Meta FAIR Paris published research on March 10, 2026, introducing neural debuggers—language models that support interactive debugging operations matching how developers actually work with code. The paper 'Towards a Neural Debugger for Python' (arXiv 2603.09951) by Maximilian Beck, Jonas Gehring, Jannik Kossen, and Gabriel Synnaeve extends traditional neural interpreters to handle stepping through functions, setting breakpoints, and inspecting program state.
Neural Debuggers Support Interactive Execution Control
While traditional neural interpreters predict line-by-line execution of entire programs, neural debuggers enable selective execution control through debugger actions. The models support stepping into, over, and out of functions; setting breakpoints at specific lines; and inspecting or modifying program state. This architecture mirrors how developers actually debug code rather than executing programs start-to-finish.
Training on 150 Billion Tokens Achieves 90% Accuracy
Researchers fine-tuned Meta's CWM-32B model on 50 billion tokens of Python execution traces and trained smaller 1.8B parameter Transformers from scratch on 150 billion tokens. Training data included execution traces with annotated debugger actions. The fine-tuned CWM-32B model achieved greater than 90% accuracy on forward next-state prediction across all debugger actions. Testing on the CruxEval benchmark demonstrated strong performance on both output and input prediction tasks.
Models Predict Both Forward and Inverse Execution
The neural debuggers support both forward execution—predicting future states and outputs—and inverse execution, which infers prior states or inputs. The models perform conditional execution modeling based on debugger actions, enabling them to simulate different debugging scenarios and execution paths.
Applications Target Agentic Coding Systems
The paper outlines future applications including world models for simulated debugging environments, execution feedback for agentic coding systems, enabling AI agents to interact with real debugging tools, and automated debugging and program understanding. The research was conducted during Maximilian Beck's internship at Meta FAIR Paris in summer 2025.
Key Takeaways
- Meta's neural debuggers extend neural interpreters with interactive debugging operations: stepping, breakpoints, and state inspection
- Fine-tuned CWM-32B achieves >90% accuracy on forward next-state prediction across all debugger actions
- Models trained on 50-150 billion tokens of Python execution traces with annotated debugger actions
- Neural debuggers support both forward execution (predicting outputs) and inverse execution (inferring inputs)
- Applications include world models for debugging, execution feedback for AI agents, and automated program understanding