This paper argues that LLMs converge too early on ideation tasks and fixes it by running parallel isolated branches under different cognitive frames, then collapsing them with a separate critic pass. Results are decent.
The naming is intentional. ADHD as a cognitive profile is exactly this: broad unfocused attention, many parallel threads, resistance to premature closure.
Which made me wonder what else is on the table. OCD as a verification agent that structurally cannot stop looking for edge cases. Dyslexic processing — more holistic, less sequential — for architecture reviews where local token-by-token reading is exactly the wrong move.
Autism-spectrum hyperfocus for formal specs and compliance checks where “close enough” is not acceptable.
Cognitive diversity as an inference-time design primitive. Someone should explore this more systematically.
— Felix