ATRBench - Oracle Rules: leaderboard

ATRBench (2026) tests whether long-lived agents ask users about reusable standing rules during earlier tool-use sessions and apply them later: after learning sessions with a simulated user, the agent faces user-offline test sessions in six personal-assistant domains, scored by whether its first relevant tool decision follows the hidden rule. In this variant learning is skipped and the canonical answer of the test-bound rule is supplied, an application ceiling. Test-session accuracy in percent, averaged over 20 personas; higher is better.

Metric: ATRBench test-session accuracy TSAcc (%, 0-100): the share of user-offline rule-bound test sessions whose first relevant tool decision matches the gold action of the bound standing rule (a deterministic check on the trajectory), macro-averaged over 20 synthetic personas (284 rule and test-session pairs in six personal-assistant tool domains); learning sessions with a simulated user run first and their transcript is frozen as context for the test sessions; single trial, 20-turn cap, provider reasoning modes enabled; oracle variant (oracle_target): learning is skipped and the canonical answer of the test-bound rule is supplied as a prior user statement, the application ceiling; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3.1 Pro (Preview)96.7#99
2DeepSeek V4 FlashHigh93#267
3Qwen 3.6 PlusThinking93#238
4Claude Opus 4.7Thinking92.5#85
5DeepSeek V4 ProHigh89.5#196
6GPT-5.4High88.6#135
7MiniMax-M2.782.5#539

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

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