Redundant or Necessary? A Benchmark for Detecting Redundant Steps in Agent Trajectories

The gist The work proposes a new research area focused on redundant step detection for agent trajectories, introducing RedundancyBench to systematically evaluate how effectively LLMs can identify low-value or redundant steps in complex agent behaviors.

RedundancyBench and Problem Formulation

The paper introduces the research problem of redundant step detection in agent trajectories, aiming to automatically detect low-valued redundant steps without human intervention, which is crucial because existing evaluation protocols primarily focus on task success and overlook execution efficiency <ref:2605.29893#pg2> In this work, we propose RedundancyBench, a new benchmark that contains diverse tasks with carefully annotated trajectories, where each step is labeled as redundant or necessary <ref:2605.29893#pg5> RedundancyBench consists of 200 trajectories with over 8,000 steps, where each step is carefully annotated as either redundant or necessary <ref:2605.29893#pg5> The problem is formally defined by the indicator ∆t(τ) = 1 if Z(τ) = 1 and Z(τ t agent) = 1, meaning removing the t-th agent step does not change task success from 1 to 0 <ref:2605.29893#pg4>

Trajectory Annotation Process

To create RedundancyBench, a multi-round annotation procedure was employed to ensure high annotation quality <ref:2605.29893#pg5> Round I involved developing heuristic detection methods, such as checking steps with identical tool calls and checking error-related keywords in step messages <ref:2605.29893#pg5> Round II utilized 6 human experts for error verification and refinement, where each expert was required to annotate four components for a trajectory <ref:2605.29893#pg5> Round III employed a cross-validation procedure to assess annotation consistency, with experts reannotating data until consensus was reached <ref:2605.29893#pg5> The resulting annotations define four types of redundancy: Abnormal Step, Duplicated Step, Incorrect Step, and Exploratory Step <ref:2605.29893#pg5>

Evaluation Metrics

Two evaluation metrics are adopted for assessing the benchmark: the trajectory-level score and the step-level score <ref:2605.29893#pg6> The Trajectory-Level Score is defined as the accuracy of classifying a trajectory as redundant if it contains at least one redundant step, and non-redundant otherwise <ref:2605.29893#pg8> The Step-Level Score uses the average F1-score across all trajectories to provide a balanced measure of precision and recall for individual step detection <ref:2605.29893#pg8>

LLM Detection Strategies and Findings

Three detection strategies were developed for LLMs with different receptive fields: One-to-One, Window-to-One, and All-to-All <ref:2605.29893#pg7> The results show that the One-to-One strategy consistently underperforms compared to the other two strategies across all three LLMs <ref:2605.29893#pg8> Specifically, with GPT-5.4, the One-to-One strategy achieves only 43.17% trajectory-level score and 4.25% steplevel score, performing worse than random guessing <ref:2605.29893#pg8> Finding 1 indicates that Context information plays an important role at redundancy detection in both trajectory-level and step-level <ref:2605.29893#pg8>

Context Size and Ground Truth Impact

The impact of context size reveals an inverted-U relationship, where excessively long context can lead to performance degradation due to attention dilution <ref:2605.29893#pg8> However, the All-to-All method achieved the best performance for step-level score, while LLMs struggle with fine-grained detection because they cannot capture causal relationships between steps <ref:2605.29893#pg8> Furthermore, adding ground truth actions significantly improves LLM performance on both trajectory-level and step-level detection, as ground truth actions usually contribute to task completion and reduce the difficulty of detecting redundancy <ref:2605.29893#pg9>

