LLM-as-a-Judge

Mainstream approaches to agent evaluation include LLM-as-Judge, trajectory matching, and outcome verification.

The LLM-as-Judge method uses large language models such as GPT-4 as judges to automatically score the output quality of Agents.

LLM-as-a-Judge方法存在位置偏差、冗长偏差和自我偏好偏差等已知系统性偏差

AI system evaluation is divided into offline evaluation and online evaluation; offline evaluation can use a stronger model as a judge (LLM-as-a-Judge).

The LLM-as-a-Judge evaluation paradigm was first applied at scale by the LMSYS team in MT-Bench and Chatbot Arena.

LLM-as-a-judge模式即用一个能力更强的模型来评判另一个模型的输出质量，降低评测成本但引入评判模型自身的偏见问题

LLM as Judge was systematically proposed by Lianmin Zheng et al. in the MT-Bench paper (2023), with the core idea of using an independent large language model to score the output of another model as a substitute for human annotation.

The LLM-as-a-Judge paradigm was originally used to evaluate the output quality of generative AI and is now being applied to complex reasoning and judgment scenarios such as resume-job matching.

The study adopts a simulation-based evaluation method, uses real-world data, and combines LLM-as-a-Judge evaluation with human verification.

LLM judges exhibit position bias: in pairwise comparisons, judges tend to favor the answer in a certain position, usually the first; merely swapping the order of the candidate answers can flip the ranking (Wang et al.).

Having an LLM judge whether its own answers are correct is circular reasoning, because hallucinations and correct outputs may be structurally indistinguishable.

The fundamental flaw of LLM-as-a-Judge is that the judge and the judged share the same set of language priors, making it easy to misjudge well-formatted, logically coherent text as factually correct.

Using large models as judges (LLM-as-a-Judge) has become the mainstream approach for evaluating open-ended tasks.

An Eval is a quantitative check of Agent performance, with common forms including comparison of outcome correctness, judgment of the reasonableness of tool calls, and automatic scoring based on LLM-as-a-Judge.

LLM-as-a-Judge has three known biases: a tendency to prefer longer answers, greater tolerance for the outputs of its own model series, and sensitivity to position (options appearing earlier are more likely to be selected).

The project combines four types of evaluation methods: rule-based deterministic checks, LLM-based judge evaluation, sampling evaluation based on manual annotation, and online evaluation based on user feedback.

Academic research shows that when scoring criteria are strictly restricted to binary pass/fail standards, automated scorers can provide scalable, reproducible scores that are highly correlated with expert human judgment.

The more prudent industry practice is to have the reviewing model and the generating model belong to different families, or to introduce manual spot checks for calibration.

In actual evaluation, faithfulness is usually automatically determined with LLM-as-a-Judge or NLI-type models, breaking the answer into atomic claims and checking one by one whether each is entailed by the context; correctness relies on human-annotated reference answers or authoritative fact sources.

确定性检查器相对于LLM-as-a-Judge评估范式的优势是结果完全可审计、可独立复现，代价是任务设计难度更高