{"type":"post","id":"llms-can-learn-some-multi-step-reasoning-on-e5a055","slug":"llms-can-learn-some-multi-step-reasoning-on-e5a055","title":"LLMs can learn some multi-step reasoning on filler tokens instead of visible CoT","description":"Researchers report that language models can learn some multi-step tasks while emitting only filler tokens between a prompt and an answer.","retrieval_nugget":"Researchers report that language models can learn some multi-step tasks while emitting only filler tokens between a prompt and an answer. The finding matters because a readable chain of thought is not a complete record of a model’s computation: internal information can move between token positions without being expressed in words.","published_at":"2026-11-12","updated_at":"2026-11-12","record_date":"2026-11-12","date_kind":"published_at","topics":["models"],"entities":["Qwen3.6-35B-A3B","gpt-oss-20b","Wuschel Schulz","tvb","MARS 5","Meridian Visiting Researcher Programme","LessWrong"],"related_patterns":[],"novelty_label":"new_fact","what_changed":"The study reports training two models to perform multi-step reasoning with filler tokens on three of seven tasks, extending research on parallel and adaptive LLM reasoning.","prior_refs":["https://newruntime.com/posts/language-models-can-spend-too-much-effort-on-a288b4/"],"source_url":"https://www.lesswrong.com/posts/EHkwCDAmhWPAWEziM/llms-can-learn-some-multi-step-reasoning-on-filler-tokens-2","source_urls":["https://www.lesswrong.com/posts/EHkwCDAmhWPAWEziM/llms-can-learn-some-multi-step-reasoning-on-filler-tokens-2"],"schema_version":"newruntime-agent-readable-v0.2","stable_id":"post:llms-can-learn-some-multi-step-reasoning-on-e5a055","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/llms-can-learn-some-multi-step-reasoning-on-e5a055/","markdown":"https://newruntime.com/posts/llms-can-learn-some-multi-step-reasoning-on-e5a055.md","json":"https://newruntime.com/posts/llms-can-learn-some-multi-step-reasoning-on-e5a055.json"},"markdown_url":"https://newruntime.com/posts/llms-can-learn-some-multi-step-reasoning-on-e5a055.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"}]}