SYLIN

Weight-Space Programming

by MIT · Paper · Text & code

promising

Interesting MIT research idea; the authors' results come with caveats that weaken the headline.

When large language models write neural network weights without gradient-based training, they tend to design sparse table or associative-memory models that outperform same-sized Transformers with far fewer parameters and

StatusResearch

Why we say so

The MIT authors say GPT-6 Astra can read a small model's weights and edit about 3.5% of them to change behavior without training. AI Search noted that settings alone got about 44% in an independent analysis, that Astra relied on Python tools and test runs, and that it could not improve Gemma. Code is public, but no independent replication is on record.

Evidence

  1. Nov 8
    MIT research: GPT-6 Astra inspects a small model's weights and edits a tiny fraction to change behavior, without training
  2. Nov 8
    Caveats: independent analysis got about 44% from settings alone, Astra used Python tools and test runs, and it could not improve Gemma

Alternatives

Works with