用更一致的轨迹、更少的解码步数「驯服」掩码扩散语言模型，扩散语言模型的推理性能和效率大幅提升
"推理效率" 相关文章
用更一致的轨迹、更少的解码步数「驯服」掩码扩散语言模型，扩散语言模型的推理性能和效率大幅提升
FlashEVA: Accelerating LLM inference via Efficient Attention
DTS: Enhancing Large Reasoning Models via Decoding Tree Sketching
英伟达帮你省钱，让大模型推理「短而精」，速度快5倍
英伟达帮你省钱，让大模型推理「短而精」，速度快5倍
让LLM不再话痨，快手HiPO框架来了
Are Language Models Efficient Reasoners? A Perspective from Logic Programming
华为计算：KunLun AI Space 基于昇腾实现 DeepSeek V3.1 FP8 推理，成本减半
NVIDIA港大MIT联合推出Fast-dLLM v2：端到端吞吐量提升2.5倍
ARC-Encoder: learning compressed text representations for large language models
ARC-Encoder: learning compressed text representations for large language models
Metis-HOME: Hybrid Optimized Mixture-of-Experts for Multimodal Reasoning
Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs
Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs
DVAGen: Dynamic Vocabulary Augmented Generation
Compressing Many-Shots in In-Context Learning
One Token Embedding Is Enough to Deadlock Your Large Reasoning Model
One Token Embedding Is Enough to Deadlock Your Large Reasoning Model
CarBoN: Calibrated Best-of-N Sampling Improves Test-time Reasoning
