Tag
This paper presents the first study of probability calibration as a mitigation for evaluator preference coupling in LLM agent feedback loops, showing that calibrated evaluator judgments reduce coupling coefficients by 20-49% and divergence by 45-67%.
Introduces Simplex-Constrained Sparse Bagging (SCSB), a post-training framework that optimizes estimator weights over the probability simplex using out-of-bag samples, achieving up to 96% ensemble compression and improved calibration.