Error Persistence in Recursive LLM Systems: Trajectory State Determines Whether Perturbations Persist or Decay

Read full text Summarise with VIRAT
Authors: Thom Pham
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This paper investigates error persistence in recursive large language model (LLM) systems and shows that it is not a property of the model itself, but of the trajectory state in which outputs are recursively reused. Using a controlled multi-agent experiment (LAB3), we compare two serialization regimes—RAW and SANITIZED—under identical conditions. An emergent perturbation arising from the system’s own dynamics leads to early trajectory divergence, followed by a controlled perturbation introduced at a fixed step. In RAW, reinforcement signals accumulate across turns, leading to early trajectory closure and persistent error amplification. In SANITIZED, reinforcement is normalized at each step, preserving an open trajectory in which the same perturbation decays and the system fully recovers. The key finding is structural: trajectory state at the moment of perturbation determines whether errors persist or decay. Serialization governs reinforcement accumulation, which determines trajectory state. The paper introduces a trajectory-level perspective on LLM failure modes and shows that error persistence arises from loss of revisability under recursive reuse, not from isolated incorrect outputs. Full experimental traces, scripts, and reproducibility data are available at:https://guardianai.fr/evidence/#lab3

Fields of Study & Subjects

Natural Language Processing Techniques Topic Modeling Adversarial Robustness in Machine Learning

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
25 March 2026
Language
EN
Document type
Preprint