What Do VLMs NOTICE? A Mechanistic Interpretability Pipeline for Gaussian-Noise-free Text-Image Corruption and Evaluation (2025.naacl-long)
Copied to clipboard
| Challenge: | Vision-Language Models (VLMs) have gained prominence due to their success in solving complex cross-modal tasks. |
| Approach: | They propose a Gaussian-Noise-free pipeline for mechanistic interpretability in VLMs that introduces Semantic Image Pairs corruption, the first visual counterpart to Symmetric Token Replacement for text. |
| Outcome: | The proposed pipeline identifies a set of “universal attention heads” in BLIP and LLaVA that consistently contribute across different tasks and modalities. |
Similar Papers
Can VLMs Actually See and Read? A Survey on Modality Collapse in Vision-Language Models (2025.findings-acl)
Copied to clipboard
| Challenge: | Vision-language models integrate textual and visual information, enabling them to process visual inputs and generate predictions. |
| Approach: | They review work on modality collapse analysis to provide insights into the reason for this unintended behavior and review probing studies for fine-grained vision-language understanding. |
| Outcome: | The proposed models can achieve competitive performance in vision-language tasks despite relying heavily on textual information and ignoring visual information. |
Spectra: A Mechanistic Interpretability Library for Vision-Language Models (2026.acl-demo)
Copied to clipboard
| Challenge: | Existing interpretability tools for visionlanguage models are limited to activation probing and saving. |
| Approach: | They propose a library specifically designed for mechanistic interpretability of visionlanguage models that provides unified abstractions for activation patching, attention pattern analysis, and meta-functions across diverse VLM architectures. |
| Outcome: | The proposed library handles architecture-specific complexities while maintaining a simple, high-level interface. |
V-SEAM: Visual Semantic Editing and Attention Modulating for Causal Interpretability of Vision-Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing work on causal interpretability focuses on large language models (LLMs) but internal mechanisms of vision-language models remain underexplored, authors say . |
| Approach: | They introduce a framework that combines visual and semantic manipulations for causal interpretation of vision-language models. |
| Outcome: | The proposed framework shows improved performance for LLAVA and InstructBLIP on three diverse benchmarks. |
Back Attention: Understanding and Enhancing Multi-Hop Reasoning in Large Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Increasing the ability of large language models to perform latent multihop reasoning is crucial for reducing the cost and deployment challenges. |
| Approach: | They propose an interpretability method that traces how logits propagate across layers and positions toward the final prediction. |
| Outcome: | The proposed method improves accuracy on five reasoning datasets. |
RedundancyLens: Revealing and Exploiting Visual Token Processing Redundancy for Efficient Decoder-Only MLLMs (2025.findings-acl)
Copied to clipboard
| Challenge: | Current decoder-only architectures achieve higher performance but lower efficiency . cross-attention-based architectures skip visual token computations . |
| Approach: | They propose a training-free framework for analyzing trained MLLMs to investigate redundancy . they propose 'probe-activated Dynamic FFN and Hollow Attention' algorithms for visual token reductions and a layer ranking algorithm for inference acceleration. |
| Outcome: | The proposed framework achieves comparable performance to or better than state-of-the-art methods while remaining compatible with them. |
SmartTrim: Adaptive Tokens and Attention Pruning for Efficient Vision-Language Models (2024.lrec-main)
Copied to clipboard
Zekun Wang, Jingchang Chen, Wangchunshu Zhou, Haichao Zhu, Jiafeng Liang, Liping Shan, Ming Liu, Dongliang Xu, Qing Yang, Bing Qin
| Challenge: | Experimental results show that SmartTrim accelerates the original model by 2-3 times with minimal performance degradation. |
| Approach: | They propose an adaptive acceleration framework which prunes redundant token representations and attention heads within each layer of the original model. |
| Outcome: | The proposed framework accelerates the original model by 2-3 times with minimal performance degradation across vision-language tasks. |
AdaV: Adaptive Text-visual Redirection for Vision-Language Models (2025.findings-acl)
Copied to clipboard
| Challenge: | Vision-language models often generate excessive visual tokens, leading to poor performance . a novel training-free visual token pruning method is proposed to improve performance despite the computational cost associated with VLMs. |
| Approach: | They propose a training-free visual token pruning method that reduces biased token pruning . they plan to open-source the code upon publication . |
| Outcome: | The proposed method reduces biased token pruning and enhances model robustness with limited visual token budget. |
Mitigating Hallucinations in VLMs: Enhancing Visual Attention via Head-Wise Perturbation (2026.findings-acl)
Copied to clipboard
| Challenge: | Vision–Language Models (VLMs) have demonstrated strong capabilities in tasks that require joint understanding of text and images. |
| Approach: | They propose a strategy that incorporates head-wise attention perturbation via continuous multiplicative noise coupled with a visual-guided loss focused on vision-sensitive text tokens to promote a more balanced attention distribution. |
| Outcome: | The proposed approach outperforms baseline models on three benchmarks and consistently outperformed the baseline model. |
Efficient Inference for Large Vision-Language Models: Bottlenecks, Techniques, and Prospects (2026.findings-acl)
Copied to clipboard
Jun Zhang, Yicheng Ji, Feiyang Ren, Yihang Li, Bowen Zeng, Zonghao Chen, Ke Chen, Lidan Shou, Gang Chen, Huan Li
| Challenge: | Large Vision-Language Models are hindered by a systemic efficiency barrier known as visual token dominance. |
| Approach: | They propose a systematic taxonomy of efficiency techniques structured around the inference lifecycle . they examine visual encoding, prefilling, and decoding to understand bottlenecks . |
| Outcome: | The proposed techniques reveal how upstream decisions dictate downstream bottlenecks . the proposed techniques include hybrid compression and modality-aware decoding . |
S2H-DPO: Hardness-Aware Preference Optimization for Vision–Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods focus on localized reasoning with pre-specified image indices, bypassing the skills of global visual search and autonomous cross-image comparison. |
| Approach: | They propose a learning framework that constructs multi-image preference data across three hierarchical reasoning levels requiring an increasing level of capabilities. |
| Outcome: | The proposed approach maintains strong single-image reasoning performance while strengthening multi-image understanding capabilities. |