Hallucinations Improve LLMs in Drug Discovery

Analysis

Why This Paper Matters

This paper challenges the prevailing assumption that hallucinations in large language models (LLMs) are always detrimental. By focusing on drug discovery, a high-stakes domain where accuracy is critical, the authors demonstrate that hallucinated text can actually enhance predictive performance in molecule property prediction. This is a paradigm shift: instead of solely trying to eliminate hallucinations, researchers might consider harnessing them as a creative or exploratory signal.

The study is timely given the rapid adoption of LLMs in scientific research. If hallucinations can be systematically identified and leveraged, they could become a valuable tool for hypothesis generation and feature enrichment. The paper's findings also have implications for how we evaluate LLMs—perhaps not all deviations from ground truth are bad, and some may even be beneficial for downstream tasks.

Technical Contributions

• Novel use of hallucinations: The paper is among the first to systematically incorporate hallucinated text into a scientific modeling pipeline, specifically for molecule property prediction.
• Comprehensive evaluation: Seven instruction-tuned LLMs are evaluated across five datasets, providing a broad comparison of how different models' hallucinations affect performance.
• Categorization of beneficial hallucinations: Over 18,000 beneficial hallucinations are identified and categorized, with structural misdescriptions emerging as the most impactful type. This provides a taxonomy that could guide future research.
• Ablation studies: The authors investigate the effects of model size and temperature, showing that larger models benefit more from hallucinations while temperature has limited impact, offering practical insights for deployment.

Results

• Falcon3-Mamba-7B outperforms all baselines when hallucinated text is included, indicating that even smaller models can benefit significantly.
• GPT-4o-generated hallucinations consistently yield the greatest gains across models, suggesting that higher-quality hallucinations (perhaps more diverse or plausible) are more beneficial.
• The identification of over 18,000 beneficial hallucinations, with structural misdescriptions being most impactful, suggests that certain types of hallucinations can increase model confidence and accuracy.
• Ablation studies reveal a positive correlation between model size and benefit from hallucinations, while temperature has a limited effect, implying that the phenomenon is robust to sampling randomness.

Significance

This paper has the potential to influence how the AI community views hallucinations. Rather than being a flaw to be eliminated, hallucinations might be repurposed as a form of data augmentation or feature engineering. In drug discovery, where molecular property prediction is a bottleneck, this approach could accelerate early-stage screening by improving model accuracy without additional experimental data.

Moreover, the findings could extend beyond drug discovery to other scientific domains where LLMs are used to interpret structured data (e.g., chemical, biological, or materials science). The categorization of beneficial hallucinations provides a foundation for developing methods to deliberately generate or select such outputs. Ultimately, this work encourages a more nuanced understanding of LLM behavior and opens new avenues for leveraging their generative capabilities in high-stakes applications.