OpenAI is rumored to have an internal control token, the "juice," which sets how "long" a model will think. Can we build models that can control their "juice"?
Idea: LLM estimates how much to think, at the beginning of its thought trace
RQ: We know we can use thinking control tokens to change generation length. However, can the model determine the “juice” ahead of time? Is there a simple training procedure where the LLM estimates the juice prior to training, and this leads to Pareto-optimal performance?
Relevant work: https://arxiv.org/abs/2410.04707
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{heineman-the-juice-estimator-2025,
author = {Heineman, David},
title = {THE JUICE ESTIMATOR},
year = {2025},
url = {https://hypogenic.ai/ideahub/idea/Ehh8Im44mMLtUg1M1tiI}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!