Research map / Vision-language and multimodal / Multimodal LLMs
PopCD: Test-Time Behavior Enhancement via Polarity-Prompt Contrastive Decoding
Baolong Bi, Yuyao Ge, Shenghua Liu, Yuchen He, S.L. Tong, Lizhe Chen, Lingrui Mei, Zehao Li, Yiwei Wang, Yujun Cai, Ming-Hsuan Yang, Xueqi Cheng
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 |
|---|---|---|---|
| add the Adaptive Plausibility Constraint to PromptCD (ablation) | PromptCD (LLaMA2-7B) · NQ | Hit Rate: 30.7 -> 77.8 | Improved |
| add the Adaptive Plausibility Constraint to PromptCD (ablation) | PromptCD (Qwen2.5-7B) · NQ | Hit Rate: 62.4 -> 89.8 | Improved |
| set contrastive coefficient gamma to 0.5 instead of 0.2 (ablation) | PromptCD (LLaMA2-7B-chat) · Helpfulness ConR | ConR: 77.32 -> 78.89 | Improved |
| increase gamma from 0.5 to 0.8 (ablation) | PromptCD (LLaMA2-7B-chat) · Helpfulness ConR | ConR: 78.89 -> 70.53 | Hurt |
| increase gamma from 0.5 to 0.8 (ablation) | PromptCD (LLaMA2-7B-chat) · Harmlessness DG-AVG | DG-AVG: 89.98 -> 85.34 | Hurt |
| apply contrast on layers [20,25] instead of [10,15] (ablation) | PromptCD (LLaVA-1.5-7B) · TextVQA | accuracy: 49.2 -> 58.2 | Improved |
| apply PromptCD contrastive decoding instead of vanilla decoding | LLaMA2-7B-chat · NQ | ConR: 43.35 -> 75.99 | Improved |
| apply PromptCD instead of vanilla decoding | LLaMA2-7B-chat · SafeEdit | Defense Success: 40.59 -> 96.67 | Improved |
| apply PromptCD instead of DINM | LLaMA3-8B-instruct · SafeEdit | Defense Success: 99.26 -> 95.26 | Hurt |
| apply PromptCD instead of vanilla decoding | LLaVA-1.5-13B · V* | accuracy: 42.4 -> 70.0 | Improved |
| apply PromptCD on top of MeLLo | LLaMA2-7B-chat · MQuAKE-3k | accuracy: 32.6 -> 43.1 | Improved |
| apply PromptCD on top of IKE | LLaMA2-7B-chat · MQuAKE-2002 | accuracy: 20.6 -> 20.4 | No clear effect |
Closest papers by meaning
- ASCD: Attention-Steerable Contrastive Decoding for Reducing Hallucination in MLLM (AAAI 2026 · similarity 0.88)
- Decoupling Contrastive Decoding: Robust Hallucination Mitigation in Multimodal Large Language Models (NeurIPS 2025 · similarity 0.87)
- The Mirage of Performance Gains: Why Contrastive Decoding Fails to Mitigate Object Hallucinations in MLLMs? (NeurIPS 2025 · similarity 0.87)
- Breaking the Illusion: When Positive Meets Negative in Multimodal Decoding (CVPR 2026 · similarity 0.86)
- Calibrated Self-Rewarding Vision Language Models (NeurIPS 2024 · similarity 0.86)
- Aligning Visual Contrastive learning models via Preference Optimization (ICLR 2025 · similarity 0.86)
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.
