Source-linked AI summary
MCPO: Modality-Contrastive Preference Optimization for Multimodal Chain-of-Thought Compression
Guangheng Yang, Zhenliang Ni, Zhenkai Wu, Han Shu, Juan Feng, Wenming Yang, Jie Hu
TL;DR
Long multimodal reasoning chains impose substantial computational costs, while existing compression methods lack fine-grained cross-modal constraints and can induce visual laziness or hallucinated reasoning. MCPO combines NCMI-based step pruning with asymmetric multimodal preference optimization, achieving up to 69.5% CoT reduction and 3.34× inference speedup while preserving accuracy.
Problem
Long M-CoT trajectories create substantial computational and KV-cache costs, while existing compression methods lack fine-grained cross-modal constraints and can induce visual laziness or hallucinated reasoning.
Method
MCPO uses step-level NCMI pruning to remove visually independent reasoning and asymmetric preference optimization to align concise trajectories across with-image and no-image contexts.
Results
69.5% maximum CoT length reduction and 3.34× maximum end-to-end speedup are achieved while preserving accuracy.
Takeaways & Limitations
MCPO combines adaptive pruning and modality-contrastive alignment to suppress visual hallucinations while retaining essential reasoning under aggressive compression.
Abstract
from arXiv · showhide
Recently, multimodal large-scale reasoning models have demonstrated remarkable capabilities in solving complex tasks through long Chains-of-Thought (M-CoT). However, excessively long reasoning trajectories incur substantial computational costs and significant KV-cache pressure. Existing CoT compression and alignment paradigms mainly rely on static rules or single-dimensional preferences, lacking fine-grained cross-modal constraints; as a result, they are prone to inducing visual laziness and hallucinatory reasoning. To address these issues, we propose Modality-Contrastive Preference Optimization (MCPO), a highly sample-efficient two-stage length-compression method that requires fewer than 900 training samples. In the compression stage, we introduce a step-level Normalized Cross-Modal Mutual Information (NCMI) pruning algorithm, which automatically identifies and removes visual-independent reasoning steps by comparing the reasoning discrepancies between with-image and no-image contexts. This significantly reduces redundancy and hallucinatory content in the reasoning chains. In the alignment stage, the model first undergoes supervised fine-tuning to achieve domain-adaptive initialization, followed by optimization using an asymmetric multimodal length-controlled preference loss. This objective adopts a highly nonlinear odds-ratio formulation that provides steep gradients in the with-image context to reinforce length constraints for preferred trajectories, while applying a scaled, flat-gradient linear difference in the no-image context to maintain modality consistency, thereby achieving stable cross-modal preference alignment. Extensive experiments on mainstream base models such as Qwen3-VL-Thinking show that our method can reduce CoT length by up to 69.5% and achieve up to 3.34x end-to-end inference speedup while preserving original accuracy.
1 Introduction
MCPO addresses the cost and cross-modal reliability problems of long multimodal reasoning chains with sample-efficient pruning and asymmetric preference alignment. It reduces reasoning length while preserving accuracy and improving inference efficiency.
- Long multimodal reasoning chains increase inference latency, KV-cache pressure, and computational cost.
- Existing compression methods rely on static heuristics or single-dimensional preferences, risking visual laziness, hallucinated reasoning, and high training overhead.
- MCPO uses step-level NCMI to compare with-image and no-image reasoning, pruning steps with low visual dependence from compact trajectories.
- MCPO applies asymmetric preference optimization with steep multimodal odds-ratio gradients and flatter unimodal constraints to reinforce concise, visually grounded reasoning.
- 69.5% maximum token reduction and 3.34× inference speedup are achieved while preserving accuracy across multimodal reasoning evaluations.
2 Related Work
Prior work compresses chain-of-thought through training-based or training-free strategies, but MCPO is positioned within multimodal reasoning compression that must separate essential reasoning from redundant steps.
- Chain-of-thought improves complex multi-step reasoning but can produce costly, verbose, and repetitive traces.
- MCPO’s framework overview combines step-level NCMI pruning with asymmetric multimodal preference alignment to preserve visual dependence while controlling length.
- Compression methods divide into training-based approaches that learn shortened policies and training-free approaches that prune or stop generation without parameter updates.
- Training-based methods: Training-based methods include supervised fine-tuning, preference optimization, reinforcement learning, skip-token policies, entropy pruning, and length-controlled objectives.
- Training-free methods: Training-free methods use entropy, confidence, semantic redundancy, or adaptive stopping signals to reduce reasoning during inference.
3 Methodology
MCPO compresses multimodal reasoning by pruning steps with weak visual dependence and aligning the resulting trajectories with asymmetric multimodal and unimodal preference objectives.
