Astra Edits Qwen Weights for Math Gains
First merged summary.
Frontier LLM Astra reads weights to identify behaviors and edits Qwen3.5-0.8B to boost math performance without gradients.
Frontier LLM Astra identifies trained behaviors from raw neural network weights and edits models directly. It correctly identifies games a network was trained on from weights alone and fixes issues in Qwen3.5-0.8B by editing just 0.016% of its weights, raising MATH-500 accuracy from 7.9% to 48.1%. The capability enables weight-space programming where LLMs reason about weights to optimize networks. However, this is not superhuman; human researchers can replicate it using steering and targeted adjustments.
- Astra identifies games from network weights alone
- Edits 0.016% of Qwen3.5-0.8B weights for math improvement
- Raises MATH-500 score from 7.9% to 48.1%
- Edits 3.5% of weights for 100% car control success
- Not superhuman; replicable by human researchers
- Generalizes to instruction following tasks
What each outlet adds
beyond the shared facts- Hacker NewsAstra game identification from weights and Qwen math edits
- AI models
- AstraQwen3.5-0.8BGPT-6 AstraQwen
Coverage
oldest first · one row per article · UTC- · 10h ago
Hacker News · AIFirst report
Astra reads and writes neural network weights
Frontier LLM Astra identifies trained behaviors from raw weights and edits Qwen3.5-0.8B weights to improve math performance without gradients.
- · 6h ago
Hacker News · AI
LLMs are not superhuman at reading and writing neural network weights
A recent paper claims frontier LLMs can work with neural network weights directly, but this is not superhuman.