Clarifying AQuA's Self-Improvement Mechanism for Local LLM Replication

A technical clarification highlights that Google's AQuA research agent achieves "recursive self-improvement" by updating its persistent research state rather than modifying the underlying language model's weights. This distinction is crucial for developers attempting to replicate the system using local LLMs. Understanding that the agent's weights remain static helps developers isolate variables when porting the system, ensuring that performance changes are attributed to the state history or prompt harness rather than the model itself. It prevents misconceptions about LLMs dynamically rewriting their own code or weights during runtime in this architecture. To properly benchmark a local port, researchers must control variables across three ablation families: swapping the agent LM under the same harness/state, resetting the research state with a fixed LM, and changing the harness/evaluator contract with a fixed LM and state. Currently, the AQuA preprint lacks a runnable implementation or sufficient detail for end-to-end reproduction.