The Perplexity Problem: How AI-Based Search Systems Favor Texts with Low Perplexity

The rapid development of large language models (LLMs) has revolutionized information retrieval. AI-based search systems based on these models offer new possibilities for precise and efficient search results. However, current research shows that these systems have an unexpected weakness: they tend to assign higher relevance scores to texts with low perplexity, regardless of their actual content quality. This phenomenon, often referred to as "Source Bias," can jeopardize the sustainable development of the information access ecosystem.

The Causes of the Perplexity Problem

A new study investigates the causes of this "perplexity problem." The researchers explain the information retrieval process using a causal graph and show that AI-based search systems learn perplexity features for relevance assessment. Simply put: the easier it is for the model to "understand" a text, the higher it is rated. This leads to texts with low perplexity, often generated by LLMs, being preferred, even if they are less relevant in terms of content than texts written by humans.

Theoretical analyses show that this phenomenon is due to the positive correlation between the gradients of the loss functions in the language modeling and information retrieval process. This means that optimizing the language model simultaneously reinforces perplexity as a relevance criterion.

Causal Diagnosis and Correction (CDC): A Solution

The study proposes a new method to address this problem: Causal Diagnosis and Correction (CDC). CDC first diagnoses the influence of perplexity on the relevance assessment and then separates this influence from the overall score. This approach makes it possible to reduce the "Source Bias" and make the search results more objective.

Impact on Practice

The results of the study show that CDC significantly improves the effectiveness of information retrieval. The preference for machine-generated texts is reduced, and the search results reflect the actual relevance of the documents. This is an important step towards ensuring the quality and reliability of AI-based search systems.

Future Developments

Research in this area is still ongoing. Further studies are necessary to fully understand the complex relationships between perplexity and relevance assessment and to develop more effective solutions for the perplexity problem. The development of robust and unbiased AI-based search systems is crucial for the future of information retrieval.

For companies like Mindverse, which specialize in the development of AI solutions, these findings are of particular importance. Integrating methods like CDC into AI-based search engines, chatbots, and other applications can help improve the quality and reliability of these systems and strengthen user trust.

Bibliography: - Wang, H., Dai, S., Zhao, H., Pang, L., Zhang, X., Wang, G., Dong, Z., Xu, J., & Wen, J.-R. (2025). Perplexity Trap: PLM-Based Retrievers Overrate Low Perplexity Documents. arXiv preprint arXiv:2503.08684. - https://openreview.net/forum?id=U1T6sq12uj - https://arxiv.org/abs/2503.08684 - https://arxiv.org/html/2503.08684v1 - https://openreview.net/pdf/620b2e897260f5a9498f330cd8ad879fadd948a9.pdf - https://x.com/_reachsumit/status/1899665309327643016 - http://paperreading.club/page?id=291091 - https://www.researchgate.net/publication/389708020_Unifying_Bias_and_Unfairness_in_Information_Retrieval_New_Challenges_in_the_LLM_Era - https://chatpaper.com/chatpaper/?id=3&date=1741708800&page=1 - https://x.com/gm8xx8/status/1899648515279569350 - https://huggingface.co/papers