Prompt Framing Distorts Count-Based Evaluation of LLM Error Detection: Evidence from Numeric Anchoring
Dekun Yang
Why It Matters
What makes this one worth your time
Understanding the limitations of count-based metrics is crucial for developing more accurate evaluation methods for LLMs, impacting both research and practical applications in error detection.
Count-based F1 scores can misrepresent LLM error detection quality due to prompt framing effects.
Summary
The paper demonstrates that count-based F1 scores can inflate without corresponding improvements in error detection accuracy, introducing a new evaluation protocol called ErrorBench to assess this phenomenon across multiple LLMs and prompt conditions.
Key contributions
- Introduction of ErrorBench, a controlled protocol for evaluating prompt-induced count distortion.
- Empirical evidence showing significant F1 Inflation across multiple LLMs and prompt conditions.
- Recommendations for using span-aware metrics alongside count-based metrics in LLM evaluations.
Notable insights
- The concept of F1 Inflation highlights a critical gap in current evaluation practices for LLMs, suggesting that traditional metrics may not fully capture model performance.
- The study's findings indicate that different LLM architectures respond variably to prompt framing, which could inform future model design and evaluation strategies.
Possible limitations
- Not stated in the abstract.
Abstract
arXiv:2607.01240v1 Announce Type: cross Abstract: Count-based F1 is widely used as a proxy for LLM error-detection quality, but this paper shows that it can rise dramatically without a corresponding improvement in span localization, a gap termed F1 Inflation. The paper introduces ErrorBench, a controlled stress-test protocol for prompt-induced count distortion. ErrorBench evaluates six contemporary LLMs under five prompt conditions over 4,290 responses from 143 CoNLL-2014 passages. Under CoNLL-2014 M2-style scoring, anchored prompts produce up to 0.79 points of F1 Inflation, and up to 0.96 under strict matching. A 100-passage replication using the official ERRANT 3.0.0 pipeline and multi-reference scoring reproduces the pattern: averaged over six models, the Blind-to-Anchored prompt shift raises Count-F1 by +0.21 while raising multi-reference ERRANT F0.5 by only +0.04. The study finds larger count responses in highly instruction-compliant GPT/Claude systems and smaller responses in the Gemini family under this stress-test protocol. The findings suggest that LLM proofreading and document-review evaluations should avoid pre-populated error counts and should report span-aware metrics alongside count-based metrics.