Integrating structural causal models into large language models: the PertAdapt framework

Abstract

To address the systemic flaw of current large-scale language models exhibiting logical deduction failures when faced with out-of-distribution test data and synthetic counterfactual interference, this paper proposes a PertAdapt causal adaptive framework integrating directed acyclic graph constraints. By constructing quantitative perturbation injection channels with orthogonal decoupling properties in the latent space and utilizing a loss residual closed-loop adjustment of local prediction edge weights with derivatives, the evolution of deep network variables is constrained to satisfy the algebraic consistency of causal probability theory. Quantitative experiments show that the system achieves an inference accuracy of 86.7% under high-order counterfactual interference, the OOD robustness index increases from 0.421 to 0.832, the system throughput remains at 16400 tokens/s, and the physical latency of single-time token generation is limited to within 485ms. The endto-end weight correction mechanism based on graph structure projection effectively truncates spurious statistical coupling responses between high-frequency random variables, achieving convergence and bounding enhancement of logical connectivity within the basic pre-training parameter space.