They propose STEP (Soft Trajectory Estimation of Prediction correctness).
Soft per-step probability scoring improves LLM correctness estimation without training
The authors show that preserving continuous confidence signals from intermediate reasoning steps outperforms binary indicators and output-only statistics.
Chinese Tech
Chenyang Yu · Zifan Zeng · Qunli Zhang · Feng Liu
Technical University of Munich · Huawei Heisenberg Research Center
Research Digest··2 min read
Chenyang Yu et al. introduce STEP, a training-free framework that scores each reasoning step of a large language model's output by re-encoding the partial trajectory and reading soft probability distributions over candidate answers.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.