acceptodds
Withdrawn from ICLR 2027

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

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

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 failure attribution to localize explicit errors in unsuccessful trajectories, which is insufficient for explaining the reason behind agent behaviors. To bridge this gap, we propose a novel framework for 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 component level, we employ temporal likelihood dynamics to identify critical interaction steps; then at the 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 covering standard tool use and subtle reliability risks like memory-induced bias, as well as on long-horizon trajectories from widely used real-world agent benchmarks (GAIA, WebShop, and ALFWorld). 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.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Agentic Attribution

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GgAJOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Agentic Attribution, a framework that explains why LLM agents take specific actions beyond traditional failure analysis. It uses temporal likelihood dynamics to identify influential trajectory components and perturbation-based analysis to locate key sentences. The framework is evaluated on Llama-3.1-70B-Instruct across agent scenarios and benchmarks (GAIA, WebShop, ALFWorld), showing reliable identification of decision drivers and improving agent transparency and accountability.

Critical strengths
  1. This paper identifies an important problem: Agentic Attribution.
  2. The authors manually annotate and analyze attribution data from 8+53 trajectories.
Critical weaknesses
  1. The proposed Agentic Attribution Framework seems like a straightforward application of previous attribution methods to agent trajectories, such as ContextCite [1] TracLLM [2], TRACE [3], etc. It also adopts similar ideas, including counterfactual / perturbation-based methods and coarse-to-fine attribution. Therefore, I remain skeptical about the novelty and contribution of this work.
  2. The experimental design makes me unconvinced about the effectiveness of the proposed method:
    • The proposed method is only evaluated based on the analysis of 8+53 trajectories, which is too limited in scale.
    • The paper does not compare against previous works, especially given the strong similarity between existing attribution methods and the proposed framework. I am curious whether previous attribution methods can be directly applied to agent trajectories.
    • The paper only uses Llama-3.1-70B-Instruct, which is too limited and somewhat outdated. I am curious whether the proposed framework can be applied with more advanced open-source models and achieve better attribution performance.

[1] ContextCite: Attributing Model Generation to Context (NeurIPS 2024)
[2] TracLLM: A Generic Framework for Attributing Long Context LLMs (SEC 2025)
[3] ContextCite: Attributing Model Generation to Context (Arxiv 2026)

Questions for authors
  1. How does the proposed Agentic Attribution Framework differ fundamentally from existing attribution methods such as ContextCite, TracLLM, and TRACE, beyond adapting them to agent trajectories?
  2. Have the authors evaluated whether existing attribution methods can be directly applied to agent trajectories as baselines? If not, why?
  3. Does the proposed framework generalize to more advanced open-source models, and how does the choice of the underlying LLM affect attribution performance?
Reviewer_eKB1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a hierarchical framework for attributing LLM agent actions to influential historical information. At the component level, it uses temporal log-likelihood changes to identify critical interaction steps, while at the sentence level, it employs perturbation-based scoring to locate specific supporting evidence. The framework is evaluated on 8 agentic case studies and 53 trajectories from GAIA, WebShop, and ALFWorld. Experimental results show improvements over several attribution baselines in identifying influential components and sentences.

Critical strengths

1. Meaningful Problem Formulation. The paper addresses an important problem by extending agent attribution beyond explicit failures to understanding the factors influencing general agent behaviors. This is relevant to improving the transparency and reliability of LLM-based agents.

2. Encouraging Empirical Results. The proposed framework consistently outperforms several component-level attribution baselines across three agent benchmarks. The evaluation also covers both benign and potentially harmful agent behaviors, demonstrating the practical relevance of the approach.

Critical weaknesses

1. Limited Methodological Novelty. The proposed framework mainly combines existing likelihood-based and perturbation-based attribution techniques. While applying these methods to agent trajectories is useful, the paper lacks substantial methodological innovation beyond adapting established approaches to temporal interactions.

2. Insufficient Validation of Attribution Faithfulness. The evaluation mainly relies on human annotations, which do not necessarily reflect the actual causal influence of historical components. Likelihood changes alone cannot establish causality, and the method may struggle with interactions among components. Controlled intervention experiments are needed to better validate the faithfulness of the proposed attribution framework.

Questions for authors

1. Methodological Novelty: What are the key technical contributions of the proposed framework beyond adapting existing likelihood-based attribution methods to agent trajectories?

2. Attribution Faithfulness: Can the authors provide intervention-based experiments to verify whether the identified components genuinely influence agent decisions, rather than merely correlating with the final action?

3. Component Interactions: How does the method handle cases where multiple components jointly influence an action or where important information is repeated across the trajectory?

Reviewer_FBaAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces a general agentic attribution framework to explain why LLM agents take specific actions, rather than focusing solely on identifying the causes of task failures. The proposed framework consists of two hierarchical levels: component-level attribution identifies the most influential components in historical interactions, while sentence-level attribution locates the key textual evidence. The authors present several case studies, but the quantitative evaluation remains limited in demonstrating the effectiveness and generalizability of the proposed framework.

Critical strengths
  • The proposed research question is meaningful by extending failure focused attribution to general agentic attributiong allowing to find the action without explicit error but high risk.

  • The proposed method is simple and clear which can reasonably identify the key historical components and text content from prompt.

  • The given case studies are interesting can support the claims quanlitatively.

Critical weaknesses
  • The experiments rely on a single LLM and a relatively small number of trajectories. Evaluations across different model families, model sizes, and agent architectures are needed to demonstrate the generalizability of the framework.

  • Unconvincing ground truth and evaluation metrics. The ground truth relies on subjective human annotations, and the Hit@k metric only measures whether key components are identified, without sufficiently assessing the quality and faithfulness of the attribution results.

The biggest concern of this paper is about evaluation. Hope the author can find more comprehensive way to demonstrate the proposed attribution framework emprically.

Questions for authors
  • How does the proposed framework perform across different LLM families, model sizes, and agent architectures?

  • Beyond Hit@k, what additional evaluation metrics could be used to assess the quality and faithfulness of the attribution results?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.