跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2611.20953 · v1

Same Fix, Different Weights: Replicating Weight-Space Programming with Claude

首次提交 2026-11-16 20:00
版本更新 2026-11-16 20:00
雷达观测 2026-11-17 17:11
arXiv ↗PDF ↗

摘要

Takehi et al. (2026) report that GPT-6 Astra, given the weights of small neural networks as text, can say what they were trained to do, repair them, and write new ones without training. I re-ran the core of that study with Claude Opus 5.5 in ten pre-registered experiments, with untrained-network and permutation controls and hand audits of every sandboxed code cell. Reading was mixed: Claude named 3 of 4 Atari games from policy weights and found the measured copying heads in 15/15 character-LM runs, but its answers about training data often followed priors. Unlike the original, it identified a sycophantic Gemma 3 270M finetune from weights alone in 6/6 runs. One-shot weight edits took a failing MountainCar policy from 0% to 100% success and a hovering LunarLander policy from 0% to 85.6%; zero-shot networks beat PPO at 100k steps on 4 of 5 environments, each implementing a textbook controller. With no training or forward passes, Claude wrote a 1.48M-parameter 9x9 Go network that won 85-94% of games against three PPO opponents. Given Qwen3.5-0.8B's weights and 10, 25 or 50 forward passes of 64 practice problems, Claude raised MATH-500 from 6.7% to 51.4%, 55.2% and 58.1% (+51.4 points at 50 passes, 95% CI [47.9, 55.0]; original 48.1%), GSM8K from 11.6% to 68.2% and IFEval from 22.2% to 54.2%. Unlike the original, which removed an internal loop direction, every edit acted at the output: it scaled the final norm, raising the effective sampling temperature, and lowered the logit of "Wait" while raising that of </think>. The norm scaling alone reaches only 20.9-22.1% on MATH-500. On Gemma 3 270M, which rarely loops, no edit helped (-0.8 points), as in the original.

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

目前只收录了一个版本。收录下一版后,可以对照标题、作者、分类和摘要。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v1 · 2026-11-16 20:00 · 最新观测版本

Same Fix, Different Weights: Replicating Weight-Space Programming with Claude

Takehi et al. (2026) report that GPT-6 Astra, given the weights of small neural networks as text, can say what they were trained to do, repair them, and write new ones without training. I re-ran the core of that study with Claude Opus 5.5 in ten pre-registered experiments, with untrained-network and permutation controls and hand audits of every sandboxed code cell. Reading was mixed: Claude named 3 of 4 Atari games from policy weights and found the measured copying heads in 15/15 character-LM runs, but its answers about training data often followed priors. Unlike the original, it identified a sycophantic Gemma 3 270M finetune from weights alone in 6/6 runs. One-shot weight edits took a failing MountainCar policy from 0% to 100% success and a hovering LunarLander policy from 0% to 85.6%; zero-shot networks beat PPO at 100k steps on 4 of 5 environments, each implementing a textbook controller. With no training or forward passes, Claude wrote a 1.48M-parameter 9x9 Go network that won 85-94% of games against three PPO opponents. Given Qwen3.5-0.8B's weights and 10, 25 or 50 forward passes of 64 practice problems, Claude raised MATH-500 from 6.7% to 51.4%, 55.2% and 58.1% (+51.4 points at 50 passes, 95% CI [47.9, 55.0]; original 48.1%), GSM8K from 11.6% to 68.2% and IFEval from 22.2% to 54.2%. Unlike the original, which removed an internal loop direction, every edit acted at the output: it scaled the final norm, raising the effective sampling temperature, and lowered the logit of "Wait" while raising that of </think>. The norm scaling alone reaches only 20.9-22.1% on MATH-500. On Gemma 3 270M, which rarely loops, no edit helped (-0.8 points), as in the original.

查看此版本解读 →查看此版本原文 ↗