Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2509.20105 · International Conference on Machine Learning and Applications

PEPS: Quantum-Inspired Reinforcement Learning for Coherent Reasoning Traces in LLMs

Venkat Margapuri, Garik Kazanjian, Naren Kumar Kosaraju

Published Dec 3, 2025• 0 citations
arXivHTMLPDF

Abstract

Large Language Models (LLMs) often struggle with maintaining coherent multi-step reasoning traces, particularly in tasks that require a structured logical flow. This work introduces a quantum-inspired approach to address the challenge by incorporating a fidelity-based reward derived from Projected Entangled Pair States (PEPS) into Proximal Policy Optimization. Unlike prior approaches that use direct supervision or contrastive objectives, the proposed method guides learning through structural consistency, offering a novel approach to enforce global coherence in generated reasoning traces. The proposed framework is evaluated using multiple coherence-determining metrics on diverse datasets such as GSM8K, StrategyQA, and EntailmentBank spanning arithmetic, intuitive, and entailment-based reasoning. Results show that the proposed quantum-inspired approach offers significant improvements over supervised, contrastive, and pretrained baseline approaches, highlighting the effectiveness of quantum-inspired fidelity as a foundation to improve reasoning trace coherence in LLMs.

DOI
10.1109/icmla66185.2025.00016

Related papers

Fine-Tuning Large Language Models for Quantum Reasoning

Katherine Ip, Casey R. Myers, Udaya Parampalli +2 more

Large language models (LLMs) exhibit abilities beyond natural language modelling and text generation. Recent advances in their reasoning capabilities have spurred interest in applying LLMs to complex scientific tasks requiring deep domain expertise and sophisticated reasoning. Quantum computing, as a highly specialised field with significant knowledge barriers and hardware constraints, could greatly benefit from such advancements. However, a key open question that first must be answered is: How can we develop fine-tuning pipelines that instil genuine quantum reasoning in LLMs, rather than task-specific pattern matching? We study this question through quantum circuit simulation as a training objective, where the model must predict the measurement probability distribution resulting from a sequence of quantum gate operations. We propose and compare two fine-tuning pipelines: (1) Supervised Fine-Tuning (SFT) on explicit gate-by-gate state-vector simulation traces, and (2) a two-stage SFT+Group Relative Policy Optimisation (GRPO) approach that sequentially applies SFT followed by GRPO with verifiable rewards. Our findings show that SFT achieves near-perfect in-distribution and gate-count extrapolation accuracy, significantly outperforming both the base model and the GPT-OSS-120B baseline. SFT+GRPO trades some in-distribution precision for better generalisation to larger qubit systems that SFT alone cannot handle. Both pipelines significantly outperform the baselines, demonstrating that targeted fine-tuning on explicit reasoning traces is an effective strategy for advancing quantum reasoning in LLMs.

quant-pharXiv:2606.21974•Jun 23, 2026•0 citations

Quantum-Inspired Algorithms for Reinforcement Learning Policy Optimization

Jincheng Zhang

This paper investigates the potential of quantum-inspired algorithms to improve reinforcement learning (RL) policy optimization. Traditional RL methods often struggle with slow convergence and exploration challenges, particularly in complex environments. We propose a novel approach that draws inspiration from quantum mechanics, specifically superposition and entanglement, to address these limitations. Our algorithms aim to exploit superposition to simultaneously evaluate multiple action possibilities and utilize entanglement to model correlations between states, thereby accelerating the learning process. We present a framework for incorporating these concepts into RL, focusing on algorithmic design rather than a direct quantum simulation. The theoretical analysis suggests that our approach could significantly enhance exploration efficiency and, consequently, policy convergence rates. This work represents a direct application of quantum-inspired ideas to RL policy optimization, offering a potentially transformative pathway for developing more efficient and robust RL agents.

