1

Mitigating Reward Hacking in RLHF via Advantage Sign Robustness
We address reward hacking in RLHF by identifying flipped advantage signs as a key cause and proposing Sign-Certified Policy Optimization (SignCert-PO), a lightweight method that down-weights non-robust completions during policy optimization.
Shinnosuke Ono, Johannes Ackermann, Soichiro Nishimori, Takashi Ishida, Masashi Sugiyama
PDF DOI
Mitigating Reward Hacking in RLHF via Advantage Sign Robustness
A Japanese Language Model and Three New Evaluation Benchmarks for Pharmaceutical NLP
This paper presents a Japanese phamaceutical-specific langauge model, JPharmatron, along with a benchmark suite consisting three new benchmarks: YakugakuQA (Japanese national pharmacist licensing exams); NayoseQA (cross-lingual synonym and terminology normalization); and SogoCheck (consistency reasoning between paired statements.)
Shinnosuke Ono, Issey Sukeda, Takuro Fujii, Kosei Buma, Shunsuke Sasaki
Cite PDF Code Dataset DOI
A Japanese Language Model and Three New Evaluation Benchmarks for Pharmaceutical NLP