04:00
2026-07-14
arxiv.org
large-language-models
Gauge dependence and structured-output corruption in sign-branched repetition penalties: measurements across models, inference stacks, and alternative repetition controls
Researchers discovered that the multiplicative repetition penalty used in LLM inference engines (HuggingFace, vLLM, llama.cpp) branches on the sign of raw logits, which is arbitrary because the softma…