New Framework Makes Hidden-State Reasoning in AI Models More Trainable and Interpretable

A team of researchers has introduced SWITCH, a switchable latent reasoning framework designed to overcome key obstacles in training and interpreting language models that reason in hidden states rather than visible text. The core innovation is a pair of explicit boundary tokens — <swi> and </swi> — that mark entry into and exit from a latent computation block, making the process compatible with the GRPO on-policy reinforcement learning objective by ensuring policy ratios are well-defined at every decision point. The model is trained using a visible-to-latent curriculum alongside a Switch-GRPO objective that propagates gradients through recurrent latent computation. Mechanistic analysis enabled by the boundary tokens yielded three notable findings: the <swi> token functions as a sharply localized, learned switching policy rather than a stylistic artifact; the latent computation it initiates is problem-specific and causally important rather than a passive placeholder; and this computation is concentrated at a single hidden-state transition upon entry. SWITCH consistently outperforms prior hidden-state-recurrence latent reasoning approaches at comparable scale, and the framework also enables direct probing of how on-policy RL improves the model's internal representations over training.