Research map / Vision-language and multimodal / Multimodal LLMs
Interpreting and Editing Vision-Language Representations to Mitigate Hallucinations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Projects VLM image-token latents to the vocabulary via logit lens to detect hallucinated objects, then erases them with a linear orthogonalization edit (ProjectAway), reducing hallucinations on COCO without training and enabling zero-shot segmentation.
Problem
VLMs hallucinate objects not present in images, and existing mitigation relies on external detectors or fine-tuning; the paper asks whether internal image representations can be interpreted and edited to reduce hallucinations.
Approach
Apply the logit lens to intermediate image-token representations to obtain an internal confidence per object, use it to separate hallucinated from present objects, and remove hallucinated objects by orthogonalizing image latents against the object's text embedding (ProjectAway), with no training.
Key results
- ProjectAway with mass removal reduces hallucinations by 25.7% on InstructBLIP-style setting on COCO (CHAIR_s 12.5 vs 14.6 beam search for InstructBLIP) and 23.8% for LLaVA vs beam search, while preserving CD object recall.
- Internal confidence improves object-hallucination classification mAP by 47.17% for InstructBLIP and 22.45% for LLaVA over the output-probability baseline.
- Zero-shot segmentation from logit-lens confidence on ImageNet-segmentation achieves mIoU 54.26 versus 39.27 for raw VLM attention, improving mAP by 8.03% over raw VLM attention.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use logit-lens internal confidence instead of maximum output-token probability to classify object presence | InstructBLIP · COCO train subset (5000 images) | mAP: 0.53 -> 0.78 | Helped |
| Use logit-lens internal confidence instead of maximum output-token probability to classify object presence | LLaVA · COCO train subset (5000 images) | mAP: 0.49 -> 0.60 | Helped |
| Detect low-confidence hallucinations and mass-remove them with ProjectAway | InstructBLIP · COCO2014 Karpathy validation | CHAIR_i: 53.4 -> 43.8 | Helped |
| Detect low-confidence hallucinations and mass-remove them with ProjectAway | LLaVA · COCO2014 Karpathy validation | CHAIR_i: 52.4 -> 42.0 | Helped |
| Use logit-lens internal confidence per patch rather than raw VLM attention for zero-shot segmentation | LLaVA · ImageNet-segmentation | mAP: 73.96 -> 79.90 | Helped |
| Orthogonalize image features against hallucinated-object features in-mass | InstructBLIP · COCO2014 training set | CHAIR_i: 15.0 -> 8.94 | Helped |
| Orthogonalize image features against hallucinated-object features in-mass | LLaVA · COCO2014 training set | CHAIR_i: 14.6 -> 11.2 | Helped |
| Orthogonalize image features against correctly detected object features in-mass | InstructBLIP · COCO2014 training set | CHAIR_i: 15.0 -> 17.3 | Hurt |
| Orthogonalize image features against correctly detected object features in-mass | LLaVA · COCO2014 training set | CHAIR_i: 14.6 -> 15.2 | Hurt |
| Edit at image layer 19 and extract text representation at layer 21 (ablation) | ProjectAway on LLaVA · COCO2014 samples with hallucinations (500 images) | Hallucination reduction: 25.7% | Helped |
| Edit at image layer 1 and extract text representation at layer 2 (ablation) | ProjectAway on InstructBLIP · COCO2014 samples with hallucinations (500 images) | Hallucination reduction: 38.5% | Helped |
| Increase the ProjectAway weight factor α (ablation) | ProjectAway · COCO2014 samples with hallucinations | Hallucination reduction | Helped |
Surprising findings
- Hallucinated objects are much easier to erase by linear editing than correctly detected objects (83.3% vs 16.2% individual removal rate for InstructBLIP).
- Mass removal of hallucinated objects slightly increases recall of correctly detected objects, possibly because removed hallucinations are replaced by objects the model is more confident in.
Limitations
- Multi-token objects introduce noise in confidence and in averaged editing embeddings.
- Editing may struggle with abstract or long sentences involving attributes or interactions; only object removal is assessed.
- Hyperparameters are chosen per model on small samples and the layer sweep is sensitive; many layer combinations erode correctly detected objects.
Experimental setup
Datasets: COCO 2014 ImageNet segmentation VQA 2.0 LLaVA-Bench
Benchmarks: COCO 2014 CHAIR ImageNet-segmentation COCO object presence classification
Metrics: CHAIR_i CHAIR_s mAP ROC AUC pixel accuracy mIoU recall hallucination reduction rate removal rate
Baselines: Greedy decoding Nucleus sampling Beam search OPERA raw attention (CLIP) TextSpan raw attention (VLM)
Builds on: logit lens OPERA CHAIR
Closest papers by meaning
- Reducing Hallucinations in Large Vision-Language Models via Latent Space Steering (ICLR 2025 · similarity 0.95)
- Image Token Matters: Mitigating Hallucination in Discrete Tokenizer-based Large Vision-Language Models via Latent Editing (NeurIPS 2025 · similarity 0.94)
- HaloProbe: Bayesian Detection and Mitigation of Object Hallucinations in Vision-Language Models (ICML 2026 · similarity 0.94)
- Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding (CVPR 2024 · similarity 0.94)
- Hallucination-aware Intermediate Representation Editing in Large Vision-Lanugage Models (ICLR 2026 · similarity 0.93)
- Mitigating Hallucinations in Large Vision-Language Models by Adaptively Constraining Information Flow (AAAI 2025 · similarity 0.93)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.