[2601.23155] SPICE: Submodular Penalized Information-Conflict Selection for Efficient Large Language Model Training

Computer Science > Machine Learning arXiv:2601.23155 (cs) [Submitted on 30 Jan 2026 (v1), last revised 8 Apr 2026 (this version, v2)] Title:SPICE: Submodular Penalized Information-Conflict Selection for Efficient Large Language Model Training Authors:Powei Chang, Jinpeng Zhang, Bowen Chen, Chenyu Wang, Chenlu Guo, Yixing Zhang, Yukang Gao, JianXiang Xiang, Yue Gao, Chaoqun Sun, Yiyi Chen, Dongying Kong View a PDF of the paper titled SPICE: Submodular Penalized Information-Conflict Selection for Efficient Large Language Model Training, by Powei Chang and 11 other authors View PDF HTML (experimental) Abstract:Information-based data selection for instruction tuning is compelling: maximizing the log-determinant of the Fisher information yields a monotone submodular objective, enabling greedy algorithms to achieve a $(1-1/e)$ approximation under a cardinality budget. In practice, however, we identify alleviating gradient conflicts, misalignment between per-sample gradients, is a key factor that slows down the decay of marginal log-determinant information gains, thereby preventing significant loss of information. We formalize this via an $\varepsilon$-decomposition that quantifies the deviation from ideal submodularity as a function of conflict statistics, yielding data-dependent approximation factors that tighten as conflicts diminish. Guided by this analysis, we propose SPICE, a conflict-aware selector that maximizes information while penalizing misalignment, and that supports...