Quantum Combinatorial Reasoning for Large Language Models

TL;DR

QCR-LLM introduces a first experimental framework for quantum-assisted reasoning in large language models by mapping multi-sample Chain-of-Thought fragments into a higher-order HUBO objective $H(\mathbf{x})$ and solving it with classical simulated annealing or quantum BF-DCQO on IBM hardware. By designing coefficients that encode fragment importance, co-occurrence, and semantic diversity, the approach selects a stable subset of reasoning fragments that re-inform prompts to improve accuracy across BIG-Bench Extra Hard tasks. Results show consistent gains over base backbones and competitive or superior performance against reasoning-native baselines, with substantial energy advantages at typical inference budgets. The work demonstrates a solver-agnostic, energy-conscious path to more coherent, interpretable, and scalable reasoning, and points toward higher-order interactions and hybrid multi-model pipelines as future avenues for quantum-accelerated AI.

Abstract

We design and implement a quantum combinatorial reasoning framework for large language models (QCR-LLM), integrating a real quantum computer in the hybrid workflow. QCR-LLM reformulates reasoning aggregation as a higher-order unconstrained binary optimization (HUBO) problem. In this sense, reasoning fragments are represented as binary variables and their interactions encode statistical relevance, logical coherence, and semantic redundancy. We tackle the resulting high-order optimization problem both classically, via simulated annealing, and quantumly through the bias-field digitized counterdiabatic quantum optimizer (BF-DCQO) executed on IBM's superconducting digital quantum processors. Experiments on BIG-Bench Extra Hard (BBEH) benchmarks demonstrate that our QCR-LLM consistently improves reasoning accuracy across multiple LLM backbones, surpassing reasoning-native systems such as o3-high and DeepSeek R1 by up to $+9\,$pp. Despite requiring multiple reasoning samples per query, our QCR-LLM remains approximately five times more energy-efficient than o3-high, owing to the low per-token energy footprint of its GPT-4o backbone. These results constitute the first experimental evidence of quantum-assisted reasoning, showing that hybrid quantum-classical optimization can efficiently enhance reasoning coherence, interpretability, and sustainability in large-scale language models. We have opened the doors to the emergence of quantum intelligence, where harder prompts require quantum optimizers at quantum-advantage level.

Figures (2)

• Figure 1: Example of the BF-DCQO optimization process for a single question from the Causal Understanding dataset.(a)Energy distributions obtained over three BF-DCQO iterations (ibm-aachen). The histogram shows convergence of the energy landscape toward lower minima; the dashed red and dotted black lines indicate respectively the minimum energy and the 25th percentile threshold used to define the low-energy subset. (b)Expected inclusion frequencies of each reason computed within this 25% lowest-energy set. The dashed magenta line represents the adjustable stability threshold: increasing it yields fewer but more consistent reasons, while relaxing it broadens the reasoning basis. Together, both panels illustrate how the BF-DCQO solver provides an interpretable energy landscape from which stable reasoning fragments are ranked and selected.
• Figure 2: Overview of the Quantum Combinatorial Reasoning (QCR-LLM) pipeline. Multiple zero-shot Chain-of-Thought completions are sampled from the base LLM, producing independent reasoning trajectories. Distinct reasoning fragments (reasons) are extracted and aggregated into a high-order binary optimization model (HUBO). The resulting Hamiltonian is solved via the Bias-field Digitized Counterdiabatic Quantum Optimization (BF-DCQO) on IBM digital quantum hardware. The selected reasoning subset is then reintroduced into the model as contextual evidence to produce the final answer.