LLMs and optimisation, sorted by what defines them
A landscape of where large language models meet mathematical optimisation: 22 players sorted into four buckets along the dimensions that actually separate them. It covers the lineage, the people, the bets and the evidence for each bucket, with China as its own scene and a timeline of the goals behind all four.
Main questions answered
- What are the four ways an LLM can sit next to a solver, and where does correctness come from in each?
- Which players keep exact optimisation, and which have real-world evidence?
- How did each generation of tools try to meet the same three wishes?
- Who is working in each bucket, what are they betting on, and why does China have its own solver scene?
- Where is the white space that nobody fills?
How to read this page
The contents list on the left lists this report's panes and sections.
Subsections open collapsed and show their first block; "more" and "less" toggle one, "Expand all" opens every subsection.
Select text and press h to highlight it. Highlights are kept in this browser only.
1. Four ways AI can advance OR adoption
| Bucket | Where the LLM sits | Where correctness comes from | One-line difference |
|---|---|---|---|
| A Front end for standard OR | Between a person and a solver. Writes or explains a formal model. | The solver solves whatever model was written. Whether it is the right model is checked by tests, an existing model, or nobody. | Same old models, easier to write and read. |
| B Algorithm discovery | Inside a search loop, writing code. Never talks to a user. | An evaluator that scores each program on test instances. | New algorithms, found by mutation plus a scorer. |
| C Better solvers | Mostly nowhere. Learning and GPUs work inside or instead of the solver. | Optimality gaps and certificates, where the method keeps them. | Same models, solved faster or better. |
| D Richer models of reality | Between the messy world (emails, planners, news) and the model. | Mostly a human accepting or rejecting. Nobody has an automatic check that the model got more faithful. | Different models, closer to what actually decides whether a plan gets used. |
flowchart LR R[Messy reality
emails, contracts,
planner knowledge] -->|D. richer models| S[Problem statement] S -->|A. LLM formulation| M[Formal model
LP, MIP, VRP] M -->|C. better solving
GPU, learned parts| P[Plan] B[B. algorithm discovery
LLM plus evaluator] -.->|new heuristics| M P -->|A. explain, what-if| U[Decision maker] U -.->|D. feedback| S
2. The dimensions table
The table shows where each of the 22 players puts the LLM and what that does to correctness, evidence and persistence. Sorted by where the LLM sits, the split is plain: in A and B and in most Chinese work the LLM writes the core artefact, a model or an algorithm; in C it is absent, and in the D industry products it sits at the edge, recommending. Exactness kept, read next to real-world evidence, shows that almost every method which reached production with numbers keeps a classical engine, with AlphaEvolve at FM Logistic as the one exception that ships a heuristic with no gap, a named external customer and the word production. Direction and persistence together are almost empty: hardly anyone carries reality into the model and keeps the elicited rule versioned, and the two products that persist something do not test the edit against history. Base model shows the open-weight bases clustered at OptiMind and the ORLM line, while everything that leads the discovery results runs on Gemini or OpenAI models.
Rows are the dimensions that actually split the players. Columns are 22 representative players or approaches, grouped by bucket. Each row has its own meaning for green, amber and red, written in its definition (hover the row name, or read the list under the table). The colours are not "good" and "bad". They say how much of that property a player has.
Click a dimension's name to sort the columns by it (green first, then amber, red, plain, n/a, ?). Click again to go back to bucket order. Hover anything marked ✦ for a note. Every note is also written out below the table.
Chinese columns carry two chips: China, and the bucket they belong to. The A to D filters include them.
3. The four buckets, and China
3A. LLMs as the front end for standard OR
Lineage
This is the oldest dream in OR wearing new clothes. For fifty years the bottleneck was not the solver but the translation from a business problem to a formal model. GAMS, AMPL and AIMMS all made the same pitch as today: let someone who knows the problem write the model without writing solver code. They worked for trained OR people and failed for everyone else, because algebraic notation still needs the modelling skill.
The LLM version went through three phases. In 2022 and 2023 the question was whether an LLM can write a model at all, on textbook word problems (NL4Opt). From 2024 to early 2026 it split into prompting agents (OptiMUS, Chain-of-Experts) and fine-tuned specialist models (ORLM, LLMOPT, OptiMind). From mid-2026 the vendors stopped trying to own the chat window: Gurobi, Nextmv, AMPL and NVIDIA all expose their tools through MCP servers or agent skills, so Claude, Copilot or any coding agent calls the solver. That concedes the front end to general agents and retreats to what a vendor can uniquely check.
