mlsanityData to 14 Oct 2026

Where MLLMs Attend and What They Rely On: Explaining Autoregressive Token Generation

arXiv 2509.22496 ↗First posted 26 Sep 2025cs.CVcited by 16

On the map: Multimodal LLMs / vision-language models → MLLM training and visual instruction tuning

Abstract (arXiv)

Multimodal large language models (MLLMs) have demonstrated remarkable capabilities in aligning visual inputs with natural language outputs. Yet, the extent to which generated tokens depend on visual modalities remains poorly understood, limiting interpretability and reliability. In this work, we present EAGLE, a lightweight black-box framework for explaining autoregressive token generation in MLLMs. EAGLE attributes any selected tokens to compact perceptual regions while quantifying the relative influence of language priors and perceptual evidence. The framework introduces an objective function that unifies sufficiency (insight score) and indispensability (necessity score), optimized via greedy search over sparsified image regions for faithful and efficient attribution. Beyond spatial attribution, EAGLE performs modality-aware analysis that disentangles what tokens rely on, providing fine-grained interpretability of model decisions. Extensive experiments across open-source MLLMs show that EAGLE consistently outperforms existing methods in faithfulness, localization, and hallucination diagnosis, while requiring substantially less GPU memory. These results highlight its effectiveness and practicality for advancing the interpretability of MLLMs.

Key facts. Cited by 16 papers in the map (data through 14 Oct 2026). Placed in MLLM training and visual instruction tuning, one of 16 directions in Multimodal LLMs / vision-language models. Builds on 30 of its 42 references that are in the map. First posted on arXiv 26 Sep 2025.

Builds on

30 of 42 references.

Closest ancestors · 7

Same direction, newest first

  1. 2025MLLMs Know Where to Look: Training-free Perception of Small Visual Details with Multimodal LLMs24 Feb
  2. 2025A Survey on Mechanistic Interpretability for Multi-Modal Foundation Models22 Feb
  3. 2024Explainable and Interpretable Multimodal Large Language Models: A Comprehensive Survey3 Dec
  4. 2024A Comprehensive Survey and Guide to Multimodal Large Language Models in Vision-Language Tasks9 Nov
  5. 2024Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution18 Sep
  6. 2024Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs11 Jan
  7. 2023Improved Baselines with Visual Instruction Tuning5 Oct

Other references · 23

  1. 2023GPT-4 Technical Reportcited by 18,396
  2. 2014Microsoft COCO: Common Objects in Contextcited by 17,608
  3. 2025Qwen2.5-VL Technical Reportcited by 5,587
  4. 2025Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilitiescited by 3,796
  5. 2025InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiencycited by 1,510
  6. 2015Microsoft COCO Captions: Data Collection and Evaluation Servercited by 1,462
  7. 2018RISE: Randomized Input Sampling for Explanation of Black-box Modelscited by 838
  8. 2023End-to-end Autonomous Driving: Challenges and Frontierscited by 490
  9. 2016Top-down Neural Attention by Excitation Backpropcited by 321
  10. 2019Smooth Grad-CAM++: An Enhanced Inference Level Visualization Technique for Deep Convolutional Neural Network Modelscited by 118
  11. 2023BadCLIP: Dual-Embedding Guided Backdoor Attack on Multimodal Contrastive Learningcited by 115
  12. 2024Interpreting and Editing Vision-Language Representations to Mitigate Hallucinationscited by 104
11 more
  1. 2024Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shiftcited by 49
  2. 2025Visual Large Language Models for Generalized and Specialized Applicationscited by 36
  3. 2025Token Activation Map to Visually Explain Multimodal LLMscited by 32
  4. 2016Guaranteed Non-convex Optimization: Submodular Maximization over Continuous Domainscited by 31
  5. 2024Interpreting Object-level Foundation Models via Visual Precision Searchcited by 20
  6. 2025Where do Large Vision-Language Models Look at when Answering Questions?cited by 15
  7. 2025SafeMobile: Chain-level Jailbreak Detection and Automated Evaluation for Multimodal Mobile Agentscited by 15
  8. 2025E2AT: Multimodal Jailbreak Defense via Dynamic Joint Optimization for Multimodal Large Language Modelscited by 12
  9. 2025DeFacto: Counterfactual Thinking with Images for Enforcing Evidence-Grounded and Faithful Reasoningcited by 10
  10. 2025RePOPE: Impact of Annotation Errors on the POPE Benchmarkcited by 9
  11. 2025Seeing It or Not? Interpretable Vision-aware Latent Steering to Mitigate Object Hallucinationscited by 3

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 2; 2026: 14.

16 papers.

Sort:

By area:

  1. 2025EgoVITA: Learning to Plan and Verify for Egocentric Video Reasoningcited by 4
  2. 2026Where Not to Learn: Prior-Aligned Training with Subset-based Attribution Constraintscited by 2
  3. 2026Can Attribution Predict Risk? From Multi-View Attribution to Planning Risk Signals in End-to-End Autonomous Drivingcited by 2
  4. 2025Did Models Learn Sufficiently? Attribution-Guided Training via Subset-Selected Counterfactual Augmentationcited by 1
  5. 2026Measuring Cross-Modal Synergy: A Benchmark for VLM Explainabilitycited by 1
  6. 2026PhaseWin: An Efficient Search Algorithm for Faithful Visual Attributioncited by 1
  7. 2026Which Modality Decides? Counterfactual Modality Attribution for Multimodal LLMscited by 1
  8. 2026Towards Reliable Medical LLMs: Benchmarking and Enhancing Confidence Estimation of Large Language Models in Medical Consultationcited by 0
  9. 2026LatentGeo: Learnable Auxiliary Constructions in Latent Space for Multimodal Geometric Reasoningcited by 0
  10. 2026PGT: Procedurally Generated Tasks for improving visual grounding in MLLMscited by 0
  11. 2026Attending to Multimodal Generation One Token at a Timecited by 0
  12. 2026A Good Initialization is All You Need for Faithful Visual Attributioncited by 0
4 more
  1. 2026Consistent Evidence, Robust Recognition: Faithful Attribution Regularization under Geometric Transformationscited by 0
  2. 2026IntentQA: Intent Question Answering in Videos by Cognitive Context Reasoningcited by 0
  3. 2026What do VLM-Based Vision-Language Navigation Models Rely on: Interpreting and Steering Policy Behaviorcited by 0
  4. 2026OmniConfess: Eliciting Token Confessions to Mitigate Omni-Modal Hallucinationcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Growing Visual Generative Capacity for Pre-Trained MLLMs
  2. 2024Accelerating Multimodal Large Language Models via Dynamic Visual-Token Exit and the Empirical Findings
  3. 2025VARGPT: Unified Understanding and Generation in a Visual Autoregressive Multimodal Large Language Model
  4. 2025Perceiving Beyond Language Priors: Enhancing Visual Comprehension and Attention in Multimodal Models
  5. 2026MLLM-Microscope: Unlocking Hidden Structure Within Multimodal Large Language Models
  6. 2026What Do Visual Tokens Really Encode? Uncovering Sparsity and Redundancy in Multimodal Large Language Models
  7. 2025Rethinking Visual Token Reduction in LVLMs Under Cross-Modal Misalignment
  8. 2026Seeing or Knowing? Visual Context Sensitivity in Multimodal Large Language Models