Cross-Modal Latent Influence: Patch-Level Training Sample Attribution for Vision-Language Models
On the map: Multimodal LLMs / vision-language models → VLM benchmarks and visual reasoning also fits: MLLM training and visual instruction tuning
Abstract (arXiv)
Explaining a vision--language model (VLM) prediction requires tracing training-example influence to the multimodal evidence that mediates it, yet existing attribution methods largely operate only at the sample level. Recent latent-mediated influence methods provide feature-level attribution in language models, but lack the multimodal and spatial structure needed to ground influence in VLMs. We introduce MOSAIC (Multimodal Orthogonalized Sparse Autoencoders for Influence Characterization). It learns a shared sparse image--text latent space whose concepts are 1) sparse within each modality, 2) supported on contiguous image regions, 3) aligned to the phrases that describe them, and 4) approximately decorrelated. It then applies EK-FAC to decompose training-example influence through these concepts and localize their contributions to specific text phrases and image regions. We evaluate MOSAIC on Qwen and Gemma across captioning, chest X-ray understanding, and visual question answering. The learned features are more localized and coherent without degrading task behavior, and ablating influence-ranked latents changes model outputs more than ablating latents ranked by activation or frequency. We find that visual evidence can substantially reorganize training-data attribution while preserving more of the shared latent structure. MOSAIC thus connects where model behavior comes from with what multimodal evidence carries that influence.
Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in VLM benchmarks and visual reasoning, one of 16 directions in Multimodal LLMs / vision-language models. Builds on 28 of its 34 references that are in the map. First posted on arXiv 9 Oct 2026.
Builds on
28 of 34 references.
Closest ancestors · 6
Same direction, newest first
- 2026Steering Vision-Language Models with Joint Sparse Autoencoders24 Jun
- 2026Extraction and Analysis of Multimodal Concepts in Vision Language Models through Sparse Autoencoders19 Jun
- 2026Cascaded Sparse Autoencoders Learn Multi-Level Visual Concepts in Multimodal LLMs15 Jun
- 2025VL-SAE: Interpreting and Enhancing Vision-Language Alignment with a Unified Concept Set24 Oct
- 2025V-SEAM: Visual Semantic Editing and Attention Modulating for Causal Interpretability of Vision-Language Models18 Sep
- 2025Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models3 Apr
Other references · 22
- 2016Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localizationcited by 6,733
- 2017Understanding Black-box Predictions via Influence Functionscited by 1,741
- 2019CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparisoncited by 1,423
- 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
- 2019Attention is not Explanationcited by 929
- 2020Quantifying Attention Flow in Transformerscited by 775
- 2019Attention is not not Explanationcited by 636
- 2023Studying Large Language Model Generalization with Influence Functionscited by 326
- 2021Generic Attention-model Explainability for Interpreting Bi-Modal and Encoder-Decoder Transformerscited by 320
- 2018Fast Approximate Natural Gradient Descent in a Kronecker-factored Eigenbasiscited by 146
- 2024Exploring scalable medical image encoders beyond text supervisioncited by 132
- 2020Captioning Images Taken by People Who Are Blindcited by 110
10 more
- 2025Interpreting CLIP with Hierarchical Sparse Autoencoderscited by 60
- 2024Do Vision & Language Decoders use Images and Text equally? How Self-consistent are their Explanations?cited by 24
- 2024Does Object Grounding Really Reduce Hallucination of Large Vision-Language Models?cited by 9
- 2025SPARC: Concept-Aligned Sparse Autoencoders for Cross-Model and Cross-Modal Interpretabilitycited by 9
- 2026Concept Influence: Leveraging Interpretability to Improve Performance and Efficiency in Training Data Attributioncited by 8
- 2025From What to How: Attributing CLIP's Latent Components Reveals Unexpected Semantic Reliancecited by 6
- 2025SAVE: Sparse Autoencoder-Driven Visual Information Enhancement for Mitigating Object Hallucinationcited by 5
- 2026Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spacescited by 1
- 2026STRIDE: Training Data Attribution via Sparse Recovery from Subset Perturbationscited by 1
- 2026When Structured Sparse Autoencoders Learn Consistent Concepts Across Modalitiescited by 1
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2026InfluenceField: A Differentiable Field with Interventionally Identifiable Causal Structure for Multimodal World Modeling
- 2025Interpreting the linear structure of vision-language model embedding spaces
- 2023Visual Explanations of Image-Text Representations via Multi-Modal Information Bottleneck Attribution
- 2026Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoders
- 2026Through the LENS: Local Geometric Decomposition of Vision-Language Model Representations
- 2026Sparse CLIP: Co-Optimizing Interpretability and Performance in Contrastive Learning
- 2026Learning to Detect Cross-Modal Negation: An Analysis of Latent Representations and an Attention-Based Solution
- 2026A Strong Baseline for Evaluating Vision Encoders in Multimodal Large Language Models