Paper: Injected </think> Doesn't Cleanly Halt LRM Reasoning

Injecting an end-of-think token to stop a reasoning model's chain-of-thought does not always work as advertised, according to a new arXiv paper by Seunghee Koh, Sungjae Choi, Minchan Kwon, Sunghyun Baek and Junmo Kim. Training-free early-exit methods force the </think> token to trigger the reasoning-to-answering transition, but the model often keeps producing reasoning-like content in what is supposed to be the answer.

The paper names this failure mode spurious CoT termination. "Answering-phase generation can continue before the model regenerates another EoT," the authors write, "with the span preceding this regenerated EoT scaling with the reasoning tokens saved by early exit and exhibiting continued reasoning behavior." The more tokens early-exit saves, the further the model reasons past the injected stop signal before eventually terminating on its own.

The team traces the problem to attention: the model does not attend strongly enough to the injected EoT. Their probe, Exit-token Attention Biasing (EAB), boosts that attention. Across four large reasoning models, five benchmarks and two early-exit methods, biasing attention toward the injected token reduces spurious termination and shortens the answering phase.

The authors close on the broader point: matching a model's think-block format from outside is not the same as controlling what it does. "Inserting the EoT token conforms to this format but does not by itself guarantee the intended reasoning-to-answering transition."