Community follow-up: EVAPO post-training for FireRedAudio

[duduke321]

• Thanks for open-sourcing FireRedAudio! We built EVAPO (https://github.com/duduke321/EVAPO) to explore a question in audio QA: how can a model learn from its mistakes when answer-based RL only tells it whether a response is correct?
• Additional audio evidence can offer richer guidance, but it may contain answer clues or mislead the teacher. It is also unavailable at inference time. EVAPO therefore first admits evidence only when, after answer clues are removed, it improves the ranking of the correct option. The teacher then evaluates the student’s own responses, adding a corrective signal only when a response has negative advantage and the teacher assigns it even lower likelihood. The student continues to answer directly from audio and the question, without privileged evidence.
• Our best results are 83.50% on MMAU test-mini and 84.46% on MMSU.