Recursive self-updates can lose value after the next update
A model update that looks useful can stop being useful when the model updates again, so teams testing self-improvement need to score the full transition, not just the first step. In 64 fresh math and code runs across five 3–8B backbones, continuation reduced utility by 1.78 percentage points in the balanced Qwen3-4B/Llama-3.2-3B comparison, with immediate-selected utility near zero. The authors recommend independent scoring and a no-update baseline.