Research map / Vision-language and multimodal / Multimodal LLMs

Spectral-Progressive Thought Flow for Lightweight Multimodal Reasoning

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Training-free token reduction for MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SpecFlow, which represents intermediate visual thoughts as a fixed-size, progressively unmasked block-DCT workspace updated by CFG-guided flow matching, interleaved with autoregressive text thoughts to cut visual token and KV-cache growth in multi-hop multimodal spatial reasoning.

Problem

Multi-hop multimodal spatial reasoning with interleaved visual thoughts causes visual token accumulation, growing attention compute and KV-cache memory with reasoning depth.

Approach

Overwrite a fixed-size visual state each hop, updated by flow matching in block-cosine (DCT) coefficient space under a time-dependent frequency mask that activates low to high bands, with classifier-free guidance from the textual trace steering the velocity field.

Key results

• On VSR, SpecFlow reaches 70.14% accuracy versus 68.88% for VoCoT and 68.82% for CODI, at 11169.74 GFLOPs and 0.41 s latency.
• On dynamic decision-making, SpecFlow averages 79.15% success versus 75.59% for DiffThinker and 63.27% for Qwen3-VL-32B SFT.
• KV-cache reduction of about 1.6x to 2.1x over baselines is reported, with memory stable around 2.1 to 3.5 GB.

Surprising findings

• A smaller AR backbone (Qwen3-VL-8B) with SpecFlow outperforms Qwen3-VL-32B fine-tuned baselines on the decision-making tasks, suggesting gains are not explained by AR scale.

Limitations

• Requires a 20B diffusion transformer for visual thoughts, so total model size is large despite the lighter inference state.
• Task-specific FM models are trained per task category, limiting generality.
• Some reported comparisons (e.g., Winoground vs VoCoT) show only marginal gains.

Experimental setup

Datasets: VSR EmbSpatial V-Star Winoground Maze MiniBehavior FrozenLake VoCoT corpus

Benchmarks: VSR V-Star EmbSpatial Winoground Maze MiniBehavior FrozenLake

Metrics: accuracy success rate FLOPs latency memory KV cache size

Baselines: VoCoT SoT LightFastV SparseVLM Heima PCCoT CODI DiffThinker MVoT

Builds on: flow matching classifier-free guidance DiffThinker MVoT Qwen-Image-Edit

Closest papers by meaning

• AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.91)
• ThinkMorph: Emergent Properties in Multimodal Interleaved Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.90)
• UniVLR: Unifying Text and Vision in Visual Latent Reasoning for Multimodal LLMs (NeurIPS 2026 · similarity 0.90)
• Look on Demand: A Cognitive Scheduling Framework for Visual Evidence Acquisition in Multimodal Reasoning (ICML 2026 · similarity 0.89)
• Visual Thoughts: A Unified Perspective of Understanding Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.89)
• See It, Say It, Sorted: An Iterative Training-Free Framework for Visually-Grounded Multimodal Reasoning in LVLMs (CVPR 2026 · similarity 0.89)