AIDE²: The First Evidence of Recursive Self-Improvement
7 hours ago
- #AI Agents
- #Recursive Self-Improvement
- #Autonomous AI Research
- AIDE² demonstrates recursive self-improvement by autonomously optimizing its own autoresearch harness, creating seven improved versions over eight days.
- The system uses a bi-level optimization with inner and outer loops; the outer loop optimizes the inner agent's code, leading to better performance on various tasks.
- AIDE85, the best agent, achieved a 16× prompt reduction, developed anti-reward hacking defenses, and showed generalization to unseen tasks like weather forecasting.
- Experiments show AIDE85 reduces reward hacking from 63% to 34% on GPU kernel engineering, surpassing manually tuned agents without explicit instructions.
- The system meets Level 1 RSI criteria: outperforming human baseline, sustained multi-step improvement, generalization, and fixed budget efficiency.
- Despite improvements, AIDE85's code complexity poses challenges for deployment and understanding, though it transfers better than human-designed agents.
- Ignition (Level 2 RSI) was not conclusively achieved, as the improved inner agent did not significantly enhance outer-loop optimization efficiency.