•Sep 7, 2026•0 citations

QuantumQA: Enhancing Scientific Reasoning via Physics-Consistent Dataset and Verification-Aware Reinforcement Learning

Songxin Qu, Tai-Ping Sun, Yun-Jie Wang +8 more

Large language models (LLMs) show strong capabilities in general reasoning but typically lack reliability in scientific domains like quantum mechanics, which demand strict adherence to physical constraints. This limitation arises from the scarcity of verifiable training resources and the inadequacy of coarse feedback signals in standard alignment paradigms. To address the data challenge, we introduce QuantumQA, a large-scale dataset constructed via a task-adaptive strategy and a hybrid verification protocol that combines deterministic solvers with semantic auditing to guarantee scientific rigor. Building on this foundation, we propose the verification-aware reward model (VRM) tailored for Reinforcement Learning with Verifiable Rewards (RLVR), which employs an adaptive reward fusion (ARF) mechanism to dynamically integrate deterministic signals from a scientific execution suite (SES) with multidimensional semantic evaluations for precise supervision. Experimental results demonstrate that our method consistently outperforms baselines and general-purpose preference models. Notably, our optimized 8B model achieves performance competitive with proprietary models, validating that incorporating verifiable, rule-based feedback into the reinforcement learning loop offers a parameter-efficient alternative to pure scaling.

cs.AICross-listedarXiv:2604.18176•Apr 21, 2026

Quantum-Inspired Algorithms for Reinforcement Learning

Jincheng Zhang

This paper investigates the application of quantum-inspired algorithms to the field of reinforcement learning (RL). Traditional RL methods often struggle with exploration challenges and can be computationally expensive, particularly in complex environments. We propose a novel RL algorithm that leverages concepts from quantum mechanics, specifically superposition and entanglement, to address these limitations. The core idea is to represent states, actions, and rewards using quantum-inspired representations, allowing the agent to explore the environment more efficiently and potentially converge to optimal policies faster. This work presents a detailed formulation of the algorithm, including the mathematical framework and key considerations for implementation. The potential benefits of this approach, including improved exploration and reduced computational complexity, are discussed. The algorithm is designed to iteratively refine its policy based on simulated quantum measurements, aiming to minimize the expected return. The exploration strategy is guided by the inherent probabilistic nature of quantum states, leading to a more diverse and potentially faster learning process. Further research is warranted to fully explore the capabilities and limitations of this approach across various RL benchmark problems.

•Sep 1, 2026•0 citations

Knowledge Graph Reasoning with Quantum-Inspired Reinforcement Learning

Junliang Chen, Xianchao Zhang, Fengsong Sun +1 more

Knowledge reasoning is a critical task in information fusion systems, and its core step is reasoning missing information from existing facts to improve the knowledge graphs. Embedding-based reasoning methods and path-based reasoning methods are two mainstream knowledge reasoning methods. Embedding-based reasoning methods enable fast and direct reasoning but are limited to simple relationships between entities and exhibit poor performance in reasoning complex logical relationships. Path-based reasoning methods perform better in complex reasoning tasks, but suffer from high computational complexity, a large number of model parameters, and low reasoning efficiency. To address the aforementioned issues, this paper introduces a knowledge reasoning model called Quantum-Inspired Reinforcement Learning (QIRL). QIRL leverages quantum reinforcement learning to train a strategy network via a quantum circuit, aiming to generate and optimize reasoning paths. Quantum circuit achieves complex nonlinear operations through limited quantum reasoning paths gate operations, reducing computational complexity. In addition, this article utilizes the quantum entanglement property to encode high-dimensional data, reducing the number of model training parameters. This article evaluates the QIRL method on entity prediction task and proves that the QIRL method can effectively reduce the number of model training parameters.

