• Web导航Agent在线学习终于跑通了，OpAgent在WebArena上达到71.6%成功率，比之前所有单体模型翻倍

• 奖励模型不一定要"正着判"。FLIP反过来推断指令，小模型比LLM-as-Judge强79.6%

• RL不止能训生成模型，也能训embedding模型的推理链，Embed-RL让跨模态检索学会了"先想再查"

阅读全文 →

• Online RL finally works for web navigation agents. OpAgent hits 71.6% on WebArena — more than doubling every previous monolithic baseline.

• Reward models don't have to judge forward. FLIP infers backward to the instruction, and a 7–9B model beats LLM-as-Judge by 79.6%.

• RL isn't just for alignment anymore — it now trains embedding reasoning chains. Embed-RL teaches cross-modal retrieval to "think before matching."

Read more →