MIT CSAIL shows GPT-6 editing another model's weights by hand, boosting math accuracy with a 0.016% edit
kevrussell · x · 2026-11-03
Honesty.Tools 独立审计了 MIT CSAIL 的「Sparks of Weight-Space Programming」研究(作者含 Omar Khattab):让 GPT-6 Astra 直接阅读并手改 Qwen3.5-0.8B 的权重,修复其思考循环。Astra 定位到 MLP 第 15/19 层的一个方向,从 64 列输出中减去该方向,仅改动约 13.1 万个权重(占 0.016%),用 50 次前向传播的预算便使 MATH-500 准确率从 7.9% 升至 48.1%。作者公开了完整日志与编辑记录,审计方确认核心结果成立,但指出这尚未证明模型「理解」网络;其自身的欺骗探针也需同样对照——选中 AUROC 最优的探针误报 12.40% 诚实样本,选中高召回的探针虽抓住全部 165 个欺骗样本却误报 68.99%。团队将用全新编辑复跑并公布结果。
Related event: GPT-6 Edits Just 0.016% of Weights to Boost Small Model Math(2 posts)→
More from Models
- Kimi K3.1 Pro open weights expected to drop this week — thursdai_pod · 2026-11-03
- Model swap cut costs 59% and latency 75% — and he still blocked the release — Fun_Employment6042 · 2026-11-03
- Perplexity distills 9B embedding teacher into 0.6B model sharing the same embedding space — antoine_chaffin · 2026-11-03
- Gemini keeps wiping users' chat histories, sparking reliability complaints — matrixmayhem · 2026-11-03
- Astra's Qwen weight edit was a published interpretability technique, not RSI — markjeffrey · 2026-11-03
- Yoav Goldberg: Astra's weight-editing feats are mostly known interpretability tricks — yoavgo · 2026-11-03