Xiaohongshu's AdaptEvo beats Gemini on content moderation, +7.5 pts over GRPO
小红书技术REDtech · wechat · 2026-10-13
Xiaohongshu's AllSpark team and its community moderation group released AdaptEvo, an agent learning framework for rule-governed contextual decision tasks (e.g., deciding if a wound image violates rules while medical exemptions apply), combining confidence-adaptive RL with evolving decision knowledge and evaluation criteria.
Method:
- CA-GRPO: a "reference confidence" built from annotator agreement plus an external support-assessment model reallocates reward weight — up to 90% outcome reward when support is strong, more process reward when weak, so evidence-gathering behaviors still yield learning signal.
- Training-Evolution alternation: after each round, the policy is frozen, cross-case failure patterns are distilled into decision knowledge (what to check) and refined process rubrics (was it checked), validated before the next round.
Results: on Qwen3.6-35B-A3B, AdaptEvo hits 61.9% exact-label accuracy and 72.2% binary accuracy on multimodal moderation, beating Gemini3.8-Flash for SOTA — +7.5/+3.7 points over GRPO. On out-of-period test sets where rules changed, gains persist (+7.6/+4.3), while GRPO's out-of-period accuracy degrades below the base model over training. Injecting evolved knowledge alone, without parameter updates, adds +3.7/+3.2 points. Paper: arxiv.org/abs/2610.11354
More from coding & agent
- Claude Code Pushes Urgent Security Update Across All Installations — umeshonai · 2026-10-23
- Claude cut a 2.6-hour video export to 20 minutes — but only after being told to rebuild the server — byhewar · 2026-10-23
- 3 flight-booking rules my AI agent checks that I'd never do by hand — Efistoffeles · 2026-10-23
- OpenJet claims Qwen3.8-27B coding runs 1.9x faster than llama.cpp on one RTX 3090 — JeremyCMorgan · 2026-10-23
- Building an autonomous SEO agent in Hermes with three skills: programmatic pages, Keywords Everywhere, DataForSEO — EXM7777 · 2026-10-23
- Building a sports betting AI agent with onchain execution and human approval by default: architecture and lessons — Dry-Comfortable-2678 · 2026-10-23