Researcher demonstrates runtime behavioral modification of frozen language models using zero-initialized adapter layers, without retraining the base model

Andy Cufari added a 2.36M-parameter adapter layer (1.9% of GPT-2's 124M parameters) at layer 6 of a frozen 12-layer GPT-2 model, trained on 20 lines of text for 4 minutes. The adapter is initialized to exact zeros, so at α = 0 the model output is bit-identical to the original (max |Δ logits| = 0.0).

At α = 0.5 (the 'sweet spot'), perplexity increases 1.18× over baseline while the model generates contextual bias ("the cat is under mayonnaise") without degrading unrelated outputs. At α = 1.0, perplexity is 7.43× worse; at α = 1.5, fluency collapses to 39.3× worse.

The overlay is portable (16 MB file), reversible (set α = 0 to restore exact original behavior), tunable (α is a continuous dial set at inference time), and does not modify the base model weights. The remaining 6 downstream layers absorb and contextualize the mid-stack perturbation, whereas injection at the final layer produces 63× worse perplexity and degenerate outputs.

Code and trained adapter are available at https://github.com/andycufari/the-cat-is-under-mayonnaise-experiment under MIT license; implementation requires Python 3.10+, ~2 GB disk, optional GPU.