# An Eval for AI Scientists That Knows the Answer Key - Ian Quigley, Leash Bio **URL:** \ **Category:** Content Creators\ **Tags:** latent-space, agents, machine-learning, benchmarking\ **Created:** [26 October 2026 15:00 UTC](https://www.artofsm.art/t/an-eval-for-ai-scientists-that-knows-the-answer-key-ian-quigley-leash-bio/26887 "2026-10-26T15:00:07Z")\ **Posts on this page:** 2\ **Page:** 1
**Author:** ![merefield](https://www.artofsm.art/user_avatar/www.artofsm.art/merefield/32/74_2.png) [@merefield](https://www.artofsm.art/u/merefield)\ **Post date:** [26 October 2026 15:00 UTC](https://www.artofsm.art/t/an-eval-for-ai-scientists-that-knows-the-answer-key-ian-quigley-leash-bio/26887/1 "2026-10-26T15:00:07Z")
[![](https://www.artofsm.art/uploads/default/original/3X/1/0/10df04a52966ed6ace268ce6e9edc5bc6cb05052.jpeg "An Eval for AI Scientists That Knows the Answer Key - Ian Quigley, Leash Bio") ](https://www.youtube.com/watch?v=vhPCo-8K78U) Ian Quigley from Leash Bio discusses their development of a virtual sandbox environment where AI agents explore chemical spaces to optimize molecules for drug discovery, enabling extensive evaluation of AI decision-making in medicinal chemistry. They plan to release their datasets and tools publicly, aiming to advance AI-driven scientific reasoning and accelerate experimental discovery through scalable, low-latency lab-in-the-loop systems. ---
**Author:** ![merefield](https://www.artofsm.art/user_avatar/www.artofsm.art/merefield/32/74_2.png) [@merefield](https://www.artofsm.art/u/merefield)\ **Post date:** [26 October 2026 15:24 UTC](https://www.artofsm.art/t/an-eval-for-ai-scientists-that-knows-the-answer-key-ian-quigley-leash-bio/26887/3 "2026-10-26T15:24:47Z")
In this episode of Lab and Space Science, Brandon and RJ catch up with Ian Quigley from Leash Bio to discuss exciting advancements in using DNA encoded libraries for drug discovery. Ian explains how Leash Bio has developed a virtual sandbox environment where AI agents, such as language models, can explore vast chemical landscapes to solve structure-activity relationship (SAR) problems. These agents are tasked with optimizing molecules to bind better to target proteins within a limited budget, simulating the decision-making process of medicinal chemists. The environment is pre-measured, allowing for extensive rollouts—up to 90,000 so far—enabling detailed evaluation of agent performance. Ian shares that their recent preprint demonstrates how different AI models perform in this environment, revealing interesting behaviors such as when agents decide to abandon unproductive chemical neighborhoods or persist with suboptimal molecules. More advanced models tend to explore more effectively and avoid wasting resources, showing a form of scientific reasoning. The team plans to release the dataset, environment packages, and thousands of agent transcripts to the public, encouraging further analysis and development in scientific decision-making AI. The conversation also touches on the challenges of representing chemical structures in AI models, particularly using SMILES notation, a string-based language for molecules. Ian notes that earlier models struggled with this, but recent advances have enabled language models to internalize chemical knowledge and make iterative improvements to molecules. This progress opens the door to more sophisticated AI-driven chemistry tasks and highlights the potential for AI to learn and apply chemical intuition. Brandon and Ian discuss the possibility of turning this environment into a game-like interface for human medicinal chemists, allowing them to interact with the chemical space in a constrained but meaningful way. Ian acknowledges the idea and mentions the need to simplify the environment for practical human use, given the complexity and pre-measured nature of the chemical libraries. They also compare their approach to other lab-in-the-loop systems, emphasizing the advantage of having a large, pre-measured dataset that enables near-instant feedback and extensive experimentation without the delays of physical lab work. Finally, Ian outlines future plans to expand the number of environments and explore other biological and scientific problems using this framework. The goal is to create scalable, low-latency lab-in-the-loop systems that can accelerate scientific discovery and improve experimental decision-making. While it remains to be seen how well reinforcement learning progress in these simulated environments will transfer to new biological challenges, the team is optimistic and actively working on integrating structural tools and decision-making frameworks to enhance AI-driven science. ## Useful Links - [Minos preprint paper by Leash Labs](https://github.com/Leash-Labs/minos-paper/blob/trunk/Minos_Leash.pdf) — Directly substantiates and explains the main AI evaluation framework and experimental results discussed in the video.