Noisy
← this week  ·  Subscribe →

The world is being quietly rearranged by people who write very long documents.

arXiv · April 8, 2026
The title they went with Pramana: Fine-Tuning Large Language Models for Epistemic Reasoning through Navya-Nyaya Noisy translates that to

Researchers teach AI to reason step-by-step using 2,500-year-old logic framework

What happened
Researchers fine-tuned smaller AI models on ancient Indian logical reasoning methods to make them more methodical. The models now work through problems in structured stages — identifying evidence, testing counterexamples, catching fallacies — instead of just pattern-matching their way to confident-sounding answers.
Why it matters
AI systems routinely hallucinate. When researchers added irrelevant details to a math problem, models failed 65% of the time, suggesting they don't actually reason but rather recognize surface patterns. This paper shows that training on explicit logical structure — a methodology developed centuries before anyone wrote code — can push models toward something resembling actual reasoning rather than fluent nonsense. The catch: 100% correctness on the logic didn't require models to follow the exact format perfectly, meaning they may be internalizing reasoning without being forced into the rigid structure. The real question is whether this holds up on anything harder than toy problems.
The signal
Whether models fine-tuned on Navya-Nyaya logic maintain higher accuracy and fewer hallucinations when deployed on real-world tasks outside the training domain, or whether the improvement was specific to the contrived problem set.
If you insist
Read the original →