ReST improves LLM translation quality on benchmarks via offline RL on self-generated data, achieving gains in a compute-efficient way compared to typical RLHF.
Figure 10 | [WMT 2020 Zh-En]:distribution of human preference and reward model scores forReST (BC, I=4, G=1 ) in side by side evaluation with supervised model
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2023 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Reinforced Self-Training (ReST) for Language Modeling
ReST improves LLM translation quality on benchmarks via offline RL on self-generated data, achieving gains in a compute-efficient way compared to typical RLHF.