DeEscalWild: leaderboard

DeEscalWild (2026) asks a language model to role-play the civilian in 150 real police-civilian interactions transcribed from public videos, replying zero-shot to each real officer utterance. A Gemini 3.1 Pro judge scores realism (behavioural plausibility, naturalness and persona adherence) from 0 to 100; higher is better.

Metric: Realism score (0-100) from a Gemini 3.1 Pro judge rating behavioural plausibility, linguistic naturalness and persona adherence of the generated civilian turns, on DeEscalWild's 150 held-out police-civilian interactions transcribed from real social-media footage: at each turn the model, prompted zero-shot with the situation and a civilian persona, receives the gold officer utterance and generates the civilian reply (autoregressive simulation); mean of three runs; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Llama 3.2 3B Instruct36.1#7241
2Qwen 2.5 3B Instruct35.4#5392
3granite-3.0-2B-instruct32.4#8277
4Falcon3-3B-Instruct30.2#8313
5Gemma 2 2B (IT)28.5#7054

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=deescalwild · How It Works · Data refreshed daily, snapshot 2026-11-21.