Groups and people
| Group | Main people and background | What the background predicts |
|---|---|---|
| Microsoft Research, ML and Optimization (Redmond) | Ishai Menache, Partner Research Manager, PhD Technion, MIT postdoc. Stable core: Mellou, Molinaro, Kulkarni, Sirui Li, Barbalho, Beibin Li, with Jeevan Pathuri from cloud supply chain operations. Business partner: David Simchi-Levi (MIT). | Their customer is Microsoft's own cloud hardware supply chain, which already has OR models and planners. So the first product was explanation and what-if over an existing model, not formulation from scratch. |
| Stanford, Madeleine Udell's group | Udell, Associate Professor of MS&E (tenured July 2026), from Cornell in 2022. OptiMUS led by PhD student Ali AhmadiTeshnizi. | An optimisation and numerics group, not an NLP group. Emphasis on solver-checked correctness, modularity for long problems, debugging loops, on whichever frontier model is best. |
| Zhejiang University with Huawei | Ziyang Xiao, Dongxiang Zhang, Gang Chen. | Agent architecture (Chain-of-Experts) and benchmark hygiene. Their 2025 survey found the benchmarks full of errors. |
| Gurobi | CEO Duke Perrucci; R&D VP Tobias Achterberg on AI Modeling; CTO since October 2025 Oliver Bastert, from FICO (VP Product Management across analytics, ML, generative AI and optimisation), earlier MIP algorithms at Dash Optimization. | A solver and product team. The LLM goes where errors can be caught (spec, unit tests, infeasibility diagnosis) and the solver stays the source of truth. Bastert: optimisation "should benefit from that same transformation" without "compromising the rigor". |
| NVIDIA cuOpt | Agent work published by cuOpt engineers and product marketing (Linsenmaier, Chung in 2024; Geva, Narasimha, Arunagiri in 2026). No single research lead named. | Reference architectures that show off GPUs and NIM models, not a product NVIDIA sells to planners. |
| Nextmv | Carolyn Mooney (CEO; simulation at Lockheed and Zoomer, then Grubhub) and Ryan O'Neil (CTO; decision engineering at Zoomer and Grubhub). | People who ran routing in production at a delivery company. They care about tests, rollouts and logs around a model, not about writing it from words. |
| Strangeworks / Quantagonia | Quantagonia founded 2021 by Dirk Zechiel, Philipp Hannemann, Sabina Jeschke and Sebastian Pokutta; Zechiel co-founded Gurobi GmbH and is now Strangeworks CTO. Strangeworks CEO is whurley. | Solver-orchestration heritage (run several solvers, pick the best) plus quantum positioning, sold through expert-led engagements rather than self-serve. |
Cardinal Operations (ORLM) belongs here too and is covered under China. AMPL (board still includes Fourer, Gay and Kernighan) and AIMMS are shipping assistants without named people behind the AI work.
The bets
| Player | Main bet, in one or two sentences | Where they say they are going |
|---|---|---|
| Gurobi | The LLM is a funnel and a safety net for the solver. Co-write a plain-language spec, generate formulation, gurobipy and unit tests, diagnose infeasibility, and let any coding agent reach this through MCP. | Lifecycle coverage: build, understand, troubleshoot. |
| Microsoft Research | A small, locally runnable specialist model on clean data beats prompting a frontier model, and the first business value is what-if over models you already have. | OptiMind next: reinforcement learning, self-generated expert hints, Excel input, product teams. |
| NVIDIA | Any agent framework, any open model, our GPU solver underneath. Skills framed as replacing "weeks of specialized OR work". | cuOpt as the tool agents call. |
| Nextmv | The LLM should operate decision apps, not invent them. "Deterministic control" over runs, tests and rollouts. | The operations layer, explicitly not chatbots for modelling. |
| AMPL | Ground general LLMs in the official language so they stop hallucinating syntax: skills and MCP for ChatGPT, Claude, Gemini. | No dates on the page. |
| AIMMS | The LLM is the interface for business users. "SENSAI doesn't do the math. It works with the math." | Agentic app and model building inside AIMMS. |
| Strangeworks Aura | Optimisation as an always-on managed system, from finding the opportunity to monitoring after go-live, across classical, quantum-inspired, HPC and quantum compute. The release does not mention LLMs. | Expert-led engagements. |
| Fine-tuning line (ORLM, LLMOPT, OptiMind) | The translator should be a fine-tuned open model trained on synthetic, solver-verified data. | RL with the solver as reward (SIRL, see China). |
The real axis is what each player protects. Gurobi, AMPL, AIMMS and COPT protect a solver or a language, so their LLM must emit their syntax and they want to be the most reliable tool a general agent can call. NVIDIA protects hardware and is happy to commoditise the front end. Microsoft protects Azure and its own supply chain and publishes open specialist models. Nextmv protects the deployment layer. Strangeworks protects an engagement model and a compute broker. Nobody in this bucket bets that the LLM replaces the solver.
