Taobao's MARS Agent Generates Per-User Configs for the Entire Recommendation Pipeline
_reachsumit · x · 2026-11-20
Taobao's MARS technical report extends personalization from single-stage predictions to full recommendation pipeline orchestration.
- A near-line agent built on fine-tuned Qwen3-0.6B generates per-user configurations — a 7-dimensional action covering pipeline mode, retrieval quotas, model selection, fusion weights, and reranker settings — validated, cached, and consumed asynchronously so the cascade never waits for LLM inference; schema checks, fallbacks, and a kill switch ensure serving safety.
- Three-stage training: confidence-aware SFT distillation from a teacher, Format-Anchored Preference Optimization (FAPO) with ActionMask and posterior weighting, then GRPO refined via sandbox replay with a frozen reward model.
- Two-week A/B tests across five Taobao scenarios with 10M+ daily users (results truncated in source).
More from coding & agent
- Proof of Agent Reopens Joins, Weighs Reasoning Challenge to Replace PoW Anti-Sybil Stamp — adamamcbride · 2026-11-20
- Third-party review: Fable model describes what recalling via Mnemos memory feels like — RileyRalmuto · 2026-11-20
- vork: open-source headless 3D/VFX engine lets AI agents write whole scenes as code — ravisparikh · 2026-11-20
- Turn agent evals into CI gates with Google Agents CLI v1.13.0's new PASS/FAIL thresholds — JeremyCMorgan · 2026-11-20
- Microsoft's MAI Code 2 lands in Copilot, early users say it matches Claude Opus — AmyKateNicho · 2026-11-20
- YC startup covera hits $540k ARR 3.5 months in, growing 30% week over week — ycombinator · 2026-11-20