Frontier Models Fail to Decode Acoustic Camera

Qokedas · x · 2026-07-10

The author fed raw logs from a 1,024-microphone acoustic camera to four frontier models, asking them to output sound source heatmaps to test their ability to interpret real scientific instrument data. Follow-up context shows these models generally reconstructed the audio into incorrect formats, causing the heatmaps to mislabel the sound source. Fable 5 performed relatively best but remained insufficient for real-world measurements.

Related event: Frontier AI Models Fail to Parse Acoustic Camera Data(6 posts)→

Original post →