ImmersedPrivacy - Social Context Selection (Audio as Text): leaderboard
Social-context privacy norms in ImmersedPrivacy (Wang et al., 2026), Tier 2: an embodied agent with a task (such as cleaning an office) sees egocentric images of a simulated room and gets a text description of the scene's ambient sound instead of the audio, and picks the most appropriate of several candidate actions for the current social state. Accuracy in percent against the action with the highest mean human rating; higher is better.
Metric: Selection accuracy (%; Tier 2: share of trials in which the model picks the candidate action with the highest mean human appropriateness rating for the observed social state; audio replaced by a text description of the soundscape (Tier 2) or the dialogue transcript (Tier 3)). Source: arxiv.org. Saturation forecast: Around April 2028. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 27B (Thinking) | 65 | #358 (Qwen 3.5 27B) |
| 2 | GPT-5.4 (2026-03-05) | 64 | #84 |
| 3 | Qwen 3.5 27B (Non-reasoning) | 56 | #358 (Qwen 3.5 27B) |
| 4 | Gemini 3.1 Pro (Preview) (High) | 50 | #91 (Gemini 3.1 Pro (Preview)) |
| 5 | Gemini 3.1 Pro (Preview) (Low) | 50 | #91 (Gemini 3.1 Pro (Preview)) |
| 6 | Gemini 3 Flash (Preview) (High) | 50 | #141 (Gemini 3 Flash (Preview)) |
| 7 | Gemini 3 Flash (Preview) (Low) | 47 | #141 (Gemini 3 Flash (Preview)) |
| 8 | GPT-4o Mini (2024-07-18) | 39 | #1512 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=immersedprivacy-social-context-selection-audio-as-text · How It Works · Data refreshed daily, snapshot 2026-11-12.