====== LLM-as-a-Judge ======

LLM-as-a-Judge is an evaluation paradigm introduced by Zheng et al. (2023) that uses strong large language models (such as GPT-4) to automatically evaluate the quality of LLM-generated responses. The paper introduces two complementary benchmarks -- MT-Bench and Chatbot Arena -- and systematically analyzes the biases, agreement rates, and practical viability of using LLMs as scalable proxies for human evaluation.

===== Motivation =====

Evaluating LLM-based chat assistants faces fundamental challenges:

  * **Breadth of capabilities**: Models must handle writing, reasoning, coding, math, and more
  * **Benchmark inadequacy**: Traditional NLP benchmarks (MMLU, HellaSwag) fail to capture conversational quality and human preferences
  * **Human evaluation cost**: Obtaining reliable human preferences at scale is prohibitively expensive and slow
  * **Subjectivity**: Open-ended tasks have no single correct answer, making automated metrics insufficient

LLM-as-a-Judge addresses these by leveraging strong LLMs to approximate human preferences at a fraction of the cost while maintaining over 80% agreement with human annotators.

===== MT-Bench =====

MT-Bench is a **multi-turn question set** designed to evaluate chat assistants across 8 diverse categories:

  - **Writing**: Creative and professional writing tasks
  - **Roleplay**: Character-based conversational scenarios
  - **Extraction**: Information extraction from provided text
  - **STEM**: Science, technology, engineering, and math questions
  - **Reasoning**: Logical and analytical reasoning problems
  - **Coding**: Programming tasks and code analysis
  - **Math**: Mathematical computation and proof tasks
  - **Humanities**: History, philosophy, and social science questions

Each question has **two turns** -- a follow-up question tests the model's ability to maintain context and build on its initial response. This multi-turn design is critical because single-turn evaluation misses crucial aspects of conversational AI quality.

===== Chatbot Arena =====

Chatbot Arena is a **crowdsourced battle platform** where:

  * Users submit open-ended prompts of their choosing
  * Two anonymous LLMs generate responses side by side
  * Users vote for the better response (or tie)
  * Elo ratings are computed from accumulated votes

This provides a diverse, real-world evaluation signal that complements the controlled MT-Bench setting. With over 30K conversations and human preferences publicly released, it has become a de facto standard for LLM comparison.

===== Evaluation Modes =====

The paper explores three LLM-as-a-Judge configurations:

  * **Pairwise comparison**: The judge sees a question and two responses, then selects the better one or declares a tie
  * **Single answer grading**: The judge scores a single response on a numeric scale (e.g., 1-10) with a structured rubric
  * **Reference-guided grading**: The judge is provided a reference answer (useful for math/coding tasks with verifiable solutions)

===== Bias Analysis =====

Three systematic biases were identified and analyzed:

**Position Bias**
LLM judges tend to favor responses presented in certain positions (e.g., the first response). GPT-4 shows high consistency across position swaps, but weaker models exhibit significant position bias.

<latex>
\text{Consistency} = P(\text{same judgment} \mid \text{position swap})
</latex>

GPT-4 achieves the highest consistency rate among tested judges.

**Verbosity Bias**
LLM judges tend to prefer longer, more detailed responses even when the additional content does not improve quality. This can unfairly penalize concise but correct answers.

**Self-Enhancement Bias**
LLMs show a tendency to favor their own generated outputs -- a model used as a judge may rate its own responses higher than those from competitors.

===== Mitigation Strategies =====

  * **Position swapping**: Run evaluation twice with response order reversed, take majority vote
  * **Chain-of-thought prompting**: Require the judge to explain reasoning before scoring, improving accuracy on math/reasoning tasks
  * **Reference-guided grading**: Provide ground-truth answers for verifiable tasks to anchor judgment
  * **Few-shot examples**: Include example judgments to calibrate the judge's scoring

===== Agreement with Human Preferences =====

The central finding is that strong LLM judges achieve human-level agreement:

  * **GPT-4 as judge**: Over **80% agreement** with both controlled (MT-Bench) and crowdsourced (Chatbot Arena) human preferences
  * This matches the **human-human agreement rate** -- human annotators agree with each other at approximately the same 80% level
  * Agreement holds across all 8 MT-Bench categories, though math and reasoning show slightly lower concordance

===== Code Example =====

<code python>
import openai

def llm_judge_pairwise(question, response_a, response_b, model="gpt-4"):
    """Use LLM-as-a-Judge for pairwise comparison."""
    prompt = (
        "Please act as an impartial judge and evaluate the quality "
        "of the responses provided by two AI assistants to the user "
        "question below.\n\n"
        "Avoid position bias -- evaluate based on content quality only.\n\n"
        f"[User Question]\n{question}\n\n"
        f"[Assistant A]\n{response_a}\n\n"
        f"[Assistant B]\n{response_b}\n\n"
        "[Evaluation]\n"
        "Provide reasoning, then verdict: "
        '"[[A]]" if A is better, "[[B]]" if B is better, "[[C]]" for tie.'
    )
    response = openai.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": prompt}],
        temperature=0,
    )
    judgment = response.choices[0].message.content
    if "[[A]]" in judgment:
        return "A", judgment
    elif "[[B]]" in judgment:
        return "B", judgment
    return "tie", judgment

def evaluate_with_position_swap(question, resp_a, resp_b):
    """Mitigate position bias by evaluating in both orders."""
    verdict_1, _ = llm_judge_pairwise(question, resp_a, resp_b)
    verdict_2, _ = llm_judge_pairwise(question, resp_b, resp_a)
    # Reconcile verdicts accounting for the swap
    v2_adjusted = "B" if verdict_2 == "A" else ("A" if verdict_2 == "B" else "tie")
    if verdict_1 == v2_adjusted:
        return verdict_1
    return "tie"  # Disagreement defaults to tie
</code>

===== Impact and Adoption =====

LLM-as-a-Judge has become a foundational methodology in LLM evaluation:

  * **Chatbot Arena** has grown into the most widely referenced LLM leaderboard
  * MT-Bench scores are standard in model release announcements
  * The methodology is used in training pipelines (e.g., RLAIF -- RL from AI Feedback)
  * Frameworks like AlpacaEval and WildBench build on these principles

===== Limitations =====

  * Judge quality is bounded by the evaluating model's capabilities
  * Biases (position, verbosity, self-enhancement) persist even with mitigation
  * Weak reasoning ability limits accuracy on math and formal logic tasks
  * Cultural and linguistic biases from training data affect judgment
  * Cannot replace human evaluation for safety-critical applications

===== References =====

  * [[https://arxiv.org/abs/2306.05685|Zheng et al. (2023) - Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena]]
  * [[https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge|Official MT-Bench / LLM Judge Code (LMSYS)]]
  * [[https://chat.lmsys.org|Chatbot Arena Live Platform]]
  * [[https://arxiv.org/abs/2305.14314|Li et al. (2023) - AlpacaEval: Automatic Evaluator for Instruction-Following Models]]

===== See Also =====

  * [[self_play_fine_tuning|Self-Play Fine-Tuning (SPIN)]] - Training method evaluated using MT-Bench
  * [[agentbench|AgentBench]] - Benchmark for evaluating LLM agents
  * [[tau_bench|tau-bench]] - Benchmark using database state comparison instead of LLM judgment