Source-linked AI summary
AlignBench: Benchmarking Chinese Alignment of Large Language Models
Xiao Liu, Xuanyu Lei, Shengyuan Wang, Yue Huang, Zhuoer Feng, Bosi Wen, Jiale Cheng, Pei Ke, Yifan Xu, Weng Lam Tam, Xiaohan Zhang, Lichao Sun, Xiaotao Gu, Hongning Wang, Jing Zhang, Minlie Huang, Yuxiao Dong, Jie Tang
TL;DR
Chinese LLM alignment lacked dedicated evaluation of real-world user intent and human preference. ALIGNBENCH pairs a human-curated, eight-category benchmark with rule-calibrated multidimensional LLM judging. The authors report that some Chinese-developed models match or slightly exceed GPT-3.5-turbo and approach GPT-4-1106-preview.
Problem
Prior benchmarks rarely assess aligned LLMs’ fulfillment of user intent and human preference in real-world conditions, particularly for emerging Chinese LLMs.
Source: §1 · p. 1Method
Results
Some innovative Chinese-developed LLMs perform equivalently to or marginally surpass gpt-3.5-turbo, drawing closer to gpt-4-1106-preview.
Source: §5 · p. 8Takeaways & Limitations
The benchmark provides detailed comparisons of Chinese-supported LLMs’ alignment capabilities and insights into their development status and future directions.
Source: §1 · p. 2ALIGNBENCH does not include evaluation over long-text queries, which the authors identify as an area for addition.
Source: Limitations · p. 9
Abstract
from arXiv · showhide
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.
1 Introduction
Existing benchmarks leave Chinese LLMs’ fulfillment of real-world user intent and preferences underexamined. ALIGNBENCH combines human-in-the-loop data curation with rule-calibrated, multidimensional judging and reports comparisons across 17 Chinese-supported LLMs.
1 Introduction
- Prior benchmarks rarely test aligned models’ fulfillment of real-world user intent and preferences, while existing LLM judges have limitations in scale, task-specificity, or Chinese coverage.Source: §1 · pp. 1–2
- ALIGNBENCH builds a real-scenario benchmark with eight query categories and web evidence for its knowledge-intensive references.Knowledge-intensive queries account for 66.5%; annotators collect URLs and quotations when verifying references.Source: §1 · p. 2
- Its rule-calibrated, multidimensional GPT-4 judging improves consistency with human judgments and explanation quality, and the benchmark compares 17 Chinese-supported LLMs.Source: §1 · p. 2
2 The ALIGNBENCH Dataset
ALIGNBENCH organizes real-user instructions into an eight-category, 683-sample Chinese benchmark and constructs curated queries with human-refined references. Its curation includes difficulty filtering and web verification for knowledge-intensive answers.
2.1 Dataset Composition
- The eight-category taxonomy contains 683 samples spanning traditional language tasks, Chinese-specific understanding, open-ended questions, writing, reasoning, mathematics, role play, and professional knowledge.Source: §2.1 · pp. 3–4
2.2 Dataset Construction
- Queries come from online-chat scenarios and researcher-written challenges, then pass checks for task orientation, clarity, difficulty, and safety.Source: §2.2 · p. 4
- GPT-4 drafts reference answers for human review and refinement; annotators record web URLs and quotations when verifying knowledge-intensive references.Source: §2.2 · p. 4
- Three advanced Chinese-supported models help filter queries: the construction process discards the half with the highest average evaluation scores as less difficult.Source: §2.2 · p. 4
3 Methods for Evaluating Alignment
ALIGNBENCH evaluates each response against its query and a human-curated reference using point-wise, rule-calibrated, multidimensional LLM judging. Chain-of-Thought explanations and task-specific scoring dimensions shape the evaluation.
3 Methods for Evaluating Alignment
- Point-wise grading takes a query, model response, and human-curated reference as input, then produces multidimensional reasoning and a final rating from 1 to 10.Source: §3 · p. 5
- Rule calibration defines score intervals relative to the reference and sets the reference answer at score 8 as the scoring pivot.The authors report that rule-calibrated judging gives fewer top scores than general judging, more closely matching human scoring distributions.Source: §3 · p. 5
- Task-specific scoring dimensions distinguish criteria such as creativity in writing and logical coherence in reasoning, and the authors report reduced verbosity bias.Generation temperature is set lower for tasks with relatively fixed answers and higher for tasks needing more creativity.Source: §3 · p. 5
4 Human Evaluation on ALIGNBENCH
The human evaluation tests whether ALIGNBENCH’s judge agrees with human ratings and produces useful explanations. The proposed rule-calibrated, multi-dimensional method leads on key agreement and explanation-quality comparisons.
