Déjà Vu in Molecules: Auditing Verbatim Retrieval Behavior of Frontier Language Models in Regression Benchmarks

Background and Context

The integration of Large Language Models (LLMs) into scientific discovery has accelerated the adoption of these systems for complex regression tasks, particularly in molecular property prediction. However, a fundamental blind spot exists within current evaluation frameworks: standard accuracy metrics fail to distinguish between genuine physical-chemical reasoning and the mechanical retrieval of published values from training corpora. This study addresses this ambiguity by systematically auditing the behavior of 22 frontier language models across 12 distinct molecular regression benchmarks. The primary objective is to determine whether high performance stems from an understanding of structure-property relationships or from memorizing data points present in public databases.

The research highlights a pervasive yet隐蔽 phenomenon termed "verbatim retrieval," where models generate answers by directly copying known numerical values rather than deriving them through logical inference. This behavior poses a significant risk to the reliability of AI-driven scientific tools, as it suggests that apparent generalization capabilities may be illusory. By introducing a rigorous auditing mechanism, the study aims to isolate the effects of memorization, thereby revealing the true cognitive level of these models in scientific computing. This work challenges the validity of existing benchmarks and compels the community to reconsider how experimental designs can verify causal reasoning capabilities in artificial intelligence systems.

Deep Analysis

To quantify verbatim retrieval, the research employed a multi-dimensional auditing strategy involving a test suite of 12 molecular regression benchmarks ranging from simple to complex property predictions. The methodology included two distinct levels of reasoning depth, allowing researchers to compare model outputs under identical molecular structures and prompts but varying levels of cognitive demand. The core technical criterion for identifying retrieval was the detection of exact matches between model outputs and known values in the training data. Additionally, the study utilized transformed SMILES strings to disrupt potential string-matching mechanisms, testing whether models relied on semantic understanding or superficial pattern recognition.

The experimental results revealed that verbatim retrieval is highly benchmark-specific rather than uniformly distributed. In five specific datasets, more than 50% of the models exhibited clear characteristics of verbatim retrieval, indicating that certain benchmarks are heavily covered by training data, thus triggering memory effects. Crucially, the study found that increasing reasoning complexity exacerbates this phenomenon. When switching from basic to advanced reasoning modes, the rate of tokens marked as verbatim retrieval increased by 89%. This data strongly suggests that deeper reasoning processes may activate more memory retrieval pathways, leading to a greater dependence on pre-stored knowledge rather than facilitating logical deduction.

Furthermore, the analysis demonstrated that even when attempts were made to disrupt retrieval through input transformation, the strongest models could still recognize combinations of transformed SMILES strings and original labels. However, when retrieval behavior was suppressed, the relative error differences between models became significantly closer. This implies that while memorization creates the illusion of superior performance, the underlying general predictive capabilities of the models are not as divergent as accuracy metrics suggest. The study confirms that general predictive capability is not solely determined by memorizing values, offering a crucial perspective for assessing the true capabilities of LLMs in scientific computing.

Industry Impact

The findings have profound implications for the open-source community and industrial applications in high-stakes fields such as drug discovery and materials science. The study warns against relying solely on accuracy metrics for evaluating scientific AI models, advocating for the inclusion of auditing mechanisms to detect memory effects and prevent "false progress." In industrial settings, models that depend on memorization rather than reasoning can lead to severe prediction biases, necessitating the development of more robust evaluation protocols that account for these vulnerabilities.

For the broader AI community, this research provides a new direction for designing memory-resistant benchmarks and developing architectural improvements to suppress unnecessary retrieval behaviors. By understanding the mechanisms behind this "déjà vu" phenomenon, developers can create more transparent and trustworthy scientific AI systems. The study emphasizes that ensuring the reliability and interpretability of AI in solving complex scientific problems requires a shift from pattern-matching experts to systems capable of genuine causal inference.

Outlook

Looking forward, this audit serves as a critical baseline for the future development of AI for Science. The identification of verbatim retrieval as a dominant behavior in frontier models underscores the need for new training strategies that prioritize reasoning over memorization. Future research should focus on balancing memory and reasoning capabilities, ensuring that models can generalize to novel molecular structures without relying on training data overlap.

The study also highlights the importance of creating diverse and challenging benchmarks that are less susceptible to data contamination. By addressing the limitations of current evaluation frameworks, the scientific community can build more reliable tools for molecular property prediction. Ultimately, this work lays the foundation for a new generation of scientific AI systems that offer not just accurate predictions, but also explainable and robust reasoning capabilities, driving the field toward deeper and more meaningful scientific discoveries.