Learning What Reinforcement Learning Can't: Interleaved Online Fine-Tuning for Hardest Questions

Abstract

ReLIFT interleaves reinforcement learning with online fine-tuning on questions beyond a model’s current capabilities, combining RL’s strength in refining existing reasoning with supervised learning’s ability to introduce new knowledge and reasoning patterns.

Xiaochen Ma

Ph.D. Student in Computer Science and Engineering

I build data infrastructure and high-performance distributed systems that turn heterogeneous data into reliable inputs for large-scale LLM training, with a focus on Ray and Apache Spark.