Title: Pramana: Enhancing Large Language Models for Epistemic Reasoning via Navya-Nyaya

Abstract

While large language models excel at generating fluent text, they frequently falter in systematic reasoning, often producing confident yet baseless assertions. Research from Apple Machine Learning demonstrated this fragility: when irrelevant context was introduced to mathematical problems, model performance plummeted by 65%, revealing that the apparent reasoning capabilities are underpinned by brittle pattern-matching. This "epistemic gap"—the failure to anchor claims in verifiable evidence—hinders AI reliability in fields that demand rigorous justification.

To address this, we present Pramana, a novel methodology that instills explicit epistemological frameworks into LLMs by fine-tuning them on Navya-Nyaya logic, an Indian reasoning tradition spanning 2,500 years. Distinct from standard chain-of-thought prompting, Navya-Nyaya mandates a structured, six-phase reasoning process: SAMSHAYA (analysis of doubt), PRAMANA (identification of evidence sources), PANCH A AVAYAVA (a five-part syllogism incorporating universal rules), TARKA (counterfactual verification), HETVABHASA (detection of fallacies), and NIRNAYA (ascertainment that distinguishes established knowledge from hypothesis). This synthesis of logic and epistemology offers cognitive scaffolding that is typically missing from conventional reasoning techniques.

We evaluated this approach by fine-tuning Llama 3.2-3B and DeepSeek-R1-Distill-Llama-8B on a dataset of 55 logically structured Nyaya problems, which included constraint satisfaction, Boolean SAT, and multi-step deduction tasks. Our first stage of evaluation yielded 100% semantic correctness on held-out test sets, even though strict format adherence was only 40%. This discrepancy suggests that models successfully internalize the underlying reasoning content despite imperfect structural enforcement. Furthermore, ablation studies indicate that format prompting and temperature settings significantly impact performance, with optimal configurations varying across different reasoning stages. To facilitate further investigation into epistemic frameworks for AI, we have made all models, datasets, and training infrastructure publicly available on Hugging Face.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC