推文

@fireandstart · 2026-10-15 02:50

The key idea here is to make progress observable even when the agent has not finished. A latent world that provides an exact answer at every step gives the environment a principled way to distinguish “made no progress” from “got partway there,” without asking a reward model or an LLM judge to invent intermediate labels. That matters for long tasks: if every rollout in a group fails, terminal-only rewards leave RL with little signal about which attempt was closer. The 12.4-minute median rollout and 1,170 verified environments make this a useful step toward training agents on work that looks more like sustained engineering than a two-minute PR. I’d be especially interested in how these gains hold up on tasks where the hidden state and grading rules are less tightly controlled.

曝光 9 · 评论 0 · 点赞 1 · 书签 0