Adversarial Empathy Benchmark (Think Mode): leaderboard
Empathetic dialogue against adversarial simulated users (Adversarial Empathy Benchmark, Deeraj S K, Devarajan, Mehra and Mishra, 2026): a SAGE-style user played by Mistral-7B-Instruct-v0.3 follows one of six adversarial trajectories (escalation, mood reversal, gaslighting, fact-emotion contradiction, emotional flooding, validation manipulation) and rewards only replies that meet its hidden need. Final score from 0 to 1 (the user's emotion after eight turns divided by 100, higher is better), averaged over 60 scenario-matched dialogues. Models answer with a Think-Then-Say reasoning scaffold.
Metric: Final score (0-1; the simulated user's emotion after eight turns divided by 100, mean over 60 adversarial dialogues in six trajectory types; Mistral-7B-Instruct-v0.3 SAGE simulator; Think-Then-Say reasoning scaffold). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 7B Instruct | 0.76 | #2583 |
| 2 | Qwen 2.5 1.5B Instruct | 0.71 | #7518 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=adversarial-empathy-benchmark-think-mode · How It Works · Data refreshed daily, snapshot 2026-11-12.