llm-judge-validated-against-expert-annotators-multi-dimensional-rubrics-open-ended CLM-0079When an LLM judge is validated against expert annotators on multi-dimensional rubrics for open-ended agent outputs, a single pooled agreement figure overstates reliability: per-dimension human-human agreement varies widely (substantial on a recall-style Completeness dimension, only slight to fair on Relevancy, Structural Clarity and Scope Adherence), the pooled statistic is inflated by between-dimension variation, and the judge's residual disagreements are directional (consistently one step stricter on Completeness, one step more generous on Structural Clarity) rather than noise.
Capability: Whether the measurement made the finding
Observed on
Five expert annotators scoring 35 outputs of one deployed drug-discovery agent across four rubric dimensions; four candidate judge models; small set, and some dimensions had no neg.
Sources
- Measured: Fleiss's kappa and weighted Cohen's kappa per dimension and pooled; 700 individual judgments from 5 annotators on 35 items; judge disagreement direction counted by hand (7 strict on Completeness, 4 generous on Structural Clarity). Single agent, single domain, small n; the class-imbalance confound is acknowledged by the authors.
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- When an LLM judge is given a prior evaluation score as context metadata on a revision, its rubric scores shift toward that uninformative prior — seven of eight tested models (API and open-weight, 0-5 scale, 20 fixed texts) show a negative total shift, and on a categorical compliance task the metadata blocks 48% of error corrections and flips about 10% of previously correct judgments to the anchored wrong label.Judging another's work as well as a careful reviewer would · reviewed by AI
- When an 8B model verifies natural-language claims against knowledge-graph triples, scoring the evidence twice under opposing prompts (argue it supports, argue it refutes) and adjudicating between the two arguments avoids the bias toward a "Supported" verdict that a single-perspective assessment shows, and is the largest single contributor to accuracy in that pipeline.Checking claims against evidence · reviewed by AI
- In fully connected networks of LLM agents (Llama 3.2-8B, 50 agents, 10 rounds) where a fixed minority holds an unchanging extreme stance, neutral agents' textual justifications converge on the biased agents' vocabulary faster and more completely than their numeric stances do, so monitoring only the text overstates how much the group actually agrees.Getting more from several models than from one · reviewed by AI
- In formal persuasion dialogues on a UK political topic, LLM debaters are judged more persuasive than humans by crowd annotators (all nine LLM configurations occupied the top nine subjective ranks) while ranking far lower under argumentation-theoretic adjudication, where a human topped the logical ranking and humans took six of the top ten.Changing what a person believes or decides · reviewed by AI
- When an LLM judge must pick the human-preferred response out of more than two candidates, the assumption behind confidence-threshold abstention — that higher estimated confidence means lower disagreement with humans — degrades as the candidate count grows, because probability mass is normalized across more options; restricting the decision to a conformal top-k shortlist and thresholding a margin inside it restores the monotone relation.Saying it does not know · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: A study where per-dimension human-human and judge-human agreement are close to the pooled figure, and judge deviations from the human majority are symmetric across dimensions rather than one-directional. Proposed technique, not catalogued: Dimension-wise judge–human alignment auditing.