SELR framework lets LLMs explain latent reasoning in plain language

Latent reasoning has been a promising alternative to text-based Chain-of-Thought (CoT), letting LLMs compress verbose thinking into compact embeddings and cut token costs significantly. But that efficiency comes with a steep trade-off: the reasoning becomes opaque. Existing methods either behave as unexplainable black boxes, like Coconut, or depend on separate decoders for explainability, like Heima—adding architectural overhead and decoupling the explanation from the actual reasoning process.

SELR, proposed by Dayuan Zhao, Shengcao Cao, Yu-Xiong Wang, and Liang-Yan Gui, solves this with a unified framework and a novel multi-task training objective. The Answer Loss optimizes the latent trajectory to produce correct final answers, while the CoT Loss forces the same model to decode its own latent representations into human-understandable reasoning steps. This dual-loss design ensures latent embeddings are both task-effective and semantically interpretable—no auxiliary models required. The team validated SELR on both LLMs and VLMs, reporting superior token efficiency and accuracy versus baselines, while uniquely offering self-contained explainability. The approach makes it possible to audit AI decisions without sacrificing the speed of latent reasoning, a critical step for real-world deployments where trust and cost both matter.