Autoregressive vision-language models unify heterogeneous perception tasks, but they are highly susceptible to compounding errors as predictions are generated step by step. A new arXiv paper introduces CA-OPD, a confidence-aware on-policy distillation method aimed at this problem for structured visual prediction.
On-policy distillation (OPD) addresses the mismatch between training and inference by training a student model on its own outputs rather than on teacher-generated sequences. The proposed approach extends this idea with a confidence-aware mechanism, though the abstract does not detail how confidence is computed or applied. The paper is described as a replacement announcement, and no experimental results are included in the source.
Because only this single source is available, the summary reflects the paper's own framing. No independent comparisons or evaluations are reported here, so the effectiveness of CA-OPD remains to be assessed from the full paper.