Evo-ERBF-LoRA: Evolutionary Exact Radial Basis Function Low-Rank Adaptation for Parametric Memory Injection in Autoregressive Transformer Models

We introduce Evo-ERBF-LoRA, a training-free, gradient-free framework for injecting parametric factual memory into frozen autoregressive transformer models. The method combines Exact Radial Basis Function (ERBF) interpolation, which guarantees perfect reconstruction of training targets, with a confidence-gated multi-layer injection mechanism whose hyperparameters are discovered automatically by a genetic algorithm (GA). Unlike conventional Low-Rank Adaptation (LoRA), which requires backpropa-gation through the model and modifies weight matrices, Evo-ERBF-LoRA operates entirely through forward hooks on MLP sublayers, leaving model weights unchanged. The genetic algorithm evolves a chromosome that jointly encodes which transformer layers to modify , the per-layer injection strength α, the per-fact gate radius σ, and the subset of generation-trajectory tokens used as anchor points. Evaluated on GPT-2 with three fictional knowledge facts, the system achieves up to 100% weighted recall on individual facts and 0.000 KL divergence on unrelated control queries in the best evolutionary configuration, demonstrating that precise non-destructive memory injection is achievable without any parameter updates.