Large Language Model Evaluation

Description

Structured evaluation of LLM-powered experiences combining rubric-based human ratings, automated metrics and task-level user outcomes such as success, hallucination and trust.

LLM evaluation assesses generative AI features at three levels: output quality (accuracy, groundedness, hallucination, safety, tone), task outcome (did the user accomplish the goal, with how many turns and corrections), and experience (trust calibration, perceived usefulness, reliance). Because outputs are open-ended and stochastic, rigour requires a curated evaluation set representing real user intents, written rubrics with anchored scale points, multiple raters with measured agreement (for example Krippendorff's alpha of 0.67 or above), repeated sampling per prompt, and separate reporting of rare but severe failures. Automated LLM-as-judge scoring can scale evaluation but must be calibrated against human ratings. UX researchers contribute the intent taxonomy, the rubric, and links between output quality and behaviour in logs.

Attributes

Loss type
Active
Data
Mixed
Effort
Weeks

Steps

  1. Build an evaluation set of 100-500 real or realistic prompts stratified by user intent, difficulty and risk, drawn from logs or research.
  2. Write a rubric with anchored levels for each dimension (for example correctness, groundedness, completeness, tone, safety).
  3. Have at least two trained raters score a calibration subset, compute inter-rater agreement and refine the rubric until agreement is acceptable.
  4. Generate multiple outputs per prompt to capture variability, and score all of them, optionally with an LLM judge calibrated against human scores.
  5. Run task-based sessions or log analysis linking output quality to user outcomes: task success, retries, edits, abandonment and expressed trust.
  6. Report scores by intent segment, track severe failure rates separately, and re-run the set on every model or prompt change as a regression suite.

Sources

  • AI for Humans
  • AI
  • Holistic Evaluation of Language Models (Liang et al., 2022)
  • Krippendorff, Content Analysis: An Introduction to Its Methodology

Also known as

  • LLM evaluation
  • AI output evaluation
  • Human evaluation of generative AI
  • LLM evals

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning