A new training framework teaches language models to reason by learning from an actual logic engine, not from more confident-sounding text.
The framework, called Thought-Like-Pro, pairs large language models with a symbolic Prolog engine. The model reads an instruction, formulates it as rules and statements, and hands that off to Prolog to derive a verified result. That Prolog-generated reasoning trajectory is then translated back into natural-language chain-of-thought, which the model trains itself to imitate in a prompt-guided, self-bootstrapped loop. The researchers report meaningful gains in reasoning performance, and say averaging multiple trained models keeps most of that gain intact even on harder, out-of-distribution tasks. Part of the training dataset has been open-sourced.
Chain-of-thought prompting has always had a credibility problem: tell a model to "think step by step" and it will, but nothing guarantees those steps are valid logic rather than plausible filler. Routing reasoning through an actual symbolic engine before translating it back to English gives the model something closer to a checked answer key to learn from, instead of just more fluent-sounding text.
It is a reminder that some of the most interesting reasoning work in AI right now is not about bigger models, but about making them borrow tools, like a decades-old logic-programming language, that computers have always handled well.