When Entropy Is Not Enough: Reclaiming Lost Semantics in LLM Output Length Prediction
Existing methods for predicting LLM output length rely only on token entropy, which measures uncertainty but ignores semantic importance. This causes core content tokens to be underweighted while filler words are overemphasized. The ESTP framework fixes this by combining token entropy with attention-based semantic weights, creating a more accurate prediction that balances uncertainty and meaning for better system performance.