MIT experiment claims GPT-6 Astra rewrote Qwen weights to boost scores; Meta researcher says it is exaggerated

On November 2, a technical experiment by a joint team from the Massachusetts Institute of Technology (MIT) set off nearly 6 million views on social networks.

The experimental setup was striking: no training data, no backpropagation, just a pure Python environment, letting OpenAI's frontier large model GPT-6 Astra directly read and rewrite the underlying high-dimensional tensors of the open-source model Qwen3.5-0.8B.

In the end, Astra changed only 0.016% of the 800 million parameters (about 130,000 weights), and the accuracy of Alibaba's Qwen3.5-0.8B on the MATH-500 math benchmark soared from 7.9% to 48.1%. In the latter part of the report, Astra even, without any training samples, directly "hand-wrote" out of thin air the weights of small neural networks capable of playing simplified versions of Atari games and 9×9 Go.

One of the blog's named authors, the well-known Stanford scholar Omar Khattab, then publicly posed a provocative rhetorical question: "If large models can directly write weights, what place is left for traditional methods like gradient descent?"

Tech enthusiasts saw this breakthrough as the starting point for "recursive self-improvement" of models, and even predicted that pre-training clusters costing tens of millions of dollars would be replaced by "weight compilers"; but on the other side, frontline researchers at Meta and major open-source communities quickly poured cold water on it, pointing out that the tweet was "clickbait."

Tech blogger Chubby, who has 150,000 followers, and others believe that Astra demonstrated the ability to diagnose and reconstruct at the level of high-dimensional parameters, meaning that higher-order models can serve as "doctors" for lower-order models, and that this is the cornerstone for achieving automated self-iteration. On social platforms, the narrative of "AI directly compiling AI" quickly captured the public's attention.

However, rigorous code archaeology soon brought a backlash.

Meta core researcher Lucas Beyer said bluntly that the project was "very cool, but oversold." He dug up the system prompt the team had made public and pointed out that the experiment did not let Astra explore blind spots on its own; humans had already given extremely explicit targeted guidance in the prompt—clearly requiring Astra to check whether the model was being truncated because it had fallen into an infinite loop, while also fixing the rules to prohibit opportunistic shortcuts through lightweight engineering tricks such as "terminating thinking early."

Researcher Yunfan Zhang further went through the complete experiment logs. He found that Astra had only done a set of on/off comparisons in long-form reasoning mode (Thinking Process), identified that Qwen's reasoning chain was going in circles, and then carried out targeted removal. "The real version of this story," Zhang quipped, "is far less dramatic than the advertised 'increasing accuracy sevenfold.'"

More engineers wrote long posts saying outright that using a large model to modify weights is like manually disassembling a compiled binary program, "not a superhuman capability," and that a smart algorithm researcher could achieve the same thing by drawing a few activation heatmaps.

Scalpel or Midas touch

In fact, Qwen3.5-0.8B's math score soaring from 7.9% to 48.1% was by no means because it learned advanced mathematics overnight, but rather a precise "circuit troubleshooting": Astra followed the signal flow to understand the internal logic and directly cut the short-circuiting wire that caused it to fall into an infinite loop. The figure below shows Astra gradually changing weights over 50 trial runs; the green line is the accuracy of the modified model on practice problems, and the gray line is the original model.

First, this is a known defect that has been explicitly acknowledged by the official developers. Tongyi Qianwen's official technical documentation had long warned that extremely small models like the 0.8B, if forced to turn on thinking mode, are very likely to fall into endless self-repetition and logical infinite loops, so it is turned off by default officially.

In MIT's experimental environment, however, the model's context limit was stretched to 32,768 tokens, and thinking mode was forcibly enabled. The result was disastrous: when the original Qwen worked on MATH-500, as many as 93% of the problems were not because it "got them wrong," but because it burned all its compute on looping nonsense until the length was exhausted and the system ruthlessly truncated it.

What Astra accomplished over multiple rounds of debugging was essentially feature ablation (Abliteration):

Comparative localization: Astra extracted Qwen's internal activation states during normal reasoning and during infinite loops, and found that in the deeper layers (layers 15 and 19) the model had a stubborn geometric direction that kept pushing signals into a dead end;

Weight projection: Astra selected 64 columns from each of the output weight matrices of these two layers, and forcibly subtracted this direction through a linear algebra transformation;

Unlocking: after the looping direction was eliminated, the model's reasoning chain, which had been locked in long text, was able to conclude smoothly, with the median length settling steadily at about 9,000 tokens.

