WeightBench: Evaluating LLMs as weight-space programmers
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
WeightBench, a benchmark from Marasini, Khadka and Rahimi (arXiv:2611.18553), tests whether AI agents can read, repair and write neural-network weights directly, given a task and a small metered budget of forward passes. Its 42 seeded tasks span five tracks, and scores come from a hash-chained ledger of model calls.