MuLA-Bench - Semantic Evidence: leaderboard

The semantic-evidence track of MuLA-Bench (Yang et al., 2026), a benchmark of open-ended questions over long in-the-wild recordings, answered from the audio alone: 3,840 questions about spoken facts and relations, 30 for every language and domain. Each answer is judged correct or wrong by GPT-5.6 Sol against the reference with rubric and timing checks; accuracy in percent, higher is better.

Metric: Accuracy (%; the 3,840 semantic-track questions about spoken facts and relations, 30 per language-domain cell; open-ended questions over in-the-wild long-form recordings (median 36.8 minutes, capped at three hours) in the source language, answered from the full recording without transcripts, audio clipped only to a local model's native input limit; each answer judged correct or not by GPT-5.6 Sol on a route-specific binary rubric with programmatic structural and 3-second timing checks; missing or invalid answers scored wrong). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.

Top models

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

Interactive version: theaggregate.ai/benchmark?slug=mula-bench-semantic-evidence · How It Works · Data refreshed daily, snapshot 2026-11-12.