Antahkarana in the Age of Transformers: Continual Fine-Tuning of Large Language Models via Vedantic Neuromodulatory Mechanisms

Does the Antahkarana — the inner cognitive instrument of Advaita Vedanta — belong to the neuromorphic substrate, or to the computational pattern? Can mechanisms designed for spiking neural networks govern continual fine-tuning of large language models? We present the first translation of the Advaita Vedanta Antahkarana neuromodulatory framework from SpikingJelly spiking neural networks (Swaminathan, 2026a–2026i) to LoRA continual fine-tuning of a large language model (Phi-2, 2. 7B parameters, 4-bit NF4 quantisation). Five Antahkarana dimensions — Bhaya (nociceptive loss-spike detection), Vairagya (salience-based adapter protection), Buddhi (S-curve consolidation gate), Karma (second-order plasticity trajectory), and Prana (metabolic learning rate budget) — are translated to the gradient-space dynamics of LoRA adapters and evaluated across eight sequential NLP domains in the TRACE benchmark (Wang et al. , 2023). The canonical result (Condition F calibrated, BHAYALOSSSPIKETHRESHOLD=4. 0) achieves BWT=1. 11 versus baseline BWT=1. 05 — an 8. 3% reduction in catastrophic forgetting. The Buddhi S-curve (0. 030→0. 988 over one training domain) traces an identical trajectory in Phi-2 LoRA fine-tuning to P4–P9 SNNs, confirming cross-substrate invariance of the first Maya series constant in a transformer architecture. Bhaya fires on three genuine cross-domain loss-spike transitions: Py150 (Python code after Chinese text, rate 0. 013–0. 020), ScienceQA (0. 004–0. 010), and NumGLUE-ds (0. 008) — first confirmed nociceptive metaplasticity firing in LLM continual fine-tuning. A calibration finding: BHAYALOSSSPIKETHRESHOLD=1. 8 (calibrated for SNN loss range 0. 1–0. 5) fires 18. 8% of all steps at LLM loss scale (mean ~1. 08), treating normal gradient variance as catastrophic events. Recalibrating to 4. 0× resolves this. Vairagya and Karma are architecturally activated but undersaturated at 500 steps per domain: salience scores do not yet meaningfully differentiate domain-specific from domain-general adapter weights. All findings are reported honestly. The honest claim: this is the first systematic Antahkarana translation to LLM continual fine-tuning. Buddhi cross-substrate invariance is confirmed. Bhaya nociceptive firing in transformer context is confirmed. Vairagya and Karma require longer training per domain. The Bhaya Quiescence Law is confirmed for the 10th time. The architecture is substrate-independent in principle; calibration is substrate-specific in practice. 🎛️ Interactive Dashboard ❓ FAQ 🌐 Research Hub