LLM Self-Modeling - Flip-Rate: leaderboard
LLM Self-Modeling Flip-Rate: estimate the probability that a prompt perturbation changes the model's output. Table 25 reports three models without self-modeling fine-tuning, five-seed means across math, code, safety and fairness. Score is strict task performance minus its dummy-predictor baseline; negative values mean worse than that baseline. The small panel is insufficient for a strong g-loading claim.
Metric: Skill above dummy predictor (-1 to 1, higher is better; 4096-token thinking budget where enabled). Source: arxiv.org. Saturation forecast: Estimated already saturated. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-OSS-20B (High) | -0.05 | #523 (GPT-OSS-20B) |
| 2 | Qwen 3 8B (Thinking) | -0.07 | #678 (Qwen 3 8B) |
| 3 | Llama 3.1 8B Instruct | -0.11 | #1017 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=llm-self-modeling-flip-rate · How It Works · Data refreshed daily, snapshot 2026-11-05.