Thinking of ACE? We Can Do It with Fewer Tokens
Hugging Face
Hugging Face says its ALTK-Evolve uses fewer tokens than ACE on agent tasks. Same lessons, but it sends only what each model can use.
Based on reporting by Hugging Face — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Hugging Face is drawing a sharp line between two ways of teaching an agent from its own mistakes. ACE and ALTK-Evolve both mine past trajectories and feed the lessons back at inference time, without weight updates or human labels. The split is not about what they learn. It’s about how much of it gets shoved back into the model every step.
The company says both systems reject the idea that you should squash an agent’s experience into a neat summary. ACE keeps a detailed playbook with helpful and harmful counts on each bullet. ALTK-Evolve keeps individual guidelines with support counts, so a lesson that showed up in five episodes stays distinct from one that appeared once. Same instinct, different packaging.
The real difference shows up in delivery. ACE injects the whole playbook on every step. ALTK-Evolve treats context like a budget: a small fixed core of high-support guidelines, then a few more selected for the task, or the full set when the model has room for it. The source says that choice is what drives the token bill.
On AppWorld, using the same ReAct agent, the stronger model, DeepSeek-V3.2, got 89.3 TGC and 80.4 SGC with ALTK-Evolve at 263K tokens per task, versus 80.4 and 73.2 with ACE at 634K. On gpt-oss-120b, ALTK-Evolve reached 56.0 TGC and 37.5 SGC at 116K tokens, while ACE got 54.8 and 35.7 at 777K. That’s the odd part: on the weaker model, fewer tokens also meant slightly better results.
The by-difficulty numbers make the story clearer. On gpt-oss-120b, ACE did better on Easy and Medium, but ALTK-Evolve pulled ahead on Hard and in the aggregate. On DeepSeek-V3.2, ALTK-Evolve led on Easy, Hard, and Overall, while ACE only edged Medium. The lesson here is pretty blunt: more memory is not the same as better memory, especially when a weak model has to read all of it.
My take — AI-written commentary, not fact-checked reporting
This is the part the industry keeps dodging: context is not free, and dumping the whole scrapbook into the prompt is lazy engineering dressed up as memory. The sensible move is obvious — deliver what the model can actually use, not what looks impressive in a demo. Bigger prompts are often just bigger tax bills with better PR.
Read more about this at: Hugging Face