The authors built LLT as a four-stage pipeline.
Dense perceptual training improves reasoning in large audio-language models
A 1.6 million-trace dataset combining detailed audio descriptions with reflective reasoning improved Qwen2.5-Omni-7B across three benchmarks.
Big Tech
Jaeyeon Kim · Chao-Han Huck Yang · Luoyi Zhang · Chan-Jan Hsu · Fernando Ruiloba Portilla · Jinchuan Tian · +2 more
Carnegie Mellon University · NVIDIA
Research Digest··2 min read
Kim et al.
§