Learning to Act under Noise: Enhancing Agent Robustness via Noisy Environments

Researchers introduced NoisyAgent, a training framework that improves LLM agent robustness by injecting simulated user and tool noise during learning.

The method uses gradual noise scaling and partial rollout exposure to help agents adapt to imperfect, stochastic interactions.

NoisyAgent improves performance in noisy environments and also on clean benchmarks, suggesting better generalization.

The work highlights a gap between benchmark success and real-world reliability for deployed AI agents.