MuSeToM - Task 2 - F1: leaderboard

MuSeToM - Task 2 - F1 — MuSeToM v1 Table1 direct prompting on human-validated textual belief revision; clean tasks1binary/2multiplechoice and noisy tasks3binary/4multiplechoice remain separate. Accuracy and Macro-F1 are distinct percentage metrics. AppendixB.1 binds the Llama/Mistral/Mixtral instruction checkpoints; Tables23and28–31bind the sole Phi row to Phi-3-Medium-4K-Instruct in the same18-model cohort. temperature0,top_p1,1024output tokens for the listed first group and10240for later reasoning models; Phi output budget is unreported because B.1omits it. Effort/API dated snapshots are unreported. Eighteen model rows; Human reference is a separate reference, not Median Human/Expert and not an AI model. Main table shorthand is resolved from this paper only; no global Base/Instruct or generation alias. Metric: Macro-F1 (%).

Metric: Macro-F1 (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 18 models tracked. Updated 15 Nov 2026.

Top models

#ModelScoreOverall rank
1Kimi K2 (Thinking)75.39
2Qwen 3.5 397B A17B74.57#251
3MiniMax-M2.574.1#584
4MiniMax-M2.173#574
5DeepSeek V3.270.68#412
6DeepSeek R169.97#478
7DeepSeek V3.165.12#502
8QwQ-32B63.64#915
9GPT-OSS-120B61.08#703
10Qwen 3 235B A22B57.12#628
11Llama 4 Maverick Instruct55.95#1163
12Llama 3.1 405B Instruct54.63#1122
13GPT-4o49.82#744
14Phi-3-medium-4k-instruct49.77#2696
15Mixtral 8x22B Instruct43.5#2097

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=musetom-task-2-f1 · How It Works · Data refreshed daily, snapshot 2026-11-16.