LLMs know a lot and sound convincing—yet in clinical diagnosis, chemical synthesis, or industrial formulation they still hallucinate, inventing reasoning that violates guidelines or even basic physics. Scientific rules guide human experts beautifully, so why can't a model actually use them?

Introducing SciDC, from @TsinghuaNLP (OpenBMB member) with Nanjing University and Northeastern University. It turns flexible scientific knowledge into decoding constraints—so a locally-deployed domain model can only generate within the feasible region defined by expert rules.

1⃣️ Knowledge becomes constraints, not prompts. A strong general LLM (GLLM) acts as a one-time knowledge compiler, parsing a natural-language knowledge document into standardized, executable rules; a smaller domain LLM (DLLM) then generates under those rules—keeping domain data local for privacy and cutting adaptation cost.

2⃣️ Three layers of control. Top-layer rules govern macro reasoning structure (enforce a step sequence, e.g. extract findings before diagnosing); middle-layer rules enforce conditional multi-hop logic with dynamic backtracking when inconsistencies appear; bottom-layer rules mask decoder logits at the token level to guarantee syntactic and numerical feasibility.

3⃣️ +12% on average, no tuning. Across industrial formulation design, clinical tumor diagnosis, and retrosynthesis, SciDC lifts overall accuracy by +11.6 for both Qwen3-4B (42.5→54.1) and Qwen3-14B (51.4→63.0)—with no parameter tuning or hand-crafted edits, just the knowledge document.

4⃣️Harder tasks, bigger gains. On real (not simulated) medical records Qwen3-14B jumps 33.5%→45.1% exact match, reaching 70%+ with light expert prompting; even the domain model ChemDFM improves retrosynthesis hit@1 from 16.2%→29.0%—showing rules and domain knowledge are synergistic, not redundant.

arXiv: https://arxiv.org/abs/2604.06603
Code: https://github.com/Maotian-Ma/SciDC
#AI #THUNLP #OpenBMB #LLM #AI4Science #Hallucination #NLP