ASCENT: Online Test-Time Training of Long-Horizon Agents via Self-Distillation of Verified Experience
What this work does
Carries LoRA updates across tasks and distills a frozen teacher conditioned on verifier-accepted trajectories, optionally removing invalid-action turns.
Connection to discovery
A concrete alternative to treating every successful or failed generated token as a direct training target.
Evidence
The protocol uses one attempt per task and contrasts unstable direct updates with hindsight-conditioned token distributions on successful experience.
Scope and limitations
A reliable terminal verifier remains essential; unsuccessful episodes do not directly become scientific learning targets. Hindsight is training-time privileged information and must be absent from the student's later decision context.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.