Self-Evolving Coding Rules for AI Coding Agents

简体中文

L1 validation-only selection after mutation; no LLM mutator is called.

Paper information

Field	Details
Paper	arXiv v1
Institution	The paper's first page does not list an affiliation for the first author.
First public release	2026-09-30 (arXiv v1)
Author code	No public author implementation found as of 2026-10-03.
Local adapter or method key	rule-evolve
Local reproduction code	src/auto_research/agent_research/latest_20261003.py

Paper method and evidence

Maintain a pool of coding rules, generate variants with a mutator, then select and persist better rules using only an isolated validation evaluator.

flowchart LR
  P[Coding rule pool] --> M[Propose mutations]
  M --> V[Score on isolated validation]
  V --> R[Write back selected rules]
  R --> P

Original paper figure

Figure 1 shows mutation, validation selection and write-back. From the original paper; copyright remains with the authors.

Core equation

\(r^*=argmax_{r\in pool\cup mutate(pool)}Score_{val}(r)\).

Reported results

Across two coding agents, four models, and three benchmarks the paper reports gains over handwritten rules and prompt-optimization baselines.

Local reproduction

Compare against the mechanism disabled or in its default state. The three-seed diagnostic is diagnostic_only=true; no formal relative gain or capability ranking.