Debiasing Multimodal Large Language Models via Penalization of Language Priors

TL;DR

The paper identifies a bias in Multimodal Large Language Models where language priors dominate outputs even with irrelevant or absent visual input. It introduces two training-free debiasing strategies—Post-Hoc Debias (calibration) and Visual Debias Decoding (contrastive decoding with adaptive token filtering)—to reduce reliance on language priors and better ground responses in visual data. Through extensive experiments across multiple datasets and backbones, the methods improve truthfulness and reasoning while highlighting the significant influence of decoding configurations on reported performance. The work emphasizes the need for robust, vision-grounded evaluation and provides practical, low-cost techniques to improve MLLM reliability in real-world use and benchmarking.

Abstract

In the realms of computer vision and natural language processing, Multimodal Large Language Models (MLLMs) have become indispensable tools, proficient in generating textual responses based on visual inputs. Despite their advancements, our investigation reveals a noteworthy bias: the generated content is often driven more by the inherent priors of the underlying Large Language Models (LLMs) than by the input image. Empirical experiments underscore the persistence of this bias, as MLLMs often provide confident answers even in the absence of relevant images or given incongruent visual inputs. To rectify these biases and redirect the model's focus toward visual information, we propose two simple, training-free strategies. First, for tasks such as classification or multi-choice question answering, we introduce a "Post-Hoc Debias" method using an affine calibration step to adjust the output distribution. This approach ensures uniform answer scores when the image is absent, acting as an effective regularization technique to alleviate the influence of LLM priors. For more intricate open-ended generation tasks, we extend this method to "Visual Debias Decoding", which mitigates bias by contrasting token log-probabilities conditioned on a correct image versus a meaningless one. Additionally, our investigation sheds light on the instability of MLLMs across various decoding configurations. Through systematic exploration of different settings, we achieve significant performance improvements--surpassing previously reported results--and raise concerns about the fairness of current evaluation practices. Comprehensive experiments substantiate the effectiveness of our proposed strategies in mitigating biases. These strategies not only prove beneficial in minimizing hallucinations but also contribute to the generation of more helpful and precise illustrations.

Figures (12)

• Figure 1: MLLM tends to generate confident answers even when presented with nonsensical or irrelevant images, thereby revealing a pronounced bias towards the learned language patterns. "None" indicates the absence of an input image, while "Noise" signifies the presence of Gaussian noise matching the image dimensions. "Zeros" indicates the input is a tensor composed entirely of zero values.
• Figure 2: Top 15 answer choices and their probabilities.
• Figure 3: Illustration of the proposed Post-Hoc Debias and Visual Debias Decoding methods. The former focuses on debiasing the prediction results, while the latter modifies the next-token distribution.
• Figure 4: A simple adjustment of the temperature from 1.0 to 0.2 results in the successful generation.
• Figure 5: Attention maps for different input images. The question is "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions. USER: $<image>$ Is there a bed in the image? Please answer this question with one word. ASSISTANT:".