Title: Spectral-Progressive Thought Flow for Lightweight Multimodal Reasoning

Abstract:

Multimodal spatial reasoning typically depends on extensive sequences of intermediate visual and textual thoughts. This approach often leads to significant computational burdens and memory overhead due to the accumulation of visual tokens and the demands of dense cross-modal attention. To mitigate these issues, we introduce Spectral-Progressive Thought Flow (SpecFlow), a new lightweight framework designed for multimodal spatial reasoning. SpecFlow encodes intermediate visual thoughts into a fixed-size discrete cosine space. Leveraging strong energy compaction, this method retains global layout and relational structures, adding high-frequency details solely when higher spatial precision is necessary.

To synchronize the evolution of visual states with linguistic intent, SpecFlow employs classifier-free guidance. This mechanism allows autoregressive textual thoughts to direct flow-based updates within the visual workspace or state, without increasing the context size. Consequently, SpecFlow ensures that the visual workspace remains bounded. Its updates are determined exclusively by the current visual state and the accumulated textual trace, facilitating long-horizon inference with consistent latency and memory consumption that does not scale with reasoning depth. Empirical evaluations demonstrate that SpecFlow delivers competitive or better reasoning performance while cutting computation and KV cache expenses by as much as 2.1 times.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC