Mixture of Probes: Learning from Privileged Modalities in Multimodal LLMs Through Probing
Quick Answer
This paper shows that The Mixture of Probes (MoP) framework enhances Multimodal Large Language Models (MLLMs) by effectively utilizing auxiliary modalities during training, achieving up to 65% relative improvement in performance across eight tasks.
Quick Take
The Mixture of Probes (MoP) framework enhances Multimodal Large Language Models (MLLMs) by effectively utilizing auxiliary modalities during training, achieving up to 65% relative improvement in performance across eight tasks. MoP employs a structured probing mechanism to disentangle modality-specific signals, enabling superior cross-modal learning even when some modalities are unavailable at inference. This approach demonstrates significant gains in real-world applications where training and inference conditions differ.
Key Points
- MoP framework disentangles modality-specific and general signals in MLLMs.
- Achieves up to 65% relative performance improvement over strong MLLM baselines.
- Introduces MoP Cross-modal Training to encourage effective cross-modal learning.
- Evaluated across eight tasks and four modalities under privileged modality settings.
- Code and model checkpoints will be publicly available for further research.
Paper Resources
📖 Reader Mode
~2 min readAbstract:Multimodal Large Language Models (MLLMs) are typically designed under the assumption that all modalities available during training will also be accessible at inference. However, many real-world settings violate this assumption, requiring models to operate under a privileged modality setting, where auxiliary modalities are available only during training. While these modalities contain valuable information, existing MLLMs largely fail to leverage them effectively, as they treat modalities as interchangeable inputs rather than sources of complementary supervision. We propose Mixture of Probes (MoP), a novel framework that disentangles modality-specific and modality-general signals within the MLLM, allowing the model to preserve modality-dependent structure while learning transferable representations across modalities. At its core, MoP achieves this through a structured probing mechanism that extracts and organizes information from intermediate representations of a shared modality encoder, rather than relying only on final-layer alignment as done in existing MLLMs. To support this disentanglement, we further introduce MoP Cross-modal Training (MoP-X), a training strategy for MoP centered around a probe disentanglement loss that prevents probe collapse and encourages cross-modal learning. We evaluate MoP across two domains spanning eight tasks and four modalities under a comprehensive evaluation protocol tailored to the privileged modality setting, where each modality is independently treated as the sole input at inference time. MoP consistently outperforms strong MLLM baselines, achieving up to 65% relative improvement, demonstrating that auxiliary modalities, even when unavailable at inference, can provide substantial gains when effectively leveraged during training. Code, model checkpoints, and evaluation protocols will be made available at this https URL.
| Comments: | Preprint (16 pages) |
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG) |
| Cite as: | arXiv:2607.08839 [cs.CV] |
| (or arXiv:2607.08839v1 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2607.08839 arXiv-issued DOI via DataCite |
Submission history
From: Dominick Reilly [view email]
[v1]
Thu, 9 Jul 2026 18:00:47 UTC (1,188 KB)
— Originally published at arxiv.org
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CV
See more →ProMoE-FL: Prototype-conditioned Mixture of Experts for Multimodal Federated Learning with Missing Modalities
ProMoE-FL introduces a Prototype-conditioned Mixture-of-Experts framework for multimodal federated learning, effectively addressing missing modalities. It outperforms existing methods on four chest X-ray datasets, demonstrating superior feature synthesis capabilities in both homogeneous and heterogeneous settings.