Research map / Vision-language and multimodal / Multimodal LLMs

Investigating Redundancy in Multimodal Large Language Models with Multiple Vision Encoders

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Instruction-tuned multimodal assistants

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.

Shows via systematic encoder masking that multi-encoder MLLMs (Eagle, Cambrian-1, I-MoF, Eagle2, DeepSeek-VL) carry substantial encoder redundancy, and proposes Conditional Utilization Rate (CUR) and Information Gap (IG) to quantify per-encoder contribution and imbalance.

Problem

Multi-encoder MLLMs assume different vision encoders give complementary visual signals, but it is unclear whether each encoder provides unique, non-redundant information, and how much this costs in training and inference.

Approach

Mask subsets of vision encoders (zero-tensor replacement) across all 2^n combinations and measure the relative accuracy change per encoder via CUR; define IG as the max-min CUR gap to quantify imbalance across encoders, and use attention KL divergence to explain dominance.

Key results

• Masking two encoders of Cambrian-1 8B yields 1.7% higher overall accuracy than the full model, and its 3-encoder subset is 3.6% higher.
• Eagle-X4 8B Plus EVA-02 reaches CUR of 92.89% on OCR & Chart; masking two encoders of Eagle-X5 7B retains about 96% of baseline performance.
• Dual-encoder Eagle-X2 7B reduces total training time by about 34% while retaining over 96% of original performance.

Surprising findings

• Some encoders have negative CUR (e.g., SigLIP at -16% on Vision-Centric in Cambrian-1 8B), meaning they are detrimental.
• Mean-masking a single encoder nearly matches full-model performance, suggesting strong redundancy among the other encoders.

Limitations

• Analysis limited to static pretrained multi-encoder MLLMs with fixed fusion; dynamic routing not covered.
• CUR/IG are post-hoc diagnostics and do not predict encoder selection a priori.
• Robustness, OOD generalization, calibration and fairness are not evaluated.

Experimental setup

Datasets: GQA MMBench MME SEED-Bench AI2D MathVista ScienceQA MMMU DocVQA ChartQA OCRBench TextVQA

Benchmarks: GQA MMBench MME SEED-Bench AI2D MathVista ScienceQA MMMU DocVQA ChartQA OCRBench TextVQA

Metrics: accuracy CUR information gap KL divergence latency training time

Baselines: Eagle-X5 7B Eagle-X4 8B Plus Cambrian-1 8B Cambrian-1 3B Cambrian-1 13B I-MoF 13B Eagle2 9B DeepSeek-VL 7B

Builds on: Eagle Cambrian-1 LLaVA MouSi

Closest papers by meaning

• Beyond Encoder Accumulation: Measuring Encoder Roles in Multi-Encoder VLMs (NeurIPS 2026 · similarity 0.93)
• CoVFT: Context-aware Visual Fine-tuning for Multimodal Large Language Models (CVPR 2026 · similarity 0.91)
• Eagle: Exploring The Design Space for Multimodal LLMs with Mixture of Encoders (ICLR 2025 · similarity 0.90)
• Divide, Conquer and Combine: A Training-Free Framework for High-Resolution Image Perception in Multimodal Large Language Models (AAAI 2025 · similarity 0.90)
• iLLaVA: An Image is Worth Fewer Than 1/3 Input Tokens in Large Multimodal Models (ICLR 2026 · similarity 0.90)
• Unleashing the Intrinsic Visual Representation Capability of Multimodal Large Language Models (CVPR 2026 · similarity 0.90)