Obsidian Metadata
| type | source |
| url | https://arxiv.org/abs/2404.12272 |
| authors | Shreya Shankar, J.D. Zamfirescu-Pereira, Björn Hartmann, Aditya G. Parameswaran, Ian Arawjo |
| published | 2024-04-18 |
Source: Who Validates the Validators? Aligning LLM-Assisted Evaluation of LLM Outputs with Human Preferences
URL: https://arxiv.org/abs/2404.12272 (UIST 2024)
LLM-generated evaluators “inherit all the problems of the LLMs they evaluate”, so they need human validation. The paper presents EvalGen, a mixed-initiative interface that proposes criteria and candidate implementations (Python assertions or LLM grader prompts), asks users to grade a subset of outputs, and selects the implementations that best align with those grades.
Key finding: Criteria Drift. Users need criteria to grade outputs, but grading outputs helps them define criteria. Some criteria are dependent on the specific outputs observed and can’t be defined a priori. This challenges approaches that assume evaluation is independent of observing outputs.
Also cited for over-trust in AI (the MIT EECS/GPT-4 self-grading incident). Most-cited paper across the FAQ.