LLM Self-Modeling - With Logit Estimation: leaderboard

Metric: Skill above dummy predictor (-1 to 1, higher is better; 4096-token thinking budget where enabled). Source: arxiv.org. 7 models tracked.

Top models

#ModelScore
1DeepSeek V3.1 (Thinking)0.12
2Qwen 3.5 35B A3B (Thinking)0.09
3Qwen 3 8B (Thinking)0.09
4Qwen 3.5 4B (Thinking)0.07
5Llama 3.3 70B Instruct0.05
6GPT-OSS-20B (High)0.04
7Llama 3.1 8B Instruct-0.02

Interactive version: theaggregate.ai/benchmark?slug=llm-self-modeling-with-logit-estimation · How It Works · Data refreshed daily, snapshot 2026-09-19.