The detailed data that followed thoroughly exposed the illusion of "intelligence improvement": among 1,152 test problems, of the 579 additional problems answered correctly after the modification, a full 576 had originally scored zero because of timeout truncation; for problems that could already be completed, accuracy after the weight modification showed no fluctuation at all.

The more fatal counter-evidence comes from the control group—Google's Gemma 3 270M. Gemma itself does not suffer from the vicious-cycle problem; its wrong answers are genuinely miscalculated arithmetic and flawed algebraic logic. Faced with a model that has no "infinite-loop structural defect," Astra repeatedly adjusted parameters, yet its math score kept struggling between 6.8% and 7.7%, with no improvement at all.

From this perspective, this is not some alchemy of conjuring a spirit out of thin air at all—it merely fixed an infinite loop along the way.

It is like a piece of code that already had its computation logic written, but got stuck in an infinite loop halfway through running, until the system forcibly timed out and killed the process. Astra did not teach the small model any new problem-solving approach; it simply added a normal exit condition, allowing the computed result to be successfully printed out.

Writing a large model from scratch?

Besides modifying Qwen, another experiment closer to "AI creating AI" was having GPT-6 Astra write neural network weights from scratch.

The team had Astra directly generate weights to beat traditional reinforcement learning (PPO) on three MinAtar Atari games (Breakout, Seaquest, and Asterix), as well as a 9×9 Go network.

In this step, the model skipped all training sets and gradient computation, directly deducing the logic in Python code, and then "compiling" it into weight tensors.

By comparison, even the smallest language model, Qwen3.5-0.8B, has 540 times the parameter count of the Go network and 40,000 times that of the game networks.

Astra's ability to "hand-write" these tiny networks presupposes that the task rules are extremely clean and the mathematical mappings highly deterministic. In Breakout, the first-layer matrix computes the ball's landing point, and the second-layer matrix links to the paddle's displacement, with the logical chain fitting together perfectly; once switched to a game with slightly fuzzier rules, its score immediately plummeted and was overtaken by traditional reinforcement learning models.

However, researchers had already made similar attempts before: having frontier large models hand-write Transformer weights to solve small algorithmic tasks such as sorting and summation, which the models completed extremely well; but as soon as the task switched to natural language understanding, where semantics are full of uncertainty, the so-called "weight generation" would instantly collapse due to evaluation leakage or dimensionality explosion.

So, translating a small game with clear rules into a toy model with twenty thousand parameters is essentially just rigidly translating a few lines of ready-made logic code into matrix format. But this is completely different from directly writing a general-purpose large model—the complex common sense of human language and the real world simply cannot be exhaustively enumerated into a few explicit rules.

What does "reading weights" really mean

Since it cannot replace traditional training, why is this report still worthy of the industry's close attention?

Because this research has torn open a long-standing blind spot in deep learning: the high-dimensional parameters of neural networks are losing their "unreadability."

For decades, humanity has treated deep neural networks like a heavily fogged black box; apart from defining loss functions and relying on stochastic gradient descent (SGD) to grope for derivatives in the dark, we have been helpless before the intricate numerical distributions inside high-dimensional matrices. But MIT's experiment demonstrated a powerful capability of large models: directly reading weights.

If general-purpose large models can directly read weights and identify the functional structures within them, there will be opportunities in the future to further use this for model auditing, backdoor detection, capability editing, and local defect repair. On the one hand is interpretability: without running the model, it may be possible to understand what it is doing from its weights. On the other hand is small-model tuning: fast, low-cost model optimization, or targeting specific domains.

The MIT team also proposed further-reaching ideas, including adding or removing capabilities without relying on large amounts of data, transferring functional modules between different models, and even recursive self-improvement.

These directions have not yet been fully validated by experiments, but they have already sketched out a model R&D approach different from traditional training.

This article comes from the WeChat public account "AI Frontline" (ID: ai-front), authors: Qinghe, Feiye, published by 36Kr with authorization.