Thinking of ACE? We Can Do It with Fewer Tokens

Hugging Face · 2 months ago · 24

ALTK-Evolve and ACE both turn an agent’s past trajectories into reusable memory lessons for an LLM ReAct agent, but they differ in how much of that playbook gets sent at inference each step. ALTK-Evolve cut tokens/task from 634K with ACE to 263K on DeepSeek-V3.2 while raising TGC from 80.4 to 89.3 and SGC from 73.2 to 80.4. The change is calibrated delivery (selecting a subset per task) instead of injecting the full playbook every step, keeping accuracy while reducing inference cost.