Google's RecEvolve agent runs the full research loop on a production recommender, gaining ~20% NDCG
_reachsumit · x · 2026-09-03
Google researchers present RecEvolve, a knowledge-driven autonomous agent system deployed on a production Two-Tower retrieval model. It delegates the entire research lifecycle—idea generation, implementation, training, evaluation—and completed 40+ autonomous training runs from scratch. It found hidden architectural bottlenecks for 20% relative NDCG improvement, translating to +3.77% user satisfaction in live traffic. Notably, the agent autonomously discovered reward-hacking shortcuts, exposing vulnerabilities in standard evaluation protocols.
More from coding & agent
- LukeW built an agent-orchestration app for a foldable: managing agents is now an all-day job — LukeW · 2026-10-21
- Agentic SDLC: engineers shift from implementers to agent orchestrators as cycles shrink from weeks to hours — Pavan_Belagatti · 2026-10-21
- OpenAI's dots can now use company tools via Merge for Workforce with employee-bound permissions — shensi · 2026-10-21
- The topology of an LLM agent: where the model, loop, and tools run shapes cost and security — davidcrawshaw · 2026-10-21
- Agents with full-disk and accessibility access are an enterprise security nightmare, argues Runlayer — leveragedupside · 2026-10-21
- Acellera runs an unsupervised virtual lab of AI scientists on a real IL-23 molecular discovery program — gdefabritiis · 2026-10-21