SpecRA: Monitor Degenerative Repetition in LLM Agents using Randomized FFT

ICLR 2026 Conference Submission15030 Authors

Keywords: Degenerative Repetition, Agent Reliability, LLM Agents, Repetition Detection, Spectral Analysis, Signal Processing

Abstract: LLM-based agents also suffer from "degenerative repetition" like chatbots, which leads to task failure and results in significant waste of computational resources and API costs until token limit is reached. Existing methods require modification of training process or customization of model deployment, and detection algorithms are brittle to approximate or structural recurrence. We therefore introduce SpecRA, a simple yet effective algorithm for detection of self-repetitions in text. Via a randomized projection from the large LLM vocabulary onto a unit-norm complex sequence, our method leverages the power of the Fast Fourier Transform (FFT) to compute the sequence's autocorrelation. Peaks in the autocorrelation function robustly reveal the underlying periodicity of the content, with tolerance to minor variations. Through an analysis of 813 repetitive samples identified from 1.13M records of anonymized agent outputs, we build a taxonomy of repetition modes in agents and show that SpecRA offers a lightweight, non-intrusive mechanism for constructing more reliable and cost-efficient LLM agents accross both standard open-source model deployments and proprietary models.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 15030