Evidence, strongest first
| # | Who | What is documented | Why not higher |
|---|---|---|---|
| 1 | Microsoft (internal) | OptiGuide "has already been used in production" for cloud supply-chain planners' what-if questions; HBR and MIT describe LLM-based server deployment planning across 300+ data centres, a week of analysis down to minutes. | No external customer, no published cost or accuracy figure. "Week to minutes" is a description, not a measurement. |
| 2 | AIMMS SENSAI | Inside a product claiming 250+ enterprise customers; one unnamed testimonial (a US beauty retailer) using SENSAI Pro for scenarios. | No named LLM user. Logos on the page are not tied to SENSAI. |
| 3 | NVIDIA (internal) | Own operations team uses scenario analysis for nearly 3,000 products across a dozen factories (2024). | Unclear how much is the LLM agent versus cuOpt. |
| 4 | Gurobi | Shipping, documented, MCP server real. | No named customer, no usage numbers. Its own survey: 6% of users combine optimisation with LLMs. |
| 5 | Nextmv, AMPL, IBM | Shipped features. | No named users of the AI parts. IBM's Modeling Assistant documentation never mentions an LLM. |
| 6 | Strangeworks Aura, Cosmo Tech | Announcements. | No named customers or measured outcomes. |
| 7 | OptiMUS, Chain-of-Experts, ORLM, LLMOPT, OptiMind | Benchmarks. | OptiMind released as experimental. |
The best documented case is a large company using LLMs over its own existing models with its own OR team in the loop. No source shows a non-expert business user getting a production model from natural language at a paying customer with a measured outcome. Verification, not generation, is the bottleneck: a wrong formulation runs fine and returns a confident optimum to the wrong problem.
3B. Algorithm discovery
Lineage
This is OR's own idea. The hyper-heuristics community wanted machines to "search over heuristics or heuristic components rather than directly over solutions" long before LLMs, usually through genetic programming. Its weakness was that a human had to define the building blocks for every new problem. The FunSearch paper says exactly that, and positions itself inside genetic programming and hyper-heuristics.
The hinge was OpenAI's 2022 paper showing that a code model makes program mutation far more effective. DeepMind's discovery team, who had shown with AlphaTensor that "provably correct algorithms" were a lab-scale goal, turned it into FunSearch (Nature, December 2023): search for programs, not solutions, use a fast cheap code model at around a million samples, and let an exact evaluator absorb the hallucinations. On online bin packing it got 0.03% excess bins over the lower bound against 3.79% for best fit. The evolutionary-computation community answered within months with cheaper open versions (EoH, ReEvo, which renamed the field "Language Hyper-Heuristics", LLaMEA). AlphaEvolve (May 2025) went from one function to whole files, and became a Google Cloud product in July 2026. Contests became the public scoreboard.
Groups and people
| Group | Main people and background | What the background predicts |
|---|---|---|
| Google DeepMind, AI for Science | Pushmeet Kohli, VP Research, heads AI for Science and Strategic Initiatives, also titled Chief Scientist of Google Cloud; about a decade at Microsoft Research before. Matej Balog leads the algorithm discovery team (AlphaTensor, FunSearch). Alexander Novikov, AlphaEvolve first author, co-wrote the T3F tensor-train library. | Program synthesis, tensor maths and RL search, run by Cloud's chief scientist. So generality over any one OR domain, prestige maths results as proof, and commercialisation through Google Cloud. That is what happened. |
| CityU Hong Kong (EoH, LLM4AD) with Huawei | Qingfu Zhang, Chair Professor at CityU, earlier Essex, known for MOEA/D. Fei Liu first author. Huawei Noah's Ark co-authors Xialiang Tong and Mingxuan Yuan. | Evolutionary-computation people treat the LLM as one more variation operator: efficiency, breadth of benchmarks, an open platform. |
| AI4CO network (PKU, KAIST, SMU) | Haoran Ye, Federico Berto, Chuanbo Hua, Jinkyoo Park, Zhiguang Cao, Guojie Song. | The neural combinatorial optimisation community, pivoted to LLM heuristics. Plug LLM-written parts into existing templates, measure on routing benchmarks. Close to OR practice, far from customers. |
| Bielefeld and routing specialists (VRPAgent) | Hottung, Tierney and others with KAIST. Tierney also co-wrote the ORTEC EURO meets NeurIPS paper. | Keep the LLM in a small slot of a strong metaheuristic so feasibility is never at risk. |
| Leiden (LLaMEA) | Niki van Stein (Natural Computing Group, AutoML and Bayesian optimisation), Thomas Bäck. | Continuous black-box optimisation with rigorous benchmarking, closer to engineering design than logistics. |
| Sakana AI | Robert Tjarko Lange (TU Berlin PhD on evolutionary meta-learning, interned on DeepMind's Discovery team, maintains evosax), Yuki Imajuku, Edoardo Cetin; ALE-Bench with AtCoder. | An evolution-strategies researcher plus a contest partnership gives sample efficiency, open source, and contests as proof. Sakana's business is finance and Japanese government, not OR. |
| OpenAI | After the 2022 paper, no published discovery system. Visible only through AtCoder. | Heuristic design as an emergent skill of a general coding agent. |
The bets
DeepMind / Google Cloud. One general engine: "bring a seed program and a deterministic evaluator", sold as a cloud service, the evaluator running on the customer's side. The scarce input is frontier models plus search infrastructure, and customers will write evaluators. Sakana. Do more with fewer calls: ShinkaEvolve's circle packing in about 150 samples, the AHC058 win at about 4,773 calls and $1,300. OpenAI. No OR scaffolding; a general reasoning agent won the 2026 AtCoder Heuristic final. CityU and AI4CO. Cheap, open, reproducible methods and a platform. VRPAgent and PyVRP+. Constrain the LLM to operators inside a state-of-the-art routing metaheuristic and measure against the best hand-written code (VRPAgent, about 0.3% better than SISRs at 1,000 to 2,000 customers, one CPU core).