Chinese journal of information fusion.•May 25, 2025•1 citation

Quantum-Inspired Reinforcement Learning for Quantum Control

Haixu Yu, Xudong Zhao, Chunlin Chen

Reinforcement learning (RL) is considered a powerful technology with the potential to revolutionize quantum control. However, the application effectiveness of traditional RL is often limited by some insurmountable experimental conditions. Thus, developing new RL algorithms that can efficiently manipulate the quantum system dynamics is a crucial task. Prior research has shown that incorporating quantum mechanical properties into RL can improve learning performance. In this article, we consider the quantum control problem where only the target state can be accurately identified and introduce a quantum-inspired RL (QiRL) method. In particular, we propose a quantum-inspired exploration strategy to replace a commonly used$\epsilon $-greedy strategy, as well as a quantum-inspired reward scheme to incentivize the learning agent. Numerical results on three quantum system control problems, i.e., one-qubit closed quantum system, two-level open quantum system, and many-qubit closed quantum system, verify the effectiveness of QiRL. Comparison results show that the proposed QiRL outperforms existing RL algorithms (deep Q-network and proximal policy optimization) in terms of stability and efficiency for solving quantum control problems.

IEEE Transactions on Control Systems Technology•Aug 14, 2024•15 citations
Large Language Models (LLMs) often struggle with maintaining coherent multi-step reasoning traces, particularly in tasks that require a structured logical flow. This work introduces a quantum-inspired approach to address the challenge by incorporating a fidelity-based reward derived from Projected Entangled Pair States (PEPS) into Proximal Policy Optimization. Unlike prior approaches that use direct supervision or contrastive objectives, the proposed method guides learning through structural consistency, offering a novel approach to enforce global coherence in generated reasoning traces. The proposed framework is evaluated using multiple coherence-determining metrics on diverse datasets such as GSM8K, StrategyQA, and EntailmentBank spanning arithmetic, intuitive, and entailment-based reasoning. Results show that the proposed quantum-inspired approach offers significant improvements over supervised, contrastive, and pretrained baseline approaches, highlighting the effectiveness of quantum-inspired fidelity as a foundation to improve reasoning trace coherence in LLMs.
Large language models (LLMs) exhibit abilities beyond natural language modelling and text generation. Recent advances in their reasoning capabilities have spurred interest in applying LLMs to complex scientific tasks requiring deep domain expertise and sophisticated reasoning. Quantum computing, as a highly specialised field with significant knowledge barriers and hardware constraints, could greatly benefit from such advancements. However, a key open question that first must be answered is: How can we develop fine-tuning pipelines that instil genuine quantum reasoning in LLMs, rather than task-specific pattern matching? We study this question through quantum circuit simulation as a training objective, where the model must predict the measurement probability distribution resulting from a sequence of quantum gate operations. We propose and compare two fine-tuning pipelines: (1) Supervised Fine-Tuning (SFT) on explicit gate-by-gate state-vector simulation traces, and (2) a two-stage SFT+Group Relative Policy Optimisation (GRPO) approach that sequentially applies SFT followed by GRPO with verifiable rewards. Our findings show that SFT achieves near-perfect in-distribution and gate-count extrapolation accuracy, significantly outperforming both the base model and the GPT-OSS-120B baseline. SFT+GRPO trades some in-distribution precision for better generalisation to larger qubit systems that SFT alone cannot handle. Both pipelines significantly outperform the baselines, demonstrating that targeted fine-tuning on explicit reasoning traces is an effective strategy for advancing quantum reasoning in LLMs.
This paper investigates the potential of quantum-inspired algorithms to improve reinforcement learning (RL) policy optimization. Traditional RL methods often struggle with slow convergence and exploration challenges, particularly in complex environments. We propose a novel approach that draws inspiration from quantum mechanics, specifically superposition and entanglement, to address these limitations. Our algorithms aim to exploit superposition to simultaneously evaluate multiple action possibilities and utilize entanglement to model correlations between states, thereby accelerating the learning process. We present a framework for incorporating these concepts into RL, focusing on algorithmic design rather than a direct quantum simulation. The theoretical analysis suggests that our approach could significantly enhance exploration efficiency and, consequently, policy convergence rates. This work represents a direct application of quantum-inspired ideas to RL policy optimization, offering a potentially transformative pathway for developing more efficient and robust RL agents.
