This paper shows a better way to train multi-turn agents:

score each turn separately, then use a self-teacher to focus learning without letting it override the reward.

Standard RL has a basic problem.

A long agent session can contain successful and failed turns, yet 1 overall reward can blur them together.

CREST fixes that by giving each turn its own verified credit, then using the same model as a teacher to put more learning weight on uncertain decisions inside that turn.

The teacher can strengthen an update, but it cannot reverse the verifier's judgment.

On Qwen3-4B-Instruct, it reaches 52.0% average BFCL V3 accuracy versus 49.25% for the strongest RL baseline.

– arxiv. org/abs/2608.13179

Title: "Teach the Magnitude, Not the Direction: Verifier-Bounded Credit Assignment for Multi-Turn Multi-step LLM Agents"