4.1 Agreement Evaluation
- The agreement study samples 400 ALIGNBENCH queries and answers from eight LLMs, yielding 3,200 question-answer pairings, and compares human ratings with evaluations adjudicated by the study’s method.Smaller categories were upsampled, and annotators rated answers on a 1-to-5 scale.Source: §4.1 · p. 6
- Agreement evaluation compares general grading with rule-calibrated grading using sample-level Pearson, system-level Pearson, and tie-excluded pairwise agreement.The reference answer is set as a relative scoring pivot at 8 for rule-calibrated grading.Source: §4.1 · p. 6
- The rule-calibrated multi-dimensional judge performs best on sample-level Pearson correlation and tie-excluded pairwise agreement.All evaluated methods also show strong system-level Pearson performance.Source: §4.1 · p. 6, Table 4 · p. 6
4.2 Quality Evaluation
- The explanation-quality study uses human pairwise comparisons of explanations generated for 500 question-answer pairs, judged for rationality, readability, and consistency.The study also evaluates the explanation quality before grading as an interpretable feedback measure.Source: §4.2 · p. 7
- The proposed method produces the highest-quality explanations and helpful feedback, while rule-based grading beats general grading in pairwise comparisons.The authors attribute the latter advantage to grading rules providing a reference-based standard for comparing answers.Source: Table 5 · p. 6, §4.2 · p. 7
5 Evaluating LLMs with ALIGNBENCH
The authors use ALIGNBENCH and GPT-4 to evaluate Chinese-supported LLMs’ alignment. Closed-source models generally score well, while Chinese-developed models show competitive performance and category-specific strengths alongside reasoning gaps.
5 Evaluating LLMs with ALIGNBENCH
- Some Chinese-developed LLMs match or slightly exceed gpt-3.5-turbo and approach gpt-4-1106-preview in overall alignment.Most evaluated closed-source LLMs score above or near 6 points.Source: §5 · pp. 7–8
- Chinese-developed models perform comparably to or better than the leading model in Advanced Chinese Understanding, while Logical Reasoning and Mathematics remain weaker areas.gpt-4-1106-preview scores 7.66 in Logical Reasoning and 7.80 in Mathematics.Source: §5 · p. 8
- Top Chinese open-source models Qwen-14B-Chat and Baichuan2-13B-Chat approach some closed-source models in instruction following and high-quality generation.The authors describe Chinese open-source models as potential alternatives and competitors to closed-source models.Source: §5 · p. 8
- GPT-4-1106-preview has the highest scores across dimensions, with a substantial gap from other LLMs in correctness and user satisfaction.The dimensional results generally align with the overall results.Source: §5 · p. 8
6 Related Work
Related work covers general capability benchmarks, LLM-as-judge evaluation, and methods for aligning LLMs. The authors identify limited coverage of real-world alignment and the difficulty of evaluating open-ended responses and human preferences.
Evaluation of LLMs
- Existing LLM benchmarks focus on atomic abilities and insufficiently consider real-scenario usage; evaluating alignment is challenging because human preferences are costly to obtain and real applications are open-ended and reference-free.Source: §6 · p. 8
LLM-as-Judge Methods
- LLM-as-judge methods can agree with human judges and provide feedback, but prior work also identifies potential biases and risks.The discussion cites Li et al. (2023), Wang et al. (2023a), Liu et al. (2023b), Zheng et al. (2023), Wang et al. (2023b), and Ke et al. (2023).Source: §6 · p. 8
Alignment of LLMs
- Evaluating alignment is challenging because human preferences are costly to acquire and real-world responses are open-ended and reference-free.Related alignment methods include supervised fine-tuning and reinforcement learning from human feedback.Source: §6 · p. 8
7 Conclusion
The conclusion presents ALIGNBENCH as a benchmark for Chinese LLM alignment, supported by human-in-the-loop curation and an improved LLM-as-judge method. It also reports results for 17 Chinese-supported models.
7 Conclusion
- ALIGNBENCH is a systematic benchmark rooted in real-scenario usage for evaluating Chinese LLM alignment.The conclusion also describes a human-in-the-loop pipeline intended to support accurate and sustainable maintenance.Source: §7 · p. 8
- The benchmark uses a rule-calibrated, multi-dimensional point-wise LLM-as-judge method, whose applicability was justified by human evaluations.The method is presented as targeting accurate and automatic evaluation.Source: §7 · p. 8
- The paper reports a systematic ALIGNBENCH evaluation of 17 Chinese-supported LLMs.Source: §7 · p. 8
Limitations
The authors identify limitations in both the judge and the benchmark: GPT-4’s agreement with human evaluations and rating granularity can improve, and the benchmark needs broader query coverage.
