arXivLLM
Base Models Can Reason: Unlocking Latent Performance with Strategic Starting Tokens
基礎模型也能推理：啟動關鍵「開頭 token」釋放隱藏實力
A new study reveals that forcing base models to start with specific token cues like 'Okay' triggers reasoning behavior comparable to RL-tuned models, tracing this effect directly to pre-training data structures.
2 min read