Pyligent Framework Improves LLM Reasoning by Teaching Failure Recovery.

Many complex reasoning tasks require more than a linear, left-to-right approach; solvers often need to explore a path, recognize a failure later on, and then backtrack to a viable point to restart. This research introduces Pyligent, a novel training and inference framework inspired by the Diligent Learner formulation, which models reasoning as a validated search process over partial solution chains. Pyligent incorporates a task validator that labels generated continuations and failures. The resulting search trees are then converted into supervised targets for three distinct actions: "continue," "finish," and "backtrack." Optionally, traces summarizing abandoned branches can also be included in the training. The framework was evaluated on a hidden directed graph task specifically designed to test delayed-failure recovery, as well as on structured reasoning domains with exact validators like 4x4 Sudoku and Blocksworld. Compared to traditional gold-only supervised fine-tuning, Pyligent demonstrated substantial improvements in solve rates. For instance, it boosted solve rates by 72.7 percentage points on hidden graphs, 17-18 points on Sudoku, and 13 points on Blocksworld. These results strongly suggest that explicit supervision of failed branches is highly effective in teaching LLMs crucial recovery behaviors beyond merely imitating perfect solution paths.