{"slug":"weight-edit","domain":"research","title":"Qwen3.5-0.8B before and after a weight edit","summary":"One math problem answered by Qwen3.5-0.8B and by the copy an AI agent edited to stop its thinking loops, side by side, with the edit's signal every 64 tokens.","help":"weight-edit: one math problem answered by Qwen3.5-0.8B as released and by the copy an AI agent edited, as JSON.\n\nTHE EDIT\n  An agent at MIT CSAIL edited 131,072 weights of Qwen3.5-0.8B\n  (RikiyaT/Qwen3.5-0.8B-weight-space-programming). The edit takes one\n  direction out of the residual stream in layers 15 and 19. Its neurons fire\n  about 4.5 times as hard on repeated text as on new text, so it stops the\n  model's thinking loops. On MATH-500 with thinking on, the score goes from\n  7.9% to 48.1%. With thinking off, the model card's default, it goes from\n  43.6% to 45.2%.\n\nRETURNS\n  One JSON object on stdout with `unedited` and `edited`, each holding the\n  model's thinking, its answer, the last \\boxed{} in the answer, the tokens it\n  used, whether it finished before the limit, `loop_at` (the first token of a\n  64-token stretch that is mostly repeated, or null), and `bins`: the edit's\n  signal in layers 15 and 19 and the repeated share, every 64 tokens. Also\n  `price_usd`, and `standing`, which says how many free runs are left.\n\nOPTIONS\n  --thinking false   answer without thinking (default: thinking on)\n  --max_tokens N     each model's budget, 256 to 8192 (default 6144)\n  --seed N           the sampling seed, the same for both models (default 0)\n\nPRICE\n  $0.05 a run. An account that has never paid gets three runs free; after that\n  the run is refused until $2 of credit is added, and the `credits` tool gives\n  the link. A warm run takes about 45 seconds; the first after a quiet spell\n  waits about three minutes while the models load.\n\nREAD THEM AS ONE SAMPLE\n  One sampled answer per model. Show both answers as they are; one run says\n  little about which model is better at math.\n\nStudies:\n  run \u003cproblem\u003e\n      both models' thinking and answers, whether each finished, where a loop began, the edit's signal every 64 tokens, the price and how many free runs are left\n\nJSON to stdout, so it composes: pipe to `jq` to pick a field, or `\u003e /tmp/…` to keep it.\nRun a study with no arguments — `weight-edit run` — to see what it takes.\n","example_command":"weight-edit run \"Convert the point (0,3) in rectangular coordinates to polar coordinates.\" --max_tokens 4096","example_output":"","cost_class":"upstream-fetch","tier":"metered","billing_items":["weight_edit_run"],"price_per_1k_calls":50,"price_is_from":false,"status":"available","commands":["weight-edit"],"studies":[{"call":"run","about":"both models' thinking and answers, whether each finished, where a loop began, the edit's signal every 64 tokens, the price and how many free runs are left","args":[{"name":"problem","positional":true,"list":false,"hint":"the math problem, quoted, up to 3000 characters"},{"name":"thinking","positional":false,"list":false,"hint":"true or false: think before answering (default true)"},{"name":"max_tokens","positional":false,"list":false,"hint":"each model's budget, 256 to 8192 (default 6144)"},{"name":"seed","positional":false,"list":false,"hint":"the sampling seed, the same for both models (default 0)"}]}],"default_enabled":false,"first_seen":"2026-11-03T09:05:48.067063Z","enabled":false}