What is MT-Bench? The multi-turn LLM judge benchmark explained
MT-Bench is a set of 80 two-turn questions that GPT-4 grades from 1 to 10. What it tests, how scoring works, and where it falls short.
MT-Bench is a benchmark of 80 multi-turn questions that a strong language model, usually GPT-4, grades on a scale of 1 to 10. Zheng and colleagues at LMSYS introduced it in a June 2023 paper, later accepted to the NeurIPS 2023 Datasets and Benchmarks Track, together with the idea that an LLM can stand in for human raters. That idea is now called LLM-as-a-judge.
What MT-Bench contains
The question set has 10 questions in each of 8 categories, according to the paper. Every question has a follow-up, so each one tests whether the model keeps track of the conversation.
| Category | What it probes |
|---|---|
| Writing | Open-ended composition |
| Roleplay | Staying in a persona |
| Extraction | Pulling facts out of text |
| Reasoning | Logic problems |
| Math | Calculation and word problems |
| Coding | Writing and fixing code |
| STEM knowledge | Science and engineering questions |
| Humanities and social science knowledge | Questions from those fields |
The LMSYS announcement of June 2023 describes it as “a challenging multi-turn question set designed to evaluate the conversational and instruction-following ability of models.”
How the judge grades
The paper describes three judging modes:
- Pairwise comparison: the judge sees two answers and picks a winner or a tie.
- Single-answer grading: the judge scores one answer.
- Reference-guided grading: the judge also gets a reference answer, which matters for math and reasoning.
The LMSYS leaderboard used single-answer grading by GPT-4. In the FastChat code that implements the benchmark, the judge gives each turn a score out of 10 and the scores are averaged across turns. The same README lists a pairwise mode against a baseline model (gpt-3.5-turbo by default) and an all-pairs mode.
How well the judge agrees with people
The authors collected about 3,000 votes from 58 expert labelers on MT-Bench answers. The table shows agreement on first-turn questions, from the paper. The first column counts ties and position-inconsistent votes as ties, where random agreement would be 33%. The second uses only non-tie votes, where random is 50%.
| Pair | With ties | Without ties |
|---|---|---|
| GPT-4 pairwise and human | 66% | 85% |
| GPT-4 single-answer and human | 60% | 85% |
| Human and human | 63% | 81% |
On the crowdsourced Chatbot Arena votes, the paper reports 87% agreement for GPT-4 pairwise judging on non-tie votes. The authors conclude that strong judges reach over 80% agreement, comparable to agreement between humans. The paper released the 80 questions, 3,000 expert votes and 30,000 conversations with human preferences.
Known problems
The same paper lists the judge’s weaknesses:
- Position bias. GPT-4 gave consistent verdicts when answer order was swapped in 65.0% of cases, and favored the first answer in 30.0%. Claude-v1 was consistent in 23.8%.
- Verbosity bias. On a test that padded an answer with a repeated list, GPT-4 was fooled in 8.7% of cases, while Claude-v1 and GPT-3.5 failed in 91.3%.
- Self-enhancement. GPT-4 favored itself with a 10% higher win rate and Claude-v1 with a 25% higher one, though the authors say limited data keeps them from settling whether the bias exists.
- Math and reasoning. On 10 math questions, GPT-4 as judge failed in 14 of 20 judgments with the default prompt, 6 with chain-of-thought and 3 with a reference answer.
Limits and biases of LLM-as-a-judge covers these and later research in more detail.
MT-Bench has limits as a benchmark too. It is a static set of 80 questions. In April 2024 the LMSYS team compared it with their Arena-Hard-Auto benchmark and measured its separability, the share of model pairs it can tell apart with confidence, at 22.6%, against 87.4% for Arena-Hard-Auto. The comparison comes from the Arena-Hard authors themselves, who also note that their newer benchmark does not evaluate multi-turn conversations, which is what MT-Bench was built for.
Other teams have extended the multi-turn idea. MT-Bench-101, an ACL 2024 paper, has 1,388 dialogues and 4,208 turns across 13 tasks.
Running it
The FastChat repository has the code and the questions. The README describes three steps: generate answers with gen_model_answer.py, generate judgments with gen_judgment.py (which needs an OpenAI API key for the default GPT-4 judge), and print scores with show_result.py. You can reuse the idea in your own evals. Write the criteria into a judge prompt, grade with a reference answer where one exists, swap answer order for pairwise checks, and compare the judge with human labels on a sample. LLM evaluators in HoneyHive take criteria written as a prompt and return a boolean, numeric or string score, on traces in production or in offline experiments. See LLM evaluation for where a judge fits among other metrics, and judge versus human evaluation for when to use a person.
FAQ
What is MT-Bench?
A benchmark of 80 multi-turn questions across 8 categories, graded by a strong LLM judge such as GPT-4. It was introduced in the June 2023 paper that proposed LLM-as-a-judge.
How is MT-Bench scored?
In the FastChat code, the default judge is GPT-4 and single-answer grading is the recommended mode. The judge scores each turn out of 10 and the scores are averaged across turns.
Is MT-Bench the same as Chatbot Arena?
No. MT-Bench is a fixed set of questions graded by a model. Chatbot Arena is a platform where people vote between anonymous answers to prompts they write. The paper introduces both. See LMArena.
How accurate is the MT-Bench judge?
On non-tie votes GPT-4 agreed with human experts 85% of the time on first-turn questions, against 81% between two humans. It still showed position, verbosity and self-enhancement bias and struggled with math.
Is MT-Bench still useful?
As a quick, cheap check of chat quality, yes. As a way to rank close models, less so: the Arena-Hard authors measured its separability at 22.6%, against 87.4% for their newer benchmark.