Combatting AI Hallucinations: GSK’s Pioneering Efforts in Drug Development

GSK, a leading pharmaceutical company, is reshaping healthcare through its innovative integration of generative AI. This advanced technology has revealed potential in key areas like drug discovery, genomic analysis, and scientific literature review. However, GSK faces a major obstacle known as AI hallucinations, which can result in producing incorrect or fabricated information. This problem is particularly concerning in the healthcare sector, where accuracy is vital. In this article, we will delve into how GSK is tackling this pressing challenge.

The AI Hallucination Challenge in Generative Healthcare

The focus on minimizing AI hallucinations has traditionally concentrated on training large language models (LLMs), the time when the model learns from its training data. However, GSK has opted for a unique approach that also addresses inference-time strategies, which come into play when the model operates in real-world scenarios. Some effective strategies include:

• Self-reflection mechanisms
• Multi-model sampling techniques
• Iterative output evaluations

Kim Branson, SVP of AI and Machine Learning at GSK, emphasizes that these methods help make AI agents “robust and reliable.” This reliability is crucial for scientists aiming to generate actionable insights swiftly. As Branson puts it, “We’re all about increasing the iteration cycles at GSK—how we think faster.”

Boosting Performance with Test-Time Compute Scaling

Enhancing the performance of generative AI solutions at inference-time—often termed test-time—requires a boost in computational capabilities. This increase allows for more sophisticated operations, including:

• Iterative output refinement
• Multi-model aggregation

Branson highlights the notable influence that scaling test-time computation has on GSK’s AI ventures. By employing techniques like self-reflection and ensemble modeling, GSK achieves results that are not only quicker but also more accurate and trustworthy.

This trend transcends healthcare, illustrating a broader movement within the industry. Branson mentions, “You’re seeing this war happening with how much I can serve, my cost per token, and time per token.” Such competitive dynamics motivate developers to innovate algorithmic strategies previously deemed impractical, accelerating the incorporation of AI agents across multiple domains.

Proven Strategies to Mitigate AI Hallucinations

To substantially address AI hallucinations within generative AI applications, GSK employs two main strategies. Both require additional computation resources during inference.

1. Self-Reflection and Iterative Output Review Techniques

One essential technique is self-reflection. In this approach, LLMs assess and challenge their own outputs, contributing to quality enhancement. The model embarks on a step-by-step process of scrutinizing its initial responses, pinpointing areas for improvement, and refining its answers as needed. GSK’s literature search tool exemplifies this method, as it compiles data from diverse sources while using self-critique to enhance its findings.

This iterative review process results in clearer and more comprehensive final outputs. Branson emphasizes the importance of this methodology, stating, “If you can only afford to do one thing, do that.” This internal evaluation mechanism aligns results with the high accuracy standards required in healthcare environments.

2. Multi-Model Sampling Techniques

The second strategy is multi-model sampling. This technique employs multiple LLMs or various configurations of a single model to validate outputs. For example, the system might initiate the same query across different temperature settings to obtain a variety of responses. It could also leverage specialized tunings of the same model for distinct domains or diverse models trained on varying datasets.

This comparative analysis assists in affirming the most consistent conclusions. Branson explains, “You can get that effect of having different orthogonal ways to come to the same conclusion.” While this approach demands increased computational power, it significantly reduces AI hallucinations and enhances overall confidence in the final output—a vital aspect in high-stakes healthcare environments.

Navigating the Ongoing Inference Wars

The strategies employed by GSK necessitate a strong infrastructure capable of managing heavier computational demands. Branson describes a scenario he refers to as “inference wars,” wherein AI infrastructure companies strive to deliver advanced hardware to improve token throughput, reduce latency, and lower costs per token.

Specialized hardware and architectures facilitate complex inferencing tasks, such as multi-model sampling and iterative self-reflection, at scale. For instance, Cerebras technology allows for processing thousands of tokens per second, enabling the practical application of advanced techniques in real-world settings. Branson notes that these innovations play a direct role in enhancing GSK’s capacity to deploy generative models effectively within healthcare.

Challenges in Scaling AI Resources

Despite substantial progress, GSK continues to face hurdles in scaling its computational resources. Lengthy inference times can hinder workflows, especially when researchers or clinicians require rapid results. Here, advanced silicon solutions become crucial.

Moreover, increased computational demands may drive up costs, necessitating careful resource management. GSK regards these trade-offs as essential for achieving higher reliability and broader functionality. Branson explains, “As we enable more tools in the agent ecosystem, the system becomes more useful for people, and you end up with increased compute usage.” Striking a balance between performance, costs, and system capacities remains a critical focus for GSK’s strategic direction.

The Future of GSK’s AI-Driven Initiatives

GSK is steadfast in its commitment to advancing AI-powered healthcare solutions, with test-time compute scaling at the helm of its priorities. By combining self-reflection, multi-model sampling, and robust infrastructure, GSK ensures that generative models can meet the exacting demands of clinical environments.

This innovative approach serves as a benchmark for other organizations, showcasing how to harmonize accuracy, efficiency, and scalability effectively. By focusing on pioneering compute innovations and sophisticated inference strategies, GSK is not only addressing present challenges but also laying the groundwork for significant advancements in drug discovery, patient care, and other healthcare applications. 🚀