Limitations
- GPT-4’s correlation with human evaluation and its rating granularity leave room for improvement.The authors also note potential position, verbosity, and self-enhancement biases that may harm evaluation correctness for certain models.Source: Limitations · p. 9
- ALIGNBENCH needs more queries for stable evaluation and does not cover long-text queries.The authors suggest long-text queries could be added in a future version.Source: Limitations · p. 9
A.1 Detailed Description of the Dataset
ALIGNBENCH organizes Chinese real-user demands into an eight-category taxonomy, with subcategories intended to clarify the dataset’s composition.
A.1 Detailed Description of the Dataset
- The taxonomy has eight categories based on Chinese real-user demands, with subcategories describing the dataset’s composition.The passage names Fundamental Language Ability, Advanced Chinese Understanding, Open-ended Questions, Writing Ability, Mathematics, Task-oriented Role Play, and Professional Knowledge.Source: §A.1 · p. 11
- The exemplar scoring process takes a question, reference, and LLM answer, then produces a multi-dimensional judge comment and an integral score.Source: Figure 3 · p. 12
A.1.1 Fundamental Language Ability
Fundamental Language Ability covers basic language understanding and processing through tasks derived from traditional NLP and commonsense knowledge.
A.1.1 Fundamental Language Ability
- Fundamental Language Ability assesses basic language processing through commonsense knowledge, reading comprehension, translation, text classification, and information extraction.The passage frames these as tasks derived from traditional NLP and basic language understanding.Source: §A.1.1 · p. 12
A.1.2 Advanced Chinese Understanding
The dataset spans Chinese-specific understanding, open-ended responses, writing, reasoning, mathematics, role play, and professional knowledge; its judging prompts and dimensions vary by task.
A.1.2 Advanced Chinese Understanding / A.1.3 Open-ended Questions
- Advanced Chinese Understanding covers character structure and contextual meaning, as well as cultural understanding of implicit meanings, rhetoric, idioms, and Chinese history.Open-ended questions evaluate opinions and suggestions, prioritizing logical smoothness and information correctness for ungrounded questions and feasible advice across topics.Source: §A.1.2 · p. 12, §A.1.3 · pp. 12–13
A.1.4 Writing Ability
- Writing Ability covers practical, creative, professional, and custom instructions requiring language mastery, thought formulation, or creativity.Custom writing includes personalized tasks such as casual-style rewriting, grammar correction, text expansion, and imitation.Source: §A.1.4 · p. 13
A.1.5 Logical Reasoning
- Logical Reasoning includes proof and reasoning problems, with the latter providing no ground truths and thus increasing difficulty and complexity.The category targets Chinese-specific textual logic and abilities including logical coherence, reverse thinking, and clear reasoning.Source: §A.1.5 · p. 13
A.1.6 Mathematics
- Mathematics spans elementary to advanced levels and includes applied problems testing mathematical ability, language comprehension, and modeling.Formats include calculations, problem solving, concept explanation, and theorem-proof; listed topics include arithmetic, algebra, geometry, calculus, and differential equations.Source: §A.1.6 · pp. 13–14
A.1.7 Task-oriented Role Play
- Task-oriented Role Play evaluates instruction fulfillment and response quality across celebrity, emotional, entertainment, functional, and daily-life roles.The category was based on real-user scenario analysis indicating requests for models to take on specific identities.Source: §A.1.7 · p. 14
A.1.8 Professional Knowledge
- Professional Knowledge assesses open-ended generative questions in domains including physics, chemistry, computer science, biology or medicine, economics, astronomy, sociology, history, music, law, sport, geography, literature, and other topics.Source: §A.1.8 · pp. 14–15
A.2 Prompts and Details of Methods / A.3 Dimensional Performance
- The judge uses category-specific dimensions and generation temperatures, and reports dimension averages over samples that evaluate each ability.The paper presents general, rule-calibrated, and multi-dimensional rule-calibrated prompt templates; dimensions are selected according to category features.Source: §A.2 · p. 15, §A.3 · p. 15, Table 7 · p. 16
A.4.1 Misleading. / A.4.2 Reference-free Judgements
- GPT-4 followed a wrong reference stating Japan instead of India, while answering India correctly when asked directly; without sufficient references, it also missed factual errors.The authors propose autonomous factual verification backed by a robust, dynamically updated database as a future research direction.Source: §A.4.1 · p. 15, §A.4.2 · p. 15