Unverified50% confidenceFactExact time
In experiment 8, Astra fixed the "looping" problem of the Qwen 3.5B (800 million parameter) model by editing its weights, after which the Math 500 score rose from slightly below 8% to slightly above 48% and GSM8K rose from 13.8% to 67.1%.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/4/2026
First Seen
Valid until: 2/2/2027
Sources
AI Directly Rewrites Model Weights: MIT Study Reveals New Black-Box Breakthrough
youtubeMatthew Berman11/4/2026
Related Entities
Related Claims
UnverifiedIn the experiment with the smallest Qwen 3.5 version (about 800 million weights), Astra modified 64 columns of numbers in each of layers 15 and 19 (about 130,000 weights), raising the standard math test score from under 8% to 48% and elementary word problems from 14% to 67%.76% similarUnverifiedGPT-6 Astra raised the MATH-500 score from 7.9% to 48.1% under the same sampling settings by natively editing 0.016% of the weights of Qwen3.5-0.8B.73% similarUnverifiedBy editing an extremely small portion of the weights, Astra raised QUEN 3.5 0.8b's solve rate on math problems from roughly 1 correct out of every 13 problems to close to half.71% similarUnverifiedOn the held-out compound recipe test set, the LoRA fine-tuned model caught 66%, while the base Qwen3.5-4B model caught 78%.67% similarUnverifiedTesting with the Qwen model yielded consistent conclusions, with the token approach achieving 67.0% accuracy versus 64.8% for real data.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1136676API
curl https://kongchang.com/api/v1/knowledge/claims/1136676MCP
get_claim(id=1136676)