Molecular Déjà Vu: Digit-Level Retrieval of Published Values in Frontier Language Models
Study finds widespread verbatim retrieval in LLMs on molecular regression benchmarks, affecting prediction accuracy.
Key Findings
Methodology
The study audits 22 frontier models on 12 regression benchmarks for verbatim retrieval, using two reasoning levels, finding that reasoning affects retrieval. Suppressing retrieval brings prediction errors closer in relative terms.
Key Results
- Over 50% of models show verbatim retrieval on five datasets.
- Retrieval flagged 89% more often at higher reasoning levels.
- Suppressing retrieval narrows prediction error gaps among models.
Significance
The study reveals memorization issues in LLMs for molecular property prediction, challenging model evaluation validity. It impacts academia and industry in model development and evaluation.
Technical Contribution
Introduces a new method to measure and suppress verbatim retrieval in LLMs, highlighting the need to consider memorization effects in model evaluation.
Novelty
First systematic revelation of verbatim retrieval in LLMs on molecular regression benchmarks, proposing methods to suppress retrieval.
Limitations
- Verbatim retrieval not fully eliminated on some benchmarks.
- Mechanism of reasoning level's influence on retrieval is unclear.
Future Work
Future research can explore more effective retrieval suppression methods and applications in other fields.
AI Executive Summary
Large language models (LLMs) excel in molecular property prediction, but their predictive capabilities may be compromised by verbatim retrieval. The study audits 22 frontier models on 12 regression benchmarks, finding widespread verbatim retrieval, especially on datasets like FreeSolv and ESOL. By suppressing retrieval, the study finds that prediction errors among models become closer in relative terms, indicating that predictive capability is not solely determined by memorized values. This study provides a new perspective on model evaluation, emphasizing the importance of considering memorization effects in model development. Future research can explore more effective retrieval suppression methods and applications in other fields.
Deep Analysis
Background
In recent years, large language models have made significant advances in natural language processing and molecular property prediction. However, their predictive capabilities may be affected by memorization of training data, challenging model evaluation.
Core Problem
Verbatim retrieval in LLMs for molecular property prediction may lead to inaccurate model evaluation, as models might be retrieving rather than predicting.
Innovation
The study systematically reveals verbatim retrieval in LLMs on molecular regression benchmarks for the first time and proposes methods to suppress retrieval.
Methodology
- �� Audit 22 frontier models on 12 regression benchmarks for verbatim retrieval
- �� Use two reasoning levels for experiments
- �� Test methods to suppress retrieval
Experiments
Experiments conducted on 22 models and 12 benchmarks, using two reasoning levels, testing the prevalence of verbatim retrieval and methods to suppress it.
Results
The study finds that over 50% of models show verbatim retrieval on five datasets. Suppressing retrieval narrows prediction error gaps among models.
Applications
The findings have significant implications for model evaluation and development, particularly in molecular property prediction and other fields.
Limitations & Outlook
While the study reveals verbatim retrieval, it is not fully eliminated on some benchmarks, and the mechanism of reasoning level's influence on retrieval is unclear.
Plain Language Accessible to non-experts
Imagine a factory where workers need to create products based on blueprints. Large language models are like these workers, but sometimes they just memorize the blueprints instead of understanding how to make the products. The study finds that models in molecular property prediction might just be retrieving known values rather than truly predicting. By adjusting the factory's workflow, the study attempts to reduce this memorization phenomenon, making the workers more focused on understanding and innovation.
ELI14 Explained like you're 14
Hey, friends! Did you know some super-smart computer models are like memory masters during exams? They can remember lots of answers! But sometimes, they just remember the answers without understanding the questions. Scientists found that these models in predicting molecular properties might just be retrieving known values rather than making real predictions. To make these models smarter, scientists are figuring out how to make them think more and remember less!
Glossary
Large Language Model (LLM)
A large-scale neural network model capable of processing and generating natural language.
Used for molecular property prediction.
Verbatim Retrieval
The phenomenon where models directly retrieve known data instead of predicting.
Found in molecular regression benchmarks.
Reasoning Level
The logical and computational complexity used by a model when generating answers.
Affects the model's verbatim retrieval phenomenon.
Memorization
The phenomenon where models remember training data during training.
Challenges the evaluation of model predictive capabilities.
Benchmark Contamination
The phenomenon where model evaluation is inaccurate due to overlap in training data.
Needs consideration in model evaluation.
Open Questions Unanswered questions from this research
- 1 How to effectively suppress verbatim retrieval in models?
- 2 What is the mechanism of reasoning level's influence on retrieval?
Applications
Immediate Applications
Molecular Property Prediction
Improve model prediction accuracy by reducing memorization effects.
Long-term Vision
Intelligent Model Evaluation
Develop smarter model evaluation methods to ensure true predictive capabilities.
Abstract
Large language models (LLMs) are increasingly evaluated on molecular property benchmarks, but accuracy cannot distinguish a model that predicts a property from one that retrieves a published number. We audit 22 frontier models on 12 regression benchmarks for verbatim retrieval and find that it is widespread but relatively benchmark-specific: on five datasets more than $50\%$ of the LLMs show verbatim retrieval, while on the remaining datasets it appears only in isolated cells. We run our experiments at two reasoning levels and find that reasoning changes retrieval. The same experiments, on the same molecules and with the same prompt, are flagged $89\%$ more often at the higher reasoning level than at the lowest one. Finally, we test a way to interrupt retrieval in our most contaminated cases, and find that the strongest models in some cases still recognise a combination of transformed SMILES strings and original labels. Furthermore, suppressing retrieval moves the prediction errors of the different models closer together in relative terms, while their differing use of verbatim retrieval spreads them apart. This indicates that the general predictive capability of an LLM is not determined solely by the amount of memorised values. This work provides an overview of the amount and depth of verbatim retrieval in molecular regression benchmarks using LLMs.