Computer-aided retrosynthesis has progressed through single-step prediction, tree search, and route scoring, according to the new arXiv preprint. But the authors argue that complex targets still require judgement over the entire route, not just local choices. ReTRACE is their proposed system: an LLM agent that operates inside a chemistry harness and uses feedback to guide its planning.
The paper identifies specific difficulties that such judgement must handle, including stereochemistry, functional-group compatibility, and cost. These factors interact across steps, making them hard to capture with stepwise scoring alone. By framing retrosynthesis as a feedback-driven agent task, ReTRACE aims to integrate these considerations into a coherent planning process.
The abstract does not provide experimental results or comparisons, so the claims are limited to the system's design and motivation. Still, the work points toward a shift from purely predictive or search-based tools toward agents that can reason about entire synthetic routes.