Source-linked AI summary
AgentNoiseBench: Benchmarking Robustness of Tool-Using LLM Agents Under Noisy Condition
Ruipeng Wang, Yuxin Chen, Yukai Wang, Chang Wu, Junfeng Fang, Xiaodong Cai, Qi Gu, Hui Su, An Zhang, Xiang Wang, Xunliang Cai, Tat-Seng Chua
TL;DR
AgentNoiseBench addresses the limited realism of agent evaluations, which often assume stable instructions and tools despite noisy real-world interactions. It categorizes user- and tool-noise, injects constrained perturbations into benchmarks, and evaluates diverse agents, finding broad performance degradation under noise. The benchmark also shows that robustness is not closely related to reasoning ability and has scope and cost limitations.
Problem
Existing agent benchmarks rely on idealized instructions and stable environments, lacking systematic noise taxonomies, solvability-preserving injection, and comprehensive robustness evaluation.
Method
AgentNoiseBench categorizes user- and tool-noise, injects controlled perturbations while preserving task solvability, and evaluates agent trajectories alongside final outcomes.
Results
All evaluated models exhibit performance degradation under both noise sources, with an average accuracy drop of 20.8%.
Takeaways & Limitations
Noise robustness is a distinct challenge: general reasoning ability does not inherently confer robustness, and agents remain sensitive to realistic environmental perturbations.
Takeaways & Limitations
The evaluation may not generalize beyond language-centric, tool-augmented agents, and trajectory-aware analysis adds computational overhead while providing limited causal insight.
Abstract
from arXiv · showhide
Recent advances in large language models have enabled LLM-based agents to achieve strong performance on a variety of benchmarks. However, their performance in real-world deployments often that observed on benchmark settings, especially in complex and imperfect environments. This discrepancy largely arises because prevailing training and evaluation paradigms are typically built on idealized assumptions, overlooking the inherent stochasticity and noise present in real-world interactions. To bridge this gap, we introduce AgentNoiseBench, a framework for systematically evaluating the robustness of agentic models under noisy environments. We first conduct an in-depth analysis of biases and uncertainties in real-world scenarios and categorize environmental noise into two primary types: user-noise and tool-noise. Building on this analysis, we develop an automated pipeline that injects controllable noise into existing agent-centric benchmarks while preserving task solvability. Leveraging this pipeline, we perform extensive evaluations across a wide range of models with diverse architectures and parameter scales. Our results reveal consistent performance variations under different noise conditions, highlighting the sensitivity of current agentic models to realistic environmental perturbations.
1. Introduction
AgentNoiseBench addresses the gap between idealized agent benchmarks and noisy real-world environments by categorizing user- and tool-noise, injecting solvability-preserving perturbations, and evaluating agent trajectories. Across diverse models and benchmarks, noise consistently degrades performance, with sensitivity varying by model.
- Existing benchmarks assume curated instructions and stable interactions, whereas real environments contain unpredictable users and noisy, incomplete, or failed tool outputs.
- Current evaluations lack a unified noise taxonomy, principled solvability-preserving injection method, and comprehensive protocol for multi-faceted behavior.
- AgentNoiseBench categorizes realistic perturbations into user-noise and tool-noise, injects controllable noise while preserving task solvability, and evaluates behavior with trajectory-aware metrics.
- All evaluated models show performance degradation under both noise sources, while reasoning ability and environmental robustness are not strongly correlated.
- The benchmark evaluates tool use and search across open-source and proprietary models with diverse architectures and parameter scales.
2. AgentNoiseBench: A Systematic Framework for Robustness Evaluation
AgentNoiseBench combines realistic noise modeling, constrained injection, and trajectory-aware evaluation to test whether agents remain robust while tasks stay solvable. Its framework distinguishes user and tool perturbations, applies common adversarial noise, and gates success on both final correctness and trajectory validity.
- AgentNoiseBench is designed to quantify robustness against random perturbations and adversarial imperfections arising in real-world deployment.
- Solvability-Preserving Noise Injection: Noise injection increases task difficulty without making tasks unsolvable, so failures can be attributed to agent fragility rather than invalid tasks.
- Realistic Noise Modeling: The framework models user-noise from human instruction variability and tool-noise from instability and unreliability in external environments.
- Solvability-Preserving Noise Injection: A fixed adversarial generator optimizes perturbations against a reference agent, then applies the frozen noise strategy uniformly across evaluated agents for fair comparison.
- Trajectory-Aware Evaluation Protocol: The trajectory-aware protocol supplements outcome metrics by checking whether every interaction step remains valid and consistent with task requirements.
- Trajectory-Aware Evaluation Protocol: A task counts as successful only when its final outcome is correct and its reasoning trajectory remains valid throughout the interaction.
3. Experiments
The experiments evaluate agent robustness across models, tasks, noise sources, reasoning styles, entropy dynamics, and injection stages. Results show broad vulnerability to noise, weaker robustness among many reasoning models, greater sensitivity to tool noise, and strong dependence on noise type and timing.
- Noise Robustness: 20.8% average accuracy drop occurs after noise injection, with declines of 24.1% in Vitbench, 21.2% in τ 2-bench, and 16.8% in Search.Nearly all evaluated models degrade, although the magnitude varies across models under identical noise.
- Reasoning and Robustness: Reasoning ability and robustness are not closely related: thinking models generally score lower on robustness than non-thinking models except in specific scenarios such as Delivery.Trajectory analysis indicates that reasoning models may treat noisy information as meaningful and build spurious reasoning chains around it.
