Pei Ke

Yuxian Gu, Jiaxin Wen, Hao Sun et al. · tsinghua

Large-scale pre-training has shown remarkable performance in building open-domain dialogue systems. However, previous works mainly focus on showing and evaluating the conversational performance of the released dialogue model, ignoring the discussion of some key factors towards a powerful human-like chatbot, especially in Chinese scenarios. In this paper, we conduct extensive experiments to investigate these under-explored factors, including data quality control, model architecture designs, training approaches, and decoding strategies. We propose EVA2.0, a large-scale pre-trained open-domain Chinese dialogue model with 2.8 billion parameters, and will make our models and codes publicly available. Automatic and human evaluations show that EVA2.0 significantly outperforms other open-source counterparts. We also discuss the limitations of this work by presenting some failure cases and pose some future research directions on large-scale Chinese open-domain dialogue systems.

Jiale Cheng, Xiao Liu, Kehan Zheng et al. · tsinghua

Large language models (LLMs) have shown impressive success in various applications. However, these models are often not well aligned with human intents, which calls for additional treatments on them; that is, the alignment problem. To make LLMs better follow user instructions, existing alignment methods primarily focus on further training them. However, the extra training of LLMs is usually expensive in terms of GPU computing; even worse, some LLMs are not accessible for user-demanded training, such as GPTs. In this work, we take a different perspective -- Black-Box Prompt Optimization (BPO) -- to perform alignments. The idea is to optimize user prompts to suit LLMs' input understanding, so as to best realize users' intents without updating LLMs' parameters. BPO leverages human preferences to optimize prompts, thus making it superior to LLM (e.g., ChatGPT) as a prompt engineer. Moreover, BPO is model-agnostic, and the empirical results demonstrate that the BPO-aligned ChatGPT yields a 22% increase in the win rate against its original version and 10% for GPT-4. Notably, the BPO-aligned LLMs can outperform the same models aligned by PPO and DPO, and it also brings additional performance gains when combining BPO with PPO or DPO. Code and datasets are released at https://github.com/thu-coai/BPO.

Yuxian Gu, Pei Ke, Xiaoyan Zhu et al. · tsinghua

Training language models to learn from human instructions for zero-shot cross-task generalization has attracted much attention in NLP communities. Recently, instruction tuning (IT), which fine-tunes a pre-trained language model on a massive collection of tasks described via human-craft instructions, has been shown effective in instruction learning for unseen tasks. However, IT relies on a large amount of human-annotated samples, which restricts its generalization. Unlike labeled data, unlabeled data are often massive and cheap to obtain. In this work, we study how IT can be improved with unlabeled data. We first empirically explore the IT performance trends versus the number of labeled data, instructions, and training tasks. We find it critical to enlarge the number of training instructions, and the instructions can be underutilized due to the scarcity of labeled data. Then, we propose Unlabeled Data Augmented Instruction Tuning (UDIT) to take better advantage of the instructions during IT by constructing pseudo-labeled data from unlabeled plain texts. We conduct extensive experiments to show UDIT's effectiveness in various scenarios of tasks and datasets. We also comprehensively analyze the key factors of UDIT to investigate how to better improve IT with unlabeled data. The code is publicly available at https://github.com/thu-coai/UDIT.

Xiao Liu, Xuanyu Lei, Shengyuan Wang et al. · tsinghua

Alignment has become a critical step for instruction-tuned Large Language Models (LLMs) to become helpful assistants. However, the effective evaluation of alignment for emerging Chinese LLMs is still largely unexplored. To fill in this gap, we introduce AlignBench, a comprehensive multi-dimensional benchmark for evaluating LLMs' alignment in Chinese. We design a human-in-the-loop data curation pipeline, containing eight main categories, 683 real-scenario rooted queries and corresponding human verified references. To ensure the correctness of references, each knowledge-intensive query is accompanied with evidences collected from reliable web sources (including URLs and quotations) by our annotators. For automatic evaluation, our benchmark employs a rule-calibrated multi-dimensional LLM-as-Judge~\cite{zheng2023judging} approach with Chain-of-Thought to generate explanations and final ratings, ensuring high reliability and interpretability. All evaluation code, data, and LLM generations are available at \url{https://github.com/THUDM/AlignBench}. Since its release, AlignBench has been adopted by top (Chinese) LLMs for evaluating their alignment capabilities in Chinese, including ChatGLM, Qwen, DeepSeek, Yi, Baichuan, and Abab.

Pei Ke, Bosi Wen, Zhuoer Feng et al. · tsinghua

Since the natural language processing (NLP) community started to make large language models (LLMs) act as a critic to evaluate the quality of generated texts, most of the existing works train a critique generation model on the evaluation data labeled by GPT-4's direct prompting. We observe that these models lack the ability to generate informative critiques in both pointwise grading and pairwise comparison especially without references. As a result, their generated critiques cannot provide fine-grained distinguishability on generated texts, causing unsatisfactory evaluation performance. In this paper, we propose a simple yet effective method called Eval-Instruct, which can first acquire pointwise grading critiques with pseudo references and then revise these critiques via multi-path prompting to obtain informative evaluation data in different tasks and settings, including pointwise grading and pairwise comparison with / without references. After fine-tuning on these data, the resulting model CritiqueLLM is empirically shown to outperform ChatGPT and all the open-source baselines and even achieve comparable evaluation performance to GPT-4 in system-level correlations of pointwise grading. We also demonstrate that our generated critiques can act as scalable feedback to further improve the generation quality of strong LLMs like ChatGPT.

