Paper finds overall LLM evaluation accuracy can obscure prompt-specific errors in interactive settings
Researchers show that high aggregate scorer performance can mask prompt-dependent mistakes, leading to unreliable comparisons between model responses under different conversational pressures.
1 source · cross-referenced
- A new arXiv paper argues that overall accuracy or F1 scores for automatic LLM evaluators can hide prompt-specific errors that distort comparisons between model responses.
A paper accepted to the Trustworthy AI for Good (AI4GOOD) Workshop at NeurIPS 2026 argues that commonly used aggregate metrics like overall accuracy or F1 scores for automatic LLM evaluators can obscure prompt-specific errors. These errors, in turn, can distort comparisons between how models respond under different conversational pressures, such as neutral requests to reconsider versus insistence that an invented entity exists.
The author demonstrates the issue using six models, 180 invented-entity questions, and six prompt conditions. In a balanced audit of 288 responses, the overall F1 of the automatic scorer was .916. However, the scorer’s prompt-specific error rates varied widely: it flagged between 2.6% and 33.3% of responses as containing unsupported claims when a primary LLM rater judged they did not. These prompt-specific estimates were also uncertain.
In a second audit comparing the automatic scorer to two separate LLM raters on 300 model-question cases under each prompt, an apparent 9.3-point increase in measured effect (from asserting certainty versus a neutral request to reconsider) shrank to 2.3 points under both raters. The primary-rater interval included zero, so the study could not establish a reliable increase. Two other prompt comparisons retained their direction under both raters, while an attempted prompt-specific correction remained too uncertain to draw firm conclusions.
The paper concludes that researchers should validate not just the scorer’s overall performance, but the exact differences they plan to report. This requires auditing both prompts on matching cases and explicitly incorporating scorer and rater uncertainty into the final results.
- Nov 6, 2026 · OpenAI — News
OpenAI releases BioBasicsBench to evaluate AI agents on wet-lab science tasks
Trust79 - Nov 3, 2026 · Hugging Face
Allen Institute releases StateBench to test long-context state tracking in language models
Trust79 - Nov 3, 2026 · Hugging Face
CWE-bench v2 expands public cybersecurity evaluation suite to 100 tasks across 91 CWEs
Trust79