- Adaptive Step Pruning via NCMI: MCPO computes step-level NCMI by comparing length-normalized step likelihoods with and without visual input.Positive NCMI indicates visual dependence, whereas near-zero NCMI identifies redundant or potentially hallucinatory steps.
- Adaptive Step Pruning via NCMI: Hierarchical segmentation divides reasoning into paragraph-level steps, falling back to sentence-level segmentation with minimum-length protection when necessary.The fallback avoids meaningless fragments while enabling local NCMI computation on unannotated traces.
- Adaptive Step Pruning via NCMI: An adaptive coefficient-of-variation threshold adjusts pruning to each sequence’s NCMI distribution, increasing pruning when visually dependent and independent steps separate clearly.The method caps CV at 3.0, bounds β between −2.0 and 1.0, and retains at least one reasoning step per sample.
- Adaptive Step Pruning via NCMI: Steps below the dynamic NCMI threshold are replaced with [SKIP], while higher-scoring steps are retained.This operationalizes removal of reasoning that contributes little visual information.
- Modality-Contrastive Preference Optimization: MCPO uses supervised fine-tuning followed by a two-part preference loss combining with-image length alignment and no-image self-contrastive regularization.The with-image term uses a nonlinear log-odds reward, while the no-image term uses a stable linear log-probability reward to regulate concise reasoning across modalities.
- Qualitative Case Study: A CLEVR case study compares a 168-token original CoT with a 42-token MCPO CoT, a 75% reduction achieved while preserving strict visual grounding.The pruned trajectory removes self-doubt loops and text-only prompt repetitions using [SKIP] tokens.
- Modality-Contrastive Preference Optimization: The asymmetric objective makes with-image chosen-probability updates dominate no-image updates, preserving a visual information gain gap while discouraging text-prior shortcuts.The no-image probability increase is damped, whereas the with-image gradient includes an accelerating odds-ratio factor.
4 Experiments
MCPO is evaluated for multimodal CoT compression, accuracy retention, and system-level efficiency across model scales, benchmarks, ablations, and difficulty levels. It substantially reduces reasoning tokens while preserving task performance and improving inference latency.
- Main Results: MCPO achieves up to 69.5% token reduction while maintaining task accuracy within 0.8% across the main comparison.The comparison covers Qwen3-VL at 8B and 4B scales against StepEntropy and REFRAIN.
- Experimental Setup: The evaluation uses four multimodal benchmarks and reports accuracy, average thinking length, and inference latency as core metrics.Experiments include MathVista_mini, ScienceQA (Image Only), MMMU_dev_val, and MMStar.
- Main Results: On MathVista-mini, MCPO reduces generated tokens from 1188 to 362 (↓69.5%) with only a 0.80 accuracy reduction.Step Entropy and REFRAIN reduce tokens less and incur larger accuracy decreases of 1.60 and 2.60.
- Main Results: MCPO achieves token pruning ratios of 58.4%, 60.3%, and 42.6% on ScienceQA, MMMU, and MMStar, with accuracy changes within ±1%.These results outperform the compared pruning methods in compression while limiting accuracy degradation.
- System-Level Efficiency: MCPO reduces 8B-model latency on MMMU from 220.30s to 65.87s, achieving 3.34× speedup and 70.1% latency reduction.It also delivers cross-scale speedups on the 4B model, including 1.73× on MathVista_mini and 1.55× on MMMU.
- Ablation Studies: The [SKIP] embedding ablation exposes a compression–accuracy trade-off between freezing and keeping the embedding unfrozen.Freezing produces stronger token compression, whereas the unfrozen embedding improves accuracy.
- Difficulty-Level Analysis: Across MMMU difficulty levels, MCPO reduces thinking-token consumption by approximately 60%, 63%, and 59% for Easy, Medium, and Hard tasks.Reasoning cost still increases with task difficulty, but compression remains stable across all three categories.
- Difficulty-Level Analysis: Contrastive weight affects accuracy differently by difficulty: α=0.1 is best on Easy, α=0.05 on Medium, and α=0.2 on Hard tasks.The reported accuracies are 62.6%, 49.0%, and 50.9%, respectively.
5 Conclusion
MCPO compresses multimodal chain-of-thought reasoning while preserving accuracy, reducing inference token usage and improving end-to-end execution speed across multiple benchmarks.
- MCPO often exceeds 50% inference-token compression while maintaining accuracy within a narrow margin of the baseline model.Experiments cover MathVista-mini, ScienceQA, MMMU, and MMStar.
- Up to 3.34× end-to-end execution speedups demonstrate MCPO’s latency-efficiency benefits.
- Combining adaptive step pruning with asymmetric modality-contrastive alignment suppresses visual hallucinations while retaining essential reasoning steps under aggressive compression.
- Ablations show stable behavior across contrastive weights, embedding configurations, and varying reasoning difficulty levels.