mSFT: Addressing Dataset Mixtures Overfitting Heterogeneously in Multi-task SFT

Multi-task supervised fine-tuning (SFT) often casts a heterogeneous data mixture as a single optimization problem, even though different tasks may reach their best generalization at different times. msft exposes this mismatch through task-wise roll-out, exclusion, and rollback, but its original formulation materializes the scheduler state as full-model checkpoints, making stage transitions costly to store, restore, and deploy. This paper introduces AuroSFT, a parameter-efficient framework that recasts the carried state of overfitting-aware multi-task SFT as a compact, mergeable adapter state. AuroSFT freezes the pretrained backbone, trains only injected adapters, rolls back adapter checkpoints at task-wise peaks, and continues on the remaining active mixture. At the layer level, each adapter applies an AuroRA-inspired adaptive nonlinear layer to a low-rank weight factor rather than to the sample representation. The resulting update remains linear in the input, rank-bounded, and exactly mergeable into the frozen projection. Under the retained-backbone comparison protocol, AuroSFT achieves 61.36% average accuracy, compared with 59.85% for the corresponding msft reference row, and obtains higher accuracy on all five backbones. Our code is available at the anonymous repository: https://anonymous.4open.science/r/AuroSFT-80D1.

Supervised fine-tuning performance for large language models depends strongly on how training budget is distributed across a heterogeneous set of tasks. In practice, mixtures are often fixed using simple heuristics (e.g., uniform or size-proportional sampling) that ignore task interactions, which can hurt transfer and waste budget on redundant sources. We introduce TaskPGM, a framework for learning continuous task mixtures via an energy-based model over tasks. Tasks form the nodes of a Markov random field: unary potentials capture per-task utility, and pairwise potentials encode inter-task relationships using behavioral divergences computed from predictive distributions of single-task fine-tuned models (e.g., Jensen--Shannon divergence and pointwise mutual information). Optimizing this objective yields mixtures that balance coverage against redundancy. We show that the resulting set function is weakly submodular under budget constraints, enabling approximation guarantees for discrete selection variants. Across multiple model families (LLaMA-7B, Qwen2-7B) and evaluation suites (BIG-Bench Hard), TaskPGM improves over standard mixing strategies and provides interpretable structure over task interactions.

Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) exhibit fundamentally different behaviors in enhancing multi-task reasoning for large language models (LLMs). Our preliminary experiments revealed a phenomenon: SFT suffers from severe task conflicts under multi-stage training, whereas RL enables stable coexistence across diverse tasks. Empirically, we trace this to the parameter level, observing that RL induces sparse and approximately orthogonal updates across tasks. We provide a theoretical explanation for this mechanism by analyzing multi-task gradient interference. Our results reveal a distinction: interference in SFT is norm-limited, scaling with the absolute gradient magnitude, whereas interference in RL is variance-limited, bounded by the gradient variance induced by advantage normalization and on-policy optimization. This small variance bound yields near-orthogonal optimization directions across tasks. Leveraging this insight, we propose Parallel-RL, a paradigm that decouples multi-task training, significantly improving efficiency and flexibility.