Long-context LLMs are constrained by the linear growth of key–value (KV) caches during autoregressive decoding, which incurs pronounced latency and memory overhead. KV eviction mitigates this issue, with existing efforts fall into offline policies with fixed eviction patterns and online policies that adaptively discard cache based on attention scores. While online eviction typically preserves accuracy under standard benchmarks, its performance can collapse in practical multi-turn dialogue scenarios where the query positions vary, and integration with pre-fill acceleration remains challenging. In contrast, offline eviction is infrastructure-friendly and generalizable but commonly sacrifices more accuracy. In this paper, we explore Supervised Fine-Tuning (SFT) for offline KV eviction and demonstrate its efficacy as a simple and powerful alternative to the design of complex online eviction metrics. We further propose Hybrid Sparse Supervised Fine-Tuning (HS-SFT) to explore the optimal offline design of KV eviction within SFT. In particular, HS-SFT employs a straight-through estimator to learn discrete local-window allocations of streaming heads across layers with budget-aware balancing loss, such that under high compression ratios—where dense-head capacity is constrained—the budget can be more effectively skewed to capture critical information. Across extensive evaluations on a wide array of LLMs and long-context tasks, HS-SFT delivers substantial performance gains over state-of-the-art eviction baselines. For example, with fewer than 4 hours of SFT of LLaMA-3-8B-1048K using a single 8-GPU node, HS-SFT achieves 5.86% and 38.3% higher average accuracy than Duo-attn on Longbench and Ruler-16K at 10% KV budget, respectively. These results position training-aware offline eviction—achieved with simple SFT—as an effective and practical path to scalable long-context inference.

Mixture-of-Experts (MoE) has emerged as the dominant paradigm for parameter-efficient scaling and cost-effective deployment of Large Language Models (LLMs). However, scaling laws developed for dense models are fundamentally inapplicable to MoE architectures, owing to three critical challenges: the multiplicity of governing factors, their intricate interdependencies, and the non-monotonic nature of their effects on model performance. These challenges collectively demand a rigorous and fine-grained investigation into MoE-specific scaling laws. In this work, we present a systematic decomposition of MoE configurations, identifying five key factors that influence model performance from both size and structural perspectives: data size (D), total model size (N), activated model size (N_a), number of active experts (G), and ratio of shared experts (S). We conduct 450 controlled experiments to characterize the marginal effects of each factor, culminating in a comprehensive and precise joint MoE scaling law. Building on this foundation, we derive the theoretically optimal and practically efficiency-aware optimal configurations for G, S and N_a/N with detailed analyses. Our findings reveal that the optimal settings for G and S are invariant to model and data size, while the optimal activation ratio N_a/N grows sparser as N increases. Our proposed MoE scaling law provides accurate and insightful guidance for future MoE model design and training.

Recent progress in large language models (LLMs) is largely driven by scaling training compute through either pre-training with next-token prediction (NTP) or post-training with reinforcement learning (RL). The former contributes to learning broad knowledge and skills from general data, while struggling with data inefficiency and catastrophic forgetting in continual learning settings. The latter incentivizes reasoning capabilities with strong generalization, but is constrained by limited data availability due to its reliance on human annotation. To alleviate these issues, we propose Reinforcement Learning on Pre-Training data (RLPT), which combines the advantages of learning from general data and RL. In particular, RLPT derives reward signals directly from general text data through a next-segment reasoning objective, rewarding the policy for correctly predicting next text segments conditioned on the prefix text. Experiments across multiple benchmarks and models demonstrate the effectiveness of . For example, RLPT yields substantial improvements in continual pre-training (+4.6%) and provides a strong foundation for post-training (+3.4%) on Qwen3-8B-Base.

Mixture-of-Experts (MoE) models rely on an external router to assign tokens to experts. This design inherently separates the routing decision from each expert’s internal capabilities, leading to suboptimal performance. In this work, we address this limitation with Union-of-Experts (UoE), an MoE variant that performs "expert-autonomous routing”. The core mechanism of UoE is to pre-designate a minute fraction of neurons within each expert as "routing neurons”. Experts autonomously select relevant tokens by comparing the activation intensity of these neurons, aligning routing decisions with each expert’s functional profile. To prevent the waste of activations from unselected experts’ routing neurons, we aggregate all routing neuron outputs and sum them into the final layer output. This aggregation acts as a novel virtual shared expert whose parameters are distributed across the individual experts, and improves overall parameter efficiency. We pre-train UoE models with up to 3B parameters, demonstrating that they outperform traditional MoEs with matched efficiency. Furthermore, our analysis of the routing neurons provides valuable insights into expert-autonomous selection and the broader routing mechanisms of MoE models.