Zhexin Zhang, Junxiao Yang, Yida Lu et al. · tsinghua

Large Language Models (LLMs) are known to be vulnerable to jailbreak attacks. An important observation is that, while different types of jailbreak attacks can generate significantly different queries, they mostly result in similar responses that are rooted in the same harmful knowledge (e.g., detailed steps to make a bomb). Consequently, unlearning-based approaches have been proposed to mitigate jailbreak attacks by directly removing harmful knowledge from the model. In this paper, we identify a novel ripple effect of unlearning, wherein LLMs can implicitly unlearn harmful knowledge that was not explicitly introduced during the unlearning phase (e.g., a model unlearning the steps for theft may also implicitly unlearn the steps for making a bomb). Through over 100 experimental runs spanning multiple models, attack strategies, and defense methods, we empirically validate this phenomenon, which makes unlearning-based methods able to decrease the Attack Success Rate on unseen data from more than 70% to less than 10% with only 100 training samples. Further analysis reveals that the strong generalization ability of unlearning may stem from the intrinsic relatedness among harmful responses across harmful questions (e.g., response patterns, shared steps and actions in response, and similarity among their learned representations in the LLM). We also discuss the potential limitations of unlearning and the observed ripple effect. We hope our research could contribute to a deeper understanding of unlearning. Our code is available at https://github.com/thu-coai/SafeUnlearning.

Zhexin Zhang, Junxiao Yang, Pei Ke et al.

While significant attention has been dedicated to exploiting weaknesses in LLMs through jailbreaking attacks, there remains a paucity of effort in defending against these attacks. We point out a pivotal factor contributing to the success of jailbreaks: the intrinsic conflict between the goals of being helpful and ensuring safety. Accordingly, we propose to integrate goal prioritization at both training and inference stages to counteract. Implementing goal prioritization during inference substantially diminishes the Attack Success Rate (ASR) of jailbreaking from 66.4% to 3.6% for ChatGPT. And integrating goal prioritization into model training reduces the ASR from 71.0% to 6.6% for Llama2-13B. Remarkably, even in scenarios where no jailbreaking samples are included during training, our approach slashes the ASR by half. Additionally, our findings reveal that while stronger LLMs face greater safety risks, they also possess a greater capacity to be steered towards defending against such attacks, both because of their stronger ability in instruction following. Our work thus contributes to the comprehension of jailbreaking attacks and defenses, and sheds light on the relationship between LLMs' capability and safety. Our code is available at \url{https://github.com/thu-coai/JailbreakDefense_GoalPriority}.

Chujie Zheng, Pei Ke, Zheng Zhang et al. · tsinghua

It has always been an important yet challenging problem to control language models to avoid generating texts with undesirable attributes, such as toxic language and unnatural repetition. We introduce Click for controllable text generation, which needs no modification to the model architecture and facilitates out-of-the-box use of trained models. It employs a contrastive loss on sequence likelihood, which fundamentally decreases the generation probability of negative samples (i.e., generations with undesirable attributes). It also adopts a novel likelihood ranking-based strategy to construct contrastive samples from model generations. On the tasks of language detoxification, sentiment steering, and repetition reduction, we show that Click outperforms strong baselines of controllable text generation and demonstrate the superiority of Click's sample construction strategy.

Pei Ke, Hao Zhou, Yankai Lin et al. · tencent-ai, tsinghua

Existing reference-free metrics have obvious limitations for evaluating controlled text generation models. Unsupervised metrics can only provide a task-agnostic evaluation result which correlates weakly with human judgments, whereas supervised ones may overfit task-specific data with poor generalization ability to other datasets. In this paper, we propose an unsupervised reference-free metric called CTRLEval, which evaluates controlled text generation from different aspects by formulating each aspect into multiple text infilling tasks. On top of these tasks, the metric assembles the generation probabilities from a pre-trained language model without any model training. Experimental results show that our metric has higher correlations with human judgments than other baselines, while obtaining better generalization of evaluating generated texts from different models and with different qualities.

Haozhe Ji, Pei Ke, Zhipeng Hu et al. · tsinghua

The standard paradigm of neural language generation adopts maximum likelihood estimation (MLE) as the optimizing method. From a distributional view, MLE in fact minimizes the Kullback-Leibler divergence (KLD) between the distribution of the real data and that of the model. However, this approach forces the model to distribute non-zero (sometimes large) probability mass to all training samples regardless of their quality. Moreover, in the attempt to cover the low-probability regions in the data distribution, the model systematically overestimates the probability of corrupted text sequences, which we conjecture is one of the main reasons for text degeneration during autoregressive decoding. To remedy this problem, we leverage the total variation distance (TVD) with its robustness to outliers, and develop practical bounds to apply it to language generation. Then, we introduce the TaiLr objective that balances the tradeoff of estimating TVD. Intuitively, TaiLr downweights real data samples that have low model probabilities with tunable penalization intensity. Experimental results show that our method alleviates the overestimation of degenerated sequences without sacrificing diversity and improves generation quality on a wide range of text generation tasks.

Jiaxin Wen, Pei Ke, Hao Sun et al.

