Gemini 4 Argon lands #3 on RSI-Exam self-improvement leaderboard
HuaxiuYaoML · x · 2026-10-26
RSI-Exam benchmarks recursive self-improvement: agents spend hours autonomously improving a task-solving method or the harness driving a frozen model, then get one final run on hidden tests across 88 tasks (35 public, 53 private) in six domains. Huaxiu Yao reports Gemini 4 Argon scored 0.4943 (Full) and 0.5153 (Public), both #3 — a big jump from Flash models, with Gemini 4 Carbon awaited. The shared leaderboard image also lists Opus 5.5 at 0.5357 and GPT-6-astra at 0.5126. Unverified: some listed model versions have no official release info.
More from Models
- Best sub-40B model on Terminal-Bench 4.0: Nanbeige4.3-33B at 23.8% accuracy — NielsRogge · 2026-10-26
- Frontier models are better judged by five minutes of weird jokes than 40 benchmark numbers — korymath · 2026-10-26
- Voice models sounding too human called 'the most braindead direction in AI' — yungcontent · 2026-10-26
- Claude free users seemingly can't pick thinking level; UI shows panel but it closes — MilkerGerk · 2026-10-26
- Hamel Husain flags poor benchmark quality: Epoch audits find Terminal-Bench, OSWorld 2.1 all flawed — HamelHusain · 2026-10-26
- Community releases layer-selective abliterated DeepSeek-V4.1 finetune on Hugging Face — AIFlow_ML · 2026-10-26