SFT Conflicts, RL Coexists: Theoretical Analysis of Multi-Task LLM Training

CASIA · hf · 2026-08-10

A paper from CASIA provides a theoretical and empirical analysis comparing Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) in multi-task reasoning for LLMs.

The study reveals that SFT suffers from severe task conflicts in multi-stage training, whereas RL enables stable coexistence across diverse tasks. At the parameter level, RL induces sparse and approximately orthogonal updates across tasks.

Theoretically, interference in SFT is norm-limited, while in RL it is variance-limited. This small variance bound yields near-orthogonal optimization directions. Leveraging this insight, the authors propose Parallel-RL, a paradigm that decouples multi-task training to significantly improve efficiency and flexibility.

Original post →

More from Research

Research channel →