Fault Injection and Anomaly-Aware RL: Making CUDA Agents Robust to Hardware and Kernel Failures

TL;DR: What if we trained CUDA Agent–like systems to expect and recover from anomalies—like hardware faults, overflows, or rare performance cliffs—by injecting faults during RL training? Initial experiments could compare robustness and recovery speed to baseline agents when facing unexpected kernel or hardware failures.

Research Question: Can training RL-driven CUDA optimizers with systematic fault injection and anomaly detection lead to agents that are more robust and self-healing in real-world, noisy deployment?

Hypothesis: RL agents exposed to systematic fault injection and anomaly-aware feedback will learn strategies that not only optimize speed, but also enhance robustness and recovery from hardware or code failures.

Experiment Plan: - Setup: Extend the RL framework to inject faults (e.g., memory errors, race conditions, rare hardware bugs) during kernel execution and provide anomaly signals as part of the reward.

• Recovery: Agents are rewarded for both optimizing performance and gracefully recovering from or mitigating failures.
• Data: Use robust-kbench and real-world CUDA kernels.
• Metrics: Evaluate fault tolerance, recovery time, and performance post-fault.
• Expected Outcome: The anomaly-aware agents should maintain higher performance and reliability under adverse conditions than traditional RL or compiler systems.

References:

• 1. Lange, R., Sun, Q., Prasad, A., Faldor, M., Tang, Y., & Ha, D. (2025). Towards Robust Agentic CUDA Kernel Benchmarking, Verification, and Optimization. arXiv.org.
• 2. Dai, W., Wu, H., Yu, Q., et al. (2026). CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation.