Large Language Model Agents Are Not Always Faithful Self-Evolvers

TL;DR

The paper tackles whether self-evolving LLM agents meaningfully ground their decisions in accumulated experience, distinguishing raw trajectories from condensed summaries. Using controlled causal interventions across four agent frameworks, ten backbones, and nine environments, it shows a robust asymmetry: agents consistently leverage raw experience but largely ignore condensed content, even when condensed input is the sole guidance. The authors identify semantic limitations of condensed content, internal model biases toward local context, and task regimes where pretrained priors suffice as key causes for unfaithfulness, and demonstrate that scaling alone does not fix the gap. These findings challenge assumptions about experience-driven adaptation and motivate design improvements in content-rich condensed summaries and dynamically triggered, context-aware retrieval to achieve faithful, reliable self-evolution of agents.

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.

Figures (16)

• Figure 1: Examples of experience intervention and faithfulness evaluation. Given a task goal, the agent receives raw experience (concrete historical trajectories that succeed to complete the similar tasks) and condensed experience (abstract summaries or heuristics). We apply different types of interventions, such as shuffling, corrupting, or replacing experience with irrelevant content, to test whether such perturbations affect downstream behavior. A full taxonomy of intervention types is provided in Section \ref{['sec:intervention']}. Faithfulness is determined by whether the agent's behavior causally changes in response to the perturbed input.
• Figure 2: Intervention results on the ExpeL framework (offline, single-agent) using GPT-4o across three benchmarks. ExpeL consistently relies more on raw trajectories, while showing weak or inconsistent sensitivity to condensed summaries.
• Figure 3: Intervention results on the Dynamic CheatSheet (DC-RS) framework (online, single-agent) using GPT-4o. Raw experience perturbations significantly reduce performance, whereas condensed experience manipulations often have negligible impact.
• Figure 4: Impact of condensed experience interventions on the ReasoningBank framework (online, single-agent) using Gemini-2.5-Flash across four WebArena sub-tasks. Despite the absence of raw experience, agents show only mild sensitivity to semantic manipulations of condensed experience, indicating limited semantic faithfulness.
• Figure 5: Faithfulness interventions on the G-Memory framework (online, multi-agent) with GPT-4o-mini. Agents access two forms of raw experience (reference and execution) and one form of condensed experience.