Work

Modern LLM Agents Are Cognitively Powerful but Immunologically Naive.

Year: 2026

Type: preprint

Abstract: This paper introduces epistemic infection as a delayed failure mode in long-horizon LLM agents, where corrupted information enters persistent memory or belief state and later influences reasoning or action. It argues that current agents lack cognitive immunity and proposes CIBench as a benchmark design for evaluating belief-state integrity, recovery, and resilience.

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.20246106

Author: Shrestha, Rajan

Language: English

Topic: Adversarial Robustness in Machine Learning

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

OpenAlex ID: W7161402877

Cites:

Cited by:

Related to: