Intermediate representations unlock LLM retrosynthesis where end-to-end fails

The paper's core claim is narrower than it appears: representation alignment failures are identifiable and fixable through task decomposition, but this only works when you can define intermediate abstractions. The finding doesn't prove LLMs are capable at retrosynthesis generally, only that the bottleneck isn't model size.

This connects directly to the evidence-value misalignment work from late September, which found that models reach correct outputs despite weak reasoning. Here we see the inverse: models fail not from incapacity but from being forced to translate across incompatible symbolic domains in one pass. Both papers suggest that accuracy metrics mask structural failures in how models handle multi-step symbolic reasoning. The QuanReview piece on annotation alignment also touches this theme, though in a different domain (human-model reconciliation rather than intra-model representation). The key insight across these stories is that LLM failures often stem from architectural choices about how information flows, not raw capability gaps.

If teams applying this decomposition approach to other symbolic domains (formal verification, constraint satisfaction, code synthesis) report similar accuracy gains within the next 6 months, that confirms the finding generalizes. If instead gains remain chemistry-specific, the result is more about domain-specific prompt engineering than a fundamental principle about LLM reasoning.