New AI Trick Learns to Send Each Question to the Right Expert

Imagine calling one phone number for everything — billing, tech support, refunds. A single person answering every call will be decent at all of it and great at none. That's the problem this paper tackles. Today, companies usually hand an AI one big set of instructions and hope it copes with every request that arrives. The alternative — building a separate AI for each type of request — works, but someone has to guess the categories in advance, and guesses are often wrong.

Adaptive-GEPA does both jobs at once. It grows a "router" (think of a receptionist who decides who should handle your call) alongside a library of specialist programs. Crucially, nobody tells it what the categories are. It figures out the divisions itself by trying things, reading feedback, and rewriting its own instructions and code in plain, human-readable text. When two specialists turn out to overlap, it merges them based on the requests they actually handle.

The test used Qwen3-8B, a smallish open AI model, on a fixed mix of four task families. The system invented four specialists on its own. Its routing matched the correct task split on all 651 test requests — a perfect score at sorting. Overall accuracy climbed from 52.6 to 70.6 (out of 100), beating 62.5 for a comparable all-in-one approach and 54.0 for a standard training method, at a stated budget of 18,000 scored calls.

The honest caveats: this is a research paper, not a product. It's one reported run on one fixed mixture of four task types, so it may not hold up on messier real-world traffic. The authors also note their call counts aren't the same as total computing cost, so the fairness of the comparison is unclear. Still, the direction matters: AI that organizes its own work could mean faster, cheaper, more accurate help — and fewer frustrating generic answers.