EchoDistill:Alignment Noisy-to-Clean Self-Distillation for Robust Audio LLMs

In three linesEchoDistill introduces an alignment-based noisy-to-clean self-distillation framework to improve Audio LLM robustness against real-world noise. A noisy student is optimized via GRPO using a frozen clean-audio teacher as semantic reference. Results: +4.18% GSR improvement under strong noise vs strongest baseline, +3.02% Acc on Qwen-Omni.

Summary generated by Claude — human-verified