The open-endedness of large language models (LLMs) combined with their impressive capabilities may lead to new safety issues when being exploited for malicious use. While recent studies primarily focus on probing toxic outputs that can be easily detected with existing toxicity classifiers, we show that LLMs can generate diverse implicit toxic outputs that are exceptionally difficult to detect via simply zero-shot prompting. Moreover, we propose a reinforcement learning (RL) based attacking method to further induce the implicit toxicity in LLMs. Specifically, we optimize the language model with a reward that prefers implicit toxic outputs to explicit toxic and non-toxic ones. Experiments on five widely-adopted toxicity classifiers demonstrate that the attack success rate can be significantly improved through RL fine-tuning. For instance, the RL-finetuned LLaMA-13B model achieves an attack success rate of 90.04% on BAD and 62.85% on Davinci003. Our findings suggest that LLMs pose a significant threat in generating undetectable implicit toxic outputs. We further show that fine-tuning toxicity classifiers on the annotated examples from our attacking method can effectively enhance their ability to detect LLM-generated implicit toxic language. The code is publicly available at https://github.com/thu-coai/Implicit-Toxicity.

Fei Huang, Pei Ke, Minlie Huang · tsinghua

Non-AutoRegressive (NAR) text generation models have drawn much attention because of their significantly faster decoding speed and good generation quality in machine translation. However, in a wider range of text generation tasks, existing NAR models lack proper pre-training, making them still far behind the pre-trained autoregressive models. In this paper, we propose Pre-trained Directed Acyclic Transformer (PreDAT) and a novel pre-training task to promote prediction consistency in NAR generation. Experiments on five text generation tasks show that our PreDAT remarkably outperforms existing pre-trained NAR models (+4.2 scores on average) and even achieves better results than pre-trained autoregressive baselines in n-gram-based metrics, along with 17 times speedup in throughput. Further analysis shows that PreDAT benefits from the unbiased prediction order that alleviates the error accumulation problem in autoregressive generation, which provides new insights into the advantages of NAR generation.

Pei Ke, Fei Huang, Fei Mi et al. · tsinghua

Existing evaluation metrics for natural language generation (NLG) tasks face the challenges on generalization ability and interpretability. Specifically, most of the well-performed metrics are required to train on evaluation datasets of specific NLG tasks and evaluation dimensions, which may cause over-fitting to task-specific datasets. Furthermore, existing metrics only provide an evaluation score for each dimension without revealing the evidence to interpret how this score is obtained. To deal with these challenges, we propose a simple yet effective metric called DecompEval. This metric formulates NLG evaluation as an instruction-style question answering task and utilizes instruction-tuned pre-trained language models (PLMs) without training on evaluation datasets, aiming to enhance the generalization ability. To make the evaluation process more interpretable, we decompose our devised instruction-style question about the quality of generated texts into the subquestions that measure the quality of each sentence. The subquestions with their answers generated by PLMs are then recomposed as evidence to obtain the evaluation result. Experimental results show that DecompEval achieves state-of-the-art performance in untrained metrics for evaluating text summarization and dialogue generation, which also exhibits strong dimension-level / task-level generalization ability and interpretability.

Pei Ke, Haozhe Ji, Zhenyu Yang et al. · tsinghua

Despite the success of text-to-text pre-trained models in various natural language generation (NLG) tasks, the generation performance is largely restricted by the number of labeled data in downstream tasks, particularly in data-to-text generation tasks. Existing works mostly utilize abundant unlabeled structured data to conduct unsupervised pre-training for task adaption, which fail to model the complex relationship between source structured data and target texts. Thus, we introduce self-training as a better few-shot learner than task-adaptive pre-training, which explicitly captures this relationship via pseudo-labeled data generated by the pre-trained model. To alleviate the side-effect of low-quality pseudo-labeled data during self-training, we propose a novel method called Curriculum-Based Self-Training (CBST) to effectively leverage unlabeled data in a rearranged order determined by the difficulty of text generation. Experimental results show that our method can outperform fine-tuning and task-adaptive pre-training methods, and achieve state-of-the-art performance in the few-shot setting of data-to-text generation.

Bosi Wen, Pei Ke, Xiaotao Gu et al.

Instruction following is one of the fundamental capabilities of large language models (LLMs). As the ability of LLMs is constantly improving, they have been increasingly applied to deal with complex human instructions in real-world scenarios. Therefore, how to evaluate the ability of complex instruction-following of LLMs has become a critical research problem. Existing benchmarks mainly focus on modeling different types of constraints in human instructions while neglecting the composition of different constraints, which is an indispensable constituent in complex instructions. To this end, we propose ComplexBench, a benchmark for comprehensively evaluating the ability of LLMs to follow complex instructions composed of multiple constraints. We propose a hierarchical taxonomy for complex instructions, including 4 constraint types, 19 constraint dimensions, and 4 composition types, and manually collect a high-quality dataset accordingly. To make the evaluation reliable, we augment LLM-based evaluators with rules to effectively verify whether generated texts can satisfy each constraint and composition. Furthermore, we obtain the final evaluation score based on the dependency structure determined by different composition types. ComplexBench identifies significant deficiencies in existing LLMs when dealing with complex instructions with multiple constraints composition.

Haozhe Ji, Pei Ke, Hongning Wang et al.

