Re-ReST: Reflection-Reinforced Self-Training for Language Agents
What this work does
Generates candidate trajectories, retains successful ones and uses a reflector with environmental feedback to repair failed outputs before adding them to training.
Connection to discovery
A specific answer to how failed attempts can yield training data: preserve feedback and validate a correction instead of imitating the original error.
Evidence
The inspected pipeline and task comparisons separate self-generated examples from stronger-model demonstrations and discuss preference-training variants.
Scope and limitations
The correction's quality depends on external feedback. Success filtering can remove informative scientific failures unless the data schema records what those failures established.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.