CACD decoding curbs drift in diffusion multimodal LLMs

A new arxiv preprint from Yikai Zhao and colleagues, accepted to EMNLP 2026, targets a specific failure mode in diffusion multimodal large language models: as outputs get longer, they drift and start repeating themselves. Two experts in our Who's Who directory have shared the preprint.

The authors trace the problem to two habits of existing decoders. Confidence scoring, they argue, looks at each masked position in isolation and ignores whether neighboring tokens actually support the choice. And block partitioning, the standard scheme for committing tokens in chunks, walls off what the paper describes as "high-readiness semantic anchors" elsewhere in the sequence.

Their fix, Context-Aware Cluster Decoding, drops the block constraint and rescores masked positions using "a multiplicative composite of softmax confidence and neighbor proximity." The method is training-free and includes what the authors call "architecture-aware calibration for confidence heterogeneity across visual integration strategies" — a per-model correction for the fact that different dMLLMs fuse image and text signals differently.

The reported experiments span three dMLLM architectures and four benchmarks, with the authors claiming "consistent quality improvements and hallucination reduction" and larger gains on longer generations. The paper is 19 pages with 12 figures and 13 tables, and the code is on GitHub.

The summary published on arxiv does not include per-benchmark accuracy numbers or a latency comparison, so the size of the gain — and what it costs at inference — is not visible from the abstract alone.