Pith. sign in

REVIEW 6 cited by

A Survey on Mechanistic Interpretability for Multi-Modal Foundation Models

Not yet reviewed by Pith; the record is open.

Re-run · record.json Download PDF Read on arXiv ↗

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2502.17516 v1 pith:PG4UEOEC submitted 2025-02-22 cs.LG cs.AI

Zihao Lin , Samyadeep Basu , Mohammad Beigi , Varun Manjunatha , Ryan A. Rossi , Zichao Wang , Yufan Zhou , Sriram Balasubramanian
show 13 more authors
Arman Zarei Keivan Rezaei Ying Shen Barry Menglong Yao Zhiyang Xu Qin Liu Yuxiang Zhang Yan Sun Shilong Liu Li Shen Hongxuan Li Soheil Feizi Lifu Huang
This is my paper · ORCID
classification cs.LGcs.AI
keywords modelsinterpretabilityfoundationmultimodalunimodallanguagellmsmechanistic
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

The rise of foundation models has transformed machine learning research, prompting efforts to uncover their inner workings and develop more efficient and reliable applications for better control. While significant progress has been made in interpreting Large Language Models (LLMs), multimodal foundation models (MMFMs) - such as contrastive vision-language models, generative vision-language models, and text-to-image models - pose unique interpretability challenges beyond unimodal frameworks. Despite initial studies, a substantial gap remains between the interpretability of LLMs and MMFMs. This survey explores two key aspects: (1) the adaptation of LLM interpretability methods to multimodal models and (2) understanding the mechanistic differences between unimodal language models and crossmodal systems. By systematically reviewing current MMFM analysis techniques, we propose a structured taxonomy of interpretability methods, compare insights across unimodal and multimodal architectures, and highlight critical research gaps.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 6 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. FairFlow: Demystifying and Mitigating Stereotype Bias in Text-to-Image Diffusion Transformers

    cs.CV 2026-07 conditional novelty 7.0 of 10

    Bias in MM-DiTs is mediated by sparse stage-wise semantic binding hubs, and sparse inference-time steering at those hubs mitigates gender, race, and intersectional stereotypes with low overhead.

  2. Through the LENS: Local Geometric Decomposition of Vision-Language Model Representations

    cs.AI 2026-08 conditional novelty 6.0 of 10

    Local low-rank Gaussian neighborhoods in VLM residual streams reveal model-specific fusion trajectories and serve as causal steering and retrieval units.

  3. Which Modality Decides? Counterfactual Modality Attribution for Multimodal LLMs

    cs.CV 2026-07 conditional novelty 6.0 of 10

    CMA uses diffusion-generated counterfactuals and two-player Shapley values to measure whether an image or the text drives a multimodal LLM's prediction, hitting 98% on synthetic biased benchmarks.

  4. How Do VLMs Fail? Vision-Operation Misalignment in Compositional VQA

    cs.CV 2026-07 reject novelty 6.0 of 10

    The paper proposes four operation-level VLM failure modes and a pathway dissociation, but the dissociation is not supported by the paper's own intervention statistics.

  5. Unsupervised Features Mining via Activation Geometry

    cs.AI 2026-07 conditional novelty 6.0 of 10

    Prefix-induced activation shifts (MAG) yield model-relative reasoning directions that predict verdicts, support matched-format steering, and select transfer datasets at 94.7% Top-1 accuracy.

  6. GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs

    cs.CL 2025-07 conditional novelty 6.0 of 10

    GrAInS uses Integrated Gradients to identify the most influential tokens, then builds layer-wise steering vectors that improve truthfulness, reduce hallucination, and preserve general capabilities in LLMs and VLMs.

Pith tools