AI Trend Research
EN

← November 3, 2026 briefing

P1 Research Paper Language modelsSafety & alignmentEfficiency & infraMITOpenAIAlibaba

GPT-6 Astra reportedly edits just 0.016% of Qwen weights without training, lifting MATH-500 from 7.9% to 48.1% (community report)

Announced November 2, 2026

Verified (fact)

A research blog post, 'Sparks of Weight-Space Programming,' published early experimental results with GPT-6 Astra, saying it can directly read and edit neural network weights without gradient descent or a training loop. By changing only 0.016% of Qwen3.5-0.8B's weights, it reduced repetitive reasoning; with sampling settings unchanged, MATH-500 scores reportedly jumped from 7.9% to 48.1% and IFEval from 21.0% to 62.3%. It identified which game or language small neural networks had learned by inspecting only their weights, without running them. In a language model with its input and output layers removed, it reportedly located the attention heads responsible for copying text. Community posts described the researchers as MIT-affiliated, but the affiliation cannot be confirmed from public sources.

Needs fact-check (claim)

The researchers claim a small neural network for MinAtar games, written in one shot without environment feedback, played better than a network trained with PPO for 100 million steps.

Why it matters (interpretation)

If LLMs can directly interpret and edit weights, it would open new paths for both optimization and interpretability research, and could intensify safety discussions about models modifying themselves.

Confidence medium · official source pending

Sources

Story thread

  1. 2026-11-03 GPT-6 Astra reportedly edits just 0.016% of Qwen weights without training, lifting MATH-500 from 7.9% to 48.1% (community report)
  2. 2026-11-03 Weight-Space Programming follow-up: critics say the Qwen improvement relied on Python execution; partly reproduced with Qwen 3.8 27B on a laptop (community reports)

More about MIT