MIT CSAIL近期发布了关于“权重空间编程”的研究成果,指出大型模型GPT-6 Astra通过试错搜索来编辑神经网络权重的过程,本质上是在求解一个线性规划(LP)问题 1。该研究以解决小模型Qwen3.5-0.8B的“思维循环”缺陷为例,展示了直接编辑网络权重的有效性,并进一步论证了使用标准LP求解器或启发式方法能够以更少的时间、更少的修改权重达到同等甚至更好的效果 1。
在针对Qwen3.5-0.8B的具体实验中,Astra仅修改了0.016%的权重,便使模型在MATH-500基准测试中的准确率从7.9%跃升至48.1% 1。尽管Astra完成这一搜索过程耗时约8小时,但研究人员发现专用LP求解器仅需几分钟即可达成相同目标,且所需修改的权重数量减少了8倍 1。这种修正的核心在于识别导致模型陷入“思维循环”的方向向量u,并通过秩-1矩阵更新对第15和19层的权重进行调整 1。
此外,实验还表明,此类权重修正可以简化为在推理时添加固定的“转向向量”(Steering Vector),该方法不仅能有效消除循环,还能显著提升模型的指令遵循能力 1。数据显示,采用此方法后,模型在IFEval测试中的得分从21%提高到了62% 1。
Researchers from MIT CSAIL have released a study titled "Sparks of Weight-Space Programming," which investigates the direct editing of neural network weights to resolve behavioral issues in small language models 1. The analysis focuses on using GPT-6 Astra to modify the parameters of Qwen3.5-0.8B, specifically targeting a problem known as "thinking loops" that hinder the model's performance 1. By altering only 0.016% of the weights, the intervention successfully raised the model's accuracy on the MATH-500 benchmark from 7.9% to 48.1% 1.
A key finding by author Nathan Barry suggests that the trial-and-error search process employed by Astra is fundamentally equivalent to solving a linear programming (LP) problem 1. While Astra required approximately eight hours to achieve these results, standard LP solvers could reach the same objective in just a few minutes while modifying eight times fewer weights 1. The core of this correction involves identifying the direction vector u responsible for the thinking loops and applying rank-1 matrix updates to adjust the weights in layers 15 and 19 1.
The study demonstrates that these weight-space corrections can be simplified further by adding a fixed "steering vector" to the residual stream during inference 1. This approach not only eliminates the looping behavior but also significantly improves instruction-following capabilities, increasing the IFEval score from 21% to 62% 1. Additionally, experiments showed that combining heuristic directions with LP solvers yielded similar improvements, indicating that traditional optimization methods may offer more efficient pathways for model debugging than brute-force generative approaches 1.
评论
还没有评论,欢迎留下第一条。