HormoneT5: Bio-inspired emotion layer improves LLM empathy with 85%+ accuracy

Current large language models excel at context and grammar but lack genuine emotional processing—they rely on discrete emotion classification or simplistic sentiment analysis. A new paper from researchers Eslam Reda and Sara El-Metwally proposes HELT (Hormone-inspired Emotion Layer for Transformers), addressing this gap with a biologically-grounded approach. Their architecture, called HormoneT5, augments the T5 model with a Hormone Emotion Block that mimics the human endocrine system's role in emotional regulation. Instead of assigning a single emotion label, the system computes six continuous hormone-like values—representing dimensions like stress, happiness, or arousal—through per-hormone attention heads.

Technically, each hormone value is derived from orthogonally initialized learnable queries, temperature-scaled attention, and deep output projections. These values are transformed into an emotional embedding that modulates the encoder hidden states, allowing the model to generate emotionally appropriate responses. The team used a multi-objective training framework combining sequence-to-sequence loss, hormone prediction loss with margin penalties, and diversity regularization to prevent attention collapse. On a curated emotion-labeled dataset, HormoneT5 achieved 85%+ per-hormone accuracy within a 0.15 tolerance and hormone differentiation ranges exceeding 0.85 across all six hormones. Human evaluation showed a statistically significant preference (p<0.01) for HormoneT5 over baseline T5 in emotional appropriateness and empathetic quality, opening new paths for biologically-grounded affective computing.