adhd

ADHD is a rare thing in the agent-tooling space: a method paper with a running implementation attached, and a clear statement of what problem it does not solve. The diagnosis is the strongest part. 'Premature convergence' names something practitioners feel but rarely articulate — the model evaluates while it generates, early tokens anchor late tokens, and you get the centroid of the training distribution presented as a recommendation. The paper is explicit that this is not a token-level bug but a task-level failure, and that it bites hardest exactly where ideation matters: architecture decisions, API and SDK design, debugging intermittent failures, refactor planning, naming and positioning. The mechanism is defensible. Fifteen cognitive frames give the branches something structural to differ about, rather than asking the same prompt to be more creative. Withholding cross-branch context is the load-bearing design choice: it is what stops branch two from being a paraphrase of branch one. Separating generator and critic into distinct LLM calls with opposing system prompts turns a prompt-engineering promise into an architectural constraint, which is the difference between a technique you can rely on and one you hope holds. The comparison work is unusually honest. The paper positions ADHD as a Tree-of-Thought variant and then names the three ways it departs from ToT: shared versus isolated branches, next-step variation versus vantage-point reframing, interleaved versus mechanically separated generator and evaluator. It also says plainly why Chain-of-Thought, self-consistency, multi-agent debate, and Mixture-of-Agents are the wrong shape here — all four optimise for correctness on a closed answer space, and ideation has no ground truth to test against. Where to be sceptical. The evaluation is six problems, one baseline, and an independent LLM-as-judge as the scoring instrument; the paper itself flags judge bias as a limitation. That is enough to justify trying the method, not enough to treat the +5.17 / +4.17 / +7.67 numbers as a settled benchmark. It is a v0.1 preprint, so expect the method and the frame library to move. Cost profile is the practical constraint. Every ideation run costs N branch calls plus a critic pass, so you are paying multiples of a single-shot prompt in both tokens and wall-clock time. That is a reasonable trade when the deliverable is a set of viable non-obvious options and an obvious answer is the expensive failure. It is a bad trade when the answer is verifiable and one call would have found it. Where it fits: teams running Claude or Codex Agent SDK workflows on design-shaped engineering problems, and anyone who wants a documented, citable structure for divergent ideation rather than an ad-hoc 'give me ten ideas' prompt. Where it does not: closed-form tasks, latency-sensitive interactive agents, and environments that cannot run the Claude or Codex Agent SDK.