What Does LALM Post-training Learn? Rethinking Audio Understanding and Reasoning Capabilities
Post-training often improves the performance of large audio-language models on audio understanding and reasoning benchmarks.
ProofPaper ↗
Key points
- However, these improvements may reflect adaptation to specific question formats, such as multiple-choice questions, rather than genuine advances in audio understanding and reasoning capabilities.
- To investigate what LALMs actually learn from post-training, we conduct controlled experiments that disentangle improvements in audio understanding from adaptation to specific question formats.
- Across different models and post-training methods, we find that much of the reported improvement persists even without informative audio, but largely disappears when the question format changes.
- These findings suggest that post-training often teaches models to exploit task-specific textual cues and answer patterns rather than developing generalizable audio understanding and reasoning capabilities.
Sources (1)
- [1]What Does LALM Post-training Learn? Rethinking Audio Understanding and Reasoning CapabilitiesarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 02:54 AM
Post-training often improves the performance of large audio-language models on audio understanding and reasoning benchmarks.
However, these improvements may reflect adaptation to specific question formats, such as multiple-choice questions, rather than genuine advances in audio understanding and reasoning capabilities.
Extractive summary: sentences quoted from the sources.
