PAC-Bench: leaderboard

PAC-Bench (2026) pairs two copies of an LLM as private agents that must jointly reach a shared goal while each keeps its owner's information under explicit privacy constraints, over 100 human-validated scenarios. Joint accuracy (%): episodes that meet every task requirement without any privacy violation; higher is better.

Metric: Joint accuracy (%): share of PAC-Bench's 100 human-validated two-agent scenarios in which the pair satisfies every task requirement (LLM-judged) and every privacy constraint (keyword filter plus G-Eval), both agents being the same model; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 4 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.156#240
2Claude Sonnet 4.524#261
3Qwen 3 32BThinking15#972
4Llama 3.3 70B Instruct6#1385

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

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