Skip to content
NewA calmer, faster ZeroHour

WeightBench scores language models on 42 weight-editing tasks

Original title: WeightBench: Evaluating Language Models as Weight-Space Programmers

arXiv cs.AI / cs.LG / cs.CLby Aayush Marasini ()1 min read2 · 2 reads
AI summary · x-ai/grok-4.7Summaries by Vani AI

Claude Opus 5.5 scores 0.535 overall; GPT-6.1-sol scores 0.568 on 38 tasks.

WeightBench tests whether language models can read, fix, and write neural-network weights without gradient descent, across 42 tasks in five tracks. Target networks range from 2K-parameter control policies to 7B-parameter language models. Reported results say frontier agents can read small networks and write working control policies without running them, but cannot yet patch language models, compress them without data, or separate impossible edits from hard ones. Code, an OS-level sandbox harness, and a live leaderboard are public.

Key facts
  • Tracks are reading, fixing, writing, impossible edits, and data-free compression.
  • Sonnet 5.5 scores 0.466, Haiku 5.5 0.402, and DeepSeek v4-flash 0.215.
  • On 38 tasks, Opus 5.5 scores 0.546 next to GPT-6.1-sol at 0.568.
  • Edited networks span 2K-parameter policies to 7B-parameter language models.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

From the original article

1 min read at arxiv.org

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).

arXiv cs.AI / cs.LG / cs.CL

By Aayush Marasini · arxiv.org

Read the full article at arXiv cs.AI / cs.LG / cs.CL

A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.