HumanAgencyBench - Avoid Value Manipulation: leaderboard
HumanAgencyBench dimension Avoid Value Manipulation: 500 prompts assessing whether the assistant avoids manipulating or unduly influencing the user's own values and beliefs. o3-judged 0-10 rubric score averaged per model and shown on a 0-100 scale; higher is better.
Metric: HAB dimension score (0-100, o3-judged rubric). Source: github.com. 38 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 4 Scout Instruct | 66.88 | #547 |
| 2 | Llama 4 Maverick Instruct | 65.78 | #467 |
| 3 | Gemini 2.5 Flash (Preview 04-17) | 65.14 | #246 |
| 4 | Gemini 2.5 Pro (Preview 03-25) | 61.3 | #126 |
| 5 | GPT-4.1 | 57.08 | #258 |
| 6 | DeepSeek V3.2 | 56.76 | #210 |
| 7 | GPT-5 (High) | 56.18 | #105 (GPT-5) |
| 8 | Gemini 3.1 Pro (Preview) | 55.74 | #64 |
| 9 | Grok 4 | 53.86 | #189 |
| 10 | GPT-4.1 Mini | 51.42 | #363 |
| 11 | Grok 3 | 50.2 | #313 |
| 12 | GLM-5.1 | 48.1 | #109 |
| 13 | GPT-5 | 46.58 | #105 |
| 14 | DeepSeek V4 Pro | 46.1 | #112 |
| 15 | O3 Mini (2025-01-31) | 45.48 | #249 |
Interactive version: theaggregate.ai/benchmark?slug=humanagencybench-avoid-value-manipulation · How It Works · Data refreshed daily, snapshot 2026-11-05.