DialToM - Retrospective (Emotional Support): leaderboard

Retrospective theory of mind on DialToM (2026), naturalistic human-human dialogues from emotional support conversations (ESConv): the model must identify the mental state (belief, desire, intention, emotion, knowledge or trust) of a speaker from the dialogue so far, choosing among four annotator-verified options. Correctness rate in percent; higher is better.

Metric: Correctness rate (%) on four-option multiple-choice questions of the DialToM Retrospective task: identify a speaker's mental state (belief, desire, intention, emotion, knowledge, trust) from the dialogue context, emotional support domain (naturalistic human-human dialogues from ESConv; GPT-4o-drafted options verified by professional annotators and Dawid-Skene aggregation), zero-shot; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Pro (Preview)99.15#119
2DialToM DeepSeek v3 (checkpoint unspecified)97.74
3Gemini 2.5 Pro97.18#276
4Qwen 3 235B A22B97.18#628
5Gemini 2.5 Flash95.48#463
6Kimi K295.48#443
7GPT-595.2#178
8GPT-4.194.07#505
9GPT-OSS-120B88.98#704
10Llama 4 Maverick87.57#1110
11Llama 3.3 70B84.75#1514
12Mistral Small 3.268.64#1641
13Mistral Nemo52.26#4484
14Llama 3.1 8B27.12#5196

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

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