Tag
This paper proposes a novel contextual bandit algorithm that explicitly models reward decay for progressive content refinement in LLMs, using EM to estimate arm-specific and decay parameters. Experiments on Sentiment Reversal and GSM8K show significant gains over strong baselines.
SocraticPO augments RL rollouts with Socratic-style natural language guidance and reward decay to improve scientific reasoning in LLMs, outperforming strong baselines on SciKnowEval benchmarks.