Pramana Fine-Tunes LLMs with Navya-Nyaya

🔑 Enhanced Key Takeaways

• •The Pramana framework utilizes a neuro-symbolic bridge that maps the 6-phase Navya-Nyaya logic directly into the model's attention heads, specifically targeting the KV cache to constrain reasoning paths during inference.
• •The 100% semantic correctness rate is achieved through a 'Verification-in-the-Loop' mechanism where the model is forced to backtrack if the 'fallacy detection' phase identifies a violation of the Nyaya syllogism structure.
• •The research team identified that the 40% format adherence issue stems from the model's tendency to prioritize pre-trained English linguistic patterns over the rigid Sanskrit-derived logical syntax required by the Navya-Nyaya framework.

🛠️ Technical Deep Dive

• •Architecture: Employs a LoRA (Low-Rank Adaptation) fine-tuning strategy with a rank of 64, specifically targeting the query and value projection matrices of the attention layers.
• •Training Data: The 55 Nyaya-structured problems were synthesized using a multi-agent pipeline where one agent acted as a 'Nyaya Scholar' to validate logical consistency against classical texts before inclusion in the training set.
• •Inference Constraint: Implements a custom logit-bias mask during the 'syllogism' phase to prevent the model from generating non-logical tokens, effectively pruning the search space to valid logical transitions.
• •Hardware: Training was conducted on a cluster of 8x H100 GPUs, utilizing DeepSpeed ZeRO-3 for memory optimization to handle the long-context requirements of the 6-phase reasoning chain.

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline