July 22, 2026
We are building reasoning models and harnesses around a pattern most systems are trained to remove: divergent, nonlinear thinking of the kind common to ADHD and other neurodivergent minds.
Start from how the model works. An autoregressive model writes one token at a time, and every token narrows what can follow. The first plausible direction it lands on becomes the direction it commits to. When there is a known answer, that is what you want. When there is not — system design, research strategy, naming, anything you cannot fully specify in advance — it is a failure mode. The model hands back the first reasonable idea instead of the best one, and a larger model mostly just says that first idea more fluently.
Chain-of-thought does not solve this. It reasons forward from the same early commitment. Tree-of-thought widens the search but carries the same anchor down every branch. The problem is not that the model thinks too little. It is that it commits too early, and then everything downstream inherits the commitment.
People who are good at hard, open problems do not work this way. A lot of the strongest lateral thinking shows up in ADHD and adjacent cognition: many threads held at once, jumps between unrelated frames, a tolerance for sitting with a half-broken idea until the good one turns up next to it. We think that is a capability to build into a system, not a defect to smooth out of it.
So we build divergence-first. Generate many candidates in parallel, each under a different frame, each kept isolated so it cannot anchor to the others. Then a separate pass scores them, cuts the weak and the misleading, and keeps what holds. Divergence is not a prompt bolted on at the end. It is the shape of the system. We are a research lab, and we publish the harness before we train the model.
The founder
Divergent Labs is founded by Udit Akhouri. He builds autonomous AI systems that run in production: today as an AI systems engineer at Emora Health, and as founder of Brane, governance and compliance infrastructure for LLMs in healthcare. An earlier project, SuperDocs, is used by 500+ developers. He studies computer science at IIT Patna. The consistent thread across that work is systems that keep producing value long after they ship, which is the same bet underneath this lab.
First research — adhd
Our first release is adhd, an open-source reasoning harness for coding agents, written up in a preprint. It runs the architecture above: N isolated branches, each under one of 15 distinct cognitive frames, no shared context while they generate, then a single critic pass that scores every idea, flags the traps, and deepens the survivors into concrete plans.
We benchmark because a claim like "reasons better" is worth nothing you cannot measure. Across six open-ended engineering problems, judged head-to-head against a single-shot baseline, adhd scored higher on the things that actually decide open-ended work: how many real options it finds, how novel they are, and whether it catches the idea that looks right and quietly breaks. It won five of six. The single loss was a problem the baseline already knew cold, which is the honest number to report. Full method and per-problem scores are in the evals.