This repository is for the realization of reconstruction-alignment on MMAda Motivation:
In the Reconstruction-Alignment(ICLR 2026) paper(https://github.com/HorizonWind2004/reconstruction-alignment), the authors compare unified multimodal large models with various architectures, including autoregressive architectures and continuous diffusion models. However, experiments on unified multimodal large models based on discrete diffusion architectures are lacking. Inspired by this, we aim to verify whether reconstruction alignment is applicable to this architecture, and therefore conduct this transfer.
The fitune weights:
The results on GenEval BenchMark.
