Title: LatentChem: Transitioning from Textual Chain-of-Thought to Latent Thinking in Chemical Reasoning

Abstract:

Current chemical large language models (LLMs) primarily depend on explicit Chain-of-Thought (CoT) methodologies to address complex reasoning challenges. However, compelling nonverbal, tacit chemical logic into discrete natural language creates a fundamental "modality mismatch." This forced translation establishes an artificial bottleneck that hinders effective reasoning. To address this, we present LatentChem, a novel reasoning interface that separates chemical logic from linguistic generation. This approach allows the model to process information through dynamic perception and continuous thought vectors.

Our research highlights a significant emergent behavior: spontaneous internalization. We define this phenomenon as the self-selected adoption of internal processes when optimization is focused solely on outcomes. Under outcome-only optimization, the model discards verbose textual derivations in favor of implicit latent computation. This suggests that the model recognizes the continuous manifold as a more native substrate for chemical logic.

This paradigm shift also demonstrates superior computational efficiency. On the rigorous ChemCoTBench, LatentChem secured a 59.88% non-tie win rate against a strong CoT baseline. Furthermore, it delivered an average reduction of 10.84$\times$ in reasoning step overhead, translating to a 5.96$\times$ improvement in wall-clock speedup, across all evaluated benchmarks. These findings offer empirical evidence that chemical reasoning is more naturally and effectively executed as continuous latent dynamics rather than discretized linguistic trajectories.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC