kNNProxy: Efficient training-free proxy alignment for black-box zero-shot LLM-generated text detection

kNNProxy is a training-free, query-efficient proxy alignment framework for black-box zero-shot detection of LLM-generated text. It uses the kNN-LM retrieval mechanism as a domain adapter for a fixed proxy LLM: a lightweight datastore is built once from the target corpus, and at inference nearest-neighbor evidence is interpolated with the proxy output, without fine-tuning the proxy or making per-token API calls. Its extension MoP routes inputs by domain to improve robustness to domain shift.