--- format: "aidr-story-markdown/v1" id: "717a4a48908c13c8fada2fe0fa3df7e4f8eae269a4ea0fe1b5b251c18d39edb1" canonical_url: "https://aidr.today/717a4a48?lang=en" title: "Replication Confirms MIT Weight Editing Study as Authors Clarify Loop Fix" lang: "en" requested_lang: "en" available_langs: ["en","vi"] translation_fallback: null fallback_fields: [] published_at: "2026-11-03T08:51:42.000Z" category: "Research" topics: ["openai","llm"] source_urls: ["https://marketbrief.now/ai/replication-confirms-mit-weight-editing-study-as-authors-clarify-loop-fi-cd8a8f15","https://x.com/rikiyatakehi/status/2117313742891327994","https://x.com/AlexKrentsel/status/2117376683724660898","https://x.com/rikiyatakehi/status/2117474915561701780","https://x.com/adam_karvonen/status/2117337487458938949","https://x.com/a1zhang/status/2117364329133621614","https://x.com/hosseeb/status/2117380109355352166","https://x.com/paul_cal/status/2117531885706359275"] summary: "An independent replication by researcher Alex Krentsel confirmed MIT CSAIL's weight space programming results and found that the ability to write functional neural network weights from scratch only emerged in the latest generation of frontier models, such as GPT-6 Astra and Claude Opus 5.5. In those tests, the newest models generated networks to play the Atari game Breakout without training or tools, scoring up to 22 times an RL-trained baseline. Lead MIT author Rikiya Takehi also released the edited Qwen3.5-0.8B model on Hugging Face Hub after it was rebuilt from published logs and scored 51.0% on a MATH-500 re-run. MIT co-author Alex Zhang clarified that GPT-6 Astra did not edit new math skills directly into Qwen3.5-0.8B, which originally scored 7.9% on MATH-500 before rising to 48.1% in the paper and jumping from 21.0% to 62.3% on instruction following. Instead, an LLM-optimized prompt contained hints to check for repetitive reasoning loops, a defect noted in Qwen's model card. Given 50 forward-pass batches, Astra wrote code using activation steering to suppress a looping signal across 131,000 weights in layers 15 and 19, or 0.016% of the 800M parameter model. Cutting those lo" --- # Replication Confirms MIT Weight Editing Study as Authors Clarify Loop Fix > [Open the canonical story]() **Published:** 2026-11-03T08:51:42.000Z **Category:** Research **Topics:** openai, llm ## Summary An independent replication by researcher Alex Krentsel confirmed MIT CSAIL's weight space programming results and found that the ability to write functional neural network weights from scratch only emerged in the latest generation of frontier models, such as GPT\-6 Astra and Claude Opus 5\.5\. In those tests, the newest models generated networks to play the Atari game Breakout without training or tools, scoring up to 22 times an RL\-trained baseline\. Lead MIT author Rikiya Takehi also released the edited Qwen3\.5\-0\.8B model on Hugging Face Hub after it was rebuilt from published logs and scored 51\.0% on a MATH\-500 re\-run\. MIT co\-author Alex Zhang clarified that GPT\-6 Astra did not edit new math skills directly into Qwen3\.5\-0\.8B, which originally scored 7\.9% on MATH\-500 before rising to 48\.1% in the paper and jumping from 21\.0% to 62\.3% on instruction following\. Instead, an LLM\-optimized prompt contained hints to check for repetitive reasoning loops, a defect noted in Qwen's model card\. Given 50 forward\-pass batches, Astra wrote code using activation steering to suppress a looping signal across 131,000 weights in layers 15 and 19, or 0\.016% of the 800M parameter model\. Cutting those lo ## Sources - [Story source]() - [Story source]() - [Story source]() - [Story source]() - [Supporting source]() - [Supporting source]() - [Supporting source]() - [Supporting source]()