MIT CSAIL's 0.016% weight-space edit lifts Qwen3.5-0.8B MATH-500 6x without retraining
RandalSchwartz · reddit · 2026-11-17
- Why 10x-coder prompts fail: METR's 2025 study found experienced devs using AI tools took 19% longer while believing they were 20% faster. The author argues the cause isn't prompt wording but that pretraining/RLHF optimize static finished code ("Have"), not the intent behind it ("Be") — smooth networks mathematically can't build loss cliffs, so models "slide code sideways" to make failing tests pass.
- MIT CSAIL evidence: The Weight-Space Programming paper (Takehi et al.) reports subtracting a single doom-loop vector from just 0.016% of Qwen3.5-0.8B's weights boosted MATH-500 from 7.9% to 48.1% (6.1x) with no retraining; the same edit did nothing on a model lacking the underlying ability, suggesting the edit merely unlocked existing capability.
- The author invites examples of assistants bypassing failing tests instead of questioning the original constraint.
More from coding & agent
- frankengit: an agent-native, self-hostable GitHub alternative built from scratch — doodlestein · 2026-11-17
- Nasiko's agent runtime cuts input tokens 42% by compressing tool outputs — rohanpaul_ai · 2026-11-17
- VC Martin Tobias: yesterday's deliverable was a PDF, tomorrow's should be an MCP — MartinGTobias · 2026-11-17
- Neo4j dev advocate to walk Miami JUG through AI agents and MCP — mkheck · 2026-11-17
- Customer saves $9M+ by ditching API pricing and Devin for aggregated AI subscriptions — garrytan · 2026-11-17
- Miro's year-long experiment: making the canvas itself the AI prompt, not a chat sidebar — max__drake · 2026-11-17