SEED teaches an agent by making it its own tutor

Long-horizon agents trained with RL have a dumb problem: you only find out at the very end whether the whole episode worked, so the model gets almost no signal about which of its hundred individual moves was good or bad. SEED, out of Tsinghua and a few other labs, closes that gap by making the agent tutor itself.

Here's the trick in plain terms. First, teach the model to look at a finished trajectory and write down, in plain language, what skill it just used. Then during RL, take the current checkpoint and run it as two things at once: an actor that collects trajectories, and an analyzer that writes hindsight skills for what it just did. Those skills become dense, token-level supervision. The model compares how likely an action was with and without the skill hint, and that difference becomes the teaching signal. No separate teacher model, no human labels. It grades its own moves while it plays.

The numbers are not subtle. On ALFWorld it hits 91.8% average success versus 75.0% for GRPO, the standard RL baseline. It also transfers to unseen task variants and needs fewer samples to get there. Web navigation and search QA show the same pattern.

This is the self-improvement thesis getting concrete: the agent that keeps score on itself keeps getting better without you in the loop. Code at github.com/jinyangwu/Seed, paper arxiv.org/abs/2607.14777