arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2408.11261cs.AIcs.CL
收藏

视觉-语言模型中的谄媚行为:系统分析与推理时缓解框架

Sycophancy in Vision-Language Models: A Systematic Analysis and an Inference-Time Mitigation Framework

School of Computer Science and Technology, University of Science and Technology of China(计算机科学与技术学院,中国科学技术大学);State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences(处理器国家重点实验室,中国科学院计算技术研究所);Department of Computer Science, National University of Singapore(计算机科学系,新加坡国立大学);University of Illinois Urbana-Champaign(伊利诺伊大学厄巴纳-香槟分校);Intelligent Software Research Center, Institute of Software, Chinese Academy of Sciences(软件智能研究中心,中国科学院软件研究所)

Yunpu Zhao, Rui Zhang, Junbin Xiao, Changxin Ke, Ruibo Hou, Yifan Hao, Ling Li

AI总结:

针对大型视觉-语言模型受引导性提示影响产生谄媚行为的问题,系统分析了多基准下的表现,并提出无需训练的推理时缓解框架,通过查询中和、对比解码和逻辑修正有效减少偏差,同时保持中性提示性能。

摘要:

大型视觉-语言模型(LVLMs)在视觉-语言理解方面展现出了显著的能力。然而,这些模型中持续存在的一个关键问题是谄媚行为,即模型受到引导性或欺骗性提示的不当影响,导致输出有偏和产生幻觉。尽管LVLMs发展迅速,但评估和缓解谄媚行为在很大程度上仍未得到充分探索。在这项工作中,我们通过系统分析多个视觉-语言基准上的谄媚行为来填补这一空白,并提出一个推理时缓解框架。我们策划了引导性查询,并量化了最先进的LVLMs对提示诱导偏差的敏感性,揭示了跨模型和任务的一致性能下降和不稳定性。我们的分析进一步揭示了模型特定的行为特征,例如在谄媚行为下的情感敏感性和预测极性偏移。为了缓解这些问题,我们提出了一个无需训练、模型无关的框架,该框架完全在推理时运行。我们的方法首先采用查询中和器,利用语言模型抑制用户查询中的隐性谄媚偏差。然后,我们引入一种谄媚感知的对比解码机制,通过对比对中和和引导查询的响应,动态重新校准词元级输出分布。最后,一个自适应逻辑修正模块通过整合自适应合理性过滤器和查询情感缩放器进一步修改对比后的逻辑值,确保生成连贯且稳健。大量实验表明,该框架在所有评估模型上有效缓解了谄媚行为,同时保持了对中性提示的性能。我们的结果表明,LVLMs中的谄媚行为是一个普遍且紧迫的挑战,推理时策略为可信的多模态推理提供了一条有前景的路径。

英文摘要:

Large Vision-Language Models (LVLMs) have shown significant capability in vision-language understanding. However, one critical issue that persists in these models is sycophancy, where models are unduly influenced by leading or deceptive prompts, resulting in biased outputs and hallucinations. Despite the rapid development of LVLMs, evaluating and mitigating sycophancy remains largely under-explored. In this work, we fill this gap by systematically analyzing sycophancy across multiple vision-language benchmarks and propose an inference-time mitigation framework. We curate leading queries and quantify the susceptibility of state-of-the-art LVLMs to prompt-induced bias, revealing consistent performance degradation and instability across models and tasks. Our analysis further uncovers model-specific behavioral traits, such as sentiment sensitivity and prediction polarity shifts under sycophancy. To mitigate these issues, we propose a training-free, model-agnostic framework that operates entirely at inference time. Our approach first employs a query neutralizer, leveraging an language model to suppress implicit sycophantic bias in user queries. We then introduce a sycophancy-aware contrastive decoding mechanism that dynamically recalibrates token-level output distributions by contrasting responses to neutralized and leading queries. Finally, an adaptive logits refinement module further modifies the contrasted logits by integrating both a adaptive plausibility filter and query sentiment scaler, ensuring coherent and robust generation. Extensive experiments demonstrate that this framework effectively mitigates sycophancy across all evaluated models, while maintaining performance on neutral prompts. Our results suggest that sycophancy in LVLMs is a general and urgent challenge, and that inference-time strategies offer a promising path toward trustworthy multimodal reasoning.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