Researchers propose Self-Explainable Latent Reasoning framework to make latent reasoning interpretable

Researchers have introduced a framework called Self-Explainable Latent Reasoning (SELR) that aims to reconcile the efficiency of latent reasoning with the interpretability of natural language explanations. Unlike prior approaches that either produce opaque latent representations or require separate post-hoc decoders to generate explanations, SELR trains a single model to perform both latent reasoning and explanation generation simultaneously.

The core innovation is a multi-task training objective that jointly optimizes two losses: an Answer Loss, which guides the latent reasoning trajectory to produce accurate final answers, and a Chain-of-Thought (CoT) Loss, which explicitly trains the model to decode its own latent representations into human-understandable reasoning steps. This dual-objective design ensures that the latent representations remain both task-effective and semantically interpretable.

The authors evaluate SELR on both large language models (LLMs) and vision-language models (VLMs), reporting that the framework achieves superior token efficiency and accuracy compared to baseline methods. Crucially, SELR provides self-contained explainability without relying on auxiliary models, addressing a key limitation of prior latent reasoning systems such as Coconut and Heima.

The work highlights a persistent trade-off in latent reasoning: systems that compress reasoning into latent space gain computational efficiency but lose interpretability, while systems that use external decoders for explanations introduce architectural overhead and decouple explanations from the actual reasoning process. SELR’s unified approach seeks to eliminate this trade-off by embedding explainability directly into the reasoning process.