AudioProcessBench - Cross-Modal Binding Errors: leaderboard
AudioProcessBench (2026) tests audio and omni-modal models as critics of audio-grounded reasoning: given an audio question and a model-written reasoning chain split into steps, the critic marks each step correct or wrong; 3,872 annotated chains from MMAR, MMSU and MMAU-Pro questions. This board is the PRMScore on a 0 to 100 scale for detecting steps with cross-modal binding errors, scored against correct steps only; higher is better.
Metric: Type-conditioned PRMScore (0-100): detection of steps with cross-modal binding errors (tying an audio observation to the wrong answer option or entity) while keeping correct steps, other error types left out; 3,872 reasoning traces (23,497 steps, 41% erroneous) from six audio and omni models on MMAR, MMSU and MMAU-Pro questions, steps labelled by two LLM annotators with human review; few-shot critic prompt with three annotated example chains, temperature 0.7; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 78.3 | #164 |
| 2 | Gemma 4 E4B | 73.5 | #2214 |
| 3 | Gemma 3n E4B | 72.4 | #3301 |
| 4 | Gemma 3n E2B | 68.1 | #6091 |
| 5 | Gemma 4 E2B | 64.9 | #3892 |
| 6 | Phi-4 Multimodal Instruct | 61.9 | #3882 |
| 7 | Qwen2.5-Omni-7B | 60.6 | #1789 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=audioprocessbench-cross-modal-binding-errors · How It Works · Data refreshed daily, snapshot 2026-11-13.