STTR Fast-Track: Combinatorial Reasoning

The broader/commercial impact of this Small Business Technology Transfer Fast-Track Pilot project lies in advancing artificial intelligence (AI) systems to think in a sophisticated and transparent way, addressing the growing challenge of errors and unreliable outputs in language models used for complex decision-making. This effort tackles issues like inaccurate reasoning which can lead to costly mistakes in fields such as finance, law, and healthcare. Blending language models with novel efficient computing methods inspired by physics leads to smarter AI that mimics human-like reasoning while being energy-efficient. This project will contribute to boost U.S. leadership in AI and will encourage further research in hardware, software and theory for optimization leading to economic competitiveness. Additional benefits include fostering sustainable computing practices and advancements in hardware and software that could create jobs and enhance AI applications in government, sustainability, and beyond, ultimately benefiting society with safer and more effective technology. This Small Business Technology Transfer Fast-Track Pilot project advances a high-risk, high-reward innovation mapping abstract reasoning into discrete, optimizable structures—a departure from purely neural or reward-driven paradigms of AI. More particularly, the project will establish the setup and empirical best practices for leveraging Ising machines, which are specialized physics-based optimizers running on high-performance computing hardware, in AI reasoning pipelines. The core technical innovation consists of a framework for sampling multiple outputs from large language models, embedding reasoning steps semantically, constructing a cost function based on frequencies and correlations, and optimizing it to select coherent reasoning chains for final AI responses. The scope encompasses designing an advanced cost function in the initial phase, incorporating higher-order correlations inspired by physics principles, and tuning accelerated Ising solvers in the subsequent phase to achieve reliable performance on reasoning benchmarks. Performance will be evaluated using metrics on diverse datasets, including confidence intervals, aiming for improvements over baselines while ensuring energy efficiency and explainability, with real-world testing in retrieval-augmented scenarios to validate commercial viability. This project will enhance AI accuracy and reduce hallucinations without the need of extensive pre-training and will result in a modular reasoning engine compatible with both proprietary and open-source large-language models. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.