AetherMind_SRL: Self-Reflective Knowledge-Distilled Transformer for Robust Natural Language Inference

AetherMind_SRL is a Self-Reflective, Knowledge-Distilled Transformer designed for robust Natural Language Inference (NLI). It integrates adversarial training on ANLI, general-domain NLI datasets (SNLI, MNLI, XNLI), and clinical reasoning using Alzheimer’s disease cognitive data (ADNI).This work combines Knowledge Distillation and a multi-stage Self-Reflective Learning (SRL) pipeline, where the model iteratively learns from its own errors through a global error buffer and a structured SMART retraining loop.The Round 12 checkpoint demonstrates strong performance across SNLI/MNLI, adversarial ANLI R1/R2/R3, and achieves perfect accuracy on ADNI cognitive claim verification. It represents a key milestone in the AetherMind project, a long-term initiative to develop adaptive reasoning systems capable of self-improvement.This Zenodo release includes the full paper, methodology details, smart error buffer construction, ADNI SRL loop, evaluation results, scripts overview, and citations.