Verbosity penalty and length bias calibrator normalizing raw model rewards against response token lengths
Verbosity penalty and length bias calibrator normalizing raw model rewards against response token lengths
Characterizing the performance of minimum Bayes risk (MBR) decoding for neural machine translation (NMT) when using various candidate generation methods by examining BLEU score, length bias, and token frequency bias. [CS287: Deep Learning for Natural Language Processing, Final Project, Fall 2021]
Data analysis for HCC screening study.
Do LLM-judge biases compose? A label-free, scale-aware test (SSIT) for position×length interaction in LLM-as-a-judge pipelines, with proofs, tie-coding and curvature diagnostics, 7-judge/6-family data, and the TMLR manuscript.
To associate your repository with the length-bias topic, visit your repo's landing page and select "manage topics."