Can Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool Use
06:00 · July 2, 2026 · arXiv cs.AI RSS
While Large Language Model (LLM) agents demonstrate proficiency in static benchmarks, their deployment in real-world scenarios is hindered by the dynamic nature of user queries, tool sets, and interaction dynamics. To address this generalization gap, we formalize OpenAgent (Tool-Use Agent in Open-World), a problem setting characterized by distributional shifts across query, action, observation, and domain dimensions. To systematically diagnose its impact, we construct a controlled sandbox environment where we define fine-grained environmental shifts across a four-tier hierarchy, Perception, Interaction, Reasoning, and Internalization, and conduct a comprehensive series of experiments. Our analysis yields a series of key insights, demonstrating that agents trained via both Supervised Fine-Tuning(SFT) and Reinforcement Learning suffer from varying degrees of performance degradation when confronting open environmental shifts. Building on these insights, we propose Perturbation-Augmented Fine-Tuning, a disturbance-based intervention strategy for SFT that lays the foundation for enhancing agent robustness and utility in realistic environments. Our code will be released at: https://github. com/LAMDA-NeSy/OpenAgent.
Summary
Large Language Model agents achieve strong results on static tool-use benchmarks yet encounter sharp performance drops once deployed in non-stationary environments where user queries, available tools, interaction outcomes, and task domains can all change. The paper formalizes this setting as OpenAgent, a problem defined by distributional shifts along four axes: query intent, action space, observation dynamics, and domain. To isolate these effects from the noise of live APIs, the authors built a controlled sandbox that injects perturbations at four diagnostic levels—Perception, Interaction, Reasoning, and Internalization—while preserving a clean closed-set baseline.
Experiments with both Supervised Fine-Tuning and Reinforcement Learning agents show consistent degradation under these shifts. SFT models tend to overfit training trajectories and anchor too rigidly to surface-level symbols, while RL agents, though better at semantic grounding, suffer from boundary blindness induced by reward structures that emphasize goal completion over robustness. These failure modes compound along multi-step trajectories because an early misstep alters subsequent observations and policy decisions.
To address the observed fragility, the authors introduce Perturbation-Augmented Fine-Tuning, a data-centric intervention that injects controlled observation anomalies and symbolic noise into SFT trajectories. The resulting models exhibit improved resilience to open-world shifts without sacrificing performance on the original static tasks. The work provides both a reproducible evaluation framework and concrete evidence that current post-training paradigms remain insufficient for realistic tool-use deployment.
Why it matters
This research is highly relevant for Dutch AI researchers and developers building autonomous agents, as it addresses critical robustness and generalization challenges in real-world tool use. Improving agent reliability aligns with the EU's focus on trustworthy AI, making the proposed fine-tuning strategies actionable for enterprise AI deployments in the Netherlands.
