Pramana: Fine-Tuning Large Language Models for Epistemic Reasoning through Navya-Nyaya

A new research paper titled "Pramana: Fine-Tuning Large Language Models for Epistemic Reasoning through Navya-Nyaya" proposes a groundbreaking method to address the core weakness of modern AI: unreliable reasoning. Authored by Sharath Sathish, the work tackles the 'epistemic gap' where models like GPT-4 and Claude produce fluent but unfounded claims, a problem highlighted by Apple research showing a 65% performance drop when irrelevant context is added. Instead of relying on brittle pattern-matching or generic chain-of-thought prompting, Pramana teaches LLMs a formal epistemology by fine-tuning them on the structured logic of Navya-Nyaya.

The Navya-Nyaya framework breaks reasoning into six distinct phases: SAMSHAYA (doubt), PRAMANA (evidence), PANCHA AVAYAVA (syllogism), TARKA (counterfactual verification), HETVABHASA (fallacy detection), and NIRNAYA (final ascertainment). This provides cognitive scaffolding absent from standard training. The researcher fine-tuned the open-source models Llama 3.2-3B and DeepSeek-R1-Distill-Llama-8B on 55 problems structured with this logic, covering constraint satisfaction and multi-step deduction. Remarkably, the fine-tuned models achieved 100% semantic correctness on held-out evaluations, even when strict format adherence was only 40%, proving the models internalized the reasoning process itself.

The research includes detailed ablation studies showing how format prompting and temperature settings critically affect performance. All models, datasets, and training code have been released on Hugging Face, enabling the community to build on this work. This integration of ancient formal logic with modern AI represents a significant shift from training on sheer data volume to instilling rigorous, traceable reasoning methodologies, potentially unlocking more trustworthy AI for critical applications in law, science, and medicine.