Coverage Before Control: Route-Instruction Grounding and Steering for Controllable Retrosynthesis

2026-09-30

AI Summary

This study addresses the challenge of balancing reaction pathway coverage and instruction controllability in retrosynthesis models by proposing the RIGS framework. Introducing a novel "coverage-before-control" paradigm, RIGS achieves instruction-guided retrosynthesis through a two-stage training procedure. First, a language projector learns preference projections to construct nested multi-support sets. Subsequently, with the generator frozen, lightweight instruction tuning is performed via residual adapters, effectively decoupling coverage capacity from control logic. Experimental results demonstrate that this approach significantly enhances both pathway diversity and instruction-following accuracy. Furthermore, the study reveals a non-monotonic relationship between model scale and coverage-control performance.

Abstract

Single-step retrosynthesis models are commonly evaluated by their ability to recover recorded reactions. In practice, chemists may need to choose among several precursor sets for the same product, for example to preserve a particular motif. Recovering a recorded answer alone does not establish this ability to follow a preference. Satisfying such requests requires both coverage of relevant alternatives and control over which alternatives are favored. We introduce Route-Instruction Grounding and Steering (RIGS), a two-stage framework for instruction-conditioned retrosynthesis. Stage A trains a language projector, teaching it which alternatives an instruction favors or discourages. Stage B uses the projector learned in Stage A to steer a frozen generative model through lightweight residual adapters. We construct nested one-to-many training supports by pairing each product with increasing numbers of candidate precursor sets. Extensive experiments demonstrate that broader support helps the model generate a wider range of alternatives, and RIGS can learn to guide generation according to instructions. The relationship between coverage and control is consistent across model scales but non-monotone.

Problem

• Retrosynthesis
• Controllability
• Coverage
• Instruction-conditioned generation
• Precursor selection

Innovation

• Controllable Retrosynthesis
• Instruction Conditioning
• Residual Adapters
• Language Projector
• Nested Training Supports