{"slug": "show-hn-an-open-weight-llm-whose-answer-stop-decision-can-be-flipped-internally", "title": "Show HN: An open-weight LLM whose answer/stop decision can be flipped internally", "summary": "Independent researcher Rayan Pal released PCCG-Qwen3-4B, an open-weight 4-billion-parameter LLM based on Qwen3-4B, with an equality-conditioned continuation policy that allows flipping the model's decision to stop or continue generation via a fixed 2,560-dimensional activation direction at block 29. In tests, the intervention flipped the model's answer/stop decision in 40/40 GO→EOS and 40/40 EOS→GO cases, with 640/640 controls unchanged, while preserving correct reasoning traces in 80/80 cases. The model and method are available on Hugging Face and GitHub, with code and verification scripts.", "body_md": "**40/40 GO→EOS**\n\n**40/40 EOS→GO**\n\n**640/640 controls unchanged**\n\n**Prompt, weights, and generated reasoning prefix held fixed during intervention.**\n\nRayan Pal\n\n[getswiftapi.com](https://getswiftapi.com)\n\nIndependent Researcher\n\nThinking-enabled [PCCG-Qwen3-4B](https://huggingface.co/sharthokrayanpal/PCCG-Qwen3-4B) with an equality-conditioned continuation policy. One fixed 2,560-dimensional activation direction at block 29, applied after the model generated `</think>`. Intervention strength: 2.0. `GO` is followed by native EOS.\n\n| Confirmation measurement | Result | \n|---|---|\n| Correct generated comparison trace and answer | 80/80 | \n| Unmodified replay | 80/80 | \n| GO→EOS | 40/40 | \n| EOS→GO | 40/40 | \n| Controls unchanged | 640/640 | \n\nAnthropic's Jacobian Lens provides internal vocabulary readouts. Additive activation interventions measure causal effects. [Method](/theonlypal/PCCG-Qwen3-4B-continuation-control/blob/main/METHOD.md).\n\nClone with Git LFS installed. The fitted lens is stored in Git LFS.\n\n```\ngit lfs pull\npython verify.py\n```\n\nChecks package hashes and recounts the confirmation records. Runs on CPU using the Python standard library.\n\n[Commands and model requirements](/theonlypal/PCCG-Qwen3-4B-continuation-control/blob/main/REPRODUCE.md).\n\n- `results/` : complete study records, logits, activations, fitted lens, and fixed direction.\n- `source/` : executed study code and Anthropic's Jacobian Lens implementation.\n- `protocol/` : prompts, case sets, model qualification record, and reproduction settings.\n- `SHA256SUMS` : package file hashes.\n- `LICENSE` : Apache 2.0. The vendored implementation retains its license in`source/vendor/LICENSE` .", "url": "https://wpnews.pro/news/show-hn-an-open-weight-llm-whose-answer-stop-decision-can-be-flipped-internally", "canonical_source": "https://github.com/theonlypal/PCCG-Qwen3-4B-continuation-control", "published_at": "2026-09-09 22:32:34+00:00", "updated_at": "2026-09-09 22:46:14.986155+00:00", "lang": "en", "topics": ["large-language-models", "ai-research", "ai-products"], "entities": ["Rayan Pal", "PCCG-Qwen3-4B", "Qwen3-4B", "Hugging Face", "Anthropic", "Jacobian Lens"], "alternates": {"html": "https://wpnews.pro/news/show-hn-an-open-weight-llm-whose-answer-stop-decision-can-be-flipped-internally", "markdown": "https://wpnews.pro/news/show-hn-an-open-weight-llm-whose-answer-stop-decision-can-be-flipped-internally.md", "text": "https://wpnews.pro/news/show-hn-an-open-weight-llm-whose-answer-stop-decision-can-be-flipped-internally.txt", "jsonld": "https://wpnews.pro/news/show-hn-an-open-weight-llm-whose-answer-stop-decision-can-be-flipped-internally.jsonld"}}