HLR-AVSceneQA
Hear, Localize, and Reason: Spatially Aware Scene Understanding for Audio-Visual LLMs
A comprehensive benchmark evaluating spatial audio-visual scene understanding through a hear-localize-reason hierarchy, assessing sound recognition, source localization, and relational reasoning.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info hlr-avsceneqa
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
9 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Dual-channel Qwen2.5-Omni | 84.5 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Hear You Are LLM | 67.7 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Qwen2.5-Omni | 57.2 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Qwen3-Omni | 55.8 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| VideoLLaMA2 | 54.5 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Gemini 2.5 Flash Lite | 38.1 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| MiniCPM-o-v2.6 | 35.3 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| GPT-Realtime-Mini | 34.2 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Video-SALMONN 2 | 32.2 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
HLR-AVSceneQA evaluates models on recognizing sounds, localizing sources in egocentric/allocentric views, and reasoning about their relation to the visual scene. It uses simulator-based pipelines to generate diverse spatial audio-visual scenes with reliable annotations.
The benchmark employs a hear-localize-reason hierarchy to systematically assess spatial grounding and relational reasoning. Experiments show existing AV-LLMs struggle with spatial tasks, while the proposed HLR-LLM improves performance via chain-of-thought supervision.