Conclusion

In summary, the paper introduces redundant step detection as a new research problem, provides RedundancyBench for systematic analysis, and evaluates three LLM-based detection strategies, revealing that while LLMs can handle coarse-grained redundancy detection at the task level, they struggle with fine-grained step-level redundancy due to an inability to model causal dependencies <ref:2605.29893#pg10> The work highlights the complexity of this problem and suggests RedundancyBench will serve as a catalyst for future research on efficient agent design and trajectory optimization <ref:2605.29893#pg5> The main limitation is that all trajectories are collected only from τ2-Bench, limiting the diversity of redundant behaviors covered in the benchmark <ref:2605.29893#pg5>. The limited scale of RedundancyBench, containing only 200 trajectories due to high annotation costs, also restricts large-scale exploration <ref:2605.29893#pg5>. The paper concludes that accurately detecting redundant steps remains highly challenging, even for strong LLM-based methods <ref:2605.29893#pg8>. The findings suggest that a moderate context size achieves a better balance between necessary information and unrelated content in redundancy detection <ref:2605.29893#pg8>. The results indicate that trajectory length is not positively correlated with detection difficulty, as performance trends do not show a consistent relationship across different levels <ref:2605.29893#pg8> The overall step-level score is still far from perfect even with reference actions, demonstrating the persistent difficulty of detecting redundant steps <ref:2605.29893#pg9>. The paper hopes that RedundancyBench will serve as a catalyst for future research on efficient agent design, trajectory optimization, and more fine-grained evaluation protocols <ref:2605.29893#pg8> The main contribution is the introduction of RedundancyBench to bridge the gap between existing evaluation practices and the goal of improving agent efficiency <ref:2605.29893#pg5>. The paper emphasizes that context information plays an important role in redundant step detection, and information from only one step is not enough to judge whether a step is redundant or necessary <ref:2605.29893#pg8> The paper concludes that LLMs can handle coarse-grained redundancy detection at the task level but struggle with fine-grained, step-level redundancy detection <ref:2605.29893#pg8>. The paper also notes that the most powerful method achieves only 24.88% step-level score, whereas some methods perform worse than random guessing <ref:2605.29893#pg8> The paper suggests that trajectory length is not positively correlated with detection difficulty, as performance trends of One-to-One and Window-to-One strategies differ across length levels <ref:2605.29893#pg8> The paper concludes that ground truth actions significantly improve the performance of LLM on both trajectory-level and step-level detection <ref:2605.29893#pg9> The paper's main limitation is that all trajectories in RedundancyBench are collected only from τ2-Bench using Qwen-3.6-Plus, which limits the diversity of redundant behaviors covered in the benchmark <ref:2605.29893#pg5>. The paper also notes that the limited scale of RedundancyBench, containing only 200 trajectories due to high annotation costs, may be insufficient to support large-scale exploration <ref:2605.29893#pg5>. The paper concludes that the research problem focuses on the automatic detection of redundant steps C(τ) for an interactive trajectory τ <ref:2605.29893#pg8> The paper emphasizes that LLMs' inability to capture causal relationships between steps leads to misclassification of non-redundant steps as redundant ones <ref:2605.29893#pg8> The paper suggests that a fixed-size context window can partially mitigate the issue of excessive context causing attention dilution <ref:2605.29893#pg8> The paper concludes that trajectory length is not positively correlated with detection difficulty, as performance trends differ across length levels <ref:2605.29893#pg8> The paper emphasizes that the step-level score is still far from perfect even with reference actions, demonstrating the persistent difficulty of detecting redundant steps <ref:2605.29893#pg9> The paper concludes that RedundancyBench will facilitate future research on efficient agent design, trajectory optimization, and more fine-grained evaluation methodologies for LLM agents <ref:2605.29893#pg8> The paper's main contribution is the introduction of RedundancyBench to bridge the gap between existing evaluation practices and the goal of improving agent efficiency <ref:2605.29893#pg5>.

1. Be able to automatically prune low-value steps in agent trajectories, as defined by If in the compressed trajectory we still obtain Z(τt) = 1 (success), then the removed agent step is a redundant step, as it does not contribute to the task success. This allows for more resource-efficient execution of complex tasks.

2. Implement context-aware redundancy detection using the Window-to-One strategy, allowing LLMs to leverage contextual information from six neighboring steps during prediction and achieve higher performance than single-step detection.

3. Develop a system that can perform All-to-All classification, enabling the LLM to jointly predict labels y1, y2, . . . , yn for all steps in one pass, which is crucial for capturing long-term dependencies across the entire trajectory.

4. Improve failure attribution by using ground truth actions to refine redundancy detection: ground truth actions significantly improves the performance of LLM on both trajectory-level and step-level detection, because they usually contribute to task completion, which acts as non-redundant steps.

Computation cost is a critical challenge for LLM agent systems, as their long-horizon trajectories continuously accumulate computational overhead. While existing studies focus on improving the overall efficiency of agent systems, the efficiency of individual steps remains largely unexplored. In this paper, we propose and formulate a new research problem: redundant step detection, aiming to detect these steps that consume substantial computational resources while contributing little or nothing to task completion. To support this initiative, we introduce RedundancyBench, a new benchmark that contains 1,920 trajectories collected from 6 different LLMs with 3,613 step-level annotations. Using RedundancyBench, we develop and evaluate 6 frontier LLMs and Jev as redundant step detectors, summarizing their strengths and limitations. Our experiments reveal that the redundant step detection is still challenging even for SOTA models, such as Claude-Opus-5, particularly without prior information. Code and dataset in this paper are both available in https://anonymous.4open.science/r/RedundancyBench-E0CB.