Non-autoregressive transformers (NATs) have demonstrated significant potential in reducing decoding latency for language generation tasks. However, “vanilla” NATs often struggle to effectively capture the sequential structure of generated text. To address this limitation, the recently proposed Directed Acyclic Transformer (DAT) introduces a graph structure into the decoder, explicitly modeling state transitions within the graph. Although DAT has achieved impressive performance, it typically requires large graph sizes to attain optimal results, which are highly memory-intensive, thereby limiting its applicability to tasks involving lengthy output sentences such as document-level machine translation. To mitigate this issue, we propose reducing DAT’s reliance on large graph sizes by coupling its state transition model with a more robust observation model—a Conditional Random Field (CRF). The CRF inherently models pairwise transitions between output tokens, enabling the model to capture dependencies without relying on large graphs. In addition, unlike other NAT-CRF models where the NAT and CRF modules operate independently, our approach is the first to jointly decode both modules, permitting joint optimality of the inferred graph and the output tokens. Experimental results on both sentence-level and document-level machine translation show that this modification substantially improves the baseline DAT in both lexical and semantic metrics, while retaining near-parity of decoding speed.