What type of study is this?

This is a Quantitative Study study.

October 20, 2025Open Access

Omni-Thinker: Scaling Multi-Task RL in LLMs with Hybrid Reward and Task Scheduling

View Full Paper

Ask AI

Bookmark

Authors

This framework combines hybrid rewards with backward-transfer-guided scheduling to enhance multi-task performance in large language models.

Improved multi-task performance with gains of 6.2% over joint training and 12.4% over model merging.
Hybrid rewards integrate rule-based signals with LLM evaluations for effective learning.
Backward transfer-guided scheduling reduces forgetting and optimizes task ordering.
Findings highlight the importance of hybrid supervision for scaling RL toward general-purpose LLMs.

Li et al. (2025) studied this question.

View Full Paper

Ask AI

Bookmark

This framework combines hybrid rewards with backward-transfer-guided scheduling to enhance multi-task performance in large language models.