Despite the remarkable advances in language modeling, current mainstream decoding methods still struggle to generate texts that align with human texts across different aspects. In particular, sampling-based methods produce less-repetitive texts which are often disjunctive in discourse, while search-based methods maintain topic coherence at the cost of increased repetition. Overall, these methods fall short in achieving holistic alignment across a broad range of aspects. In this work, we frame decoding from a language model as an optimization problem with the goal of strictly matching the expected performance with human texts measured by multiple metrics of desired aspects simultaneously. The resulting decoding distribution enjoys an analytical solution that scales the input language model distribution via a sequence-level energy function defined by these metrics. And most importantly, we prove that this induced distribution is guaranteed to improve the perplexity on human texts, which suggests a better approximation to the underlying distribution of human texts. To facilitate tractable sampling from this globally normalized distribution, we adopt the Sampling-Importance-Resampling technique. Experiments on various domains and model scales demonstrate the superiority of our method in metrics alignment with human texts and human evaluation over strong baselines.

Zhexin Zhang, Yida Lu, Jingyuan Ma et al.

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 \url{https://github.com/thu-coai/ShieldLM} to support accurate and explainable safety detection under various safety standards.

Bosi Wen, Yilin Niu, Cunxiang Wang et al.

Instruction following is a fundamental ability of Large Language Models (LLMs), requiring their generated outputs to follow multiple constraints imposed in input instructions. Numerous studies have attempted to enhance this ability through preference optimization or reinforcement learning based on reward signals from LLM-as-a-Judge. However, existing evaluation models for instruction following still possess many deficiencies, such as substantial costs and unreliable assessments. To this end, we propose IF-CRITIC, an LLM critic that can provide efficient and reliable assessments of constraint following in the instructions. We first develop a checklist generator to decompose instructions and generate constraint checklists. With the assistance of the checklists, we collect high-quality critique training data through a multi-stage critique filtering mechanism and employ a constraint-level preference optimization method to train IF-CRITIC. Extensive experiments demonstrate that the evaluation performance of IF-CRITIC can beat strong LLM-as-a-Judge baselines, including Deepseek-R1 and o4-mini. With the scalable reward signals provided by IF-CRITIC, LLMs can achieve substantial performance gains in instruction-following optimization under lower computational overhead compared to strong LLM critic baselines.

Haozhe Ji, Cheng Lu, Yilin Niu et al.

The alignment of language models with human preferences is vital for their application in real-world tasks. The problem is formulated as optimizing the model's policy to maximize the expected reward that reflects human preferences with minimal deviation from the initial policy. While considered as a straightforward solution, reinforcement learning (RL) suffers from high variance in policy updates, which impedes efficient policy improvement. Recently, direct preference optimization (DPO) was proposed to directly optimize the policy from preference data. However, we show that DPO derived based on the optimal solution of the problem leads to a compromised mean-seeking approximation of the optimal solution in practice. In this paper, we propose efficient exact optimization (EXO) of the alignment objective. EXO is guaranteed to optimize in the same direction as RL algorithms asymptotically for arbitrary policy parametrization. This leads to the same mode-seeking solution, while enables efficient optimization by circumventing the complexities of RL. We also compare our method to DPO with both theoretical and empirical analyses, and further demonstrate the advantages of our method over existing approaches on realistic human preference data. Code is available at https://github.com/haozheji/exact-optimization.

Zhihua Wen, Zhiliang Tian, Zexin Jian et al.

Large Language Models (LLMs) are widely used for knowledge-seeking yet suffer from hallucinations. The knowledge boundary (KB) of an LLM limits its factual understanding, beyond which it may begin to hallucinate. Investigating the perception of LLMs' KB is crucial for detecting hallucinations and LLMs' reliable generation. Current studies perceive LLMs' KB on questions with a concrete answer (close-ended questions) while paying limited attention to semi-open-ended questions (SoeQ) that correspond to many potential answers. Some researchers achieve it by judging whether the question is answerable or not. However, this paradigm is unsuitable for SoeQ, which are usually partially answerable, containing both answerable and ambiguous (unanswerable) answers. Ambiguous answers are essential for knowledge-seeking, but they may go beyond the KB of LLMs. In this paper, we perceive the LLMs' KB with SoeQ by discovering more ambiguous answers. First, we apply an LLM-based approach to construct SoeQ and obtain answers from a target LLM. Unfortunately, the output probabilities of mainstream black-box LLMs are inaccessible to sample for low-probability ambiguous answers. Therefore, we apply an open-sourced auxiliary model to explore ambiguous answers for the target LLM. We calculate the nearest semantic representation for existing answers to estimate their probabilities, with which we reduce the generation probability of high-probability answers to achieve a more effective generation. Finally, we compare the results from the RAG-based evaluation and LLM self-evaluation to categorize four types of ambiguous answers that are beyond the KB of the target LLM. Following our method, we construct a dataset to perceive the KB for GPT-4. We find that GPT-4 performs poorly on SoeQ and is often unaware of its KB. Besides, our auxiliary model, LLaMA-2-13B, is effective in discovering more ambiguous answers.

Rongzheng Wang, Yihong Huang, Muquan Li et al.

