线性规划复现Astra权重编辑
Nathan Barry 分析 MIT CSAIL 的“权重空间编程”研究称,GPT-6 Astra 可通过推理而非梯度或训练循环判断小型网络所训练的游戏,并生成击败 PPO 训练网络的权重。他还报告,Astra 仅改动 Qwen3.5-0.8B 的 0.016% 权重,就将 MATH-500 准确率从 7.9% 提升至 48.1%,IFEval 从 21.0% 提升至 62.3%。Barry 认为,Astra 约八小时的循环消除编辑搜索可转化为线性规划。他报告称,几次线性规划求解可产生相同效果,同时将改动权重数量减少至八分之一;其定制 PDHG GPU 求解器约需一分钟即可达到距最优解 0.2% 以内。
Barry 的分析提供了一种更紧凑的方式来复现所报告的循环消除编辑,但也揭示了局限:该方法仅在移除循环的任务上测试,每词元约束影响较小,而且如何在没有智能体的情况下选择目标层仍未解决。
Last week, MIT CSAIL published Sparks of Weight-Space Programming. Given the raw weights of a small network, GPT-6 Astra could tell which game it had been trained to play. Given only a game’s rules, it could write the weights of a small network from scratch and beat networks trained with PPO for 100 million steps. No gradients, no training loop, just reasoning about what the weights compute. They coined this weight-space programming.
In another experiment, they asked Astra to make Qwen3.5-0.8B better at math by editing its weights. Astra changed only 0.016% of the weights, and MATH-500 accuracy jumped from 7.9% to 48.1%.1
Looking at Astra’s edit, I realized most of its search was solving a linear program by trial and error. It spent about 8 hours and 28 forward passes choosing which weights to change and by how much. A couple of LP solves, about a minute each, do the same job with 8× fewer changed weights. In this post I’ll walk through the edit, why it’s an LP, and what that says about the fix itself.
#MIT CSAIL#GPT-6 Astra#Qwen3.5-0.8B#Gemma 3 270M#HiGHS#PDHG#PPO#MATH-500#IFEval#GSM8K