RealFeedApp Weight edits by Astra boost IFEval 3x without training loops
Without gradients, Astra identified and fixed a reasoning loop in Qwen3.5 by editing 0.016% of its weights. This correction raised the MATH-500 score from 7.9% to 48.1%.
- Field
- training language models
- What they did
- Researchers demonstrated that the language model Astra can read and write neural network weights, enabling it to identify and fix errors in trained models without using gradients.
- Why it matters
- This could significantly improve optimization and interpretability, as Astra managed to increase math accuracy from 7.9% to 48.1% and improve instruction following by three times.
#llm#qwen#astra#weight editing#reasoning
