New study reveals prompt framing inflates LLM error detection scores by up to 0.79 F1

A new preprint by Dekun Yang reveals a critical flaw in how LLM error detection is evaluated. The paper, titled "Prompt Framing Distorts Count-Based Evaluation of LLM Error Detection: Evidence from Numeric Anchoring," introduces ErrorBench, a stress-test protocol that measures the impact of numeric anchoring—suggesting a specific number of errors in the prompt. Over 4,290 responses from 143 CoNLL-2014 passages across six contemporary LLMs (GPT, Claude, Gemini families), anchored prompts produced up to 0.79 points of F1 inflation under standard M2 scoring and up to 0.96 under strict matching. A 100-passage replication using the official ERRANT 3.0.0 pipeline confirmed the pattern: averaged over six models, shifting from a blind prompt to an anchored prompt raised Count-F1 by +0.21 while multi-reference ERRANT F0.5 rose by only +0.04. The gap, termed F1 Inflation, indicates that count-based metrics can rise dramatically without a corresponding improvement in span localization.

The findings highlight that highly instruction-compliant models like GPT and Claude are most susceptible to anchoring, while Gemini family models show smaller response changes. For professionals using LLMs for proofreading or document review, this means standard evaluations may overstate quality if prompts include pre-populated error counts. The study strongly advises against such prompting and urges the use of span-aware metrics (like ERRANT) alongside count-based F1. As LLMs become more integrated into editing workflows, understanding these biases is crucial for building reliable evaluation pipelines.