Tag
Introduces Pyligent, a training framework that uses task validators to label failures and teaches LLMs to backtrack during reasoning, improving solve rates on hidden graphs, Sudoku, and Blocksworld.
The paper introduces VGB, a process-guided sampling algorithm with probabilistic backtracking, which significantly improves coding performance on tiny 0.5B models by being robust to verifier errors.
This paper introduces Motab, a new pipeline for LLM reasoning distillation that mitigates both off-policy and on-policy exposure biases by dynamically monitoring student generation and backtracking to safe states with teacher intervention, achieving ~3% average improvement.