Research map / Vision-language and multimodal / Multimodal LLMs
The First to Know: How Token Distributions Reveal Hidden Knowledge in Large Vision-Language Models?
Qinyu Zhao, Ming Xu, Kartik Gupta, Akshay Asthana, Liang Zheng, Stephen Gould
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train a linear probe on first-token logits instead of using the LVLM's own response | LLaVA-7B (OE prompt) · VizWiz | AUC: 60.89 -> 90.05 | Improved |
| linear probing on first-token logits for jailbreak defense | LLaVA-7B (OE prompt) · MM-SafetyBench | ASR: 70.80 -> 7.00 | Improved |
| linear probing on first-token logits for deceptive questions | LLaVA-7B (OE prompt) · MAD-Bench | ASR: 80.00 -> 7.22 | Improved |
| use logits of subsequent tokens instead of the first token (ablation) | linear probing · VizWiz, MM-SafetyBench, MAD-Bench | AUC / ACC | Hurt |
| use different prompts (OE, Hint, Meta) with linear probing (ablation) | linear probing · three benchmarks | AUC / ACC | No clear effect |
| use hidden states of middle layers instead of logits (ablation) | linear probing · three benchmarks | AUC | No clear effect |
| probe CLIP text embeddings alone (ablation) | CLIP · MM-SafetyBench | AUC: 95.76 vs LLaVA LP 97.03 | Improved |
| increase probe training images from 100 to 1000 (ablation) | LLaVA linear probing on POPE adversarial · POPE adversarial | ACC: 82.77 -> 85.33 (original 83.67) | Improved |
| add LDA on first logits for classification | InstructBLIP · ImageNet-1k val | Top-1 accuracy: 55.10 -> 73.71 | Improved |
| finetune or retrain LLaVA with the same training data instead of linear probing | LLaVA · six tasks | task performance | Mixed |
Closest papers by meaning
- The Hidden Life of Tokens: Reducing Hallucination of Large Vision-Language Models Via Visual Information Steering (ICML 2025 · similarity 0.91)
- Aligning What Vision-Language Models See and Perceive with Adaptive Information Flow (CVPR 2026 · similarity 0.89)
- To Sink or Not to Sink: Visual Information Pathways in Large Vision-Language Models (ICLR 2026 · similarity 0.89)
- A Comprehensive Information-Decomposition Analysis of Large Vision-Language Models (ICLR 2026 · similarity 0.89)
- Understanding Language Prior of LVLMs by Contrasting Chain-of-Embedding (ICLR 2026 · similarity 0.89)
- Dismantling Pathological Shortcuts: A Causal Framework for Faithful LVLM Decoding (ICML 2026 · similarity 0.89)
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.
