#activation-steering on Hacker News

2 stories tagged activation-steering, newest first.

Screenshot of arxiv.org: "As a Language Model": Chat Template Switches LLM Self-Referential Voice

"As a Language Model": Chat Template Switches LLM Self-Referential Voice (arxiv.org)

92 points by yu3zhou4 29 days ago | 96 comments

arXiv paper showing that chat templates toggle LLM disclaimer versus experiential self-referential voice, and that an activation direction can steer this behavior.

llm chat-templates activation-steering interpretability

Screenshot of blog.madhukaraphatak.in: Dynamic Abliteration: Non-Destructive Refusal Suppression via Engram Steering

Dynamic Abliteration: Non-Destructive Refusal Suppression via Engram Steering (blog.madhukaraphatak.in)

96 points by phatak-dev 1 month ago | 32 comments

Blog post describing dynamic abliteration, using PyTorch forward hooks to inject multi-layer residual steering in Qwen3-4B to suppress refusals without modifying weights.

abliteration activation-steering refusal-suppression qwen

#activation-steering on Hacker News

"As a Language Model": Chat Template Switches LLM Self-Referential Voice(arxiv.org)

Dynamic Abliteration: Non-Destructive Refusal Suppression via Engram Steering(blog.madhukaraphatak.in)

"As a Language Model": Chat Template Switches LLM Self-Referential Voice (arxiv.org)

Dynamic Abliteration: Non-Destructive Refusal Suppression via Engram Steering (blog.madhukaraphatak.in)