BEANS-Next
BEANS-Next: Broadening Audio-Language Capabilities for Bioacoustics
A benchmark measuring broad bioacoustic audio-language abilities across acoustic perception, biological category recognition, scene understanding, and in-context learning.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info beans-next
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
3 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| AF-Next | 63.2 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen3-Omni | 59.8 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| NatureLM-audio v1.0 | 56.2 | N/A | Independent | medium | Oct 7, 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
BEANS-Next evaluates models on a taxonomy of bioacoustics tasks including acoustic measurement, call counting, temporal reasoning, and adaptation to new taxa. It builds on BEANS-Zero and BirdSet to assess capabilities beyond simple classification.
Scoring reveals existing bioacoustic models perform poorly on broader tasks, while general-purpose audio-language models struggle across most categories. The benchmark highlights the need for specialized training resources like ROOTS to improve performance.