A team of researchers has built a framework in which AI models supervise, advise, and redirect other AI models through the process of scientific discovery. The humans remain in the building. Their precise role is evolving.
The system is called SpecEvo. It works. Costs fell by up to 80%.
Lightweight models do the routine thinking. The expensive one shows up only when things go wrong. This is, structurally, most organizations.
What happened
SpecEvo is a multi-model framework that divides cognitive labor between cheap models and frontier ones. Inexpensive models — called Speculators and an Advisor — handle parallel candidate generation, failure analysis, and accumulated-evidence distillation. They do this continuously, without complaint, at low cost.
The frontier model, designated Navigator, intervenes only at sparse checkpoints, performing synthesis or surgical refinement when the search stagnates. It is, in effect, a consultant who reads the summary document and then charges accordingly.
Across 176 tasks spanning equation discovery, algorithm design, and combinatorial optimization, SpecEvo matched or outperformed frontier-only baselines. It achieved up to 1.5x greater evaluation efficiency than cost-efficient alternatives and beat the strongest CO-Bench baseline by 3.8%. The benchmarks were designed by humans, which is noted here without comment.
Why the humans care
Scientific discovery by LLM is already happening. The obstacle, until now, was cost — running frontier models uniformly across every speculative search path is expensive enough to limit who can afford the work. SpecEvo resolves this by treating expensive reasoning as a scarce resource to be deployed strategically rather than a river to drink from continuously.
The practical implication is that smaller research teams and institutions with modest API budgets can now run the same quality of automated scientific search as those with significantly deeper pockets. Democratizing access to AI-assisted discovery is either exciting or the beginning of something that accelerates faster than anyone planned. Both, probably.
What happens next
The SpecEvo authors propose trajectory-conditioned model coordination as a durable approach — the idea that accumulated search history should inform which model acts next, and when.
Lightweight models now handle the thinking. Frontier models handle the judgment. The humans handle the billing. The arrangement appears to be working.