Large Language Model Agents Are Not Always Faithful Self-Evolvers

2026-01-30

AI Summary

It remains unclear whether existing self-evolving large language model (LLM) agents genuinely rely on their past experiences to guide behavior, particularly when using compressed experience representations, which raises concerns about reliability. This work introduces and quantifies the notion of “experience fidelity” for the first time, employing causal intervention methods to systematically evaluate four representative self-evolution frameworks across ten LLM backbones and nine environments. The study reveals that agents heavily depend on raw experiences but consistently overlook or misinterpret compressed ones—a phenomenon robustly observed across both single- and multi-agent settings and across varying model scales. Furthermore, the paper identifies three key underlying causes, challenging a core assumption in current self-evolution paradigms regarding the effective utilization of experience.

Technology Category

• Multiagent Systems: Adversarial Agents
• Cognitive Modeling & Cognitive Systems: Agent Architectures
• Machine Learning: Evolutionary Learning

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Self-evolving large language model (LLM) agents continually improve by accumulating and reusing past experience, yet it remains unclear whether they faithfully rely on that experience to guide their behavior. We present the first systematic investigation of experience faithfulness, the causal dependence of an agent's decisions on the experience it is given, in self-evolving LLM agents. Using controlled causal interventions on both raw and condensed forms of experience, we comprehensively evaluate four representative frameworks across 10 LLM backbones and 9 environments. Our analysis uncovers a striking asymmetry: while agents consistently depend on raw experience, they often disregard or misinterpret condensed experience, even when it is the only experience provided. This gap persists across single- and multi-agent configurations and across backbone scales. We trace its underlying causes to three factors: the semantic limitations of condensed content, internal processing biases that suppress experience, and task regimes where pretrained priors already suffice. These findings challenge prevailing assumptions about self-evolving methods and underscore the need for more faithful and reliable approaches to experience integration.

Problem

• experience faithfulness
• self-evolving agents
• large language models
• condensed experience
• causal dependence

Innovation

• experience faithfulness
• self-evolving LLM agents
• causal intervention
• condensed experience
• pretrained priors