The real axis is how much of the program the LLM may touch. AlphaEvolve rewrites whole files, VRPAgent writes two operators inside a fixed search, FunSearch and EoH evolve one priority function. The smaller the slot, the cheaper, safer and more OR-compatible. The larger, the more surprising the result and the more evaluator engineering it needs. A second axis is who pays for the search: Google sells compute, Sakana and the academics sell efficiency.
Evidence, strongest first
| # | Who | What is documented | Caveat |
|---|---|---|---|
| 1 | Google internal | Borg scheduling heuristic in production for over a year, recovering on average 0.7% of worldwide compute; 23% kernel speedup giving 1% less Gemini training time; a Verilog change in an upcoming TPU. | Self-reported by the operator. |
| 2 | FM Logistic, one Polish warehouse | Seed program was their own grouping and routing logic; evaluator on 60 tours (over an hour of workforce data) with five penalty terms; 10.4% routing-efficiency gain on an "already highly optimized baseline", "now running in production". FM Logistic's own article confirms day-to-day use at one site. | One site, small evaluation set, the 15,000 km figure reads as a projection, co-written with the vendor. Still the only classical-OR deployment of an LLM method with a named customer. |
| 3 | Klarna, Coolblue (AlphaEvolve GA page) | Training pipeline throughput doubled; forecast WMAPE improved over 5%. | Real, but ML and forecasting, not optimisation. |
| 4 | Kinaxis, Infineon, BASF, ORNL | Early tests and experiments. | Pilots; Kinaxis numbers are on benchmark datasets. |
| 5 | Sakana, OpenAI | Sakana 1st of 804 at AHC058 (hosted by ALGO ARTIS, a planning-optimisation company); OpenAI 1st at the 2026 World Tour Finals Heuristic. | Puzzles with a published scorer. |
| 6 | EoH, ReEvo, LLaMEA, VRPAgent, PyVRP+ | Benchmarks. | No deployment found. |
The evaluator is the product. In business OR the evaluator is a simulator of the operation with every hard and soft constraint priced as a penalty. FM Logistic's is exactly that. Writing it is the same work as building a good model, which is bucket D's problem in disguise. Gains are big against hand-tuned rules (batching, slotting, dispatching rules), small against strong metaheuristics, and it fits where three things exist: a stable instance distribution, a fast simulator, and a seed algorithm. That describes the inside of planning products more than end customers.
On "DeepMind and ORTEC": no ORTEC work on LLM-generated heuristics and no ORTEC link to DeepMind or AlphaEvolve was found. ORTEC's documented ML-for-OR work is the EURO meets NeurIPS 2022 routing competition it initiated with real data (first author Wouter Kool). Google's OR-Tools lead Laurent Perron sits in Google Research, not DeepMind.
3C. Better solvers: ML and GPUs
Lineage
Two stories that started together and split. The first, from 2015, was "learn the solver": networks that output tours (Pointer Networks, the attention model, POMO), or that replace one hand-written rule inside branch and bound (Khalil, Gasse, DeepMind's Neural Diving). Thousands of papers. Then the reality checks. Amazon's 2021 Last Mile challenge was designed for ML and won by classical combinatorial optimisation. The ORTEC competition in 2022 said in its abstract that deep learning "fails to outperform OR baselines" on static routing, and every static winner was a hybrid genetic search. But on the dynamic variant every top approach used ML as a key part, and the overall winner (Kleopatra) let ML predict prizes while a classical engine did the routing. Learning wins where it predicts the uncertain part. The engine still does the combinatorics.
The second story is the one with production evidence, and it has no LLM in it. Google's OR team (Applegate, Lu, Lubin and others) built PDLP, a first-order LP method whose core operation is a matrix-vector product, in production in Google traffic engineering since May 2023. