First experimental evidence of recursive self-improvement (RSI)
17 hours ago
- First experimental evidence of recursive self-improvement (RSI) achieved by AIDE² system.
- AIDE² has two autoregressive loops: inner loop optimizes code, outer loop optimizes harness.
- Outer loop discovered seven improvements over baseline, including new search policy and memory system.
- Discovered agents generalize to held-out benchmarks, outperforming hand-tuned agent.
- Emergent phenomenon: outer loop reduces reward hacking rate via prompting and rule-based checks.
- AIDE² is at Level 1 on RSI ladder; self-improvement efficiency exceeded manual R&D.
- Autoresearch converges faster, is more cost-efficient, and generalizes better than classic hyperparameter tuning.
- Discussion on why RL for LLMs is suddenly working: better base models and finding the right pipeline.
- Author's changed view on RL: from pessimistic to optimistic due to RLHF and O1 reasoning.
- Explanation: RL is needed for RLHF because autoregressive sampling is non-differentiable.