Sparse Mixture-of-Experts (SMoE) architectures have enabled a new frontier in scaling Large Language Models (LLMs), offering superior performance by activating only a fraction of their total parameters during inference. However, their practical deployment is severely hampered by substantial static memory overhead, as all experts must be loaded into memory. Existing post-training pruning methods, while reducing model size, often derive their pruning criteria from a single, general-purpose corpus. This leads to a critical limitation: a catastrophic performance degradation when the pruned model is applied to other domains, necessitating a costly re-pruning for each new domain. To address this generalization gap, we introduce Mosaic Pruning (MoP). The core idea of MoP is to construct a functionally comprehensive set of experts through a structured ``cluster-then-select" process. This process leverages a similarity metric that captures expert performance across different task domains to functionally cluster the experts, and subsequently selects the most representative expert from each cluster based on our proposed Activation Variability Score. Unlike methods that optimize for a single corpus, our proposed Mosaic Pruning ensures that the pruned model retains a functionally complementary set of experts, much like the tiles of a mosaic that together form a complete picture of the original model's capabilities, enabling it to handle diverse downstream tasks.Extensive experiments on various MoE models demonstrate the superiority of our approach. MoP significantly outperforms prior work, achieving a 7.24\% gain on general tasks and 8.92\% on specialized tasks like math reasoning and code generation.

The design of Large Language Models (LLMs) has long been hampered by a fundamental conflict within their core attention mechanism: its remarkable expressivity is built upon a computational complexity of O(H·N²) that grows quadratically with the context size (N) and linearly with the number of heads (H). This standard implementation harbors significant computational redundancy, as all heads independently compute attention over the same sequence space. Existing sparse methods, meanwhile, often trade information integrity for computational efficiency. To resolve this efficiency-performance trade-off, we propose SPAttention, whose core contribution is the introduction of a new paradigm we term Principled Structural Sparsity. SPAttention does not merely drop connections but instead reorganizes the computational task by partitioning the total attention workload into balanced, non-overlapping distance bands, assigning each head a unique segment. This approach transforms the multi-head attention mechanism from H independent O(N²) computations into a single, collaborative O(N²) computation, fundamentally reducing complexity by a factor of H. The structured inductive bias compels functional specialization among heads, enabling a more efficient allocation of computational resources from redundant modeling to distinct dependencies across the entire sequence span. Extensive empirical validation on the OLMoE-1B-7B and 0.25B-1.75B model series demonstrates that while delivering an approximately two-fold increase in training throughput, its performance is on par with standard dense attention, even surpassing it on select key metrics, while consistently outperforming representative sparse attention methods including Longformer, Reformer, and BigBird across all evaluation metrics. Our work demonstrates that thoughtfully designed structural sparsity can serve as an effective inductive bias that simultaneously improves both computational efficiency and model performance, opening a new avenue for the architectural design of next-generation, high-performance LLMs.

Large language models (LLMs) have shown strong performance on hard reasoning and general instruction-following tasks.However, when sampling multiple outputs for the same prompt, they often produce highly homogeneous, repetitive responses, resulting in inefficient exploration.This limits the gains from test-time scaling and constrains the upper bound of reinforcement learning (RL) training.We attribute this issue in part to supervised fine-tuning (SFT): when a single prompt is paired with multiple reference responses, the model is trained to generate diverse outputs under the same prior condition, which induces optimization interference and can lead to diversity collapse.To address this, we propose Prefix-Conditioned SFT (P-SFT), a simple yet effective method that constructs semantically consistent yet distributionally distinct prior contents to different responses, thereby projecting the instruction into distinct latent regions to establish diverse prior distributions and decouple the one-to-many mapping.Experiments on large reasoning language models show that our approach improves absolute performance by 5.3% on reasoning benchmarks and increases generation diversity by 198.3% on average, while substantially enhancing output diversity and test-time scaling.Notably, even without any additional training, our prefixing strategy can be applied at inference time alone and still yields significant gains in both diversity and reasoning performance for instruction-tuned LLMs and reasoning-enhanced models.

Wentao Hu, Yanbo Zhai, Xiaohui Hu, Mingkuan Zhao, Shanhong yu, Xue Liu, Kaidong Yu, Shuangyong Song, Xuelong Li. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.