paper web signal

WeightBench scores Claude Opus 5.5 at 0.535 on weight-space edits

TL;DR

  • Claude Opus 5.5 reaches a macro score of 0.535 on 42 weight-space programming tasks; GPT-6.1-sol scores 0.568 on the 38 tasks it was run on.
  • The benchmark spans five tracks covering reading, fixing, writing, impossible-edit recognition, and data-free compression, on networks from 2K parameters to 7B parameters.
  • Frontier agents can write working control policies without ever running them but cannot yet patch language models or recognize impossible edits.

Claude Opus 5.5 scores 0.535 across 42 tasks that ask a model to read, repair, and write other neural networks' weights without any gradient descent, according to a new benchmark posted to arXiv by researchers at the University of Southern Mississippi.

The paper introduces WEIGHTBENCH and calls the capability it measures Weight-Space Programming. The tasks span five tracks: reading, fixing and writing weights, recognizing edits that are impossible, and data-free compression. The networks the models operate on range from 2K-parameter control policies up to 7B-parameter language models.

On the full 42 tasks, Sonnet 5.5 scores 0.466, Haiku 5.5 0.402, and DeepSeek v4-flash 0.215. On the 38 tasks it was run on, GPT-6.1-sol reaches 0.568 against Opus's 0.546 on the same subset.

The abstract is blunt about the ceiling. 'Frontier agents can already read small networks and write working control policies without ever running them, but cannot yet patch language models, compress them without data, or tell an impossible edit from a hard one,' the authors write.