An AI system that can generate ideas, write code, run GPU experiments, and analyze the results — without a human steering every step — just got a real-world trial at a major online travel company.
The team built what they call an AI Co-Scientist framework, wiring LLM agents directly to cloud compute so the full research cycle runs end-to-end. For routine work, a single model handles the task; for higher-stakes decisions, the system convenes a panel of three frontier models — GPT-5.2, Gemini Pro 3, and Claude Opus 4.5 — and goes with consensus. Layered on top of a human-designed transformer baseline, the automated loop added +0.083 percentage points in offline ranking quality over roughly one extra week of wall-clock time, bringing the combined gain over a pre-transformer baseline to +0.201 points.
The most useful proposals the AI surfaced — unified long-sequence layouts, slot-type embeddings, and multi-phase learning-rate schedules — are standard practice in NLP and computer vision but had never made it into the team's production stack. That gap is the real story: the system didn't invent new ML; it acted as a cross-disciplinary connector, bridging knowledge the field already has but that hadn't reached this particular team. For organizations running specialized ML groups, that reframes the pitch from "AI replaces researchers" to "AI reads the papers your team hasn't gotten to."
At the scale of a major travel search engine, 0.083 percentage points in ranking quality can move real booking revenue — so the gain is less trivial than it reads. Whether the multi-LLM consensus architecture was necessary to find it, or just expensive overhead, is a question the paper raises but doesn't fully settle.