mlsanityData to 14 Oct 2026

Causal Probing for Internal Visual Representations in Multimodal Large Language Models

arXiv 2605.05593 ↗First posted 7 May 2026cs.AIcited by 0

On the map: Multimodal LLMs / vision-language models → VLM benchmarks and visual reasoning

Abstract (arXiv)

Despite the remarkable success of Multimodal Large Language Models (MLLMs) across diverse tasks, the internal mechanisms governing how they encode and ground distinct visual concepts remain poorly understood. To unravel these mechanisms, we propose a causal framework based on activation steering to actively probe and manipulate internal visual representations. Through systematic intervention across four visual concept categories, our results reveal a divergence in concept encoding: entity knowledge is distinctively localized, whereas abstract concepts are globally distributed across the network. Critically, this divergence uncovers a mechanistic driver of scaling laws: increasing model depth is indispensable for encoding distributed and complex abstract concepts, whereas entities maintain a consistently high degree of localization. Furthermore, reverse steering uncovers that blocking explicit output triggers a surge in latent activations, exposing a compensatory mechanism between perception and generation. Finally, by extending our analysis to visual reasoning, we expose a disconnect between perception and reasoning: although MLLMs successfully recognize geometric relations, they treat them merely as static visual features, failing to trigger the procedural execution necessary for solving problems.

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 44 of its 56 references that are in the map. First posted on arXiv 7 May 2026.

Builds on

44 of 56 references.

References · 44

  1. 2019Sentence-BERT: Sentence Embeddings using Siamese BERT-Networkscited by 9,170
  2. 2017A Unified Approach to Interpreting Model Predictionscited by 6,555
  3. 2016"Why Should I Trust You?": Explaining the Predictions of Any Classifiercited by 5,070
  4. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  5. 2025Gemma 3 Technical Reportcited by 1,824
  6. 2016Understanding intermediate layers using linear classifier probescited by 1,353
  7. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,336
  8. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,302
  9. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,165
  10. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,085
  11. 2019BERT Rediscovers the Classical NLP Pipelinecited by 1,059
  12. 2023Towards Understanding Sycophancy in Language Modelscited by 1,040
32 more
  1. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 1,032
  2. 2022Mass-Editing Memory in a Transformercited by 1,001
  3. 2019Attention is not Explanationcited by 929
  4. 2023Steering Language Models With Activation Engineeringcited by 922
  5. 2019Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inferencecited by 876
  6. 2021Probing Classifiers: Promises, Shortcomings, and Advancescited by 876
  7. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 724
  8. 2023The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasetscited by 712
  9. 2019Attention is not not Explanationcited by 636
  10. 2023MagicBrush: A Manually Annotated Dataset for Instruction-Guided Image Editingcited by 624
  11. 2022Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Taskcited by 431
  12. 2021Causal Abstractions of Neural Networkscited by 422
  13. 2023Towards Best Practices of Activation Patching in Language Models: Metrics and Methodscited by 359
  14. 2025Persona Vectors: Monitoring and Controlling Character Traits in Language Modelscited by 345
  15. 2017Don't Just Assume; Look and Answer: Overcoming Priors for Visual Question Answeringcited by 339
  16. 2023Emergent Linear Representations in World Models of Self-Supervised Sequence Modelscited by 319
  17. 2024Programming Refusal with Conditional Activation Steeringcited by 206
  18. 2020Causal Mediation Analysis for Interpreting Neural NLP: The Case of Gender Biascited by 191
  19. 2024Improving Instruction-Following in Language Models through Activation Steeringcited by 160
  20. 2022On the Paradox of Learning to Reason from Datacited by 93
  21. 2023EmoSet: A Large-scale Visual Emotion Dataset with Rich Attributescited by 82
  22. 2023Igniting Language Intelligence: The Hitchhiker's Guide From Chain-of-Thought Reasoning to Language Agentscited by 72
  23. 2024Semantics-Adaptive Activation Intervention for LLMs via Dynamic Steering Vectorscited by 63
  24. 2024Reducing Hallucinations in Vision-Language Models via Latent Space Steeringcited by 59
  25. 2022CEBaB: Estimating the Causal Effects of Real-World Concepts on NLP Model Behaviorcited by 55
  26. 2025OmniConsistency: Learning Style-Agnostic Consistency from Paired Stylization Datacited by 32
  27. 2024Beyond Single Concept Vector: Modeling Concept Subspace in LLMs with Gaussian Distributioncited by 27
  28. 2024The Quest for the Right Mediator: Surveying Mechanistic Interpretability Through the Lens of Causal Mediation Analysiscited by 24
  29. 2025Activation Steering for Chain-of-Thought Compressioncited by 24
  30. 2025Analyzing Finetuning Representation Shift for Multimodal LLMs Steeringcited by 16
  31. 2024How Reliable are Causal Probing Interventions?cited by 11
  32. 2025LaSM: Layer-wise Scaling Mechanism for Defending Pop-up Attack on GUI Agentscited by 6

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025ISO-Bench: Benchmarking Multimodal Causal Reasoning in Visual-Language Models through Procedural Plans
  2. 2026See2Think: Do Multimodal Models Really Use Intermediate Visual States?
  3. 2026Visual Instruction Tuning Aligns Modalities through Abstraction
  4. 2025From Perception to Cognition: A Survey of Vision-Language Interactive Reasoning in Multimodal Large Language Models
  5. 2026Why Vision Fails as a Universal Bridge: Rectifying Modality Asynchrony in Multilingual MLLMs
  6. 2025Look-Back: Implicit Visual Re-focusing in MLLM Reasoning
  7. 2026When VLMs Destroy What They Know: Spatial Representations Fall Below Lexical Priors
  8. 2026The Geometry of Representational Failures in Vision Language Models