Back to AI Research

AI Research

GRUET: Quantifying Uncertainty of Agentic Reasoning... | AI Research

Key Takeaways

  • GRUET: Quantifying Uncertainty of Agentic Reasoning-and-Acting Processes Autonomous agents powered by Large Language Models (LLMs) have become increasingly p...
  • Agents have attracted considerably increasing attention due to the power of executing both Reasoning and Acting (ReAct) in open and dynamic environments.
  • The ReAct process typically exhibits a multi-turn trajectory in which one drives Large Language Models (LLMs) to generate both reasoning chains and task-specific actions in an interleaved manner.
  • Empirical evaluations across nine LLMs and five benchmarks validate the effectiveness of our proposed GRUET in terms of selective generation performance, measured by AUROC, AUPRC, and AUARC.
  • GRUET: Quantifying Uncertainty of Agentic Reasoning-and-Acting Processes
Paper AbstractExpand

Agents have attracted considerably increasing attention due to the power of executing both Reasoning and Acting (ReAct) in open and dynamic environments. The ReAct process typically exhibits a multi-turn trajectory in which one drives Large Language Models (LLMs) to generate both reasoning chains and task-specific actions in an interleaved manner. However, agents often suffer from significant uncertainty, where identical tasks yield divergent trajectories; trajectories with higher uncertainty often produce incomprehensible behaviors, severely undermining agent credibility. This work conjectures that such trajectory-level uncertainty frequently stems from cumulative turn-level reasoning uncertainty induced by LLMs; the latter often exhibits a collection of branches of divergent reasoning chains and their resulting actions. Built upon this, we present the Graph-based Reasoning UncErtainty in Trajectories (GRUET) method for the uncertainty quantification of ReAct, comprising turn-level reasoning uncertainty quantification and trajectory-level uncertainty aggregation; the former precisely quantifies reasoning uncertainty via modeling the reasoning space spanned by potential reasoning branches as a graph and then approximating the reasoning space complexity with graph complexity, while the latter employs simple aggregation strategies for quantifying the overall trajectory credibility. Empirical evaluations across nine LLMs and five benchmarks validate the effectiveness of our proposed GRUET in terms of selective generation performance, measured by AUROC, AUPRC, and AUARC.

GRUET: Quantifying Uncertainty of Agentic Reasoning-and-Acting Processes
Autonomous agents powered by Large Language Models (LLMs) have become increasingly popular for their ability to perform complex tasks by reasoning and taking actions in dynamic environments. However, these agents often struggle with uncertainty, where the same task can lead to wildly different outcomes. When an agent’s reasoning becomes inconsistent, its behavior can become incomprehensible, which undermines user trust. This paper introduces GRUET, a method designed to quantify this uncertainty by analyzing the reasoning process at both the individual turn level and the overall trajectory level. The ai agents story also surfaces in Google AI Introduces EnvHarness for Adaptive..., adding another angle.

Modeling Reasoning as a Graph

The core idea behind GRUET is that trajectory-level uncertainty is the result of cumulative errors during the reasoning process. To capture this, the researchers model the agent's reasoning as a Directed Acyclic Graph (DAG). In this graph, nodes represent specific reasoning steps or actions, and edges represent the flow from one step to the next. By generating multiple potential reasoning paths and merging semantically equivalent nodes, the method creates a map of the "reasoning space." The complexity of this graph serves as a proxy for the agent's uncertainty: a more complex, branching graph indicates a less confident and potentially more error-prone reasoning process.

Quantifying Turn-Level Uncertainty

To calculate the uncertainty of a single turn, GRUET combines token-level statistical data with the topological structure of the reasoning graph. It calculates the uncertainty of individual tokens and aggregates them into "node uncertainty" for each reasoning step. By treating the reasoning graph like a neural network, the method propagates these uncertainty scores through the graph to an auxiliary node. This allows the system to derive a precise numerical score that reflects how much the agent is "struggling" at that specific point in its decision-making process. The ai agents story also surfaces in Arm unveils AI-native mobile platform for..., adding another angle.

Aggregating Results for Trajectory Credibility

Once the turn-level uncertainty is calculated, the agent needs to determine the overall credibility of the entire trajectory. The researchers found that the most informative moments for identifying failures often occur at the beginning, the end, or during the most uncertain turns of a process. Based on this, they developed several aggregation strategies—such as focusing on the first turn, the last turn, or the most uncertain turns—to produce a final trajectory-level uncertainty score. These strategies help distinguish between successful and failed task executions.

Performance and Validation

The effectiveness of GRUET was tested across nine different LLM models and five distinct benchmarks, including code generation tasks. The results show that GRUET consistently outperforms traditional uncertainty quantification methods, such as those based on simple perplexity or reflexive self-judgment. By providing a more accurate measure of agent confidence, GRUET helps developers better identify when an agent is likely to produce an unreliable or incorrect result, offering a more robust way to evaluate agent performance in real-world applications. The ai agents story also surfaces in Andrew Ng Launches OpenWorker to Deliver..., adding another angle. as detailed in the full paper on Arxiv

More from Franklin

When Agents DisagreeresearchAn Empirical Study of Harness Design...researchClaude autonomously improved models across 10...newsArm unveils AI-native mobile platform for...newsGoogle AI Introduces EnvHarness for Adaptive...newsAndrew Ng Launches OpenWorker to Deliver...news

Comments (0)

No comments yet

Be the first to share your thoughts!