Interpreting and Editing Vision-Language Representations to Mitigate Hallucinations

2024/10/03 by Jiang, Nick, Anish Kachinthaya, Suzie Petryk +4 · 25 citations
Arts and Humanities · Psychology · #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Mental Health and Psychiatry #Psychedelics and Drug Studies

paper · pdf · doi:10.48550/arxiv.2410.02762

openalex publication_date 2024/10/03 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

We investigate the internal representations of vision-language models (VLMs) to address hallucinations, a persistent challenge despite advances in model size and training. We project VLMs' internal image representations to their language vocabulary and observe more confident output probabilities on real objects than hallucinated objects. We additionally use these output probabilities to spatially localize real objects. Building on this approach, we introduce a knowledge erasure algorithm that removes hallucinations by linearly orthogonalizing image features with respect to hallucinated object features. We show that targeted edits to a model's latent representations can reduce hallucinations by up to 25.7% on the COCO2014 dataset while preserving performance. Our findings demonstrate how a deeper understanding of VLMs' latent representations can enhance reliability and enable novel capabilities, such as zero-shot segmentation.

Cited by

• Revealing Perception and Generation Dynamics in LVLMs: Mitigating Hallucinations via Validated Dominance Correction
• Journey Before Destination: On the importance of Visual Faithfulness in Slow Thinking
• SAVE: Sparse Autoencoder-Driven Visual Information Enhancement for Mitigating Object Hallucination
• Too Late to Recall: Explaining the Two-Hop Problem in Multimodal Knowledge Retrieval
• ESMC: MLLM-Based Embedding Selection for Explainable Multiple Clustering
• Boosting Reasoning in Large Multimodal Models via Activation Replay
• Tell Model Where to Look: Mitigating Hallucinations in MLLMs by Vision-Guided Attention
• TRANSPORTER: Transferring Visual Semantics from VLM Manifolds
• PAS : Prelim Attention Score for Detecting Object Hallucinations in Large Vision--Language Models
• Prior Directions: Why GUI Grounding Gets Locked in the Past
• ChainMPQ: Interleaved Text-Image Reasoning Chains for Mitigating Relation Hallucinations
• Visual Representations inside the Language Model
• Zero-Shot Fine-Grained Image Classification Using Large Vision-Language Models
• From Perception to Cognition: A Survey of Vision-Language Interactive Reasoning in Multimodal Large Language Models
• Uncovering Grounding IDs: How External Cues Shape Multimodal Binding
• Where MLLMs Attend and What They Rely On: Explaining Autoregressive Token Generation
• Explaining multimodal LLMs via intra-modal token interactions
• Exposing Hallucinations To Suppress Them: VLMs Representation Editing With Generative Anchors
• SAEmnesia: Erasing Concepts in Diffusion Models with Supervised Sparse Autoencoders
• Reading Images Like Texts: Sequential Image Understanding in Vision-Language Models
• Interpreting Attention Heads for Image-to-Text Information Flow in Large Vision-Language Models
• GLSim: Detecting Object Hallucinations in LVLMs via Global-Local Similarity
• Learning to Steer: Input-dependent Steering for Multimodal LLMs
• CircuitProbe: Dissecting Spatiotemporal Visual Semantics with Circuit Tracing
• A Survey of Multimodal Hallucination Evaluation and Detection

Related

• Reducing Hallucinations in Vision-Language Models via Latent Space Steering
• Investigating and Mitigating Object Hallucinations in Pretrained Vision-Language (CLIP) Models
• Watch Closely: Mitigating Object Hallucinations in Large Vision-Language Models with Disentangled Decoding
• Plausible May Not Be Faithful: Probing Object Hallucination in Vision-Language Pre-training
• VCE: A Zero-Cost Hallucination Mitigation Method of LVLMs via Visual Contrastive Editing