Huxley-Godel Machine: Human-Level Coding Agent Development
A self-improving coding-agent method that guides search by estimating the improvement potential of agent descendants rather than only current benchmark score.
Does descendant-based metaproductivity predict durable self-improvement beyond the benchmarks and search depths used to select it?
HGM sharpens an important point: an agent that scores best now is not necessarily the agent with the best descendants. The paper calls this a metaproductivity-performance mismatch.
Technical move
The proposed CMP metric aggregates descendant performance as a proxy for an agent’s self-improvement potential. Search is then guided toward lineages that are productive, not only toward nodes that look best under the current benchmark.
ASI relevance
Recursive improvement is a lineage problem. If the objective only selects for near-term benchmark wins, the search may discard systems that would have become better improvers after a few generations.