The ability to make new substances is a key pillar of modern medicine and advanced materials. It helped move humanity from an era of survival to one of abundance through agrochemistry, gave us new ways of storing and using energy, and enabled much of the infrastructure we use to understand the universe. These advances are not trivial: they often rely on human intuition, heuristics, and serendipity, making computation an appealing route for advancing the discovery and synthesis of such materials. Yet many computational approaches ignore much of the complexity that makes chemistry both useful and challenging: they reduce chemical reactions to graphs and strings. In practice, decisions in chemistry are shaped by context, including synthetic goals, reaction mechanisms, failed experiments, preferences, strategic insight, and precedent. This thesis argues that natural language provides a powerful infrastructure for preserving, expanding, and operationalizing this chemical context. Large language models make this possible at scale.
The thesis first makes a historical argument, positioning computational chemistry across three categories of autonomy: from fully inert but extremely powerful chemical environments to systems with operational agency in laboratory settings. It then establishes the methodological foundations used throughout the work: LLM training and post-training, scaling, relevant inference techniques, and a short introduction to traditional retrosynthesis planning.
In the following chapters, the idea of context expansion in chemistry is explored in greater detail across several levels. First, context can be implicitly encoded in evaluation metrics, setting requirements or desiderata for models within a context- and use-case-aware framework. It can also be approached from the perspective of data: richer, contextually enriched data extracted from synthesis papers can recover reaction networks and contextual information that are often ignored in existing datasets.
With the goal of automating scientific experimentation, ChemCrow shows how LLMs can orchestrate chemical software, databases, literature search, and execution of robotic labs to achieve automated synthesis of small molecules. BOLLaMa illustrates how natural language interfaces can democratize access to optimization tools. In chemical reasoning, Synthegy uses LLMs as evaluators within grounded search procedures for strategy-aware synthesis planning and reaction mechanism elucidation. Finally, MiST studies how smaller open models can be prepared for chemical reasoning, showing that symbolic competence in scientific languages such as SMILES and CIF is a prerequisite for effective reasoning-oriented post-training.
These works support a view of LLMs as context-expansion engines for chemistry. Their value is not only in providing interfaces to other tools, but in connecting sparse molecular representations to the broader scientific situations in which chemists actually reason and act. T
