Answer generation stabilizes before explanation unfolds in AI systems
Two Clocks in Diffusion MLLMs: When Answers Stabilize Before Rationales Unfold
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Some AI models give answers that settle quickly while their explanations keep developing. The authors studied this timing difference, finding that most of the explanation remains unwritten when the answer appears stable. Changing how the model processes information affects both how complete the answers are and how much context it uses. Their work shows there's a difference between when answers become fixed and when the reasoning behind them is fully formed.
What this means in practice
- •For ai system developers: Improve timing control in AI models to better balance answer stability with rationale generation for visual question-answering tasks.
- •For product designers: Design interfaces that show AI answers early while rationale explanations continue updating, enhancing user trust and understanding.
Authors
Keuntae Kim, Yong Suk Choi
Abstract
An answer candidate in a masked diffusion MLLM can stabilize while its rationale is still unfolding. We distinguish retrospective stabilization of the logged candidate from token commitment, and examine these two clocks relative to rationale generation. Analyzing our results across three visual question-answering benchmarks, we find that 89.4-98.1% of the rationale-side canvas remains unwritten at stabilization in single-block, EOS-suppressed LaViDa runs. On V*Bench, reducing block length from 128 to 8 changes this fraction from 89.4% to 1.7%, together with answer coverage and the eligible observation window. Under EOS-enabled prompting, direct instructions improve Nemotron's overall accuracy by 15.0 and 19.5 percentage points on M3CoT and ScienceQA, but reduce LaViDa/V*Bench accuracy by 11.0 points. A symmetric decomposition associates the larger absolute component of each change with coverage rather than conditional accuracy. Matched-canvas image ablations measure visual sensitivity alongside answer stabilization, separating the two temporal readouts. Together, these measurements distinguish answer stabilization, rationale unfolding, and visual sensitivity, and identify coverage as the larger component of the prompting differences.