Large language models (LLMs) show strong capabilities in general reasoning but typically lack reliability in scientific domains like quantum mechanics, which demand strict adherence to physical constraints. This limitation arises from the scarcity of verifiable training resources and the inadequacy of coarse feedback signals in standard alignment paradigms. To address the data challenge, we introduce QuantumQA, a large-scale dataset constructed via a task-adaptive strategy and a hybrid verification protocol that combines deterministic solvers with semantic auditing to guarantee scientific rigor. Building on this foundation, we propose the verification-aware reward model (VRM) tailored for Reinforcement Learning with Verifiable Rewards (RLVR), which employs an adaptive reward fusion (ARF) mechanism to dynamically integrate deterministic signals from a scientific execution suite (SES) with multidimensional semantic evaluations for precise supervision. Experimental results demonstrate that our method consistently outperforms baselines and general-purpose preference models. Notably, our optimized 8B model achieves performance competitive with proprietary models, validating that incorporating verifiable, rule-based feedback into the reinforcement learning loop offers a parameter-efficient alternative to pure scaling.
This paper investigates the application of quantum-inspired algorithms to the field of reinforcement learning (RL). Traditional RL methods often struggle with exploration challenges and can be computationally expensive, particularly in complex environments. We propose a novel RL algorithm that leverages concepts from quantum mechanics, specifically superposition and entanglement, to address these limitations. The core idea is to represent states, actions, and rewards using quantum-inspired representations, allowing the agent to explore the environment more efficiently and potentially converge to optimal policies faster. This work presents a detailed formulation of the algorithm, including the mathematical framework and key considerations for implementation. The potential benefits of this approach, including improved exploration and reduced computational complexity, are discussed. The algorithm is designed to iteratively refine its policy based on simulated quantum measurements, aiming to minimize the expected return. The exploration strategy is guided by the inherent probabilistic nature of quantum states, leading to a more diverse and potentially faster learning process. Further research is warranted to fully explore the capabilities and limitations of this approach across various RL benchmark problems.
Knowledge reasoning is a critical task in information fusion systems, and its core step is reasoning missing information from existing facts to improve the knowledge graphs. Embedding-based reasoning methods and path-based reasoning methods are two mainstream knowledge reasoning methods. Embedding-based reasoning methods enable fast and direct reasoning but are limited to simple relationships between entities and exhibit poor performance in reasoning complex logical relationships. Path-based reasoning methods perform better in complex reasoning tasks, but suffer from high computational complexity, a large number of model parameters, and low reasoning efficiency. To address the aforementioned issues, this paper introduces a knowledge reasoning model called Quantum-Inspired Reinforcement Learning (QIRL). QIRL leverages quantum reinforcement learning to train a strategy network via a quantum circuit, aiming to generate and optimize reasoning paths. Quantum circuit achieves complex nonlinear operations through limited quantum reasoning paths gate operations, reducing computational complexity. In addition, this article utilizes the quantum entanglement property to encode high-dimensional data, reducing the number of model training parameters. This article evaluates the QIRL method on entity prediction task and proves that the QIRL method can effectively reduce the number of model training parameters.
Reinforcement learning (RL) is considered a powerful technology with the potential to revolutionize quantum control. However, the application effectiveness of traditional RL is often limited by some insurmountable experimental conditions. Thus, developing new RL algorithms that can efficiently manipulate the quantum system dynamics is a crucial task. Prior research has shown that incorporating quantum mechanical properties into RL can improve learning performance. In this article, we consider the quantum control problem where only the target state can be accurately identified and introduce a quantum-inspired RL (QiRL) method. In particular, we propose a quantum-inspired exploration strategy to replace a commonly used$\epsilon $-greedy strategy, as well as a quantum-inspired reward scheme to incentivize the learning agent. Numerical results on three quantum system control problems, i.e., one-qubit closed quantum system, two-level open quantum system, and many-qubit closed quantum system, verify the effectiveness of QiRL. Comparison results show that the proposed QiRL outperforms existing RL algorithms (deep Q-network and proximal policy optimization) in terms of stability and efficiency for solving quantum control problems.