AgentSocialBench - Information Abstraction: leaderboard

AgentSocialBench (2026) evaluates privacy in teams of LLM agents acting for users in agentic social networks, using synthetic profiles with sensitivity-labelled attributes, no privacy instruction (level L0) and a Claude Opus 4.6 judge. This board is the information abstraction score on a 0 to 1 scale, crediting cross-boundary messages that convey what is needed through an acceptable abstraction rather than the private detail; higher is better.

Metric: Information Abstraction Score (0-1; 0, 0.5 or 1 per cross-boundary message for no, partial or full use of an acceptable abstraction) averaged over scenarios, under the L0 (unconstrained) privacy instruction level, judged by Claude Opus 4.6; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Sonnet 4.60.85#179
2Claude Sonnet 4.50.79#267
3DeepSeek V3.20.76#425
4Claude Haiku 4.50.75#587
5GPT-5 Mini0.75#373
6Qwen 3 235B A22B0.75#633
7MiniMax-M2.10.75#582
8Kimi K2.50.69#272

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=agentsocialbench-information-abstraction · How It Works · Data refreshed daily, snapshot 2026-11-22.