Tag
This paper formulates a multi-item capacitated lot-sizing problem with stochastic demand timing as a discrete-time MDP and proposes a genetic algorithm to solve it, demonstrating efficiency on benchmark instances.
This paper introduces a learning-based method to accelerate Mixed-Integer Linear Programming solvers by predicting early-to-final solution consistency, achieving substantial performance gains with Gurobi and zero-shot transfer to SCIP.
This article discusses the parallels between Soviet economic planning and modern data science practices, focusing on issues like resource allocation and simplifying assumptions based on historical books.
This paper proposes a training-free, uncertainty-aware inference framework for using large language models in operations research. The method uses short lookahead simulations and importance resampling to improve the coherence of mathematical formulations, outperforming standard baselines on OR benchmarks.
This paper introduces a decision-aware weak-to-strong (W2S) learning framework that uses limited labeled data to train a weak model, which then generates soft supervision on unlabeled data to train a strong model for improved contextual stochastic optimization. Theoretical bounds and empirical experiments show that abundant unlabeled data can reduce downstream decision risk when the correlation between weak and strong feature representations is small.
This paper presents SLAI T-Rex, a full-parameter post-training optimization framework for trillion-parameter MoE models on Ascend NPU SuperPOD, achieving 34.22% MFU and outperforming GPT-5.4-Mini on Operations Research tasks by 3.98 percentage points.
JOR-Bench is a collection of five Japanese-language benchmarks for evaluating large language models on operations research problem formulation, translated from existing English benchmarks. Evaluation shows overall language-neutral performance with minor cross-lingual differences.
A detailed benchmark comparing Claude Fable 5 and GPT-5.6 Sol on a tough NP-hard fiber-network design problem, finding Fable 5 significantly outperforms and that /goal mode is not a game-changer.
This paper introduces COOPA, a modular LLM agent architecture for operations research problems that combines iterative confidence-based modeling, element-level provenance, and multi-solver routing. Evaluated across eight LLM backbones and four baselines, COOPA achieves the best macro-average accuracy on six backbones and improves over the strongest baseline by up to 6.7 percentage points.
本文介绍ORAgentBench，一个用于评估LLM代理在端到端运筹学任务中表现的执行基准，包含107个经过人工审查的任务。实验表明，当前最佳代理仅通过35.51%的任务，揭示了在可靠决策制定方面的重大不足。
This paper introduces SkillChain-Gym, a benchmark specification for reskilling-aware production-inventory control that models worker skill dynamics, training actions, and disruption scenarios to evaluate policy regimes.
This paper introduces Bellman-Taylor Score Decoding, a method to handle state-dependent feasible action sets in Markov decision processes, addressing a key challenge in applying deep reinforcement learning to operations research problems.
Researchers from Beihang University and Baidu propose 'constraint injection,' a dual verification method for LLM-based optimization modeling that detects spurious or omitted constraints beyond objective equivalence. They develop VRPCoder, an 8B model for translating natural-language vehicle routing problems into Gurobi scripts, achieving 93% average Pass@1 and outperforming Claude Sonnet and prior OR-LLMs by large margins.
This paper introduces a totally unimodular linear programming reformulation for alignment-based conformance checking, which complements A* search by providing speedups for long traces with deviations. The approach achieves 38.6% average runtime savings with 96% selection accuracy.
The article argues that current benchmarks for Constraint Acquisition are inadequate and introduces MPMMine, a benchmark suite designed to assess algorithms that discover, validate, and enhance mathematical programming models using diverse domain knowledge artifacts.
OR-Space is a benchmark for evaluating large language model agents in industrial operations research workflows, focusing on multi-stage task lifecycles and persistent workspaces beyond simple text generation.
TriVAL introduces a tri-validation framework that performs explicit validation at three stages of automatic optimization modeling (semantic specification, mathematical formulation, code generation) to improve faithfulness, and also presents NL4COP, a new benchmark for combinatorial optimization problems.
This paper presents a hybrid approach combining dynamic programming and constraint programming to solve the Partial Shop Scheduling Problem, demonstrating the viability of integrating both paradigms despite not outperforming pure CP solvers.
This paper introduces Text2Opt-Bench, a scalable benchmark for text-to-optimization, and identifies that LLMs struggle with 'binding' (grounding problem data) rather than 'modeling' (choosing optimization structure). The authors propose BIND, a simple inference-time method that externalizes numeric data, significantly improving accuracy across models.
This paper presents a constraint programming approach to determine NHL playoff clinching scenarios with n-day lookahead, using tree search and preprocessing techniques.