Science rules move from the prompt into decoding: layered constraints lift average accuracy 12% across three scientific tasks
Synopsis
Large language models still fabricate reasoning that contradicts domain rules in specialized scientific tasks. SciDC has a general model compile natural-language scientific knowledge into three layers of rules, top-level reasoning steps, middle-level conditional checks and bottom-level token masks, which a locally deployed domain model enforces during decoding. Across industrial formulation design, thyroid cancer TNM staging and single-step retrosynthesis, average accuracy rises 12% over vanilla generation, and ablations show removing the top or bottom layer sharply lowers formulation accuracy.
Interpretation
Scientific knowledge is embedded directly in the decoding process rather than appearing only as a prompt or a post-hoc filter. Prompt-based and tuning-based methods did not fully use scientific knowledge to guarantee logical correctness, and post-processing filters assume the model can produce useful candidates without knowledge guidance. Across industrial formulation design, clinical tumor diagnosis and retrosynthesis planning, average accuracy improves 12% over vanilla generation.
Rules are split by granularity into three layers governing the reasoning skeleton, conditional logic and token legality. Top-layer rules fix the macro order of reasoning steps, middle-layer rules check cross-token segments and trigger local regeneration on inconsistency, and bottom-layer rules rewrite decoder logits to zero out tokens outside the valid set. In the Qwen3-14B formulation ablation, the full method reaches 68.3% accuracy, dropping to 24.0% without top-layer rules, 25.3% without bottom-layer rules and 38.6% without middle-layer rules.
A general model compiles the knowledge once, and a locally deployable domain model does the actual generation. Rule compilation needs strong generalization, while downstream generation needs subject grounding and privacy protection, so the two demands are assigned to two models. With Claude-3.5-Sonnet as the knowledge compiler, Qwen3-4B formulation validity rises from 56.8% without knowledge to 71.0%, and Qwen3-14B from 18.4% to 75.5%.
Perspective
The demonstrated scope covers industrial formulation design, thyroid cancer TNM staging and single-step retrosynthesis, plus hearsay-evidence classification on LegalBench; all reported results come from the fully automated pipeline, with expert-in-the-loop verification used only as diagnostic analysis.
The authors note that tacit domain knowledge is hard to compile; in one formulation failure case the plasticizer ratio was raised without adjusting other components, pushing total mass fraction above 100%. A preliminary test of automatic scientific hypothesis generation did not reach satisfying results, and beyond expert review there is no other good verification method for generated hypotheses.