Training scaffold boosts agent scores 18.6%, then deletes itself

A new training method called PATS lifts agent task scores 18.6%, then deletes the scaffolding that produced the gain, per a paper posted to arXiv.

That inverts the standard assumption behind agent skill files, that a curated skill belongs permanently in the system prompt. PATS treats that guidance as scaffolding to remove once training's done, and the deployed model still runs 32.1% leaner on prompt tokens.

PATS builds that scaffolding by turning rollout groups from the agent's current policy into evidence cards fed back in as adjusted context. The method beats strong baselines on ALFWorld and WebShop and holds its own on search-augmented QA, per the same arXiv paper.

What the paper doesn't say is whether this holds up outside ALFWorld and WebShop, in messier real-world tool use. Search-augmented QA is the closest proxy, and the paper only calls the result competitive, not a win.

Watch for

The paper doesn't say how much compute the evidence-card generation adds during training, the real cost against a static skills marketplace. If that scaffolding needs a policy-refresh loop most teams can't run, the deployed skill file stays the practical default. It's the less elegant answer, but it's the one most teams can actually run.