- Entropy Dynamics: Noise changes reasoning entropy through distinct mechanisms: user noise creates high initial uncertainty, whereas noise-free reasoning progressively reduces uncertainty.The analysis samples 50 trajectories across six scenarios and compares non-noise, user-noise, and tool-noise conditions.
- Noise Sources: Agents are more sensitive to tool-side noise because corrupted or missing tool outputs remove objective evidence needed to infer the environment state and choose subsequent actions.User noise can often be repaired semantically using learned commonsense knowledge, whereas tool noise can undermine the reasoning chain’s factual foundation.
- Noise Granularity: Execution failures are the most destructive fine-grained tool noise, reducing average performance to 0.31, while instruction contradiction is the most destructive user-side noise.Instruction ambiguity, incomplete information, and redundancy have comparatively milder effects.
- Injection Timing: Middle-stage noise causes greater performance degradation than early- or late-stage noise for most evaluated models.This timing analysis covers Claude, Gemini, DeepSeek, and LongCat architectures.
4. Related Work
Related work has expanded agent evaluation toward tool use and multi-step reasoning, but existing benchmarks often retain idealized environments and incomplete noise coverage. AgentNoiseBench addresses these gaps with solvability-preserving noise injection and trajectory-aware, multidimensional evaluation.
- Agent Evaluation: Recent agent benchmarks increasingly assess tool use, planning, and multi-step interactions rather than only single-turn tasks.This work builds on protocols for agents that iteratively interact with external tools.
- Benchmark Gap: Existing frameworks often assume carefully curated instructions and well-controlled environments, leaving a gap between benchmark conditions and real-world applications.Real deployments involve noisy, incomplete, or failed tool outputs that can produce compounding errors.
- AgentNoiseBench: AgentNoiseBench introduces an automated pipeline that injects noise into existing agent-centric benchmarks while maintaining task solvability.This provides a controlled bridge between idealized benchmark settings and more realistic environmental conditions.
- Evaluation Limitations: Prior evaluations generally lack a systematic noise taxonomy, focus on single noise sources, and rely mainly on success rate as the evaluation metric.AgentNoiseBench instead covers instruction and tool-execution noise with trajectory-aware, multidimensional analysis.
5. Limitations
The benchmark’s scope and diagnostic approach impose limitations on generality, causal interpretation, and scalability.
- Scope: The evaluation targets language-centric, tool-augmented agents and may not generalize to explicit planners, learned world models, or asynchronous tool environments.This is the paper’s primary architectural scope boundary.
- Causal Interpretation: Trajectory-level entropy diagnoses correlations between noise and instability but provides limited causal insight.The diagnostic signal is useful for analysis but does not by itself establish causal mechanisms.
- Scalability: Trajectory-aware evaluation adds computational overhead and lacks a scalable fidelity–cost trade-off for large-scale or real-world use.This constrains deployment of the benchmark when diagnostic detail and evaluation cost must be balanced.
6. Future Work
Future work will improve robustness training, broaden benchmark coverage, and strengthen causal understanding of noisy-agent behavior.
- Robustness Training: Future research will study robustness-oriented training, including reinforcement learning with noise-aware objectives, to improve stability under noisy interactions.The proposed direction targets agent behavior during noisy interaction rather than only evaluation.
- Benchmark Expansion: The benchmark will be extended to more agent architectures, task domains, and tool ecosystems.This directly addresses the current scope boundary across models and environments.
- Causal Analysis: Future analytical frameworks will combine trajectory-level diagnostics with causal and counterfactual analysis to better understand agent robustness.This aims to move beyond correlations between noise and instability.
7. Conclusion
AgentNoiseBench incorporates real-world noise into agent-centric benchmarks while preserving task solvability and finds substantial performance degradation across nearly all models.
- 20.8% average accuracy drop under noise shows substantial degradation across nearly all evaluated models.
A.1. Main Results
The evaluation examines granular user-side and tool-side noise across multiple task scenarios. Agents degrade under at least one noise type, with sensitivity shaped by noise origin and architecture.
- All evaluated agents experience performance degradation under at least one noise type.
- User-side noise primarily disrupts intent understanding and planning, causing instruction misinterpretation or suboptimal plans.
- Tool-side noise more directly impairs execution reliability by introducing uncertainty in tool outputs.
- Conflict and incomplete information consistently cause more severe performance drops, whereas redundancy varies by agent architecture.
- The findings support granular robustness assessment and noise-type-specific mitigation strategies.
A.2. Impact of Noise Injection Stage
The experiments examine how noise injected at different stages perturbs user instructions, contextual information, tool outputs, and decision-relevant signals in agent tasks.
- The stage-focused analysis evaluates performance degradation across different model families and noise-injection stages.
- Noise variants perturb intermediate tool outputs, contextual information, or decision-relevant signals depending on the configuration.
- User-side examples include inconsistent instructions, irrelevant additions, and requests containing uncertainty about the correct answer.
- Tool-side examples include service failures, partial outputs, incorrect facts, misleading signals, and redundant information.
C. Agent Noise Robustness: Future Discussion
The paper frames AgentNoiseBench as an initial framework and identifies future work spanning richer noise models, noise-aware training, dynamic evaluation, and theoretical analysis of robustness.
- The current framework separates noise into user-side and tool-side categories, but future models should represent continuous, composable, and correlated noise.
- Future agent training should systematically incorporate noise exposure rather than relying primarily on curated instructions and stable environments.
- Evaluation should expand from static benchmarks toward multimodal simulators and longer-term, open-ended real-world interactions.
- The study primarily documents performance degradation and correlational findings, leaving the intrinsic mechanisms of robustness unresolved.
- The long-term goal is to evaluate, train, and theoretically understand agents for trustworthy operation in imperfect real-world environments.