Researchers have shown that large language models (LLMs) can directly read and edit the weights of another model, a capability the article describes as a new recursive self-improvement capability. In a new paper co-authored by Alex Zhang, a sufficiently smart model was able to identify what a network does from its weights, fix bugs in one step given failed traces, and write complete network weights from scratch that beat some independently trained models.
The findings build on what AI researcher Andrej Karpathy earlier called 'autoresearch', the prospect of training a model autonomously without human intervention. That is usually done through an involved process of writing code and monitoring training runs. The new method, dubbed weight space programming, treats neural network weights as raw data that an LLM can interpret and manipulate directly, akin to reading a compiled program's machine code rather than its source language.
The source notes that improvement in a general direction over 50 steps could be overstated and that the approach is right now unlikely to scale to more advanced tasks than simple games and RL policies. An independent reproduction seems to have already found the capability.
The paper marks the first concrete step toward using LLMs themselves to improve other models without human-written code, a capability the article says was, in theory, bound to happen if LLMs can fit and simulate everything given enough parameters.
Analysis
Why This Matters
- This represents the first concrete demonstration of a key mechanism for recursive self-improvement in AI, where a model directly enhances another model without human-coded training pipelines.
- It could drastically reduce the cost and time of training, as weight editing may be far more efficient than gradient descent in certain cases.
- The approach also opens a novel angle for mechanistic interpretability, allowing models to 'read' what raw weights represent rather than inferring it.
Background
Weight space programming builds on the idea of autoresearch, popularised by AI researcher Andrej Karpathy, in which a sufficiently advanced model runs its own experiments to train another model without human oversight. Previous approaches required writing PyTorch or JAX code, tracking loss curves, and optimising kernels. The new method cuts that loop by having the model operate directly on the weight data. The paper was co-authored by researcher Alex Zhang and colleagues, and follows months of speculation that such a capability would eventually emerge as models grow more capable.
What to Watch
- Whether the technique can be scaled to networks larger than toy tasks, and whether it can match or beat traditional training methods.
- Independent validation and further reproductions: the claimed results must hold up under scrutiny.
- Adoption by frontier labs: if weight space programming is integrated into production training, it could accelerate the pace of AI improvement significantly.
