CoTA++: Solving the Repeat Curse in Diffusion-based Multimodal LLMs
In brief
- dMLLMs suffer from token repetition caused by approximate caching during decoding.
- CoTA++ reorganizes attention and adjusts predictions to minimize repetitions.
- The framework improves cache reliability without additional training or performance loss.
Diffusion-based multimodal large language models (dMLLMs) often repeat tokens during decoding due to approximate caching. CoTA++, a training-free framework, enhances cache reliability by re-anchoring attention, refreshing states adaptively, and adjusting predictions based on contextual entropy.
This summary gives you the essentials. The full article, its figures and quotes are on cs.AI updates on arXiv.org.
Read the original articleAI-generated summary of the source, checked by rules. Report an error