LogicTrack audits LLM reasoning with formal solvers

Correct answers reached through flawed reasoning are the most dangerous failure mode in LLM applications, making formal intermediate verification essential.

• –Auto-formalizing reasoning steps into symbolic representations lets automated theorem provers catch hallucinated logic before final conclusions are reached.
• –Solver-Based Backtracking Reward (SBR) provides actionable, step-wise supervisory signals during tree search instead of crude end-to-end outcome scoring.
• –Incorporating solver-verified backtracking traces into SFT allows models to internalize step-by-step logical self-correction.
• –Across 8 benchmarks and 7 models, the framework proves neuro-symbolic auditing simultaneously boosts reasoning soundness and task pass rates.