Latent Reasoning Faithfulness Varies During AI Training

Latent reasoning methods allow AI models to perform multi-step inference within their hidden states, promising efficiency. However, the faithfulness of these opaque reasoning steps—whether they genuinely drive the final answer—is a critical question. Previous work examined this at converged model checkpoints, revealing unfaithful behaviors, but the evolution of faithfulness during the training process remained unexplored. This research tracks how faithfulness changes across various training checkpoints for different latent reasoning paradigms. By applying verifiable counterfactual edits to inputs and noise-ablation activation patches to latent reasoning steps, the study uncovered several key findings. It was observed that while latent reasoning methods might appear similarly unfaithful at convergence under counterfactual edits, their faithfulness trajectories during training diverge qualitatively. Crucially, the causal contribution of latent reasoning steps to the final answer tends to decay over the course of training, particularly for binary choice tasks, while it may rise for open-ended decoding. These findings underscore that latent reasoning faithfulness is not static but is highly dependent on the training stage and the specific answer format.