Korea University Team Presents LLM-Qubo Framework for Multi-Drone Assignment Optimisation

A new framework combines large language models with quantum optimisation to efficiently assign tasks to multi-drone fleets, overcoming computational limitations of traditional computers. Junyeop Bang and colleagues at Korea University translate natural language instructions into quantum constraints and use a novel method for partitioning and compressing the problem for current quantum hardware. The architecture achieves optimal solutions in all tested ideal scenarios and maintains a 96.3% success rate with realistic quantum sampling, representing a key step towards scalable, natural-language-guided drone task allocation.

Quantum optimisation enables scalable multi-drone task allocation via constraint embedding

A framework has been developed for structurally strong Quadratic Unconstrained Binary Optimisation (QUBO) constraints without false negatives. To address qubit limits of near-term quantum devices, the framework features a novel constraint-preserving graph partitioner and a compressed separator-based dynamic programming (DP) merge. Encoding constraints via W-state initialisation and XY-mixers in Conditional Value-at-Risk Quantum Approximate Optimisation (CVaR-QAOA) keeps the pipeline highly compact.

Empirical results demonstrate that this architecture circumvents classical scaling walls, recovering the global optimum on 100% of idealized oracle cases and 96.3% under real QAOA sampling. This enables natural-language-guided task allocation at previously intractable scales. Deploying multi-drone systems for complex spatial tasks, such as regional surveillance, disaster response, and logistics, requires efficient and dynamic zone assignment. Traditionally, human operators translate high-level mission goals into precise programmatic constraints or mathematical formulations, a time-consuming and error-prone process requiring specialised expertise.

The advent of Large Language Models (LLMs) presents a growing opportunity to bridge this gap by enabling robots to parse free-form, natural-language instructions directly into executable task specifications. Converting natural language into structured mathematical models is exceptionally challenging, as even minor logical omissions can lead to critical optimisation failures. Recent advances have successfully demonstrated the automated translation of unstructured problem specifications into formal Quadratic Unconstrained Binary Optimisation (QUBO) models, using a systematic fine-tuning strategy that integrates Supervised Fine-Tuning (SFT) and Direct Preference Optimisation (DPO) to generate structured constraints with high structural accuracy and logical consistency.

An important strength of this pipeline is its exceptional durability in eliminating False Negatives, ensuring that the critical boundaries of the multi-drone task allocation problem are fully captured from the user’s natural language input. As the fleet size and the number of candidate zones increase, the multi-drone assignment problem triggers an exponential combinatorial explosion (O(M N)), quickly rendering classical exact solvers and exhaustive search methods intractable within real-time operational windows. To overcome this classical barrier, quantum optimisation algorithms, specifically the Quantum Approximate Optimisation Algorithm (QAOA), have emerged as a major alternative, offering the potential to explore massive state spaces concurrently.

However, physical hardware limitations currently block the direct application of QAOA to large-scale robotics problems. Real-world Noisy Intermediate-Scale Quantum (NISQ) devices are restricted by strict qubit budgets and gate error rates, meaning a monolithic LLM-generated QUBO instance will quickly overwhelm available quantum capacity. This work presents an end-to-end integration that bridges a strong language front-end with a highly optimised, domain-specific quantum-classical back-end.

This approach is tailored specifically to the multi-drone spatial assignment problem, rather than simply chaining existing computational tools. A domain-specific sub-QUBO encoding eliminates the severe qubit overhead typically associated with one-hot assignment penalty terms. Mapping the drone-zone constraints directly into W-state initializations and Hamming-weight-preserving XY-mixers within CVaR-QAOA creates a highly compact representation that structurally enforces hard constraints at the quantum circuit level.

Classical graph partitioning and separator-based dynamic programming are adapted to the quantum domain, establishing a scalable architecture that decomposes massive, LLM-generated assignment problems into hardware-compatible sub-QUBOs and recombines them while rigorously preserving the global optimum. Through this tailored integration, the framework successfully bypasses the classical exhaustive-search wall, demonstrating a practical blueprint for translating high-level human intent into scalable, globally optimal quantum execution. Traditional multi-robot task allocation (MRTA) relies on rigid, hand-coded formal specifications like MILP. While recent frameworks (e.g., SayCan, Code as Policies) enable LLMs to parse natural-language instructions, they mostly target single-robot tasks.

Extending LLMs to multi-drone systems via heuristic prompting lacks optimisation guarantees and is prone to hallucinated constraints. Quantum optimisation has emerged as a promising quantum alternative for solving NP-hard QUBO problems. Classical decomposition methods (e.g., ADMM) incur heavy communication overheads. Recent quantum-classical partitioners fit subproblems within NISQ limits but rely on greedy merges that destroy global context. This is overcome using a constraint-preserving partitioner coupled with a compressed separator-DP merge that provably recovers the global optimum.

As multi-drone fleets scale, zone assignment rapidly evolves into an intractable NP-hard combinatorial problem that overwhelms classical exhaustive search. While quantum optimisation offers a potential solution, mapping complex spatial tasks from human intent to restricted quantum hardware remains a severe challenge. To address this, a framework integrates a fine-tuned Large Language Model (LLM) with a highly scalable quantum-classical backend.

