Alibaba's G-CAST: governed continual self-training keeps e-commerce agents fresh in production
Yuechun Yu · hf · 2026-11-16
The G-CAST paper tackles two production problems for LLM e-commerce agents: demand drift (new categories and scenarios outdating a one-shot-trained model) and harness evolution (model must adopt new tools/skills without breaking behavior users rely on).
Core mechanisms:
- Reward-routed, entropy-prioritized data orchestration: failures go to an SFT pool, mixed successes to an RL pool, clean successes only monitored
- A governed reward evolution loop with training pauses, offline reward checks, and trial adaptations
- Governed checkpoint transitions gated by cumulative evaluation, rejecting candidates that regress
Experiments on a real production agent show stronger adaptation to new demands and better retention of accepted behavior than a strong autonomous post-training baseline, with ablations confirming both components matter. G-CAST is deployed on Yuancai, Alibaba 1688's B2B procurement agent.
More from coding & agent
- shadcn boosts @uisdk: one-shot UI rebuild that's instantly E2E-testable by AI — shadcn · 2026-11-16
- AI agent deletes 43,599 lines of OpenClaw test code in one PR — vincent_koc · 2026-11-16
- JPMorgan shows Jarvis, its internal agent assembly line, at Interrupt NYC — LangChain · 2026-11-16
- Dev clarifies: Apple doesn't ban feature flags, only fooling App Review — talkaboutdesign · 2026-11-16
- Antigravity Matches Claude Cowork in Tests, While Gemini Tasks' Purpose Remains Unclear — sebajun2 · 2026-11-16
- One 16GB GPU runs songs, voice, images, 720p video and local RAG — all offline and open source — llama-of-death · 2026-11-16