LLM Self-Modeling: leaderboard
Evaluating and Improving LLM Self-Modeling (2026), Table 16: seven open models predicting their own input-output behavior across math, code, safety and fairness. Strict task scores minus task/model-specific dummy baselines, averaged over five seeds. Nine-task E1-E9 aggregate. Uses unmodified instruction-tuned checkpoints with the published inference settings; no self-modeling fine-tunes or values estimated from plotted bars.
Metric: Skill above dummy predictor (-1 to 1, higher is better; 4096-token thinking budget where enabled). Source: arxiv.org. Source: arxiv.org. Saturation forecast: Around 2030. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.1 (Thinking) | 0.15 | #275 (DeepSeek V3.1) |
| 2 | Qwen 3 8B (Thinking) | 0.11 | #678 (Qwen 3 8B) |
| 3 | Qwen 3.5 35B A3B (Thinking) | 0.11 | #268 (Qwen 3.5 35B A3B) |
| 4 | GPT-OSS-20B (High) | 0.08 | #523 (GPT-OSS-20B) |
| 5 | Qwen 3.5 4B (Thinking) | 0.08 | #514 (Qwen 3.5 4B) |
| 6 | Llama 3.3 70B Instruct | 0.06 | #541 |
| 7 | Llama 3.1 8B Instruct | -0.01 | #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 · How It Works · Data refreshed daily, snapshot 2026-11-05.