LLM Agent Optimization: Token Efficiency through Calibrated Delivery

Carte mentaleMind map

IBM Research's paper introduces ALTK-Evolve, an approach to improve the reliability and efficiency of LLM agents in executing complex tasks. The main problem with LLM agents is not a lack of knowledge, but an inability to reliably use their APIs, a skill that can be learned from their own past trajectories.

ALTK-Evolve, like ACE (Agentic Context Engineering), uses agentic memory to transform past experiences into reusable "lessons," without requiring weight updates or human labeling. The key difference lies in how these lessons are consolidated and, more importantly, delivered to the model during inference. While ACE injects a complete and static "playbook" at each step, ALTK-Evolve adopts a calibrated approach, providing only a relevant subset of directives, tailored to the model and the task.

This selective delivery allows ALTK-Evolve to achieve equal or superior accuracy to ACE, while significantly reducing inference costs (up to one-seventh of the tokens used for weaker models). Efficiency is particularly noticeable on difficult tasks, where the relevant selection of lessons outweighs sending an overly voluminous context that can overwhelm the model.