LatentChem: From Textual CoT to Latent Thinking in Chemical Reasoning

TL;DR

The paper identifies a fundamental mismatch between discrete linguistic CoT and the continuous, structural nature of chemical reasoning. It introduces LatentChem, a latent-thinking interface that decouples chemical computation from language, enabling multi-step reasoning in continuous latent space with a dynamic perception loop powered by ChemAdapter, ChemUpdater, and LatentProjector. A four-stage training protocol culminating in a Group Relative Policy Optimization phase reveals an emergent phenomenon: models spontaneously internalize reasoning into latent space, achieving superior performance and large inference-speedups over explicit CoT baselines across multiple chemical benchmarks. This work demonstrates that chemical reasoning can be more naturally and efficiently realized through continuous latent dynamics, offering a substantial practical impact for faster, more reliable chemical AI systems and prompting future exploration of hybrid cognitive architectures that can expose latent traces when needed.

Abstract

Chemical large language models (LLMs) predominantly rely on explicit Chain-of-Thought (CoT) in natural language to perform complex reasoning. However, chemical reasoning is inherently continuous and structural, and forcing it into discrete linguistic tokens introduces a fundamental representation mismatch that constrains both efficiency and performance. We introduce LatentChem, a latent reasoning interface that decouples chemical computation from textual generation, enabling models to perform multi-step reasoning directly in continuous latent space while emitting language only for final outputs. Remarkably, we observe a consistent emergent behavior: when optimized solely for task success, models spontaneously internalize reasoning, progressively abandoning verbose textual derivations in favor of implicit latent computation. This shift is not merely stylistic but computationally advantageous. Across diverse chemical reasoning benchmarks, LatentChem achieves a 59.88\% non-tie win rate over strong CoT-based baselines on ChemCoTBench, while delivering a 10.84$\times$ average inference speedup. Our results provide empirical evidence that chemical reasoning is more naturally and effectively realized as continuous latent dynamics rather than discretized linguistic trajectories.

Figures (14)

• Figure 1: Conceptual illustration of the continuity–discretization gap in chemical reasoning. (a) The intrinsic chemical property landscape is continuous and high-dimensional. (b) We posit that a continuous latent space can theoretically offer a smoother optimization surface akin to the property landscape, avoiding the jagged trajectories of discrete tokens. (c) Linguistic tokenization fragments chemical state transitions into discrete symbolic steps, inducing staircase-like landscapes and inefficient reasoning paths.
• Figure 2: Comparison of reasoning paradigms in chemical LLMs. (Top) Explicit CoT relies on discrete linguistic steps, forcing high-dimensional chemical intuition into a constrained textual bottleneck. (Middle) Generic latent reasoning (e.g., Coconuthao2025traininglargelanguagemodels) shifts reasoning to a continuous latent space but treats molecular embeddings as a static context, limiting the ability to focus on different substructures during reasoning. (Bottom) LatentChem (Ours) introduces a dynamic perception-reasoning loop. Leveraging the ChemUpdater mechanism, latent thoughts actively re-query and refine the molecular representation at each step, ensuring structure-aware reasoning. Depending on the complexity of the query and the allocated latent thinking budget, the model exhibits dynamic reasoning behaviors, ranging from fully latent computation (no textual CoT) to hybrid reasoning with varying lengths of textual CoT.
• Figure 3: The overview of LatentChem architecture. The system decouples reasoning from generation via a dedicated latent thinking phase. (1) A ChemAdapter aligns molecular features with the LLM space. (2) During the reasoning phase, the model generates a sequence of continuous latent thought vectors. (3) The ChemUpdater allows these thoughts to recursively re-query the molecular encoder, refining the representation based on the current reasoning state before the final response is decoded.
• Figure 4: Case study: spontaneous internalization vs. explicit CoT. While the standard CoT model generates verbose reasoning chains that fail to execute the planned modification, LatentChem spontaneously internalizes these logics. It bypasses textual output, utilizing the latent space to perform structural optimization.
• Figure 5: Causal necessity analysis. We measure task performance (success rate, scaffold similarity, or correct rate) as the first $k$ latent tokens are replaced with Gaussian noise. The monotonic degradation observed in both Molecule Optimization (left) and Understanding (right) confirms that early latent states encode critical precursors for the final solution rather than redundant noise.

Theorems & Definitions (7)

• Definition A.1: Chemical Manifold
• Definition A.2: Optimal Reasoning Path
• Definition A.4: Modality Representation Trajectory
• Definition A.5: Valid Discrete Approximation
• Definition A.6: Effective Curvature in Representation Space
• Definition A.7: Intrinsic Semantic Resolution (Path-Advance)
• Theorem A.9: Curvature--Resolution Trade-off (Qualitative Lower Bound)