Judges and verification
When there is no exact answer, use an LLM-as-judge or a debate between models — carefully. A judge is a heuristic, and it inherits the biases of the models it grades.
Many agent outputs have no exact answer: a summary, a plan, a diagnosis. For these you cannot write a test, so you use a judge — often another model — to score the result against a rubric. It is the only scalable option, and it is a heuristic that can be wrong in systematic ways.
Start here
A judge is a model, so it has the same biases as the model it grades: it favours longer answers, its own style, and confident-sounding text. Use it, but treat its scores as noisy evidence — and anchor it with rubrics, references, and as much mechanical checking as you can.
One model, five benchmarks — same checkpoint, five different stories. Click each for its caveat.
Multiple-choice: format-coachable. Widely cited, easy to game via training data overlap.
Scores are illustrative; the per-benchmark caveats are the lesson. Always ask what a benchmark measures — and whether the model saw its answers during training.
LLM-as-judge
- Pointwise — score one output against a rubric. Cheap; drifts between runs unless the rubric is tight.
- Pairwise — show two outputs and ask which is better. More reliable than absolute scores, which is why arenas use it.
- Known biases — position bias (favouring the first option), length bias (longer looks better), and self-preference (favouring its own family's style).
Debate and verifiers
- Multi-agent debate — several models answer, critique each other, and converge; the agreement is often more reliable than any single answer.
- Verifier models — a separate model trained to check a specific property (does this code compile, does this claim follow). Narrow and mechanical beats general and vague.
- Wherever you can, verify mechanically — a test, a schema, a tool that disagrees. A ground-truth check beats any judge.
Careful
Judges are gameable. Optimising an agent against a judge teaches it to please the judge, not to be right — the reward-hacking story from the LLM domain, applied to evaluation itself. Keep a human-audited sample, measure the judge against human labels, and never let a judge be the only thing standing between a change and production.
Check yourself
Eduspheria wiki · Agentic AI, Evaluation & observability
0 / 4 answered
Next: seeing what an agent actually did — tracing.