Valiant's new preprocessing method enables efficient, trustworthy LLM reasoning

Turing Award winner Leslie Valiant has published a paper proposing a principled method for efficient reasoning in large language models that addresses a core weakness: while LLMs produce fluent text, they lack trustworthy reasoning. Current wisdom holds that adding reasoning is computationally prohibitive. Valiant's method challenges this with a two-stage approach: first, data is preprocessed into a "Unary Relational Integracode" that explicitly encodes relationships between objects described in the text. Second, a standard (but potentially streamlined) machine learning process learns to predict these relationships. Crucially, this recoding makes the task of learning a core subset of relational rules polynomial-time learnable, with the polynomial depending on rule complexity. The method can be viewed as realizing a world model, and it retains much of the existing software and hardware base, making it practical for current LLM deployments.

Valiant articulates the advantages using "Robust Logic," a system for principled chaining on learned, uncertain information. The preprocessing has the surprising property that while succinct, it enables sound reasoning both within a single call of the learned classifier and across multiple calls. The approach extends beyond natural language to vision and actions — for example, bringing together multiple properties of an object referenced in an input, rather than leaving them distributed. This work, submitted to arXiv on May 13, 2026, falls at the intersection of AI, computational complexity, and machine learning. For professionals relying on LLMs for high-stakes decisions, Valiant's method offers a path to more reliable AI without requiring entirely new infrastructure or exponential compute.