Reasoning-capable large language models (LLMs) achieve strong performance on complex tasks but often exhibit overthinking after distillation, generating unnecessarily long chain-of-thought (CoT) reasoning even for simple inputs and incurring high inference cost. However, naively shortening reasoning length can degrade reasoning accuracy, as concise reasoning may be insufficient for certain inputs and lacks explicit supervision. We propose Auto Long-Short Reasoning (AutoL2S), a distillation framework that empowers non-reasoning LLMs to think thoroughly but only when necessary. AutoL2S first learns a lightweight switching token with verified long-short CoTs to enable instance-wise long-short reasoning selection. Then it leverages long-short reasoning rollouts induced by switching tokens within a GRPO-style loss to improve reasoning efficiency while maintaining accuracy. Experiments demonstrate that AutoL2S effectively reduces reasoning length up to 71% with minimal accuracy loss, yielding markedly better trade-off in token length and inference time while preserving accuracy. The code is available at https://github.com/amandaluof/AutoL2S.

Diffusion language models (dLLMs) have recently shown strong potential in various tasks with competitive decoding efficiency, but their bidirectional attention incurs quadratic compute and memory complexity. Many prior works rely on KV cache optimizations to accelerate inference. However, the full-sequence BF16 KV cache remains the dominant memory bottleneck. We propose SampKV, a Saliency-aware Mixed-Precision KV cache for dLLMs that stores most history in 2 bits while promoting a small budget of attention-salient tokens to high precision. A customized Triton kernel operates directly on the mixed-precision cache and avoids materializing full-precision KV tensors, alleviating the storage and traffic bottleneck. Experiments on LLaDA and Dream models show that the memory bottleneck is substantially alleviated: up to 54% peak memory reduction at matched workload, or up to 2.06x throughput improvement at matched peak memory compared to the BF16 baseline, while maintaining the accuracy.

The emergence of Mixture of Experts (MoE) LLMs has significantly advanced the development of language models. Compared to traditional LLMs, MoE LLMs outperform traditional LLMs by achieving higher performance with considerably fewer activated parameters. Despite this efficiency, their enormous parameter size still leads to high deployment costs. In this paper, we introduce a two-stage compression method tailored for MoE to reduce the model size and decrease the computational cost. First, in the inter-expert pruning stage, we analyze the importance of each layer and propose the Layer-wise Genetic Search and Block-wise KT-Reception Field with the non-uniform pruning ratio to prune the individual expert. Second, in the intra-expert decomposition stage, we apply the low-rank decomposition to further compress the parameters within the remaining experts. Extensive experiments on Qwen1.5-MoE-A2.7B, Deepseek-V2-Lite, and Mixtral-8×7B, demonstrate that our proposed methods can both reduce the model size and enhance inference efficiency while maintaining performance in various zero-shot tasks.