DALM: A Domain-Algebraic Language Model via Three-Phase Structured Generation

Researcher Chao Li has introduced DALM (Domain-Algebraic Language Model), a novel AI architecture designed to solve the core problem of knowledge contamination in large language models. Unlike standard LLMs that compress all information into a single parameter space—where facts from medicine, law, and history can interfere—DALM imposes a rigorous algebraic structure. It replaces unconstrained token-by-token generation with a three-phase process of structured denoising over a mathematically defined 'domain lattice.' This forces the model to first resolve which domain (e.g., chemistry, finance) a query belongs to, then the relevant relations within that domain, and finally the specific concepts.

The framework requires three core components: a lattice of domains with computable operations (meet, join, implication), a typing function for relations to control inheritance, and a 'fiber partition' that localizes knowledge to domain-specific subsets. This design confines generation to a chosen domain 'fiber,' structurally preventing cross-domain contamination in closed-vocabulary mode and making it auditably bounded in open-vocabulary mode. Li demonstrates the concept by instantiating DALM with the CDC knowledge representation system and outlines training on domain-annotated crystal libraries. The work fundamentally reframes language generation from unconstrained decoding to algebraically constrained, auditable reasoning.