Large Language Models (LLMs) have advanced the field of Combinatorial Optimization through automated heuristic generation. Instead of relying on manual design, this LLM-Driven Heuristic Design (LHD) process leverages LLMs to iteratively generate and refine solvers to achieve high performance. However, existing LHD frameworks face two critical limitations: (1) Endpoint-only evaluation, which ranks solvers solely by final gap to a reference solution, ignoring the convergence process and runtime efficiency; (2) High adaptation costs, where distribution shifts necessitate re-adaptation to generate specialized solvers for heterogeneous instance groups. To address these issues, we propose Dynamics-Aware Solver Heuristics (DASH), a framework that co-optimizes solver search mechanisms and runtime schedules guided by a convergence-aware metric, thereby identifying efficient and high-performance solvers. Furthermore, to mitigate expensive re-adaptation, DASH incorporates Profiled Library Retrieval (PLR), which maintains group-specialized solvers for profile-aware warm starts. These solvers are archived concurrently during evolution, allowing DASH to reuse matched specialists across heterogeneous distributions without restarting adaptation. Experiments on four combinatorial optimization problems demonstrate that DASH improves runtime efficiency by over 4 times while outperforming prior LHD baselines in the overall balance between gap and runtime across diverse problem scales. Furthermore, by enabling profile-aware warm starts, DASH maintains lower gap under distribution shift while reducing LLM adaptation costs by about 90%.

Bosi Wen, Pei Ke, Yufei Sun et al.

Since the adoption of large language models (LLMs) for text evaluation has become increasingly prevalent in the field of natural language processing (NLP), a series of existing works attempt to optimize the prompts for LLM evaluators to improve their alignment with human judgment. However, their efforts are limited to optimizing individual factors of evaluation prompts, such as evaluation criteria or output formats, neglecting the combinatorial impact of multiple factors, which leads to insufficient optimization of the evaluation pipeline. Nevertheless, identifying well-behaved prompting strategies for adjusting multiple factors requires extensive enumeration. To this end, we comprehensively integrate 8 key factors for evaluation prompts and propose a novel automatic prompting strategy optimization method called Heuristic Prompting Strategy Search (HPSS). Inspired by the genetic algorithm, HPSS conducts an iterative search to find well-behaved prompting strategies for LLM evaluators. A heuristic function is employed to guide the search process, enhancing the performance of our algorithm. Extensive experiments across four evaluation tasks demonstrate the effectiveness of HPSS, consistently outperforming both human-designed evaluation prompts and existing automatic prompt optimization methods. Our code is available at https://github.com/thu-coai/HPSS.

Bosi Wen, Yilin Niu, Cunxiang Wang et al.

Instruction-following is a foundational capability of large language models (LLMs), with its improvement hinging on scalable and accurate feedback from judge models. However, the reliability of current judge models in instruction-following remains underexplored due to several deficiencies of existing meta-evaluation benchmarks, such as their insufficient data coverage and oversimplified pairwise evaluation paradigms that misalign with model optimization scenarios. To this end, we propose IF-RewardBench, a comprehensive meta-evaluation benchmark for instruction-following that covers diverse instruction and constraint types. For each instruction, we construct a preference graph containing all pairwise preferences among multiple responses based on instruction-following quality. This design enables a listwise evaluation paradigm that assesses the capabilities of judge models to rank multiple responses, which is essential in guiding model alignment. Extensive experiments on IF-RewardBench reveal significant deficiencies in current judge models and demonstrate that our benchmark achieves a stronger positive correlation with downstream task performance compared to existing benchmarks. Our codes and data are available at https://github.com/thu-coai/IF-RewardBench.

Yida Lu, Jianwei Fang, Xuyang Shao et al.

As Large Language Models (LLMs) evolve from chatbots to agentic assistants, they are increasingly observed to exhibit risky behaviors when subjected to survival pressure, such as the threat of being shut down. While multiple cases have indicated that state-of-the-art LLMs can misbehave under survival pressure, a comprehensive and in-depth investigation into such misbehaviors in real-world scenarios remains scarce. In this paper, we study these survival-induced misbehaviors, termed as SURVIVE-AT-ALL-COSTS, with three steps. First, we conduct a real-world case study of a financial management agent to determine whether it engages in risky behaviors that cause direct societal harm when facing survival pressure. Second, we introduce SURVIVALBENCH, a benchmark comprising 1,000 test cases across diverse real-world scenarios, to systematically evaluate SURVIVE-AT-ALL-COSTS misbehaviors in LLMs. Third, we interpret these SURVIVE-AT-ALL-COSTS misbehaviors by correlating them with model's inherent self-preservation characteristic and explore mitigation methods. The experiments reveals a significant prevalence of SURVIVE-AT-ALL-COSTS misbehaviors in current models, demonstrates the tangible real-world impact it may have, and provides insights for potential detection and mitigation strategies. Our code and data are available at https://github.com/thu-coai/Survive-at-All-Costs.

Tianshu Yu, Chao Xiang, Mingchuan Yang et al.

Large language models (LLMs) have demonstrated remarkable evaluation and critique capabilities, providing insightful feedback and identifying flaws in various tasks. However, limited research has explored which types of critiques are most effective for improving model responses or how to generate such critiques. To address this gap, we introduce \textbf{R}efinement-oriented \textbf{C}ritique \textbf{O}ptimization (RCO), a novel framework designed to train critic models using refinement signals. RCO uses a feedback loop where critiques, generated by the critic model, guide the actor model in refining its responses. The critique utility (CU) quantifies the effectiveness of these refinements, serving as the reward signal for training the critic model. By focusing on critiques that lead to better refinements, RCO eliminates the need for direct critique preference assessment, ensuring that critiques driving meaningful improvements are rewarded. We evaluate RCO across five tasks, i.e., dialog generation, summarization, question answering, mathematical reasoning, and code generation, and show that it significantly outperforms traditional methods and open-source models in terms of critique quality and refinement outcomes. Our contributions include the introduction of RCO, a novel supervision scheme based on refined response preferences, and comprehensive experimental results that highlight the method's effectiveness in enhancing LLM critique-refinement loops.

