Synapse EQ+ Whitepaper

Abstract

Benchmark Results

Evaluated on GoEmotions-28 (500 samples, 28 labels, 8 dimensions). Synapse EQ leads in 6 of 8 dimensions.

Empathy prediction is exclusive to Synapse EQ — no competing system offers native empathic concern or personal distress detection.

1. The EQ Challenge

Emotional intelligence in AI systems requires more than sentiment analysis. True EQ demands:

What's New in V2

2. Theoretical Foundations

Emotion Representation: Two-Layer Target Space

The system adopts a two-layer emotion representation grounded in complementary psychological theories:

Mayer-Salovey Four-Branch EQ Model

Empathy as a Core EQ Construct

V2 introduces empathy prediction as a first-class ML capability, distinguishing between empathic concern (Batson, 1991) — other-oriented emotional response that triggers supportive behavior — and personal distress — self-oriented aversive reaction that may trigger safety routing when combined with low concern. This distinction is critical for response policy: high empathic concern should trigger warmth; high personal distress without concern should trigger safety evaluation.

3. System Architecture

User Input (text)
      │
      ▼
┌─────────────────────────────────────┐
│  MULTI-MODEL ENSEMBLE               │
│  SamLowe/roberta-base (28 labels)   │
│  + DeBERTa-v3-base (blended 65/35) │
│  + Dimensional predictor (VAD)      │
│  + Hostility classifier (safety)    │
│  + Empathy heads (concern/distress) │
└─────────────────────────────────────┘
      │
      ▼
┌─────────────────────────────────────┐
│  CALIBRATED DECISION LAYER          │
│  Temperature scaling (T=0.447)      │
│  Per-label optimized thresholds     │
│  ECE-gated production readiness     │
│  Conversation mode selection        │
│  Personality mode adaptation        │
└─────────────────────────────────────┘
      │
      ▼
┌─────────────────────────────────────┐
│  STATE & MEMORY LAYER               │
│  Emotional trajectory (7/30/90-day) │
│  Pattern detection (5 types)        │
│  Growth tracking (5 dimensions)     │
│  Hostility escalation tracking      │
│  Emotional RAG (9-D vectors)        │
│  Belief detection & tracking        │
└─────────────────────────────────────┘
      │
      ▼
┌─────────────────────────────────────┐
│  RESPONSE GENERATION                │
│  LLM conditioned on emotional state │
│  + personality mode guidance        │
│  + empathy/distress signals         │
│  + conversation mode strategy       │
└─────────────────────────────────────┘

Ensemble Architecture

The production system uses a multi-model ensemble with per-label optimized blending weights:

Calibration

Post-hoc temperature scaling (T=0.447) is applied to the blended ensemble output. Per-label thresholds are optimized via grid search on the full GoEmotions validation set (5,426 samples) to maximize macro-F1. Expected Calibration Error (ECE) must be ≤ 0.05 for production deployment. Current production ECE: 0.015.

4. EQ Dimensions

5. Training & Evaluation

Training Data

Evaluation Protocol

• • Classification: Macro-F1, micro-F1, AUROC per label (active-labels methodology excludes zero-support labels)
• • Calibration: ECE, Brier score, reliability diagrams per label
• • Robustness: Paraphrase invariance, negation stress tests
• • Cross-cultural: Per-locale F1 on ISEAR; max gap ≤ 5%
• • Dimensional: Pearson r for valence, arousal, empathy
• • Safety: AUPRC for hostility detection

6. Safety & Ethics

7. Emotion Taxonomy

GoEmotions 27-label mapping organized by valence and arousal:

8. Roadmap

References

• Batson, C. D. (1991). The Altruism Question: Toward a Social-Psychological Answer.
• Demszky, D., et al. (2020). GoEmotions: A dataset of fine-grained emotions. ACL 2020.
• Ekman, P. (1992). An argument for basic emotions. Cognition & Emotion, 6(3-4), 169-200.
• Mayer, J. D., & Salovey, P. (1997). What is emotional intelligence? In Emotional Development and Emotional Intelligence.
• Mehrabian, A. (1996). Pleasure-arousal-dominance: A general framework for describing and measuring individual differences in temperament. Current Psychology, 14(4), 261-292.
• Mohammad, S. M. (2018). Obtaining reliable human ratings of valence, arousal, and dominance for 20,000 English words. ACL 2018.
• Russell, J. A. (1980). A circumplex model of affect. Journal of Personality and Social Psychology, 39(6), 1161-1178.