Tag
PCCG-2 presents an EOS gate study where a frozen Qwen3-4B answer computation is controlled by a learned 101-parameter gate that only alters the native termination score, alongside cross-vendor evidence of zero-byte 'Void' responses across 11 frontier models from OpenAI, Anthropic, Google, and Moonshot.