大语言模型不应被简单理解为"下一个Token预测器"

大语言模型的真实能力超越了单纯的序列预测。1虽然在预训练阶段，大语言模型通过学习训练数据中的实际Token序列来运作，但在后训练阶段，引入强化学习与可验证奖励（RLVR）使模型的功能发生了根本性变化。1此时模型不再仅预测训练数据中已有的序列，而是开始探索生成新序列并从奖励信号中学习。1

模型的内部编码已经包含了两层含义：一是"有用助手的模拟"，二是"通过探索发现的知识"。1这一根本区别可以用国际象棋引擎来类比——从学习大师棋局的"下一步预测"，到通过穷尽探索学习获胜策略，两者的能力本质上截然不同。1因此，仅将大语言模型定义为下一个Token预测器忽视了其在后训练阶段获得的超越预测本身的真实能力。1

Large language models should not be understood solely as next-token prediction systems, despite their autoregressive architecture that generates tokens sequentially.1 While LLMs operate through learning actual token sequences in training data during the pretraining phase, the introduction of post-training techniques fundamentally transforms their capabilities.1 Specifically, reinforcement learning with verifiable rewards enables models to explore new sequences and learn from reward signals rather than merely reproducing patterns from their training corpus.1

This distinction carries profound implications for how we conceptualize model capabilities.1 Beyond their structural reliance on token-by-token generation, post-trained models develop internal representations that encode simulated helpful assistant behavior and knowledge discovered through exploration—characteristics that extend far beyond simple prediction.1 The transformation mirrors the difference between learning to predict the next move in master chess games and acquiring winning strategies through exhaustive search and analysis.1