Academic

PromptCD: Test-Time Behavior Enhancement via Polarity-Prompt Contrastive Decoding

arXiv:2602.20696v1 Announce Type: new Abstract: Reliable AI systems require large language models (LLMs) to exhibit behaviors aligned with human preferences and values. However, most existing alignment approaches operate at training time and rely on additional high-quality data, incurring significant computational and annotation costs. While recent work has shown that contrastive decoding can leverage a model's internal distributions to improve specific capabilities, its applicability remains limited to narrow behavioral scopes and scenarios. In this work, we introduce Polarity-Prompt Contrastive Decoding (PromptCD), a test-time behavior control method that generalizes contrastive decoding to broader enhancement settings. PromptCD constructs paired positive and negative guiding prompts for a target behavior and contrasts model responses-specifically token-level probability distributions in LLMs and visual attention patterns in VLMs-to reinforce desirable outcomes. This formulation ext

B
Baolong Bi, Yuyao Ge, Shenghua Liu, Yuchen He, Siqian Tong, Lizhe Chen, Lingrui Mei, Zehao Li, Yiwei Wang, Yujun Cai, Ming-Hsuan Yang, Xueqi Cheng · · 1 min read · 10 views
#cs.AI

arXiv:2602.20696v1 Announce Type: new Abstract: Reliable AI systems require large language models (LLMs) to exhibit behaviors aligned with human preferences and values. However, most existing alignment approaches operate at training time and rely on additional high-quality data, incurring significant computational and annotation costs. While recent work has shown that contrastive decoding can leverage a model's internal distributions to improve specific capabilities, its applicability remains limited to narrow behavioral scopes and scenarios. In this work, we introduce Polarity-Prompt Contrastive Decoding (PromptCD), a test-time behavior control method that generalizes contrastive decoding to broader enhancement settings. PromptCD constructs paired positive and negative guiding prompts for a target behavior and contrasts model responses-specifically token-level probability distributions in LLMs and visual attention patterns in VLMs-to reinforce desirable outcomes. This formulation extends contrastive decoding to a wide range of enhancement objectives and is applicable to both LLMs and Vision-Language Models (VLMs) without additional training. For LLMs, experiments on the "3H" alignment objectives (helpfulness, honesty, and harmlessness) demonstrate consistent and substantial improvements, indicating that post-trained models can achieve meaningful self-enhancement purely at test time. For VLMs, we further analyze contrastive effects on visual attention, showing that PromptCD significantly improves VQA performance by reinforcing behavior-consistent visual grounding. Collectively, these results highlight PromptCD as a simple, general, and cost-efficient strategy for reliable behavior control across modalities.

Executive Summary

The article introduces Polarity-Prompt Contrastive Decoding (PromptCD), a test-time behavior control method that enhances the reliability of AI systems by leveraging a model's internal distributions. PromptCD constructs paired positive and negative guiding prompts to reinforce desirable outcomes, applicable to both large language models (LLMs) and Vision-Language Models (VLMs) without additional training. Experimental results demonstrate substantial improvements in alignment objectives and VQA performance, indicating PromptCD's potential as a simple, general, and cost-efficient strategy for behavior control across modalities. While the approach shows promise, its limitations and potential applications warrant further exploration.

Key Points

  • ▸ PromptCD is a test-time behavior control method that generalizes contrastive decoding to broader enhancement settings.
  • ▸ PromptCD constructs paired positive and negative guiding prompts to reinforce desirable outcomes.
  • ▸ PromptCD is applicable to both LLMs and VLMs without additional training.

Merits

Strength

PromptCD offers a simple and cost-efficient approach to behavior control, leveraging a model's internal distributions without requiring additional training data or computational resources.

Demerits

Limitation

The article focuses primarily on narrow behavioral scopes and scenarios, limiting the generalizability of PromptCD to broader real-world applications.

Limitation

The approach relies on the availability of paired positive and negative guiding prompts, which may not always be feasible or practical in certain scenarios.

Expert Commentary

The article makes a significant contribution to the field of AI behavior control, introducing a novel approach that leverages a model's internal distributions to enhance reliability. While PromptCD shows promise, its limitations and potential applications warrant further exploration. The approach's reliance on paired positive and negative guiding prompts raises questions about its generalizability and practical feasibility. Nevertheless, PromptCD's potential to enhance the reliability of AI systems is significant, and its development has important implications for industries and policymakers.

Recommendations

  • ✓ Further research is needed to explore the generalizability and scalability of PromptCD across different AI applications and domains.
  • ✓ Developing methods to generate paired positive and negative guiding prompts in a more efficient and practical manner is essential for widespread adoption of PromptCD.

Sources

Related Articles