Large Language Model Agents Are Not Always Faithful Self-Evolvers

Weixiang Zhao, Yingshuo Wang, Yichen Zhang, Yang Deng, Yanyan Zhao, Wanxiang Che, Bing Qin, Ting Liu
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:162243-162268, 2026.

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 13 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.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-zhao26aw, title = {Large Language Model Agents Are Not Always Faithful Self-Evolvers}, author = {Zhao, Weixiang and Wang, Yingshuo and Zhang, Yichen and Deng, Yang and Zhao, Yanyan and Che, Wanxiang and Qin, Bing and Liu, Ting}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {162243--162268}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/zhao26aw/zhao26aw.pdf}, url = {https://proceedings.mlr.press/v306/zhao26aw.html}, 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 13 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.} }
Endnote
%0 Conference Paper %T Large Language Model Agents Are Not Always Faithful Self-Evolvers %A Weixiang Zhao %A Yingshuo Wang %A Yichen Zhang %A Yang Deng %A Yanyan Zhao %A Wanxiang Che %A Bing Qin %A Ting Liu %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-zhao26aw %I PMLR %P 162243--162268 %U https://proceedings.mlr.press/v306/zhao26aw.html %V 306 %X 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 13 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.
APA
Zhao, W., Wang, Y., Zhang, Y., Deng, Y., Zhao, Y., Che, W., Qin, B. & Liu, T.. (2026). Large Language Model Agents Are Not Always Faithful Self-Evolvers. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:162243-162268 Available from https://proceedings.mlr.press/v306/zhao26aw.html.

Related Material