Judging LLM-as-a-judge with MT-Bench and Chatbot Arena
By Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuo-Han Li, Zhuohan Li, Dacheng Li and 6 more. 2023
9,887 citations recorded by the source corpus.
Evaluating large language model (LLM) based chat assistants is challenging due to their broad capabilities and the inadequacy of existing benchmarks in measuring human preferences. To address this, we explore using strong LLMs as judges to evaluate these models on more open-ended questions. We examine the usage and limitations of LLM-as-a-judge, including position, verbosity, and self-enhancement biases, as well as limited reasoning ability, and propose solutions to mitigate some of them. We then verify the agreement between LLM judges and human preferences by introducing two benchmarks: MT-bench, a multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our results reveal that strong LLM judges like GPT-4 can match both controlled and crowdsourced human preferences well, achieving over 80% agreement, the same level of agreement between humans. Hence, LLM-as-a-judge is a scalable and explainable way to approximate human preferences, which are otherwise very expensive to obtain. Additionally, we show our benchmark and traditional benchmarks complement each other by evaluating several variants of LLaMA and Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with human preferences are publicly available at https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
Subfields: Artificial Intelligence, Political Science and International Relations, Prompting & Reasoning, Evaluation & Benchmarks
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Artificial Intelligence
- Mohit Bansal
- Jianfeng Gao
- Percy Liang
- Luke S. Zettlemoyer
- Yejin Choi
- Quoc V. Le
- Yoshua Bengio
- Graham Neubig
In Political Science and International Relations
- Yejin Choi
- Noah A. Smith
- Matei A. Zaharia
- Percy Liang
- Minlie Huang
- Dan Jurafsky
- Maosong Sun
- Ion Stoica
In Prompting & Reasoning
In Evaluation & Benchmarks
- Yejin Choi
- Mohit Bansal
- Jianfeng Gao
- Noah A. Smith
- Douwe Kiela
- Ronan Le Bras
- Greg Durrett
- Hannaneh Hajishirzi
Citations in the corpus
- Mixtral of Experts (2024, background): “Mixtral – Instruct reaches a score of 8.30 on MT-Bench [33] (see Table 2), making it the best open-weights model as of December 2023.”
- From Crowdsourced Data to High-Quality Benchmarks: Arena-Hard and BenchBuilder Pipeline (2024, background): “Recent developments in benchmark evaluation [50, 38] have made it feasible to evaluate more open-ended prompts, yet generating high-quality prompts at scale remains challenging.”
- Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference (2024, methodology): “Firstly, the questions within these benchmarks are not open-ended, hindering the ability to capture the flexible and interactive use found in real-world settings (Zheng et al., 2023b).”
- MAmmoTH: Building Math Generalist Models through Hybrid Instruction Tuning (2023, background): “5 (Zheng et al., 2023b), Tulu (Wang et al., 2023g), Platypus-2 (Lee et al., 2023) and Guanaco (Dettmers et al., 2023).”
- UltraFeedback: Boosting Language Models with High-quality Feedback (2023, neutral): “Although many researchers pointed out that GPT-4 is capable of providing meticulous judgments and feedback (Dubois et al., 2023; Bai et al., 2022b; Lee et al., 2023; Zheng et al., 2023a), and exhibits high agreement with human preference, we still concern whether the GPT-4 preferences are qualified.”
- LLM-QAT: Data-Free Quantization Aware Training for Large Language Models (2023, neutral): “For example, Vicuna (Zheng et al., 2023) utilized user-uploaded ShareGPT data for instruction fine-tuning, while Alpaca (Taori et al., 2023) relied on predefined human prompts with careful balance in each category to ensure diversity.”
- RouteLLM: Learning to Route LLMs with Preference Data (2024, neutral): “Evaluation benchmarks : We evaluate our routers on three widely-used academic benchmarks: MMLU [16] consisting of 14,042 questions across 57 subjects, MT Bench [27] with 160 open-ended questions using LLM-as-a-judge, and GSM8K [11] with over 1,000 grade school math problems.”
- GPTFUZZER: Red Teaming Large Language Models with Auto-Generated Jailbreak Prompts (2023, methodology): “From the detailed ASR, we can get consistent results that human-written jailbreak templates are effective towards Vicuna-7B and ChatGPT, while Llama-2-7B-Chat is the most robust.”
- From Local to Global: A Graph RAG Approach to Query-Focused Summarization (2024, background): “Many benchmark datasets for open-domain question answering exist, including HotPotQA (Yang et al., 2018), MultiHop-RAG (Tang and Yang, 2024), and MT-Bench (Zheng et al., 2024).”
- LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding (2023, methodology): “Although using LLM as examiner may reduce this problem (Bai et al., 2023; An et al., 2023), the runtime overhead for evaluation may be high, and LLM also has bias when used as an evaluation metric (Zheng et al., 2023a).”
- ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate (2023, neutral): “This methodology is dubbed LLM-as-a-judge (Zheng et al., 2023).”
- DreamLLM: Synergistic Multimodal Comprehension and Creation (2023, methodology): “Our D REAM LLM is built with a causal decoder-only LLM F θ as the model foundation, i.e. , Vicuna (Chiang et al., 2023) based on LLaMA (Touvron et al., 2023a) trained on ShareGPT (Zheng et al., 2023).”
- Skywork-Reward: Bag of Tricks for Reward Modeling in LLMs (2024, methodology): “Generative Models While discriminative models are widely adopted, generative models offer an alternative approach by directly using LLM-generated outputs to evaluate preference data (Zheng et al., 2023).”
- DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code Intelligence (2024, neutral): “On the other hand, DeepSeek-V2 Chat demonstrates slightly better results in benchmarks such as MT-bench (Zheng et al., 2023), AlpacaEval 2.0 (Dubois et al., 2024), and AlignBench (Liu et al., 2023c).”
- V*: Guided Visual Search as a Core Mechanism in Multimodal LLMs (2023, methodology): “In this work, we implement the VQA LLM in our SEAL framework with Vicuna-7B[61] as the language model.”