Preprint claims polynomial-time relational reasoning for language models
A new arXiv paper proposes recoding training data into explicit object relationships before training, aiming to make core relational reasoning tractable without inflating compute costs.
A preprint posted to arXiv on May 15 argues that large language models can reason more reliably without ballooning compute costs. The paper, Enhanced and Efficient Reasoning in Large Learning Models (arXiv:2605.14036v1), introduces a two-stage method: first, recode training data into what the authors call Unary Relational Integracode, which makes relationships among objects explicit rather than leaving them scattered across token sequences; second, run standard machine learning on that recoded corpus.
The central technical claim is that this recoding renders a core subset of relational rules polynomial-time learnable, with the polynomial's degree tied to rule complexity. That would mean the model can learn sound reasoning patterns within a single classifier call and across multiple calls, addressing what the authors frame as the lack of "principled basis to justify trust in the content" of current LLM outputs. The authors frame the approach through Robust Logic, a system for chaining inferences over uncertain learned facts, and claim it applies beyond text to vision and action domains where multiple properties of an object need to be unified. The method is designed to work with existing software and hardware stacks, positioning it as a preprocessing layer rather than a full architecture overhaul.