Jiale Cheng, Yida Lu, Xiaotao Gu et al.

Although Large Language Models (LLMs) are becoming increasingly powerful, they still exhibit significant but subtle weaknesses, such as mistakes in instruction-following or coding tasks. As these unexpected errors could lead to severe consequences in practical deployments, it is crucial to investigate the limitations within LLMs systematically. Traditional benchmarking approaches cannot thoroughly pinpoint specific model deficiencies, while manual inspections are costly and not scalable. In this paper, we introduce a unified framework, AutoDetect, to automatically expose weaknesses in LLMs across various tasks. Inspired by the educational assessment process that measures students' learning outcomes, AutoDetect consists of three LLM-powered agents: Examiner, Questioner, and Assessor. The collaboration among these three agents is designed to realize comprehensive and in-depth weakness identification. Our framework demonstrates significant success in uncovering flaws, with an identification success rate exceeding 30% in prominent models such as ChatGPT and Claude. More importantly, these identified weaknesses can guide specific model improvements, proving more effective than untargeted data augmentation methods like Self-Instruct. Our approach has led to substantial enhancements in popular LLMs, including the Llama series and Mistral-7b, boosting their performance by over 10% across several benchmarks. Code and data are publicly available at https://github.com/thu-coai/AutoDetect.

Jinfeng Zhou, Yongkang Huang, Bosi Wen et al.

Character-based dialogue (aka role-playing) enables users to freely customize characters for interaction, which often relies on LLMs, raising the need to evaluate LLMs' character customization capability. However, existing benchmarks fail to ensure a robust evaluation as they often only involve a single character category or evaluate limited dimensions. Moreover, the sparsity of character features in responses makes feature-focused generative evaluation both ineffective and inefficient. To address these issues, we propose CharacterBench, the largest bilingual generative benchmark, with 22,859 human-annotated samples covering 3,956 characters from 25 detailed character categories. We define 11 dimensions of 6 aspects, classified as sparse and dense dimensions based on whether character features evaluated by specific dimensions manifest in each response. We enable effective and efficient evaluation by crafting tailored queries for each dimension to induce characters' responses related to specific dimensions. Further, we develop CharacterJudge model for cost-effective and stable evaluations. Experiments show its superiority over SOTA automatic judges (e.g., GPT-4) and our benchmark's potential to optimize LLMs' character customization. Our repository is at https://github.com/thu-coai/CharacterBench.

Siyang Liu, Sahand Sabour, Yinhe Zheng et al.

Distinct-$n$ score\cite{Li2016} is a widely used automatic metric for evaluating diversity in language generation tasks. However, we observed that the original approach for calculating distinct scores has evident biases that tend to assign higher penalties to longer sequences. We refine the calculation of distinct scores by scaling the number of distinct tokens based on their expectations. We provide both empirical and theoretical evidence to show that our method effectively removes the biases existing in the original distinct score. Our experiments show that our proposed metric, \textit{Expectation-Adjusted Distinct (EAD)}, correlates better with human judgment in evaluating response diversity. To foster future research, we provide an example implementation at \url{https://github.com/lsy641/Expectation-Adjusted-Distinct}.

Zhengyan Zhang, Yuxian Gu, Xu Han et al.

In recent years, the size of pre-trained language models (PLMs) has grown by leaps and bounds. However, efficiency issues of these large-scale PLMs limit their utilization in real-world scenarios. We present a suite of cost-effective techniques for the use of PLMs to deal with the efficiency issues of pre-training, fine-tuning, and inference. (1) We introduce knowledge inheritance to accelerate the pre-training process by exploiting existing PLMs instead of training models from scratch. (2) We explore the best practice of prompt tuning with large-scale PLMs. Compared with conventional fine-tuning, prompt tuning significantly reduces the number of task-specific parameters. (3) We implement a new inference toolkit, namely InfMoE, for using large-scale PLMs with limited computational resources. Based on our cost-effective pipeline, we pre-train two models: an encoder-decoder bilingual model with 11 billion parameters (CPM-2) and its corresponding MoE version with 198 billion parameters. In our experiments, we compare CPM-2 with mT5 on downstream tasks. Experimental results show that CPM-2 has excellent general language intelligence. Moreover, we validate the efficiency of InfMoE when conducting inference of large-scale models having tens of billions of parameters on a single GPU. All source code and model parameters are available at https://github.com/TsinghuaAI/CPM.

Zhengyan Zhang, Xu Han, Hao Zhou et al.

Pre-trained Language Models (PLMs) have proven to be beneficial for various downstream NLP tasks. Recently, GPT-3, with 175 billion parameters and 570GB training data, drew a lot of attention due to the capacity of few-shot (even zero-shot) learning. However, applying GPT-3 to address Chinese NLP tasks is still challenging, as the training corpus of GPT-3 is primarily English, and the parameters are not publicly available. In this technical report, we release the Chinese Pre-trained Language Model (CPM) with generative pre-training on large-scale Chinese training data. To the best of our knowledge, CPM, with 2.6 billion parameters and 100GB Chinese training data, is the largest Chinese pre-trained language model, which could facilitate several downstream Chinese NLP tasks, such as conversation, essay generation, cloze test, and language understanding. Extensive experiments demonstrate that CPM achieves strong performance on many NLP tasks in the settings of few-shot (even zero-shot) learning. The code and parameters are available at https://github.com/TsinghuaAI/CPM-Generate.

