Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models
What this work does
Alternates policy sampling with reward annotation and reward-weighted likelihood training, instantiated with correctness-filtered math/code solutions.
Connection to discovery
A simple self-training baseline that isolates whether expensive multi-turn RL is needed beyond verified trajectory reuse.
Evidence
Inspected the terminal-reward objective, generate/improve pseudocode and evaluation protocol using answer checks and executable tests.
Scope and limitations
Its binary verified-reward experiments differ from noisy, delayed scientific evidence. Training and validation leakage controls must be redesigned for test-time use.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.