In the strongest example of "weight space programming", as the technique is called, Astra’s intervention raised Qwen3.5-0.8B’s MATH-500 score from 7.9% to 48.1% while touching approximately 0.016% of ...
Scores jumped without learningOn November 2, an experiment published on a research blog by a team at MIT (Massachusetts ...
Conclusion: The 0.016% weight edit fixed "premature termination due to thinking loops," and even unedited models can reach an ...
MIT CSAIL researchers say GPT-6 Astra edited 0.016% of Qwen3.5-0.8B weights to fix reasoning loops, lifting MATH-500 from 7.9% to 48.1% without training.
Copyright (c) 2015-2025 The Brenwill Workshop Ltd. MoltenVK provides the ability to configure and optimize MoltenVK for your particular application runtime ...
GPT-6 Astra edited Qwen’s weights and raised MATH-500 from 7.9% to 48.1% in author-run tests. How weight-space programming works—and its limits.
A new MIT CSAIL blog post is making the rounds in AI circles this week, and it describes something that sounds almost ...
Astra identifies game policies from raw weight matrices and fixes reasoning loops in Qwen3.5-0.8B by editing 0.016% of parameters, lifting MATH-500 scores from 7.9% to 48.1% without training loops.
The MIT CSAIL team has proposed a "weight-space programming" method that enables GPT-6 Astra to directly read and modify ...
Article by Shenchao TechFlowNo training required, no data needed, no gradient descent.Researchers at MIT fed the weights of ...
MIT CSAIL's "weight-space programming" has a frontier LLM read and edit neural network weights directly, with no gradient descent. Surgical edits ended Qwen's reasoning loops, but the results come ...
Simon Willison ran the 17-gigabyte open-weight model locally and fed it floating-point arrays from four reinforcement-learning policies. Without calculators or code execution, Qwen identified Breakout ...
Some results have been hidden because they may be inaccessible to you
Show inaccessible results