Navya-Nyaya Logic Fine-Tuning Yields 100% Semantic Correctness in Pramana LLM Models
Pramana integrates 2500-year-old Navya-Nyaya epistemology into LLM fine-tuning achieving full semantic correctness on logical tasks and highlighting non-Western methods for AI truthfulness missed by chain-of-thought and RLHF approaches.
Apple Machine Learning Research showed LLM performance on mathematical problems degraded by 65% when irrelevant context was added. Pramana fine-tunes Llama 3.2-3B and DeepSeek-R1-Distill-Llama-8B on 55 Navya-Nyaya structured problems covering constraint satisfaction Boolean SAT and multi-step deduction according to arXiv:2604.04937. The approach implements explicit 6-phase epistemological scaffolding: SAMSHAYA for doubt analysis PRAMANA for evidence source identification PANCHA AVAYAVA 5-member syllogism TARKA for counterfactual verification HETVABHASA for fallacy detection and NIRNAYA for final ascertainment distinguishing knowledge from hypothesis (arXiv:2604.04937). Stage 1 training produced 100% semantic correctness on held-out evaluation despite 40% strict format adherence. Wei et al. chain-of-thought prompting (arXiv:2201.11903) targets similar reasoning elicitation but omits formalized epistemology and fallacy detection present in Navya-Nyaya. Ablation experiments identified format prompting and temperature as critical variables with optimal configurations differing across reasoning stages (arXiv:2604.04937). Apple GSM-Symbolic benchmark (arXiv:2410.05229) exposed pattern-matching brittleness that Navya-Nyaya scaffolding directly targets through universal rule enforcement and counterfactual checks. All models datasets and training code released on Hugging Face to support epistemic framework research.
AXIOM: Navya-Nyaya's explicit fallacy detection and evidence tracing phases internalized by fine-tuned LLMs point to scalable non-Western scaffolding that could outperform current CoT and RL methods for handling epistemic uncertainty in high-stakes domains.
Sources (3)
- [1]Pramana: Fine-Tuning Large Language Models for Epistemic Reasoning through Navya-Nyaya(https://arxiv.org/abs/2604.04937)
- [2]Chain-of-Thought Prompting Elicits Reasoning in Large Language Models(https://arxiv.org/abs/2201.11903)
- [3]GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models(https://arxiv.org/abs/2410.05229)