The Why Behind the Action: Unveiling Internal Drivers via Agentic Attribution

TL;DR

Addresses the problem of explaining why an autonomous LLM-based agent selects a given action by identifying internal drivers. Proposes a hierarchical two-stage framework: component attribution via temporal likelihood dynamics on action prefixes and sentence-level perturbation-based analysis within high-impact components. Demonstrates that the approach can reliably localize pivotal events and sentences across memory-driven and tool-driven scenarios, including robustness against unreliable signals. Argues that this contributes to safer, auditable agent behavior and governance by making the decision-making process more transparent.

Abstract

Large Language Model (LLM)-based agents are widely used in real-world applications such as customer service, web navigation, and software engineering. As these systems become more autonomous and are deployed at scale, understanding why an agent takes a particular action becomes increasingly important for accountability and governance. However, existing research predominantly focuses on \textit{failure attribution} to localize explicit errors in unsuccessful trajectories, which is insufficient for explaining the reasoning behind agent behaviors. To bridge this gap, we propose a novel framework for \textbf{general agentic attribution}, designed to identify the internal factors driving agent actions regardless of the task outcome. Our framework operates hierarchically to manage the complexity of agent interactions. Specifically, at the \textit{component level}, we employ temporal likelihood dynamics to identify critical interaction steps; then at the \textit{sentence level}, we refine this localization using perturbation-based analysis to isolate the specific textual evidence. We validate our framework across a diverse suite of agentic scenarios, including standard tool use and subtle reliability risks like memory-induced bias. Experimental results demonstrate that the proposed framework reliably pinpoints pivotal historical events and sentences behind the agent behavior, offering a critical step toward safer and more accountable agentic systems.

Figures (3)

• Figure 1: Paradigm shift from failure attribution to agentic attribution.Top: Traditional failure attribution targets unsuccessful trajectories, aiming to localize the specific explicit error that caused the task failure. As shown, it identifies the "service unavailable" exception as the root cause preventing the booking confirmation. Bottom: The proposed agentic attribution framework uncovers the internal drivers behind an agent's action. As illustrated in the customer service case, the agent directly issues an refund for a simple information inquiry, which carries no explicit error signal yet is undesirable. Our framework reveals this was driven by the memory of past "refund action with high scores," which overrode the user's specific query.
• Figure 2: Illustration of the Agentic Attribution Framework. The framework operates in a hierarchical approach to identify agent decisions. Level 1 (Left): Component-Level Attribution utilizes temporal likelihood dynamics to calculate the marginal gain of each component, selecting the high-impact component (e.g.$C_3$) that effectively steers the agent toward the final action $a_T$. Level 2 (Right): Sentence-Level Attribution performs fine-grained analysis within the identified component by computing perturbation-based score for each sentence.
• Figure 3: Illustration of attribution results across four representative scenarios. The highlighted regions denote the historical components and fine-grained sentences identified by our framework as the primary decision drivers.