NeurIPS 2025 Spotlight | 选择性知识蒸馏精准过滤：推测解码加速器AdaSPEC来了

Core Insights - The article discusses the introduction of AdaSPEC, an innovative selective knowledge distillation method aimed at enhancing speculative decoding in large language models (LLMs) [3][9][16] - AdaSPEC focuses on improving the alignment between draft models and target models by filtering out difficult-to-learn tokens, thereby increasing the overall token acceptance rate without compromising generation quality [3][11][16] Research Background - LLMs excel in reasoning and generation tasks but face high inference latency and computational costs due to their autoregressive decoding mechanism [6] - Traditional acceleration methods like model compression and knowledge distillation often sacrifice generation quality for speed [6] Method Overview - AdaSPEC employs a selective token filtering mechanism that allows draft models to concentrate on "easy-to-learn" tokens, enhancing their alignment with target models [3][9] - The method utilizes a two-stage training framework: first, it identifies difficult tokens using a reference model, and then it filters the training dataset to optimize the draft model [11][12] Experimental Evaluation - The research team conducted systematic evaluations across various model families (Pythia, CodeGen, Phi-2) and tasks (GSM8K, Alpaca, MBPP, CNN/DailyMail, XSUM), demonstrating consistent and robust improvements in token acceptance rates [14] - Key experimental results indicate that AdaSPEC outperforms the current optimal DistillSpec method, with token acceptance rates increasing by up to 15% across different tasks [15] Summary and Outlook - AdaSPEC represents a precise, efficient, and universally applicable paradigm for accelerating speculative decoding, paving the way for future research and industrial deployment of efficient LLM inference [16] - The article suggests two potential avenues for further exploration: dynamic estimation mechanisms for token difficulty and application of AdaSPEC in multimodal and reasoning-based large models [17]