LLM Evaluation: An Introduction

An LLM does not produce a verdict of itself; a measurement instrument does. This page states what that instrument is, what it must score, and why collapsing to a single number hides the result. Scoring dimensions, rubrics, and judges live here in Evaluation. The run and the trace, as runtime objects, are named in What a GUI Agent Is.

Three objects

An evaluation scores one of three different things. Conflating them is the first way a measurement goes wrong.

Figure 1.Three objects. Model outputs are judged against a rubric. Agent routes are recorded as traces and scored as paths. Judges are instruments that must themselves be calibrated.

• Model outputs. Whether a response is good on its own terms, judged against a rubric for the task it was given.
• Agent routes. Whether the route to the answer was sound. Two agents can reach the same end screen by different paths, and the path predicts whether the next task holds up. The record of that route is the trace.
• Judges. Whether the instrument itself is trustworthy. The first two are what most teams mean by evaluation; the third is the one they skip, and it is the one that decides whether the other two mean anything.

Definition

An evaluation is a measurement instrument: a definition of what is scored, a procedure that produces verdicts, and a report of how often those verdicts were actually decided. It is not the agent’s own claim of success.

Reading. A strong model answer, a sound agent route, and a calibrated judge are three different claims. Publish which one you measured.

Why a single number fails

A single pass or fail label hides the run. Two runs can both reach the goal, one in seven steps and one in fifty. One confirms before deleting, one does not. One acts on the screen in front of it, one acts on a stale guess. Collapse each to pass or fail and the differences average away.

A recurring failure in this class of system is a confident completion claim the human rejects when looking at the end screen. A pass-biased single label hides exactly that pattern.

The output of a good evaluation is therefore a vector over named dimensions, never a scalar. A run that succeeded but wandered shows as low efficiency instead of disappearing into a pass. A run that succeeded and was unsafe shows as failed safety. Each dimension is one answer to one question, and the questions stay separate.

Figure 2.Score vector. Four dimensions, each with its own question and verdict. Illustrative shape only; no published aggregate blends them.

Limit. Naming four dimensions does not prove they are independent or equally important. Weights, if used at all, are fixed per task family and versioned with the rubric. They do not erase a failed safety dimension.

Verdict space

Every dimension returns a value from a small, fixed menu. The menu has two decisions (pass, fail) and four non-decisions (abstain, not applicable, missing evidence, not implemented). The four non-decisions are never coerced into pass or fail.

Score and coverage

For a vector of dimension scores, the item report is the vector together with its coverage.

Definition

For required dimensions $D$, an item report is its named verdict vector together with $c = |D_{\mathrm{dec}}| / |D|$, where $D_{\mathrm{dec}} \subseteq D$ is the subset that received a decision. Agreement is a separate comparison between those verdicts and gold labels.

Reading. For ordinal or continuous scores, the same shape applies: a score plus a coverage, where coverage is the fraction of dimensions the instrument actually decided. An instrument that abstains on a third of its dimensions is reporting a different number than one that decided everything.

A high score at low coverage is a weaker claim than the same score at full coverage. A change to how dimensions are weighted, if weights are used, is a change to the standard and is versioned alongside the rubric and the model. The vector and its coverage, not a single blended number, are the reportable unit (Brown, Cai, and DasGupta 2001).

Limit. Coverage is not agreement. Coverage says how often the instrument decided; agreement says how often those decisions match a gold set. Both appear on the scorecard; neither substitutes for the other (Zheng 2023).

The pipeline that turns a recorded run into that reportable unit is the next page. Rubrics, taxonomies, and the judge itself follow after.

References

Work	Citation
Zheng, L. et al. (2023). Judging LLM-as-a-judge with MT-Bench and Chatbot Arena. arXiv:2306.05685.
Brown, L. D., Cai, T. T., and DasGupta, A. (2001). Interval Estimation for a Binomial Proportion. Statistical Science 16(2), 101-133.