EXG: Self-Evolving Agents with Experience Graphs

2026-05-18 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, medium

Summary

EXG is a novel experience graph framework designed for self-evolving large language model (LLM)-based agents, addressing the limitation of static agents that do not systematically improve from execution experience. Unlike existing approaches relying on ad hoc reflection or unstructured memory, EXG organizes accumulated successes and failures into a structured, relational representation. It supports both online, real-time graph growth for immediate cross-task experience reuse and offline reuse as an external memory module. This plug-and-play design allows EXG to integrate with existing self-evolving agents, enhancing solution quality and resource efficiency. Experiments on code generation and reasoning benchmarks demonstrate that EXG achieves more favorable performance-efficiency trade-offs compared to reflection- and memory-based baselines in both online and offline evaluations.

Key takeaway

For NLP Engineers developing self-evolving LLM agents, adopting a structured experience management system like EXG can significantly improve agent performance and resource efficiency. You should consider integrating EXG to move beyond static agent behaviors, enabling your agents to systematically learn from deployment experiences and achieve better solution quality and faster adaptation across tasks.

Key insights

EXG uses a structured experience graph to enable LLM agents to systematically learn and improve from past interactions.

Principles

• Structured experience improves agent learning.
• Online and offline experience reuse is critical.
• Relational representation enhances transferability.

Method

EXG organizes agent successes and failures into a structured, relational graph, supporting real-time growth for immediate cross-task reuse and offline consolidation as an external memory module.

In practice

• Integrate EXG as a plug-and-play component.
• Apply EXG to code generation tasks.
• Use EXG for reasoning benchmarks.

Topics

• Self-Evolving Agents
• Experience Graphs
• LLM Agents
• Structured Memory
• Cross-Task Learning

Best for: Research Scientist, NLP Engineer, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Escaping the Self-Confirmation Trap: An Execute-Distill-Verify Paradigm for Agentic Experience Learning — Decoupling LLM agent experience learning into execute, distill, and verify stages prevents self-confirmation errors and improves reliability.
• How AI Agents Navigate Massive Skill Graphs — Skill-DAG enables LLMs to dynamically interpret and evolve a typed skill graph for scalable, context-aware skill selection, moving beyond semantic similarity.
• SEAGym: An Evaluation Environment for Self-Evolving LLM Agents — Self-evolving LLM agent evaluation requires comprehensive metrics beyond isolated scores to understand harness updates.
• Does RL Expand the Capability Boundary of LLM Agents? A PASS@(k,T) Analysis — RL expands LLM agent capabilities for compositional tool use, unlike its role in static reasoning.
• Why I Stopped Using One Agent and Built a Multi-Agent Pipeline Instead — Single LLM agents struggle with complex, multi-faceted tasks by attempting too many competing mental models in one context.
• From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms — LLM agent memory evolves through stages of storage, reflection, and experience, driven by consistency, dynamic environments, and continual learning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.