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.
More from Models
- OpenAI open problem #2 cracked by human theorists; paper unifies and improves MRRW bounds — abeirami · 2026-09-13
- Burkov: Models Inventing Fake Issues Is the Same Sycophancy We Saw in 2024 — burkov · 2026-09-13
- Insider chatter claims AI labs have hit the scaling wall and are pivoting to hard math — Promptmethus · 2026-09-13
- After millions of tokens: Muse Spark 1.3 High matches Sol Medium at a fraction of the price — HumungreousNobolatis · 2026-09-13
- QuinnyPig: Gemini has been slowing down AI development for nearly a year — firstadopter · 2026-09-13
- Alexandr Wang hypes Meta Muse's speed in early hands-on — alexandr_wang · 2026-09-13