List your company

Audio-dependent question answering on ADQA (dev)

65.65Strict AccuracySimPO
53.804456.879759.95563.0303Jun 20, 2026

Evaluation results

MethodLinks
SimPOmode=continuation, steps=5002026.0665.6566.96
Strong+10% cat-balanced mix2026.0665.21—
Empty5 repeatseed=seed3 best2026.0665.21—
SFT-init DAPO-litesteps=1002026.0665.2166.4
DPOmode=continuation, steps=5002026.0665.1566.46
SFT-init GRPOsteps=1002026.0665.0966.21
Empty-aware GRPOsteps=1002026.0665.0966.46
Empty-aware DAPO-litesteps=1002026.0664.7265.9
Empty5 repeatseed=seed2 best2026.0664.16—
Category-balanced2026.0664.09—
Answer-only Qwen SFTsteps=9k2026.0662.85—
Silent CoT SFTsteps=3k2026.0661.05—
Explicit CoT SFTsteps=3k2026.0658.49—
AudioFlamingo3mode=rerun2026.0654.32—
FunAudioChatvariant=base2026.0654.26—