Uncertainty Quantification for LLM Agents via Semantic Abstraction Trajectories

Nikita Kazeev, Andrey E Ustyuzhanin

Keywords: Uncertainty Quantification, Selective Prediction, Structured Inference, LLM Agents, Probabilistic Ranking Models

TL;DR: A black-box uncertainty quantification method that predicts LLM agent failures by analyzing the "shape" of their reasoning trajectories through different levels of semantic abstraction.

Abstract: Quantifying uncertainty in agentic LLM outputs is essential for safe deployment, yet existing approaches require access to model internals or ignore the reasoning trace entirely. We propose a black-box uncertainty signal derived from \emph{structured probabilistic inference over an agent's reasoning trace}. We define Semantic Level of Detail (SLoD), a continuous macro-to-micro abstraction axis, and recover per-chunk abstraction scores via a Plackett--Luce ranking model fit to pairwise LLM judgments -- a latent-variable inference step that requires only a non-reasoning judge. The resulting SLoD trajectory is a structured time series whose shape serves as a selective-prediction signal. A LightGBM classifier over interpretable trajectory-shape detectors predicts SWE-agent failures at ROC-AUC $0.89\pm0.02$ and, used as an out-of-fold scorer to abstain from low-quality attempts, lifts Pass@1 on FrontierScience Olympiad from $58.2\%$ to $68.0\%$ over five DeepSeek-V3.2 attempts (recovering $39.5\%$ of the Pass@5 oracle gap); the coverage--accuracy curve dominates a length-only baseline at every operating point. The same features classify hallucination categories on AgentHallu/OpenManus at macro-AUC $0.71$. We also report negative results on GPQA-Diamond and Qwen3-30B traces, delineating the conditions under which the signal is informative.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 148