Project Aletheia: Verifier-Guided Distillation of Backtracking for Small Language Models

Aradhya Dixit, Tianxi Liang, Jai Telang

Keywords: language model reasoning, verifier-guided training, backtracking, symbolic reasoning, small language models, constraint satisfaction, SAT reasoning

Abstract:

Small Language Models (SLMs, under 10B parameters) are
attractive for private, on-device deployment, yet they fre-
quently fail on strict constraint-satisfaction problems due to
linear, overconfident reasoning traces that do not recover
from early mistakes. We introduce Verifier-Guided Distilla-
tion, a training protocol that transfers the process of error
repair—explicit conflict detection and backtracking—rather
than only correct final answers. By training a 7B model
on verified reasoning traces that include mistakes and self-
corrections, we show that latent verification behavior can
emerge in small models, enabling them to occasionally stop,
detect contradictions, and revise earlier assumptions.

Submission Number: 107