LatentChem: From Textual CoT to Latent Thinking in Chemical Reasoning

Current chemical large language models (LLMs) predominantly rely on explicit Chain-of-Thought (CoT) to solve complex reasoning problems. However, forcing nonverbal tacit chemical logic into discrete natural language imposes a fundamental ``modality mismatch,'' creating an artificial bottleneck for reasoning. We introduce LatentChem, a reasoning interface that decouples chemical logic from linguistic generation, enabling the model to process information via continuous thought vectors and dynamic perception. Our investigation reveals a pivotal emergent behavior: spontaneous internalization, defined here as self-selected under outcome-only optimization. When optimized for task success, the model abandons verbose textual derivations in favor of implicit latent computation, suggesting that it identifies the continuous manifold as a more native substrate for chemical logic. This paradigm shift also proves to be a superior computational strategy: LatentChem achieves a 59.88\% non-tie win rate against the strong CoT baseline on the rigorous ChemCoTBench, while delivering a broad 10.84$\times$ average reduction in reasoning step overhead (5.96$\times$ wall-clock speedup) across all evaluated benchmarks. Our results provide empirical evidence that chemical reasoning is more naturally and effectively realized as continuous latent dynamics rather than discretized linguistic trajectories.

AI tools can help chemists design molecules, but many current systems solve chemistry problems by writing out long step-by-step explanations. This can be awkward because chemistry often depends on shapes, patterns, and smooth changes that are hard to squeeze into words. We built LatentChem, a system that lets an AI model do much of its chemical reasoning in compact internal signals before writing the text response. Surprisingly, after training, the model often stopped writing long explanations and chose these internal steps on its own. This made it both faster and stronger: it beat a strong step-by-step system on challenging chemistry benchmarks and needed about ten times fewer reasoning steps on average. The result suggests that future scientific AI systems may work better when they can reason in forms closer to the structure of the science, while still explaining themselves when people need to check the answer.