MIT researchers report that OpenAI's GPT-6 Astra can natively read and modify neural network weights, raising a small model's math benchmark score sixfold without gradient-based training3,5.
The work, described by researcher Rikiya Takehi under the label "Weight-Space Programming," demonstrates that GPT-6 Astra ("Astra") edited just 0.016% of the parameters inside Alibaba's Qwen3.5 0.8B ("Qwen3.5 0.8B") to lift its MATH-500 accuracy from 7.9% to 48.1%. The same weight correction generalized beyond math: instruction-following performance on IFEval rose roughly threefold2.
How the edit worked
Astra used no training loops, no environment feedback, and no gradient computation. Instead, the researchers report, it reasoned about what each part of the target network computes and altered weights accordingly. In the Qwen3.5 0.8B case, Astra identified an issue in the smaller model's weights that caused reasoning loops, a problem the model's own model card had flagged, and corrected it by modifying the narrow slice of parameters.
Beyond weight editing, the paper reports that Astra can determine what a network was trained to do solely by inspecting raw weights, and can write small game-playing networks from scratch that outperform networks trained with PPO for 100 million steps.
Skepticism on the "superhuman" framing
Not all observers accept the paper's strongest claims. Sean Goedecke, writing on his personal site, argued that the results do not demonstrate superhuman capability4. He contended that a skilled human ML researcher could perform the same tasks, noting that the game-identification experiments benefited from labeled input channels that map directly to an 8×8 grid, making the inference less opaque than it first appears.
The researchers themselves frame the capability more cautiously, stating that frontier LLMs perform these tasks "well past what we'd expect from humans or conventional algorithms".
Recursive-improvement implications
Goedecke flagged the safety-relevant dimension: the reason to pay attention, he wrote, is that superhuman ability to improve language models is precisely the kind of capability that could cause recursive self-improvement. The paper's own framing echoes this, noting that throughout deep learning history, optimization has relied on simple symbolic optimizers such as Adam or Muon, and weight-space programming could change that dynamic.
ANALYSIS Weight-space programming, if it scales, would represent model improvement achieved not through additional compute or data but through direct parameter surgery by a more capable model. Separately, GPT-6 Astra has already figured in distillation controversies involving DeepSeek and Moonshot AI[3], and this paper adds a qualitatively different use of the same model, one where Astra acts not as a training-data source but as an optimizer.
The MATH-500 jump from 7.9% to 48.1% on Qwen3.5 0.8B was achieved by touching 0.016% of the model's weights. Whether the technique transfers to larger models or harder benchmarks is not addressed in the available reporting.