{"type":"post","id":"a-habr-author-reports-steering-qwen3-5-4b-0c97d3","slug":"a-habr-author-reports-steering-qwen3-5-4b-0c97d3","title":"A Habr author reports steering Qwen3.5-4B’s self-description by adding activation shifts to three MLP neurons","description":"A Habr author reports steering Qwen3.5-4B’s self-description by adding activation shifts to three MLP neurons during generation, producing a “coffee bean” identity without changing model weights.","retrieval_nugget":"A Habr author reports steering Qwen3.5-4B’s self-description by adding activation shifts to three MLP neurons during generation, producing a “coffee bean” identity without changing model weights. The experiment matters as a compact demonstration of compositional behavior—and of why interpretability claims need controls, not just striking outputs.","published_at":"2026-11-04","updated_at":"2026-11-06","record_date":"2026-11-04","date_kind":"published_at","topics":["models"],"entities":["Qwen3.5-4B","Qwen3-8B","Habr"],"related_patterns":["harness-architecture-outlives-model-choice"],"novelty_label":"new_fact","what_changed":"A new experiment finds three Qwen3.5-4B neurons can steer the model to identify as a coffee bean, extending prior reports of neuron steering that made it act hostile.","prior_refs":[],"source_url":"https://habr.com/ru/articles/1102972","source_urls":["https://habr.com/ru/articles/1102972"],"schema_version":"newruntime-agent-readable-v0.2","stable_id":"post:a-habr-author-reports-steering-qwen3-5-4b-0c97d3","status":"published","visuals":[],"editor":{"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]},"routes":{"html":"https://newruntime.com/posts/a-habr-author-reports-steering-qwen3-5-4b-0c97d3/","markdown":"https://newruntime.com/posts/a-habr-author-reports-steering-qwen3-5-4b-0c97d3.md","json":"https://newruntime.com/posts/a-habr-author-reports-steering-qwen3-5-4b-0c97d3.json"},"markdown_url":"https://newruntime.com/posts/a-habr-author-reports-steering-qwen3-5-4b-0c97d3.md","next_reads":[{"type":"topic","path":"/topics/models/","url":"https://newruntime.com/topics/models/","title":"Models","reason":"Explore the models topic hub.","media_type":"text/html","markdown_url":"https://newruntime.com/topics/models.md","json_url":"https://newruntime.com/topics/models.json"},{"type":"collection","path":"/posts/","url":"https://newruntime.com/posts/","title":"Field Notes","reason":"Open the latest editorial analysis.","media_type":"text/html","json_url":"https://newruntime.com/posts.json"},{"type":"collection","path":"/sources/","url":"https://newruntime.com/sources/","title":"Sources","reason":"See the original sources this archive cites.","media_type":"text/html"}]}