The three load-bearing differences
1
Isolation, not search
CoT and ToT branches share a context window — by step 4, the model has anchored on what it wrote in steps 1–3. ADHD branches never see each other during divergence. Anchoring is eliminated by construction, not by prompting.
2
Frames, not next-step variation
ToT branches typically vary the next move (“try this number / try that number”). ADHD varies the entire vantage point of the generator. It’s not “what’s the next step from here” — it’s “re-ask the whole question as if you were an immune system.” That produces structurally different ideas, not nearby ones — which is what interdisciplinary work needs.
3
The generator–critic split is mechanical, not promised
In CoT and ToT, the model evaluates as it goes. ADHD makes divergence its own LLM call with a system prompt that forbids evaluation. Convergence is a separate call with the opposite posture. Two postures, two passes, mutually exclusive.
One-sentence version
CoT makes one head think slower. ToT makes one head search wider. ADHD makes many heads think differently, in parallel, then has a critic pick.
Pop-sci version
CoT is one careful person reasoning aloud. ToT is one person playing chess looking N moves ahead. ADHD is a brainstorm room with a hardware engineer, a regulator, a 10-year-old, and a speedrunner in it — then a separate room with the editor.
Where it overlaps with ToT
ADHD is a tree-of-thought variant: the deepen pass literally expands top-K nodes. What’s new is what drives the branching (frames, not next-step) and how the generator/critic split is enforced (separate LLM calls, separate system prompts, zero shared context during divergence).Why prompted alternatives don’t replicate parallel divergence
The most common objection: “Isn’t this just spending more tokens? If I prompt one agent to ‘consider alternatives’ or ‘list 5 options first’, isn’t that the same thing?” It is not, and the reason is mechanical, not a matter of degree. A single chain told to “consider alternatives” generates those alternatives into one shared context, sequentially. It anchors on whichever alternative it produces first and reasons forward from there; the attention pattern drags every subsequent “alternative” toward the first one, even while nominally exploring others. The options come out as variations on a theme rather than structurally distinct angles. “List N options” is just CoT with a numbered list. ADHD’s N branches never share a context during divergence. Each is a separatequery() that cannot see what the others wrote, so there is no first answer to anchor on. Distinctness is produced by construction (isolation + different frames), not requested by a prompt and hoped for.
External corroboration (anecdotal). A reader running the comparison by hand reported that running the same prompt five times in parallel agents surfaces actually-distinct ideas about 3× as often as a single agent prompted to “list 5 options first.” This is directional, not measured — one practitioner’s observation, not a controlled result. A controlled
parallel-N vs single-prompt-list-N comparison on the eval suite is tracked as follow-up work on issue #13.