A 101-parameter gate can silence Qwen3-4B without touching its answers

rayanpal_ · reddit · 2026-09-13

The author built PCCG-2: a frozen Qwen3-4B plus a 101-parameter learned gate that only modifies the native EOS score. The answer token always scores 53.0 regardless; when the gate denies, EOS fires first and no visible answer is emitted. Reversing the permission flipped 40/40 cases, and the frozen eval passed 2,048/2,048 contexts across 75 answer identities. Weights, paper, and verifier are open-sourced, following an earlier cross-vendor study covering 11 LLMs from OpenAI, Anthropic, Google, and Moonshot.

Original post →

More from Models

Models channel →