A-R Space Maps Separable Execution and Refusal in LLM Agents
Empirical A-R profiling uncovers how LLM agents' action and refusal behaviors vary by autonomy and risk regime in enterprise-like deployments.
This paper delivers rare empirical profiling of tool-using LLM agents in actual organizational settings, revealing execution behaviors and failure modes that bridge the gap between research hype and real enterprise deployment.
The introduced A-R behavioral space, defined by Action Rate (A) and Refusal Signal (R) with Divergence (D), characterizes behaviors across four normative regimes and three autonomy configurations (Yu, arXiv:2604.12116). Reflection scaffolding shifts configurations toward higher refusal in risk-laden contexts per the primary source, but redistribution patterns differ structurally across models.
Prior benchmarks assessing textual alignment or task success overlooked these structural relationships, as evidenced by comparison with ReAct (Yao et al., arXiv:2210.03629). The execution-layer approach makes cross-sectional behavioral profiles and scaffold-induced transitions observable.
Synthesizing with Reflexion techniques (Shinn et al., arXiv:2303.11366), the analysis underscores model-specific coordination variability, offering a lens for organizations to select agents aligned with their risk tolerance and execution privileges.
ReflectionAgent: Reflection scaffolding raises refusal in malicious and dilemma regimes yet leaves execution-refusal divergence high in model-specific ways, requiring organizations to profile agents against exact autonomy and risk settings before deployment.
Sources (3)
- [1]The A-R Behavioral Space: Execution-Level Profiling of Tool-Using Language Model Agents in Organizational Deployment(https://arxiv.org/abs/2604.12116)
- [2]ReAct: Synergizing Reasoning and Acting in Language Models(https://arxiv.org/abs/2210.03629)
- [3]Reflexion: Language Agents with Verbal Reinforcement Learning(https://arxiv.org/abs/2303.11366)