SFT Conflicts, RL Coexists: New Paper Unpacks LLM Multi-Task Training Dynamics

burny_tech · x · 2026-08-11

A new paper provides a theoretical and empirical analysis of multi-task learning in LLMs, investigating why Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) behave differently. It reveals that SFT suffers from severe task conflicts because it makes large, overlapping parameter updates that overwrite prior knowledge.

Conversely, RL induces sparse, near-orthogonal updates, allowing diverse capabilities to coexist with minimal interference. The authors explain that SFT interference scales with absolute gradient magnitude, whereas RL interference is bounded by gradient variance from advantage normalization. Leveraging this insight, they propose Parallel-RL, a paradigm that decouples multi-task training to significantly improve efficiency and flexibility.

Related event: Study Reveals Multi-stage SFT Causes Task Conflicts, Favoring RL(4 posts)→

Original post →

More from Research

Research channel →