Think in Latent, Explain in Language: Self-Explainable Latent Reasoning

August 17, 2026 | Source: arXiv AI

Tags: latent-reasoning, interpretability, chain-of-thought, LLM, VLM, explainability

Self-Explainable Latent Reasoning (SELR) trains a single model to reason efficiently in latent space while decoding its own latent representations into human-readable Chain-of-Thought steps—outperforming both efficiency-only (Coconut) and separate-decoder (Heima) approaches on accuracy, token efficiency, and interpretability.

Details

Latent reasoning approaches like Coconut compress Chain-of-Thought reasoning into embeddings, improving efficiency but producing opaque outputs. Existing explainability fixes—like Heima—add separate decoder architectures, decoupling the explanation from actual reasoning and adding architectural overhead. SELR resolves this tension with a multi-task training objective: an Answer Loss trains the latent reasoning trajectory to produce accurate answers, while a CoT Loss explicitly trains the same model to decode its own latent representations into human-readable reasoning steps. The key is that both objectives train the same model—the latent representations are grounded in interpretable structure by design, not retrofit. Validated on both Large Language Models and Vision-Language Models, SELR outperforms baselines on accuracy and token efficiency while providing self-contained explainability. No external decoder is needed to inspect the reasoning: the model decodes its own latent traces on demand. The regulatory relevance is direct: as AI explainability requirements grow in healthcare, finance, and other regulated domains, architectures that can produce auditable reasoning traces without separate infrastructure have a practical advantage. The project page is publicly available.