Latent chemical reasoning works more like soft prompting than true reasoning
Reasoning or Just Prompting? Probing the Mechanism of Latent Chemical Reasoning
Machine LearningComputation and Language
Summary
The authors study a type of chemical prediction where a model does reasoning in a hidden continuous space instead of step-by-step with words or symbols. They find that changing or mixing up these hidden steps doesn't change results much, suggesting the hidden states don't encode detailed logical steps. Instead, they act more like instructions that guide the model’s next moves, similar to soft prompts used in AI models. This challenges the idea that the model is performing real reasoning internally and suggests it uses simpler guiding cues.
What this means in practice
- •For chemical data scientists: Use soft prompt techniques to design models that predict chemical properties without complex step-wise reasoning chains.
- •For natural language processing engineers: Incorporate fixed or learned soft prompts in language models to steer multi-step prediction tasks with less sensitivity to intermediate state changes.
Authors
Kaiyuan Lu, Jiatong Li, Yatao Bian
Abstract
Latent reasoning performs intermediate computation in continuous space rather than through explicit tokens, which is particularly appealing in chemistry, where errors in explicit reasoning can propagate to the final answer. Yet if latent states truly constitute an implicit reasoning chain, they should be causally important. We find otherwise: swapping or moderately perturbing intermediate latent states has little effect on performance, while removing them entirely sometimes causes a substantial score drop. Thus, the latent states matter, but their molecule-specific content appears to matter far less. This points to a simpler interpretation. Rather than encoding a step-by-step reasoning chain, latent states may primarily steer subsequent model computation, much like soft prompts. Indeed, a single fixed prefix shared across molecules reproduces LatentChem's performance, and a learned static soft prompt performs comparably to per-input latent generation in controlled models. The tolerance of latent sequences to substitution and moderate perturbation further supports this connection. We therefore suggest that latent chemical reasoning may be better viewed as soft prompting than as a reasoning chain hidden in continuous space.