mlsanityData to 7 Oct 2026

Faster but Different: Diagnosing and Controlling Content Drift in Accelerated Multimodal Diffusion Language Models

arXiv 2607.29079 ↗First posted 31 Jul 2026cs.CLcited by 0

On the map: Multimodal LLMs / vision-language models → Image editing and quality assessment

also in 1 more
Abstract (arXiv)

Training-free acceleration makes diffusion-based multimodal large language models (dMLLMs) more deployable, but it may silently change generated content. We study this serving-time consistency problem on 300 real images, comparing Fast-dLLM outputs with the same model's unaccelerated outputs. Across the mild parallelism induced in our long-form setting (1.05--1.25 committed tokens per step), confidence-threshold tuning changes decoding behavior but not baseline agreement. State-refresh ablations and an image-swap intervention instead identify stale visual and generated-text states as contributors to drift. For the tested Fast-dLLM implementation, shortening the KV-cache refresh interval yields a monotonic speed--agreement frontier and near-exact agreement at a measured 1.3x speedup. The initial diagnosis also appears with dLLM-Cache and LaViDa, although dLLM-Cache recovers agreement only after both caches are tightened, which removes its speed advantage. Independent prompts and images reproduce the threshold-insensitivity and refresh recovery. A targeted audit finds genuine content substitution in half of 50 low-agreement pairs. In a separate blinded two-annotator evaluation, the pooled accelerated-minus-baseline factual-error difference is 0.00 (95% CI [-0.17,+0.17]); this sample detects no difference but does not establish factual equivalence. Finally, none of the tested adaptive or smoothed-refresh variants beats the fixed interval at matched compute. Our contribution is a paired diagnostic and an implementation-scoped consistency control, not an accuracy or safety guarantee.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Image editing and quality assessment, one of 16 directions in Multimodal LLMs / vision-language models. Builds on 7 of its 12 references that are in the map. First posted on arXiv 31 Jul 2026.

Builds on

7 of 12 references.

References · 7

  1. 2023MME: A Comprehensive Evaluation Benchmark for Multimodal Large Language Modelscited by 1,624
  2. 2024GPT-4V(ision) is a Generalist Web Agent, if Groundedcited by 547
  3. 2025Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decodingcited by 404
  4. 2025LLaDA-V: Large Language Diffusion Models with Visual Instruction Tuningcited by 145
  5. 2025Accelerating Diffusion LLMs via Adaptive Parallel Decodingcited by 108
  6. 2025The Hidden Life of Tokens: Reducing Hallucination of Large Vision-Language Models via Visual Information Steeringcited by 62
  7. 2026Fast-dLLM++: Fr'echet Profile Decoding for Faster Diffusion LLM Inferencecited by 2

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs
  2. 2025A Comprehensive Study on Visual Token Redundancy for Discrete Diffusion-based Multimodal Large Language Models
  3. 2026Dynamic-dLLM: Dynamic Cache-Budget and Adaptive Parallel Decoding for Training-Free Acceleration of Diffusion LLM
  4. 2026Accelerating Masked Diffusion Large Language Models: A Survey of Efficient Inference Techniques
  5. 2026Seeing the End at Step Zero: Accelerating Diffusion MLLMs via MLP Sparsity-Aware Truncation
  6. 2025Sparse-dLLM: Accelerating Diffusion LLMs with Dynamic Cache Eviction
  7. 2025Discrete Diffusion in Large Language and Multimodal Models: A Survey
  8. 2025dKV-Cache: The Cache for Diffusion Language Models