Agent Evolving Learning: A Two-Timescale Framework for LLM Agents in Open-Ended Environments
A new framework called Agent Evolving Learning (AEL) has been developed by researchers to facilitate LLM agents in learning from their experiences across numerous sequential episodes in open-ended settings. This framework tackles the challenge of not only determining what information to retain but also how to effectively utilize it, including the selection of retrieval policies, interpretation of past results, and strategy adjustments. On a rapid timescale, a Thompson Sampling bandit identifies the appropriate memory retrieval policy for each episode. Meanwhile, on a slower timescale, LLM-driven reflection analyzes failure patterns and integrates causal insights into the agent's decision-making process. The framework was evaluated using a sequential portfolio benchmark featuring 10 diverse tickers over 208 episodes, and the findings were published on arXiv with the paper number 2604.21725.
Key facts
- AEL is a two-timescale framework for LLM agents.
- Fast timescale uses Thompson Sampling bandit for memory retrieval policy.
- Slow timescale uses LLM-driven reflection for causal insights.
- Tested on sequential portfolio benchmark with 10 tickers over 208 episodes.
- Published on arXiv with ID 2604.21725.
- Addresses how to use remembered information, not just what to remember.
- Framework enables agents to convert past experience into better future behavior.
- AEL stands for Agent Evolving Learning.
Entities
Institutions
- arXiv