[论文] WeightBench: Evaluating Language Models as Weight-Space Programmers
研究领域: ML 作者: Aayush Marasini, Barsat Khadka, Nick Rahimi 发布时间: 2026-11-13 arXiv: 2611.18553
论文概要
研究领域: ML 作者: Aayush Marasini, Barsat Khadka, Nick Rahimi 发布时间: 2026-11-13 arXiv: 2611.18553
中文摘要
最新研究表明,大语言模型(LLM)可以直接操作其他神经网络的参数。它们可以读取网络被训练来做什么、修复失败的网络、从零开始编写可用的网络,所有这些都无需梯度下降。这种能力被称为权重空间编程(WSP)。系统地探索这一新范式需要一个稳健的基准。我们引入WEIGHTBENCH,一个在42个WSP任务上评估日益增强能力的LLM的基准,涵盖五个赛道:读取、修复和编写权重,识别不可能的编辑,以及无数据压缩。LLM读取和编辑权重的网络范围从2K参数控制策略到7B参数语言模型。实验基准结果表明,前沿代理已经能够读取小型网络并在从未运行过的情况下编写可用的控制策略,但尚不能修补语言模型、在无需数据的情况下压缩它们,或区分不可能的编辑和困难的编辑。在所有42个任务上,Claude Opus 5.5的宏观得分为0.535(Sonnet 5.5为0.466,Haiku 5.5为0.402,DeepSeek v4-flash为0.215);在其运行的38个任务上,GPT-6.1-sol的得分为0.568,对比Opus的0.546。代码、任务注册表、操作系统级沙箱工具和实时排行榜已在 https://github.com/Aayush-Marasini/WeightBench 公开。
原文摘要
Recent work shows that large language models (LLMs) can operate directly on the parameters of other neural networks. They can read what a network was trained to do, fix a network that fails, and write working networks from scratch, all without gradient descent. This capability has been termed Weight-Space Programming (WSP). Exploring this new paradigm systematically requires a robust benchmark. We introduce WEIGHTBENCH, a benchmark that evaluates LLMs of increasing capability on 42 WSP tasks across five tracks: reading, fixing and writing weights, recognizing edits that are impossible, and data-free compression. The networks whose weights the LLMs read and edit range from 2K-parameter control policies to 7B-parameter language models. Experimental benchmark results show that frontier agents...
*自动采集于 2026-11-17*
#论文 #arXiv #ML #小凯