SEAR
Spoofing Evidence-Grounded Audio Reasoning Benchmark for Audio Language Models
A four-task audio question-answering benchmark evaluating audio language models on deepfake detection using acoustic evidence identification and quantification.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info sear
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
6 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Gemini-Flash | 48.6 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| Qwen2-Audio | 39.1 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| Qwen2.5-Omni | 25.2 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| MOSS-Audio | 23.5 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| MiniCPM-o | 13.8 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| GPT-Audio | 13.2 | N/A | Independent | medium | Sep 30, 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
Content & Construction: SEAR is a benchmark constructed from ASVspoof datasets, utilizing 35 statistical acoustic features to generate reference evidence. It includes tasks for binary classification (verdict), feature identification/quantification, and open-ended forensic rationale generation.
Scoring & Findings: The benchmark assesses the gap between plausible rationales and verifiable evidence. Key findings show that tool-augmented agents improve detection performance when grounded in reliable acoustic anomalies.