Definition
Plain language
Salvaging failed AI training trajectories by training only on the parts where the agent was actually making progress.
As stated in the literature
A supervised fine-tuning method that uses an LLM to estimate per-step value along failed teacher trajectories and trains the student only on segments preceding the critical mistake.
Why it matters: It rescues training signal from runs that would otherwise be thrown away, while keeping the student from imitating the steps that actually caused the failure.
For example, a teacher trajectory fails on turn 18, and the student is trained only on turns 1 through 14, where an LLM judge says the agent was still on track.
Heard on the show
“The first is called credit-assignment SFT, and it's about the problem of failed trajectories.”Episode 047 — When Agent Benchmarks Lie: The Harness Problem in Open-Source AI