Reliable Reasoning with Large Language Models via Preference-Based Maximum Satisfiability

Researchers propose a hybrid reasoning system that combines Large Language Models with preference-based Maximum Satisfiability solvers to tackle complex optimization problems with multiple constraints. The approach achieves over 80% correctness rates on preference-based reasoning tasks, substantially outperforming traditional LLM baselines that rarely produce feasible solutions.

This research addresses a fundamental limitation of current Large Language Models: their difficulty with constraint-satisfaction and optimization problems that require precise logical reasoning. While LLMs demonstrate strong natural language understanding, they frequently fail at tasks requiring simultaneous satisfaction of multiple constraints and user preferences—common requirements in robotics, scheduling, and resource allocation domains.

The proposed hybrid approach leverages each technology's strengths by having LLMs generate Python code that translates natural language problem descriptions into MaxSAT formulations, which dedicated solvers then optimize. This externalization of reasoning enables verification of solutions against canonical encodings, ensuring correctness independent of the specific code generated. The researchers tested both open-source and proprietary LLMs, comparing their MaxSAT pipeline against direct answers, chain-of-thought reasoning, and program-of-thought approaches using identical models.

Results demonstrate dramatic improvements: while baseline methods rarely produced feasible solutions, the MaxSAT pipeline consistently achieved 80%+ acceptance rates. This capability matters for industries where constraint violations carry significant costs—incorrect robot task scheduling, invalid logistics routing, or infeasible resource allocation can be costly or dangerous. The verification mechanism addresses a critical trust issue in AI systems.

Looking forward, this framework suggests a broader pattern: LLMs may prove most valuable not as direct problem-solvers, but as interfaces that transform human intent into formal specifications that specialized systems can verify and optimize. As AI systems integrate into safety-critical domains, combining language models with verifiable reasoning engines may become standard practice.

• →LLM-generated code combined with MaxSAT solvers achieves over 80% correctness on preference-based reasoning tasks versus baseline methods that rarely succeed.
• →The approach enables independent verification of solutions against canonical encodings, addressing correctness concerns in AI-driven optimization.
• →Hybrid reasoning systems pairing language models with specialized solvers outperform pure LLM approaches on constraint-satisfaction problems.
• →This method applies to robotics, scheduling, and logistics domains where multiple constraints and user preferences must be simultaneously satisfied.
• →The research demonstrates LLMs work better as problem translators than direct optimizers, suggesting broader architectural implications for AI systems.