LangBridge: Interpreting Image as a Combination of Language Embeddings¶
Basic Information¶
- Conference: ICCV 2025
- arXiv: 2503.19404
- Code: Project Page
- Area: Information Retrieval
- Keywords: Vision-language alignment, adapter transfer, vocabulary embedding projection, MLP analysis, cross-model reuse
TL;DR¶
LangBridge achieves interpretable vision-language alignment by explicitly decomposing visual features into linear combinations of LLM vocabulary embeddings, and supports pretraining-free adapter transfer across different LLMs.
Background & Motivation¶
Mainstream large vision-language models (LVLMs) follow the LLaVA paradigm, employing an MLP to project visual features into the text embedding space of an LLM. While effective, this approach has two core issues:
Opaque MLP alignment mechanism: The underlying mechanism by which the MLP bridges the modality gap remains poorly understood, with few studies systematically analyzing its fundamental alignment behavior.
Non-transferable adapters: Switching the LLM backbone requires retraining the MLP adapter from scratch due to input dimension mismatches and feature distribution discrepancies, incurring substantial computational cost.
The authors first conduct a systematic investigation of MLP adapter behavior and identify two key findings: - Visual embeddings exhibit strong correlation with semantically related text tokens (e.g., high cosine similarity between apple image patches and the text "Green Apple"). - The projection capability of the MLP develops progressively during training, gradually learning to project visual features into the corresponding text embedding subspace.
Method¶
Overall Architecture¶
The core idea of LangBridge is Language Basis Vector Projection: representing each visual embedding as a weighted linear combination of LLM vocabulary embeddings:
where \(\beta_k\) is a probability distribution over vocabulary tokens. The method consists of three stages:
Key Designs¶
Stage 1: Visual Feature Extraction
A Vision Transformer (CLIP-ViT-L/14@336px) is used to extract patch-level visual features from the input image:
Stage 2: Probability Computation
A two-layer MLP projects visual features into the LLM text embedding space, followed by a linear layer that produces a probability distribution over the vocabulary:
where \(\mathbf{W} \in \mathbb{R}^{T \times D}\) and \(T\) denotes vocabulary size.
Stage 3: Linear Combination of Text Embeddings
The probability distribution is used as coefficients to linearly combine the LLM's vocabulary embeddings:
Vocabulary Selection Strategy¶
Directly using the full vocabulary embedding matrix is computationally prohibitive (~1B parameters). The authors merge the vocabularies of LLaMA and Qwen, retaining only tokens shared by both, then compute token frequencies over the ShareGPT4V and LLaVA-CC3M-Pretrain-595K datasets to select the top-19,200 high-frequency tokens as a compact shared vocabulary.
Cross-LLM Adapter Transfer¶
LangBridge learns only the linear combination relationship between visual patches and vocabulary embeddings (i.e., the probability distribution), rather than a direct dimensional mapping. During transfer:
The probability distribution \(P\) can be directly applied to weight the vocabulary embeddings of any target LLM without repretraining.
Key Experimental Results¶
Main Results: Same-Architecture Transfer¶
Results of LangBridge pretrained on Qwen2-0.5B and directly transferred to larger models:
| SFT-LLM | Connector | GQA | TextVQA | MME | MMBench | MMVeT | POPE | SciQA |
|---|---|---|---|---|---|---|---|---|
| Qwen2-7B | 7B-Pretrain-MLPs | 62.92 | 57.24 | 1938 | 72.7 | 35.5 | 87.8 | 79.44 |
| Qwen2-7B | 0.5B-Pretrain-LB | 63.03 | 57.25 | 1886 | 71.7 | 34.1 | 88.2 | 79.23 |
| Qwen2.5-14B | 14B-Pretrain-MLPs | 63.71 | 61.32 | 2038 | 78.2 | 37.7 | 88.1 | 85.59 |
| Qwen2.5-14B | 0.5B-SFT-LB | 63.92 | 62.02 | 1990 | 77.4 | 38.4 | 87.6 | 84.77 |
Key finding: After transferring a LangBridge connector pretrained on a 0.5B model to a 14B model, performance on TextVQA (+1.14%) and MMVeT (+1.86%) even surpasses the baseline.
Ablation Study: Vocabulary Size¶
| Vocab Size | GQA | TextVQA | MME | MMBench | MMVeT | POPE | SciQA |
|---|---|---|---|---|---|---|---|
| 19,200 | 63.15 | 57.34 | 1904 | 71.0 | 31.6 | 88.3 | 79.25 |
| 25,600 | 63.13 | 57.58 | 1842 | 71.8 | 32.9 | 87.9 | 79.01 |
| 32,000 | 63.11 | 57.19 | 1832 | 72.7 | 33.2 | 88.6 | 79.11 |
A vocabulary size of 19,200 significantly outperforms larger vocabularies on MME (by 6–8%) and achieves the best overall performance.
Other Key Findings¶
- Cross-architecture transfer: Transferring a LangBridge connector pretrained on Qwen2-0.5B to LLaMA3-8B yields a +9.68% gain on MMVeT with overall performance improvements.
- Standard setting comparison: LangBridge trained directly on LLaMA3-8B outperforms the MLP baseline on GQA, TextVQA, MME, and MMBench.
- Computational cost: Training overhead increases by only ~10% (4.273 vs. 3.876 s/iter), while cross-LLM pretraining can be entirely skipped.
Highlights & Insights¶
- In-depth mechanistic analysis: Visualization-based analysis reveals the progressive process by which the MLP adapter learns to project visual features into the text embedding subspace, providing theoretical grounding for the proposed design.
- Elegant design philosophy: Transforming "implicit projection" into "explicit linear combination" renders the vision-language alignment process interpretable.
- Practical transferability: A single pretraining run on a small model enables transfer to multiple large models, substantially reducing the cost of multi-model deployment.
- Shared vocabulary strategy: Cross-model compatible tokens are selected via frequency statistics, achieving cross-architecture compatibility with minimal parameter overhead.
Limitations & Future Work¶
- Vocabulary selection relies on frequency statistics over specific training datasets and may not generalize to other data distributions.
- Validation is currently limited to the LLaVA-style framework and has not been extended to other architectures such as Flamingo or BLIP-2.
- A vocabulary size of 19,200 may be insufficient to cover all fine-grained visual concepts.
- On certain benchmarks (e.g., MME), transfer still incurs a 2–4% performance drop.
Related Work & Insights¶
- Ovis employs a visual embedding table for structured alignment but does not support cross-LLM transfer.
- The MLP adapters in the LLaVA series, while simple and effective, lack interpretability.
- The "vocabulary embeddings as basis vectors" paradigm proposed in this paper may inspire further research on cross-modal alignment grounded in the intrinsic representations of LLMs.
Rating¶
⭐⭐⭐⭐ — Mechanistic analysis is insightful, the method design is elegant, and the cross-LLM transfer offers high practical value; however, some performance degradation persists on certain benchmarks after transfer.