Alibaba's E-GenRM: generative reward model for open-ended shopping agents hits 87.4% accuracy
_reachsumit · x · 2026-11-05
An Alibaba paper introduces E-GenRM, a generative reward model for open-ended e-commerce shopping agents.
- Data: builds long-horizon conversational trajectories from naturally occurring user behavior in the shopping app, beyond existing usage patterns
- Modeling: SFT on behavior-grounded evaluation chains, then criterion-guided RL rewarding correct preference judgments and evidence-supported claims; personalized hard-negative training improves sensitivity to implicit preferences
- An active evidence retrieval extension acquires missing e-commerce info for evaluation
- On the self-built E-GenRM-Bench, E-GenRM-27B achieves 87.40% preference accuracy, beating the strongest baseline
More from coding & agent
- Founder says three weeks of doing all work inside Claude Code made him 'unreasonably effective' — jacob_posel · 2026-11-05
- Coder of 35+ years: the real sloppy code was written by humans, not AI — facontidavide · 2026-11-05
- Visual Token Routing cuts VLM agent visual tokens by up to 5.8x at same accuracy — wellecks · 2026-11-05
- CI costs scale directly with token prices, dev observes — aniketmaurya · 2026-11-05
- Helay MCP + Claude generates motion graphics narrated in your own voice — aziz4ai · 2026-11-05
- Codex Computer Use Slammed: 20-Second Waits Per Click Make the Loop Unbearable — flowersslop · 2026-11-05