Google DeepMind's SIM method lets AI self-improvement experience transfer across models

智东西 · wechat · 2026-11-19

Researchers from Google Cloud AI, Google DeepMind and Penn State published a paper proposing Self-Improvement Memory (SIM), a method enabling continual recursive self-improvement (CRSI). Unlike prior RSI work that only keeps modified harness code, SIM records each change's rationale, expected and actual outcomes—like a searchable experiment log—so optimization experience survives model and task changes.

In 10 controlled coding experiments, SIM-enabled runs won 9 of 10, averaging 2.36 points higher (p=0.003). When a harness optimized for Claude was migrated to Gemini, SIM-enabled optimization topped all four domains, reaching 84.75 on office tasks versus 73.42 without SIM. But inherited experience isn't always a win—it trailed by 0.8 points on Gemini 4 Argon.

Notably, model weights stayed frozen throughout; the AI improves the external harness, not the model itself. First author Xin Yu, a Nankai University statistics graduate now at Penn State, completed the work during a Google internship.

Original post →

More from AGI Musings

AGI Musings channel →