The front-end uses Supervised Fine-Tuning and Direct Preference Optimisation to translate natural language into structurally robust Quadratic Unconstrained Binary Optimisation (QUBO) constraints without false negatives. To overcome qubit limits, the framework features a constraint-preserving graph partitioner and a compressed separator-based dynamic programming merge. By encoding constraints via W-state initialisation and XY-mixers in Conditional Value-at-Risk Quantum Approximate Optimisation (CVaR-QAOA), the pipeline remains highly compact.

The entry point is a free-form instruction. A fine-tuned LLM (Qwen1.5-1.8B with LoRA) is employed to translate instructions into structured constraints (force/forbid assignments, minimum coverage). Supervised Fine-Tuning (SFT) fixes the output schema, while Direct Preference Optimisation (DPO) maximizes structural validity and eliminates missing constraints. A deterministic layer then validates these extracted constraints and emits the global QUBO formulation.

To address qubit limitations in quantum devices, a constraint-preserving graph partitioner is used to divide the zone assignment problem into smaller sub-problems. A Large Language Model translates natural language instructions into Quadratic Unconstrained Binary Optimisation constraints, avoiding missed constraints. The system excludes any drone whose detour to a zone exceeds a penalty threshold, safely removing assignments that cannot improve the overall objective.

This targeted approach ensures the partitioning process does not eliminate the globally optimal solution set. Subproblems are sized to fit quantum hardware limitations. The qubit footprint is reduced before quantum execution. A Large Language Model translates natural language instructions into Quadratic Unconstrained Binary Optimisation constraints without omitting any requirements. Constraint-preserving graph partitioning and dynamic programming then manage the remaining variables.

Quantum optimisation enables high-performance multi-drone task allocation with natural language

The team at Korea University has achieved a 96.3% success rate in multi-drone zone assignment using quantum optimisation, a substantial improvement over the 100% success achieved in idealised simulations. This breakthrough crosses a vital threshold for practical application, as previously, scaling these complex task allocations to larger drone fleets was impossible for both classical and standalone quantum computers. Their new framework integrates a Large Language Model with a quantum-classical backend, translating natural language instructions into actionable constraints for drones and then optimising assignments using quantum techniques.

This combination of approaches has created a system capable of handling previously intractable problems in multi-drone fleet assignment, which rapidly becomes a complex NP-hard problem overwhelming classical search methods. At Korea University, a framework integrating a fine-tuned Large Language Model (LLM) with a quantum-classical backend has demonstrated the ability to recover the global optimum in 100% of idealised cases and 96.3% under real Quantum Approximate Optimisation Algorithm (QAOA) sampling. The framework employs Supervised Fine-Tuning and Direct Preference Optimisation to translate natural language instructions into structurally robust Quadratic Unconstrained Binary Optimisation (QUBO) constraints, without missing any requirements during conversion. The system also utilises a novel method of dividing the complex problem into smaller parts and then recombining the solution, allowing it to work within the limitations of current quantum computers.

Natural language control of drone swarms via quantum-inspired task allocation

Deploying multi-drone systems demands increasingly sophisticated task allocation, yet current methods often rely on hand-coded instructions or limited heuristic prompting. The researchers acknowledge that their framework successfully bridges natural language and quantum computation, but currently operates within the constraints of Noisy Intermediate-Scale Quantum (NISQ) devices. This presents a tension; scaling to genuinely large drone fleets may require fundamentally different quantum architectures, potentially moving beyond the current reliance on partitioning and merging techniques.

However, acknowledging that genuinely large drone fleets will likely necessitate more advanced quantum hardware is important, yet dismissing this work as premature would be a mistake. The team’s method demonstrably expands the scale at which natural language can direct complex drone tasks, overcoming limitations of existing systems which struggle with even moderately sized groups. The team at Korea University has created a system enabling natural language instructions to directly control multi-drone task allocation, a significant step beyond manually programmed assignments.

This framework integrates a fine-tuned Large Language Model, which translates natural language instructions into Quadratic Unconstrained Binary Optimisation (QUBO) constraints, with a quantum-classical backend to address multi-drone zone assignment. By partitioning large problems into smaller segments and then recombining the solution, the system circumvents limitations imposed by current quantum hardware. Empirical results show a 96.3% success rate under real Quantum Approximate Optimisation Algorithm sampling, enabling natural-language-guided task allocation at previously intractable scales.

The research demonstrated a new framework for controlling multi-drone systems using natural language instructions. This approach combines a Large Language Model with quantum-inspired algorithms to translate human intent into task assignments, overcoming limitations of traditional methods. The system achieved a 96.3% success rate in allocating tasks under realistic quantum conditions, enabling control at larger scales than previously possible. Researchers acknowledge that future scaling may require advancements in quantum hardware, but this work represents a significant step towards more intuitive and flexible drone swarm management.