cd /news/large-language-models/show-hn-an-open-weight-llm-whose-ans… · home topics large-language-models article
[ARTICLE · art-125227] src=github.com ↗ pub= topic=large-language-models verified=true sentiment=· neutral

Show HN: An open-weight LLM whose answer/stop decision can be flipped internally

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.

read1 min views1 publishedSep 9, 2026
Show HN: An open-weight LLM whose answer/stop decision can be flipped internally
Image: Michielbdejong (auto-discovered)

40/40 GO→EOS

40/40 EOS→GO

640/640 controls unchanged

Prompt, weights, and generated reasoning prefix held fixed during intervention.

Rayan Pal

getswiftapi.com

Independent Researcher

Thinking-enabled 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.

Confirmation measurement Result
Correct generated comparison trace and answer 80/80
Unmodified replay 80/80
GO→EOS 40/40
EOS→GO 40/40
Controls unchanged 640/640

Anthropic's Jacobian Lens provides internal vocabulary readouts. Additive activation interventions measure causal effects. Method.

Clone with Git LFS installed. The fitted lens is stored in Git LFS.

git lfs pull
python verify.py

Checks package hashes and recounts the confirmation records. Runs on CPU using the Python standard library.

Commands and model requirements.

  • results/ : complete study records, logits, activations, fitted lens, and fixed direction.
  • source/ : executed study code and Anthropic's Jacobian Lens implementation.
  • protocol/ : prompts, case sets, model qualification record, and reproduction settings.
  • SHA256SUMS : package file hashes.
  • LICENSE : Apache 2.0. The vendored implementation retains its license insource/vendor/LICENSE .
── more in #large-language-models 4 stories · sorted by recency
── more on @rayan pal 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/show-hn-an-open-weig…] indexed:0 read:1min 2026-09-09 ·