MIT CSAIL: GPT-6 Astra edits 0.016% of Qwen3.5-0.8B weights — MATH-500 7.9% → 48.1%
Major
Primary
On November 2, 2026, MIT CSAIL researchers published Sparks of Weight-Space Programming: frontier LLMs natively reading and writing neural-network weights. With the same sampling settings and no training loops, they report GPT-6 Astra raising Qwen3.5-0.8B MATH-500 from 7.9% to 48.1% by editing 0.016% of the weights, with IFEval rising about 3×.
At a glance
Figures are from the MIT CSAIL authors’ November 2, 2026 Weight-Space Programming post.
Lab
MIT CSAIL
Date
November 2, 2026
MATH-500
7.9% → 48.1% after Astra weight edit (authors)
Weights edited
0.016% of Qwen3.5-0.8B (authors)
IFEval
About 3× after the same edit (authors)
Training loops
Disallowed for Astra in the reported setup
How to read this claim
Measurement — Weight-Space Programming: Astra edits 0.016% of Qwen3.5-0.8B — MATH-500 7.9% → 48.1% (MIT CSAIL authors’ figure; not a fine-tune, not…
Bill — A score is not a price. If you will call a model next, list input and output per million tokens on the vendor page.
Door — How far Weight-Space Programming scales beyond the reported model sizes and tasks is not established by this post alone.
MIT CSAIL published Weight-Space Programming results where GPT-6 Astra edits Qwen3.5-0.8B weights (0.016%) and raises MATH-500 from 7.9% to 48.1%.
What it introduces
Weight-Space Programming — LLMs natively reading and writing neural-network weights.
Who benefits Anyone quoting “Astra fixed Qwen’s math by editing weights,” and anyone conflating weight-file audits (synthetic tensors on Hugging Face) with this kind of LLM weight editing.
Try this
When a slide says MATH-500 jumped from 7.9% to 48.1%, ask: which base model (Qwen3.5-0.8B), who edited (GPT-6 Astra per the authors), what fraction of weights (0.016%), and whether training loops were disallowed. Separate this from synthetic weight-file claims like Numbat.
What we do not know
How far Weight-Space Programming scales beyond the reported model sizes and tasks is not established by this post alone.
Exact IFEval before/after percentages beyond the stated ~3× rise are not extracted as a full table from the page text used here.
Whether third parties can reproduce the Astra weight edits without the authors’ logging setup is not stated as a one-click public harness.
Evidence
PrimaryWeight-Space Programming post. MATH-500, IFEval, weight-edit share, and PPO comparisons are from the MIT CSAIL authors’ November 2, 2026 Weight-Space Programming post.
Independent MATH-500 and IFEval deltas are the authors’ measurements with GPT-6 Astra. The post is a research report, not a vendor scorecard you can buy as a product.
Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab (MIT CSAIL) published Sparks of Weight-Space Programming on November 2, 2026. They report that GPT-6 Astra can identify what a network was trained to do from raw weights alone, write small game-playing networks that outplay PPO-trained policies after 100 million steps without environment feedback during writing, and — asked to improve Qwen3.5-0.8B at math without gradients — edit 0.016% of the weights to raise MATH-500 from 7.9% to 48.1% under the same sampling settings, with IFEval rising about 3×. The post says Astra is not allowed to write training loops; it reasons about what each part should compute and changes weights accordingly. Logs and an edited Qwen checkpoint are linked from the authors’ pages.
Why it matters
If frontier models can read and patch weights in natural language, optimization and interpretability start to share a door that is not “run Adam longer.” Keep the measurement labeled: these MATH-500 and IFEval deltas are the authors’ Astra experiments on Qwen3.5-0.8B, not a vendor model card.
The post also reports Astra naming which game a 17,253-weight policy was trained on from weights alone, and one-shot edits that raise a failing control policy from 0% to 100% success by changing 3.5% of weights.
Edited weights and experiment logs are linked from the authors’ site (Hugging Face checkpoint and GitHub logs).