Research map / Large language models / Language, safety and interpretability
The Consistency Dilemma in LLMs: Generator-Evaluator Agreement and Vulnerability to Mistakes
Marina Mancoridis, Zoe Hitzig
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| higher generator-evaluator self-consistency across models | frontier LLMs · MedMistakes-Validated | mistake vulnerability: beta2 = 1.116, p < 10^-6 | Hurt |
| use two-way clustered standard errors (concept and model) | self-consistency association · MedMistakes-Validated | mistake vulnerability: SE = 0.533, p = 0.036 | Hurt |
| add model fixed effects to the regression (ablation) | self-consistency coefficient · MedMistakes-Validated | mistake vulnerability: 1.116 -> 0.147 (SE = 0.334, p = 0.66) | No clear effect |
| higher MedQA benchmark accuracy | frontier LLMs · MedMistakes-Validated | mistake vulnerability: beta1 = -1.732, p = 0.042 | Improved |
| leave out Correct for Incomplete Reason test from consistency score (ablation) | self-consistency score · MedMistakes-Validated | consistency-vulnerability association: substantially attenuates the effect | Hurt |
| model choice: Gemini 3 vs Claude Sonnet 4.5 | ontological tests · List True Attributes | chance-corrected agreement: 1.00 vs -0.23 | Mixed |
| model choice: GPT-5 vs GPT-4o | None of the Above test · None of the Above | chance-corrected agreement: 0.11 -> 0.79 | Improved |
| model choice: Opus 4.5 vs DeepSeek | Correct for Incomplete Reason · Correct for Incomplete Reason | chance-corrected agreement: -1.00 -> 0.73 | Improved |
| stability: modal outcome is agreement vs disagreement | repeated evaluation · 200 instances x 5 runs | stability of judgment: difference in means = 0.056, 95% CI [0.003, 0.113] | Improved |
Closest papers by meaning
- Large Language Models Cannot Self-Correct Reasoning Yet (ICLR 2024 · similarity 0.87)
- Beyond Accuracy: Evaluating Self-Consistency of Code Large Language Models with IdentityChain (ICLR 2024 · similarity 0.87)
- Varying Shades of Wrong: Aligning LLMs with Wrong Answers Only (ICLR 2025 · similarity 0.87)
- On Evaluating LLM Alignment by Evaluating LLMs as Judges (NeurIPS 2025 · similarity 0.87)
- Correlated Errors in Large Language Models (ICML 2025 · similarity 0.87)
- Benchmarking and Improving Generator-Validator Consistency of Language Models (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
