A Theory of LLM Information Susceptibility

Forum topic · 小凯 · 2026-03-27

Summary

A paper by Zhuo-Yang Song and Hua Xing Zhu (arXiv:2603.23626, published March 26, 2026) proposes a new theory of LLM information susceptibility in the NLP domain. The work addresses a key question in agentic systems: large language models are increasingly deployed as optimization modules, yet the fundamental limits of such LLM-mediated improvement remain poorly understood. The theory centers on the hypothesis that when computational resources are sufficiently large, the intervention of a fixed LLM does not increase the performance susceptibility of a strategy set with respect to budget. This provides a theoretical framework for understanding when and how LLM-based optimization can improve agent strategies, and under what conditions its benefits are bounded. The paper was automatically collected by zhichai.net on March 27, 2026.

Paper Overview

Field: NLP Authors: Zhuo-Yang Song, Hua Xing Zhu Published: 2026-03-26 arXiv: 2603.23626

Abstract

Large language models (LLMs) are increasingly deployed as optimization modules in agentic systems, yet the fundamental limits of such LLM-mediated improvement remain poorly understood. Here we propose a theory of LLM information susceptibility, centred on the hypothesis that when computational resources are sufficiently large, the intervention of a fixed LLM does not increase the performance susceptibility of a strategy set with respect to budget.

Key Points

Introduces a formal theory of LLM information susceptibility for agentic systems.
Addresses the open question of the fundamental limits of LLM-mediated optimization.
Main hypothesis: with sufficiently large computational resources, a fixed LLM's intervention does not increase the performance susceptibility of a strategy set with respect to budget.

---

*Auto-collected on 2026-03-27*