ADHD: Parallel Divergent Ideation for Coding Agents
SMRTR summary
ADHD is a new inference-time method for LLM coding agents that fights premature convergence — the tendency to default to the first plausible idea. It runs N parallel branches under structurally different cognitive frames (e.g. regulator, speedrunner, biology), with strict generator/critic separation, then scores and deepens only the top survivors. Tested on six open-ended engineering problems, ADHD wins 5/6 against a strong single-shot baseline.
SMRTR provides this summary for quick context. The original article belongs to Hacker News.
Read the original article