Menu

Training-free early exit for reasoning models: a linear probe folded into the row of the output layer [p]

Reading 0:00
15s threshold

Reasoning models often reach the right answer early and then keep double-checking it for thousands of tokens. I got the model to stop at that point by changing a single row of weights. .There's no training and no extra code at inference time; the edited model is an ordinary weights file that runs anywhere.

Processing img 5yk991rjlt1i1...

Background,

Reasoning models (DeepSeek, Qwen, and so on) "think" in a scratchpad before answering. The scratchpad ends when the model writes a special token, </think>.

To pick each next token, the model first computes a summary of everything so far: a long list of numbers called the hidden state. The last layer, the output layer, has one row of weights per word in the vocabulary.…

Continue reading — create a free account

Join HashtagPLUS to read full articles, follow hashtags, vote, and join the conversation.