{"rewrite":{"id":"r_35e63c9d22fbab9d5a5c61f3","clusterId":"c_c086b3cf0229f88bb1e8b9cd","slug":"mit-team-reports-gpt-6-astra-weight-edit-lifts-qwen3-5-0-8b-math-accuracy-to-48-1","model":"deepseek-v4-1-flash","headline":"MIT Team Reports GPT-6 Astra Weight Edit Lifts Qwen3.5-0.8B Math Accuracy To 48.1%","summary":"A research team led by Rikiya Takehi at MIT's Computer Science and Artificial Intelligence Laboratory published a report on November 2, 2026, describing Weight-Space Programming, in which a large language model reads and edits another network's weights directly. Using GPT-6 Astra, the team changed 0.016% of the parameters in Alibaba's Qwen3.5-0.8B, lifting math benchmark accuracy from 7.9% to 48.1% without additional training. The report has drawn criticism that its framing is misleading.","whyItMatters":"If a large model can read intent out of raw weights and rewrite them without training, the cost of adapting a model shifts from data and compute to a single inference pass, which is why the framing of the 7.9% to 48.1% result is being argued over so closely.","webCardHtml":"\u003cp\u003eThe headline number comes from 131,072 changed weights inside a model of about 800 million parameters. Astra found the cause of Qwen3.5-0.8B\u0026#39;s repeated output in a specific direction written by a specific layer, after using up to 50 evaluation runs on math practice problems to locate it.\u003c/p\u003e\u003cp\u003eEarlier experiments in the same report had Astra identify which of four Atari games a network was trained on (Breakout, Pong, Pac-Man, Freeway), all sharing the same 17,253 weights, without being told the names. In a driving task, it rewrote 3.5% of weights so the car reversed before climbing, and the fixed network finished all 1,024 trials. Skeptics argue the report\u0026#39;s presentation overstates what was shown.\u003c/p\u003e","blueskyPost":"GPT-6 Astra touched 0.016% of Qwen3.5-0.8B's parameters. The 48.1% math score comes without retraining, which is why critics call the framing misleading: the result is real, the comparison to training is not.","twitterPost":"GPT-6 Astra rewrote 0.016% of Qwen3.5-0.8B's weights. The 7.9% to 48.1% jump needs a baseline, and critics say the report's framing hides one.","threadsPost":"The number worth studying in the MIT report is 0.016%: that is how much of Qwen3.5-0.8B GPT-6 Astra actually edited. A jump from 7.9% to 48.1% on math benchmarks without training is striking, and the criticism that the framing misleads turns on what that small edit is allowed to claim.","newsletterBlurb":"A MIT CSAIL team led by Rikiya Takehi published a report on November 2, 2026 describing Weight-Space Programming, where a large model reads and edits another network's weights directly. GPT-6 Astra changed 0.016% of Qwen3.5-0.8B's parameters and lifted math accuracy from 7.9% to 48.1% with no additional training. Critics call the explanation misleading.","attributionJson":"[{\"source\":\"GIGAZINE\",\"url\":\"https://gigazine.net/news/20261104-gpt-6-astra-qwen-weight-space-programming/\",\"title\":\"Research results claim 'GPT-6 Astra rewrote the weights of Qwen3.5-0.8B to improve math performance about 6x,' but critics call it 'misleading'\"}]","lintFlagsJson":null,"lintHits":0,"costUsd":0,"inputTokens":5572,"outputTokens":931,"status":"published","repairAttempts":0,"nextRepairAt":null,"factsAttemptedAt":1793765997,"createdAt":"2026-11-04T04:15:36.000Z","publishedAt":"2026-11-04T04:18:53.000Z","updatedAt":"2026-11-04T04:18:53.000Z"},"cluster":{"id":"c_c086b3cf0229f88bb1e8b9cd","canonicalTitle":"「GPT-6 AstraがQwen3.5-0.8Bの重みを書き換えて数学性能を約6倍に改善した」との研究結果、ただし「誤解を招く」との批判も","representativeArticleId":"a_7d65f8de25093e90b6f41886","sourceCount":1,"writtenSourceCount":1,"writeAttempts":0,"isSolo":true,"entitiesJson":"{\"anime_titles\":[],\"manga_titles\":[],\"work_titles\":[\"GPT-6 Astra\",\"Qwen3.5-0.8B\"],\"studios\":[\"MIT\"],\"people\":[],\"type\":\"news\",\"domain\":\"other\",\"is_roundup\":false}","contentType":"news","status":"published","firstSeenAt":"2026-11-04T03:30:00.000Z","lastSeenAt":"2026-11-04T03:30:00.000Z","updatedAt":"2026-11-04T04:18:52.000Z"},"attribution":[{"source":"GIGAZINE","url":"https://gigazine.net/news/20261104-gpt-6-astra-qwen-weight-space-programming/","title":"「GPT-6 AstraがQwen3.5-0.8Bの重みを書き換えて数学性能を約6倍に改善した」との研究結果、ただし「誤解を招く」との批判も"}],"entities":{"anime_titles":[],"manga_titles":[],"work_titles":["GPT-6 Astra","Qwen3.5-0.8B"],"studios":["MIT"],"people":[],"type":"news","domain":"other","is_roundup":false},"keyFacts":["Rikiya Takehi of MIT's Computer Science and Artificial Intelligence Laboratory led the research, published November 2, 2026.","GPT-6 Astra changed 0.016% of the weights in Qwen3.5-0.8B, raising math benchmark accuracy from 7.9% to 48.1%.","Astra changed 131,072 weights in Qwen3.5-0.8B, a model of about 800 million parameters.","In a driving task, Astra rewrote 3.5% of a network's weights and the fixed network completed all 1,024 trials.","Astra identified which of four Atari games a network was trained on, all sharing 17,253 weights, without being told the game names."],"pageLinks":{"healthSlugs":["gpt-6-astra","qwen3-5-0-8b","mit"],"storySlugs":[],"entities":[{"slug":"gpt-6-astra","supported":4,"disputed":0,"contradicted":0,"unverified":7,"overall":"green","story_count":9,"has_dossier":true},{"slug":"qwen3-5-0-8b","supported":0,"disputed":0,"contradicted":0,"unverified":0,"overall":"unverified","story_count":1,"has_dossier":false},{"slug":"mit","supported":0,"disputed":0,"contradicted":0,"unverified":1,"overall":"unverified","story_count":1,"has_dossier":false}],"stories":[],"entitiesOk":true,"storiesOk":true}}