Source
Thinking of ACE? We Can Do It with Fewer Tokens
IBM Research compares the ALTK-Evolve and ACE systems, which allow LLM agents to learn from past experiences. The systems differ primarily in how they deliver these stored lessons to the model during inference.
Why it matters
Reducing token usage lowers the operational cost of running AI agents. Selective delivery also ensures that less powerful models are not overwhelmed by excessive information.
The details
- ALTK-Evolve selectively retrieves guidelines rather than injecting a comprehensive playbook every step.
- ALTK-Evolve used 116K tokens versus ACE's 777K on the gpt-oss-120b model.
- Both systems maintain itemized lessons to avoid brevity bias and context collapse.
Get the weekly recap
The stories like this one, picked and explained — once a week, straight to your inbox.