NL4Opt — Ramamonjison et al., 2023.
NER + LP-generation pipeline that turns NL word problems into solvable LPs; stage decomposition exposes how span errors compound into LP errors.
LM4OPT — Ahmed & Choudhury, 2024.
Large LMs match the NL4Opt pipeline on F1 but lose the diagnosability the pipeline's stage decomposition provided.
FLEx — Avsian et al., 2026.
Mine LLM failure modes from training data, distill into prefix hints injected at inference; aligns representational shift to failure shift.
Model → Human
OptiGuide — Li et al., 2023.
LLM-mediated what-if queries on a deployed MILP via code-gen + re-solve, kept inside an enterprise enclave; interrogation-only.
OptiChat — Chen et al., 2025.
Five-query taxonomy (diagnose, retrieve, sensitivity, what-if, why-not); only why-not requires LLM-generated counterfactual code.
Democratizing Optimization with Generative AI — Simchi-Levi et al., 2025.
Research-agenda paper for non-experts to query and steer optimization models; positions the bidirectional channel without fully specifying it.
Multi-Human / Social Choice
Habermas Machine — Tessler et al., 2025.
AI-mediated group-preference statements rated above unaided mediator output; large-scale LLM test in deliberative settings.
SFUSD School Choice — Allman et al., 2022.
Mechanism design for diversity in San Francisco school assignment; concrete deployment of social-choice theory in public policy.
Conversation Agents for OR — Drossman et al., 2026.
LLM agents that converse with users to elicit objectives for downstream optimization; front-end to OR models.