Research map / Large language models / Agents, code and math

Automated Reformulation of Robust Optimization via Memory-Augmented Large Language Models

NeurIPS 2026 · Large language models › Agents, code and math · approach: LLMs for optimization and heuristics · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces AutoRO-Bench for LLM-based robust optimization reformulation and AutoREM, a tuning-free memory-augmented framework that builds a verified textual experience memory offline via dual-check commit and validation-based acceptance.

Problem

LLMs struggle to automatically reformulate uncertain (robust) LP/MILP models into tractable deterministic robust counterparts, and no benchmark exists for evaluating this; existing memory methods use additive-only updates without quality verification, which harms exact mathematical reasoning.

Approach

A tuning-free offline memory construction procedure with unit-level experiences, structured add/update/delete operators, a step-level dual-check commit on locked validation batches, and epoch-level validation-based rollback, plus a generator producing solver-verified RO reformulation instances.

Key results

• AutoREM reaches 97.4% accuracy on the in-distribution Random RO reformulation set, versus 87.5% for the base LLM and 87.0% for ACE.
• On the out-of-distribution Hard set AutoREM scores 94.8% against 83.3% for the best baseline, while using 23% fewer output tokens than the base LLM.
• On the end-to-end Application dataset AutoREM achieves 81.3% versus 75.0% for AlphaOPT; memory transfers to GPT-5.4 (+6.3%) and Qwen3.6-Plus (+7.3%) without re-adaptation.

Surprising findings

• ReasoningBank and ACE fail to improve over the base LLM on the Random reformulation set (85.4% and 87.0% vs 87.5%).
• Auto-generated memory from AutoREM outperforms the expert-written RO prompt.

Limitations

• Offline adaptation relies on solver-verified objective values as supervision, which may not be available for all problem classes.
• Limited to LP-based robust counterparts; not extended to more complex RO.
• Large-scale variable-count split shows all methods degrade, indicating long-context bottleneck.

Experimental setup

Datasets: AutoRO-Bench Random (generated) IndustryOR OptiBench

Benchmarks: AutoRO-Bench Random AutoRO-Bench Hard AutoRO-Bench Large AutoRO-Bench Application

Metrics: accuracy output tokens

Baselines: Base LLM Max Thinking Expert Prompt ReasoningBank ACE AlphaOPT LEAN-LLM-OPT OptiTree

Builds on: ACE ReasoningBank chain-of-thought

Closest papers by meaning

• Autoformulation of Mathematical Optimization Models Using LLMs (ICML 2025 · similarity 0.88)
• Large Language Models to Enhance Bayesian Optimization (ICLR 2024 · similarity 0.88)
• OptMATH: A Scalable Bidirectional Data Synthesis Framework for Optimization Modeling (ICML 2025 · similarity 0.88)
• LLMOPT: Learning to Define and Solve General Optimization Problems from Scratch (ICLR 2025 · similarity 0.87)
• Unleashing the Potential of Large Language Models as Prompt Optimizers: Analogical Analysis with Gradient-based Model Optimizers (AAAI 2025 · similarity 0.87)
• MGUP: A Momentum-Gradient Alignment Update Policy for Stochastic Optimization (NeurIPS 2025 · similarity 0.87)