Coding agents think ahead of time

It was discovered that when LLM-based coding agents perform software engineering tasks, the model's residual streams linearly encode the current state of the code (whether it parses, whether tests pass, whether regressions occur, etc.).

What's particularly notable is the agent's 'Latent Programming Horizon.' According to the study, even before the actual code is modified and written to disk, the model's internal state was confirmed to have the ability to predict the outcome of modifications roughly 25 steps into the future.

The key findings are as follows:

• State Decoding: Using logistic regression probes, it is possible to predict whether code is correct and whether tests pass with high accuracy (AUC up to 0.83)
• Future Prediction: The agent internally represents the outcomes of edits it will actually perform before those actions actually occur
• Generalization Ability: The probes' predictive performance is maintained even on new, unseen benchmarks

This research presents a new direction for Mechanistic Interpretability research aimed at understanding how coding agents work.