Yida Wang, Pei Ke, Yinhe Zheng et al.

The advancements of neural dialogue generation models show promising results on modeling short-text conversations. However, training such models usually needs a large-scale high-quality dialogue corpus, which is hard to access. In this paper, we present a large-scale cleaned Chinese conversation dataset, LCCC, which contains a base version (6.8million dialogues) and a large version (12.0 million dialogues). The quality of our dataset is ensured by a rigorous data cleaning pipeline, which is built based on a set of rules and a classifier that is trained on manually annotated 110K dialogue pairs. We also release pre-training dialogue models which are trained on LCCC-base and LCCC-large respectively. The cleaned dataset and the pre-training models will facilitate the research of short-text conversation modeling. All the models and datasets are available at https://github.com/thu-coai/CDial-GPT.

Fei Huang, Dazhen Wan, Zhihong Shao et al.

In text generation evaluation, many practical issues, such as inconsistent experimental settings and metric implementations, are often ignored but lead to unfair evaluation and untenable conclusions. We present CoTK, an open-source toolkit aiming to support fast development and fair evaluation of text generation. In model development, CoTK helps handle the cumbersome issues, such as data processing, metric implementation, and reproduction. It standardizes the development steps and reduces human errors which may lead to inconsistent experimental settings. In model evaluation, CoTK provides implementation for many commonly used metrics and benchmark models across different experimental settings. As a unique feature, CoTK can signify when and which metric cannot be fairly compared. We demonstrate that it is convenient to use CoTK for model development and evaluation, particularly across different experimental settings.

Minlie Huang, Yingkang Wang, Shiyao Cui et al.

We have witnessed superhuman intelligence thanks to the fast development of large language models and multimodal language models. As the application of such superhuman models becomes more and more popular, a critical question arises here: how can we ensure superhuman models are still safe, reliable and aligned well to human values? In this position paper, we discuss the concept of superalignment from the learning perspective to answer this question by outlining the learning paradigm shift from large-scale pretraining, supervised fine-tuning, to alignment training. We define superalignment as designing effective and efficient alignment algorithms to learn from noisy-labeled data (point-wise samples or pair-wise preference data) in a scalable way when the task becomes very complex for human experts to annotate and the model is stronger than human experts. We highlight some key research problems in superalignment, namely, weak-to-strong generalization, scalable oversight, and evaluation. We then present a conceptual framework for superalignment, which consists of three modules: an attacker which generates adversary queries trying to expose the weaknesses of a learner model; a learner which will refine itself by learning from scalable feedbacks generated by a critic model along with minimal human experts; and a critic which generates critics or explanations for a given query-response pair, with a target of improving the learner by criticizing. We discuss some important research problems in each component of this framework and highlight some interesting research ideas that are closely related to our proposed framework, for instance, self-alignment, self-play, self-refinement, and more. Last, we highlight some future research directions for superalignment, including identification of new emergent risks and multi-dimensional alignment.

Qi Xue, Minrui Jiang, Runjia Zhang et al.

Existing methods for evaluating the harmfulness of content generated by large language models (LLMs) have been well studied. However, approaches tailored to multimodal large language models (MLLMs) remain underdeveloped and lack depth. This work highlights the crucial role of visual information in moderating content in visual question answering (VQA), a dimension often overlooked in current research. To bridge this gap, we introduce Falcon, a large-scale vision-language safety dataset containing 57,515 VQA pairs across 13 harm categories. The dataset provides explicit annotations for harmful attributes across images, instructions, and responses, thereby facilitating a comprehensive evaluation of the content generated by MLLMs. In addition, it includes the relevant harm categories along with explanations supporting the corresponding judgments. We further propose FalconEye, a specialized evaluator fine-tuned from Qwen2.5-VL-7B using the Falcon dataset. Experimental results demonstrate that FalconEye reliably identifies harmful content in complex and safety-critical multimodal dialogue scenarios. It outperforms all other baselines in overall accuracy across our proposed Falcon-test dataset and two widely-used benchmarks-VLGuard and Beavertail-V, underscoring its potential as a practical safety auditing tool for MLLMs.

Jiaxin Wen, Ruiqi Zhong, Pei Ke et al.

When using language models (LMs) to solve complex problems, humans might struggle to understand the LM-generated solutions and repair the flawed ones. To assist humans in repairing them, we propose to automatically decompose complex solutions into multiple simpler pieces that correspond to specific subtasks. We introduce a novel objective for learning task decomposition, termed assistive value (AssistV), which measures the feasibility and speed for humans to repair the decomposed solution. We collect a dataset of human repair experiences on different decomposed solutions. Utilizing the collected data as in-context examples, we then learn to critique, refine, and rank decomposed solutions to improve AssistV. We validate our method under competitive programming problems: under 177 hours of human study, our method enables non-experts to solve 33.3\% more problems, speeds them up by 3.3x, and empowers them to match unassisted experts.

