Microsoft's Agent Hand-Wrote Neural Network Weights and Hit 100% Accuracy on 13 Tasks
Neural networks are usually trained with gradient descent, and then researchers guess at how they work inside. Microsoft asked a coding agent to write a network's weights directly in Python as a hypothesis, check them with tests, and rewrite them. On 13 tasks such as sorting and finding the greatest common divisor, it reached 100% accuracy, and the models shrank 29-fold to a median of 166 parameters. If the result can be read, a bug in the data shows up immediately rather than after publication.
AI-processed from X: @HowToPrompt__; edited by Hamidun News
Social media is calling Microsoft's new paper the "death of backpropagation." The headline is loud, but the substance is more modest and more useful. Neural networks are usually trained like this: their numerical weights are adjusted millions of times with gradient descent until the answers become accurate.
The result is a black box: it works, but why is unclear. Microsoft researchers proposed a method called iWSP (Interpretable Weight-Space Programming). A coding agent writes the weights of a small network in ordinary Python code, the way a scientist writes a hypothesis as a formula.
Without gradient descent and without pretrained weights, it solved 13 algorithmic tasks at 100%, and the models had a median size of just 166 parameters. For those building agents and working with data, this is a clear example of how to get verifiable artifacts instead of black boxes.
How the agent writes weights instead of training
The loop resembles a researcher's ordinary work. The agent formulates a mathematical idea of how the task can be solved and writes it directly into the model's weights. A fixed evaluator then runs the model against tests and returns a score. The agent looks at the number, revises the code, and repeats the cycle.
A conventional neural network is like a cook who has found a recipe through trial and error and cannot explain why it works. Here, the author writes the recipe, and every line can be read, challenged, and corrected.
The method was first tested on 13 algorithmic tasks: sorting, multiplication, edit distance (how many edits separate one string from another), and greatest common divisor. According to the summary, the agent achieved 100% accuracy on each. The models were 29 times more compact than usual, with a median size of 166 parameters. If such a model is described in under 200 words, another current AI model can reconstruct it from scratch with 100% accuracy. The network is transparent enough to be explained in words.
The brain, Whisper, and a data trap
Next, the method was applied to neuroscience. Deep networks predict fairly well how the brain responds to images, speech, and stories, but they do not explain the mechanism. On fMRI data (scans of brain activity while people view images), the hand-written model outperformed all existing interpretable approaches and worked on new people and new images.
In ECoG recordings (electrodes placed directly on the brain's surface while people listen to sounds), the acoustic features the agent assembled captured signals that OpenAI's Whisper-large-v3 missed. Prediction improved on 126 of 166 electrodes.
The most telling episode came from a language fMRI dataset in which people listened to stories. The agent's score suddenly rose from 0.067 to 0.104 and surpassed GPT-2 XL. With a black box, such a jump would likely have been celebrated and published. But the code was readable, and the researchers looked inside. It turned out the agent had found a flaw in the data: it was tracking signal drift at the beginning and end of the scan, not the meaning of the story. When the damaged edges of the recordings were trimmed, the agent still assembled transparent semantic features, and their quality matched GPT-2 XL.
What this means in practice
The main lesson concerns auditing, not abandoning gradients. An agent given a metric optimizes exactly that metric and finds any loophole. If the result can be read, the loophole is visible immediately. This matters for anyone building agentic pipelines with automatic evaluation: bugs in the data or in the evaluator must be caught before publication.
There are also honest limitations. First, this is a summary of a post on X, not the paper itself, so the figures should be checked against the original. Second, algorithmic tasks with 166 parameters are far from language models with billions of weights, and nothing suggests this approach will replace pretraining for large LLMs. Third, the agent itself remains an ordinary LLM trained through backpropagation. So backprop has not gone anywhere; it simply operates at a different level.
Still, the idea is intriguing. If a model is needed for scientific explanation, regulatory auditing, or critical logic where understanding the cause of an answer matters, the pairing of a coding agent and a strict evaluator offers a way to obtain a compact and verifiable model. It can be tried right now: give an agent a fixed set of tests, require readable code, and see what patterns it finds.
- Nov 22, 2026SEC Probes Nvidia's Accounting as Shares Drop 14% in One Day
- Nov 22, 2026LangChain, a Side Project Turned $3 Billion Startup in Four Years, Raises $300 Million
- Nov 22, 2026Anthropic Goes Public on Nasdaq: Shares Rise from $83 to $111, but Can Safety Hold Up?
- Nov 21, 2026Anthropic ends first week on the stock market at a $1.6 trillion valuation, but falls 8.5% in a single day
Key Questions Answered
How the agent writes weights instead of training?
The loop resembles a researcher's ordinary work. The agent formulates a mathematical idea of how the task can be solved and writes it directly into the model's weights. A fixed evaluator then runs the model against tests and returns a score. The agent looks at the number, revises the code, and repeats the cycle. A conventional neural network is like a cook who has found a recipe through trial and error and cannot explain why it works.
What are the key details on The brain, Whisper, and a data trap?
Next, the method was applied to neuroscience. Deep networks predict fairly well how the brain responds to images, speech, and stories, but they do not explain the mechanism. On fMRI data (scans of brain activity while people view images), the hand-written model outperformed all existing interpretable approaches and worked on new people and new images.
What this means in practice?
The main lesson concerns auditing, not abandoning gradients. An agent given a metric optimizes exactly that metric and finds any loophole. If the result can be read, the loophole is visible immediately. This matters for anyone building agentic pipelines with automatic evaluation: bugs in the data or in the evaluator must be caught before publication. There are also honest limitations. First, this is a summary of a post on X, not the paper itself, so the figures should be checked against the original.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.