The authors conducted controlled experiments on LALMs (including Qwen2-Audio and SALMONN) using different post-training methods (SFT, RFT, GRPO).
Post-training gains in audio models may reflect format adaptation, not true audio understanding.
Controlled experiments show that improvements persist with silence and disappear when multiple-choice options are removed.
Big Tech
Hua-Hsuan Chen · Xuanjun Chen · Wei-Chung Lu · Chih-Kai Yang · Yu-Xiang Lin · Sung-Feng Huang · +2 more
National Taiwan University · NVIDIA · Independent Researcher · NTU AI-Core
Research Digest··2 min read
The authors investigated whether post-training of large audio-language models (LALMs) actually improves audio understanding or merely adapts to question formats.
Why this paper
From NVIDIA and 3 others
In one line
Post-training gains in audio-language models come from exploiting text cues and question formats, not genuine audio understanding.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks (2 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§