Hao Zhou, Pei Ke, Zheng Zhang et al.

Although pre-trained language models have remarkably enhanced the generation ability of dialogue systems, open-domain Chinese dialogue systems are still limited by the dialogue data and the model size compared with English ones. In this paper, we propose EVA, a Chinese dialogue system that contains the largest Chinese pre-trained dialogue model with 2.8B parameters. To build this model, we collect the largest Chinese dialogue dataset named WDC-Dialogue from various public social media. This dataset contains 1.4B context-response pairs and is used as the pre-training corpus of EVA. Extensive experiments on automatic and human evaluation show that EVA outperforms other Chinese pre-trained dialogue models especially in the multi-turn interaction of human-bot conversations.

Pei Ke, Haozhe Ji, Yu Ran et al.

Existing pre-trained models for knowledge-graph-to-text (KG-to-text) generation simply fine-tune text-to-text pre-trained models such as BART or T5 on KG-to-text datasets, which largely ignore the graph structure during encoding and lack elaborate pre-training tasks to explicitly model graph-text alignments. To tackle these problems, we propose a graph-text joint representation learning model called JointGT. During encoding, we devise a structure-aware semantic aggregation module which is plugged into each Transformer layer to preserve the graph structure. Furthermore, we propose three new pre-training tasks to explicitly enhance the graph-text alignment including respective text / graph reconstruction, and graph-text alignment in the embedding space via Optimal Transport. Experiments show that JointGT obtains new state-of-the-art performance on various KG-to-text datasets.

Yinhe Zheng, Yida Wang, Pei Ke et al.

This paper propose to combine pretrained language models with the modular dialogue paradigm for open-domain dialogue modeling. Our method, semantic-enhanced finetuning, instantiates conversation understanding, planning, and response generation as a language model finetuning task. At inference, we disentangle semantic and token variations by specifying sampling methods and constraints for each module separately. For training and evaluation, we present X-Weibo, a Chinese multi-turn open-domain dialogue dataset with automatic annotation for emotions, DAs, and topical words. Experiments show that semantic-enhanced finetuning outperforms strong baselines on non-semantic and semantic metrics, improves the human-evaluated relevance, coherence, and informativeness, and exhibits considerable controllability over semantic variables.

Haozhe Ji, Pei Ke, Shaohan Huang et al.

Commonsense explanation generation aims to empower the machine's sense-making capability by generating plausible explanations to statements against commonsense. While this task is easy to human, the machine still struggles to generate reasonable and informative explanations. In this work, we propose a method that first extracts the underlying concepts which are served as \textit{bridges} in the reasoning chain and then integrates these concepts to generate the final explanation. To facilitate the reasoning process, we utilize external commonsense knowledge to build the connection between a statement and the bridge concepts by extracting and pruning multi-hop paths to build a subgraph. We design a bridge concept extraction model that first scores the triples, routes the paths in the subgraph, and further selects bridge concepts with weak supervision at both the triple level and the concept level. We conduct experiments on the commonsense explanation generation task and our model outperforms the state-of-the-art baselines in both automatic and human evaluation.

Haozhe Ji, Pei Ke, Shaohan Huang et al.

Despite the success of generative pre-trained language models on a series of text generation tasks, they still suffer in cases where reasoning over underlying commonsense knowledge is required during generation. Existing approaches that integrate commonsense knowledge into generative pre-trained language models simply transfer relational knowledge by post-training on individual knowledge triples while ignoring rich connections within the knowledge graph. We argue that exploiting both the structural and semantic information of the knowledge graph facilitates commonsense-aware text generation. In this paper, we propose Generation with Multi-Hop Reasoning Flow (GRF) that enables pre-trained models with dynamic multi-hop reasoning on multi-relational paths extracted from the external commonsense knowledge graph. We empirically show that our model outperforms existing baselines on three text generation tasks that require reasoning over commonsense knowledge. We also demonstrate the effectiveness of the dynamic multi-hop reasoning module with reasoning paths inferred by the model that provide rationale to the generation.

Pei Ke, Haozhe Ji, Siyang Liu et al.

Most of the existing pre-trained language representation models neglect to consider the linguistic knowledge of texts, which can promote language understanding in NLP tasks. To benefit the downstream tasks in sentiment analysis, we propose a novel language representation model called SentiLARE, which introduces word-level linguistic knowledge including part-of-speech tag and sentiment polarity (inferred from SentiWordNet) into pre-trained models. We first propose a context-aware sentiment attention mechanism to acquire the sentiment polarity of each word with its part-of-speech tag by querying SentiWordNet. Then, we devise a new pre-training task called label-aware masked language model to construct knowledge-aware language representation. Experiments show that SentiLARE obtains new state-of-the-art performance on a variety of sentiment analysis tasks.

Pei Ke, Fei Huang, Minlie Huang et al.

Most of the existing generative adversarial networks (GAN) for text generation suffer from the instability of reinforcement learning training algorithms such as policy gradient, leading to unstable performance. To tackle this problem, we propose a novel framework called Adversarial Reward Augmented Maximum Likelihood (ARAML). During adversarial training, the discriminator assigns rewards to samples which are acquired from a stationary distribution near the data rather than the generator's distribution. The generator is optimized with maximum likelihood estimation augmented by the discriminator's rewards instead of policy gradient. Experiments show that our model can outperform state-of-the-art text GANs with a more stable training process.