A Comprehensive Information-Decomposition Analysis of Large Vision-Language Models
arXiv:2603.29676v1 Announce Type: new Abstract: Large vision-language models (LVLMs) achieve impressive performance, yet their internal decision-making processes remain opaque, making it difficult to determine if the success stems from true multimodal fusion or from reliance on unimodal priors. To address this attribution gap, we introduce a novel framework using partial information decomposition (PID) to quantitatively measure the "information spectrum" of LVLMs -- decomposing a model's decision-relevant information into redundant, unique, and synergistic components. By adapting a scalable estimator to modern LVLM outputs, our model-agnostic pipeline profiles 26 LVLMs on four datasets across three dimensions -- breadth (cross-model & cross-task), depth (layer-wise information dynamics), a
View PDF HTML (experimental)
Abstract:Large vision-language models (LVLMs) achieve impressive performance, yet their internal decision-making processes remain opaque, making it difficult to determine if the success stems from true multimodal fusion or from reliance on unimodal priors. To address this attribution gap, we introduce a novel framework using partial information decomposition (PID) to quantitatively measure the "information spectrum" of LVLMs -- decomposing a model's decision-relevant information into redundant, unique, and synergistic components. By adapting a scalable estimator to modern LVLM outputs, our model-agnostic pipeline profiles 26 LVLMs on four datasets across three dimensions -- breadth (cross-model & cross-task), depth (layer-wise information dynamics), and time (learning dynamics across training). Our analysis reveals two key results: (i) two task regimes (synergy-driven vs. knowledge-driven) and (ii) two stable, contrasting family-level strategies (fusion-centric vs. language-centric). We also uncover a consistent three-phase pattern in layer-wise processing and identify visual instruction tuning as the key stage where fusion is learned. Together, these contributions provide a quantitative lens beyond accuracy-only evaluation and offer insights for analyzing and designing the next generation of LVLMs. Code and data are available at this https URL .
Comments: Accepted at ICLR 2026. Project page: this https URL
Subjects:
Machine Learning (cs.LG); Computation and Language (cs.CL); Computer Vision and Pattern Recognition (cs.CV)
Cite as: arXiv:2603.29676 [cs.LG]
(or arXiv:2603.29676v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2603.29676
arXiv-issued DOI via DataCite (pending registration)
Submission history
From: Lixin Xiu [view email] [v1] Tue, 31 Mar 2026 12:32:29 UTC (684 KB)
Sign in to highlight and annotate this article
Conversation starters
Daily AI Digest
Get the top 5 AI stories delivered to your inbox every morning.
More about
modellanguage modeltraining
Choosing an AI Agent Orchestrator in 2026: A Practical Comparison
Running one AI coding agent is easy. Running three in parallel on the same codebase is where things get interesting — and where you need to make a tooling choice. There's no "best" orchestrator. There's the right one for your workflow. Here's an honest comparison of five approaches, with the tradeoffs I've seen after months of running multi-agent setups. The Options 1. Raw tmux Scripts What it is: Shell scripts that launch agents in tmux panes. DIY orchestration. Pros: Zero dependencies beyond tmux Full control over every detail No abstractions to fight You already know how it works Cons: No state management — you track everything manually No message routing between agents No test gating — agents declare "done" without verification Breaks when agents crash or hit context limits You become
How AI Is Changing the Way We Build Online Businesses
Not long ago, building an online business meant: months of development hiring developers large upfront costs Today? AI has completely changed the game. Now, one person can go from idea → to revenue faster than ever before. And this shift is just getting started. ⚠️ The Old Way vs The New Way Before AI: Build everything from scratch Spend weeks on infrastructure Launch slowly Iterate even slower With AI: Build faster Automate key tasks Launch quickly Iterate in real time The difference is massive. 🧠 AI Is Reducing the Cost of Building One of the biggest changes: 👉 Building is no longer the bottleneck AI helps with: generating content writing code automating workflows handling repetitive tasks What used to take weeks… 👉 now takes days ⚙️ Infrastructure Is No Longer the Hard Part Another s
Knowledge Map
Connected Articles — Knowledge Graph
This article is connected to other articles through shared AI topics and tags.
Discussion
Sign in to join the discussion
No comments yet — be the first to share your thoughts!