MAYA RESEARCH SERIES · POST-SERIES LLM EXTENSION

Maya-LLM FAQ

Antahkarana in the Age of Transformers · Continual Fine-Tuning via Vedantic Mechanisms

Venkatesh Swaminathan · ORCID 0000-0002-3315-7907 · Nexus Learning Labs, Bengaluru

8.3%

BWT IMPROVEMENT

Condition F calibrated vs Condition A baseline

0.988

BUDDHI S-CURVE

Cross-substrate invariant — SNN P4–P9 and LLM identical

3 domains

BHAYA FIRED ON

Py150, ScienceQA, NumGLUE-ds — genuine loss spikes

Scale

CALIBRATION NEEDED

SNN thresholds don't transfer directly to LLM loss scale

⚠ CALIBRATION NOTE — KEY FINDING

Condition F with SNN-calibrated hyperparameters (Bhaya threshold=1.8×) performs worse than baseline (BWT=1.42 vs 1.05) because Bhaya fires on 18.8% of all steps at LLM loss scale — treating normal variance as catastrophic events. Recalibrating threshold to 4.0× (matched to LLM mean loss ~1.08) achieves the best result (BWT=1.11, 8.3% improvement). This is not a failure — it is a quantified cross-substrate scaling finding with a clear path forward.

MAYA RESEARCH SERIES · P1–P9 + POST-SERIES

Nociceptive

66.6% ↑ velocity

DOI

Maya-OS

OS arbitration

DOI

Maya-CL

AA=62.38%

DOI

Maya-Smriti

AA=31.84%