset and codes will be open-sourced to facilitate future research.

Large language models have demonstrated exceptional capability in natural language understanding and generation. However, their generation speed is limited by the inherently sequential nature of their decoding process, posing challenges for real-time applications. This paper introduces Lexical Unit Decoding (LUD), a novel decoding methodology implemented in a data-driven manner, accelerating the decoding process without sacrificing output quality. The core of our approach is the observation that a pre-trained language model can confidently predict multiple contiguous tokens, forming the basis for a lexical unit, in which these contiguous tokens could be decoded in parallel. Extensive experiments validate that our method substantially reduces decoding time while maintaining generation quality, i.e., 33% speed up on natural language generation with no quality loss, and 30% speed up on code generation with a negligible quality loss of 3%. Distinctively, LUD requires no auxiliary models and does not require changes to existing architectures. It can also be integrated with other decoding acceleration methods, thus achieving an even more pronounced inference efficiency boost. We posit that the foundational principles of LUD could define a new decoding paradigm for future language models, enhancing their applicability for a broader spectrum of applications. All codes are be publicly available at https://github.com/tjunlp-lab/Lexical-Unit-Decoding-LUD-.