AIAny
Icon for item

SFT Conflicts, RL Coexists: A Theoretical and Empirical Analysis of Multi-Task Learning for LLMs

Analyzes why supervised fine-tuning (SFT) causes severe task conflicts under multi-stage multi-task training while reinforcement learning (RL) enables stable coexistence, attributing the effect to sparse, near-orthogonal RL parameter updates and proposing Parallel-RL to decouple multi-task training.

Introduction

Why this matters

Large language models are increasingly trained or adapted on many tasks in stages, but SFT-style pipelines often suffer catastrophic interference across tasks while RL-based adaptation appears to accumulate capabilities. The core insight: RL produces much smaller, sparser, and near-orthogonal parameter updates across tasks, which empirically and theoretically reduces gradient interference and enables stable multi-stage multi-task training.

Key Findings
  • Parameter-scale and sparsity: RL update vectors have a much smaller average L2 norm (~3×10^-2) versus SFT (~7.4), and only ~20% of parameters exceed a small magnitude threshold in RL compared to ~93% in SFT — implying RL changes are minimal and localized. This reduces destructive overwriting across tasks.

  • Directional decorrelation: Pairwise cosine similarity of parameter updates across RL tasks is effectively negligible (on the order of 10^-5), whereas SFT updates show high similarity (10^-1 to 1.0) and occasional opposing directions. Near-orthogonality explains why RL updates interfere far less.

  • Theoretical view of interference: SFT interference scales with absolute gradient norms (norm-limited), so large overlapping updates cause catastrophic forgetting; RL interference is variance-limited due to advantage normalization and on-policy optimization, yielding a small bound on cross-task interference and hence robustness in sequential training.

  • Empirical consequences & method: Multi-stage SFT can produce large average performance drops (reported average decline ~23%), while RL yields cumulative gains and robustness (reported gains ~25% in some settings). Building on these mechanisms, the paper proposes Parallel-RL, a training paradigm that decouples tasks to improve efficiency and flexibility.

Who it's for and tradeoffs

Great fit if you are researching or engineering multi-task adaptation of LLMs and need to understand failure modes of SFT versus RL, or if you want a principled approach (Parallel-RL) to reduce interference in staged pipelines. Look elsewhere if your environment cannot afford RL's practical costs: on-policy RL methods can be sample- and compute-inefficient, require careful reward/advantage design, and may need more infrastructure than straightforward SFT. The paper offers concrete empirical metrics (norms, sparsity, cosine similarities) and a theoretical framing to decide which approach suits your constraints.

Information

  • Websitearxiv.org
  • AuthorsKejian Zhu, Zhuoran Jin, Shangqing Tu, Hongbang Yuan, Yushi Bai, Kang Liu, Juanzi Li, Jun Zhao
  • Published date2026/08/04

More Items

Supervises audio reasoning by generating per-sample, audio-grounded rubrics that evolve with model rollouts and serve as reinforcement-learning rewards, improving perception and adaptive multi-step reasoning while avoiding reward saturation.

Analyzes how to build effective training environment distributions for multimodal agents and proposes Ability-aware Environment Selection (AES) and Hierarchical Difficulty Curriculum (HDC) to improve diversity and difficulty scheduling, yielding large relative gains in experiments.

Replaces external environment interaction in agentic RL training with 'world rehearsal': the policy alternates between making tool calls and simulating their environment responses, jointly optimizing both roles so the agent internalizes environment dynamics and improves long-horizon tool use and transfer.