EnactToM (Hard) - Literal ToM Probes: leaderboard

EnactToM (2026) literal Theory of Mind on the hard task split, generated from seeds that frontier models failed: after each embodied multi-agent episode, agents answer explicit questions about what another agent knows. Percent of probes answered correctly, averaged over three runs of 40 tasks; higher is better.

Metric: Literal ToM probe accuracy (%, Avg: mean per-run success over three runs of each task): at the end of each embodied multi-agent episode every agent is asked explicit questions about what another agent knows, derived from the task's epistemic goal operators, over the overall scope (22 cooperative and 18 mixed-motive tasks) of the matched EnactToM hard subset (seed-task failure ratio 0.9); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1EnactToM Gemini-Pro (checkpoint unspecified)63.3
2O352.5#242
3GPT-5.444.2#138
4Kimi K2.544.2#272
5EnactToM Gemini-Flash (checkpoint unspecified)42.5
6DeepSeek V3.236.7#425
7GPT-5.4 Mini31.7#409

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

Interactive version: theaggregate.ai/benchmark?slug=enacttom-hard-literal-tom-probes · How It Works · Data refreshed daily, snapshot 2026-11-22.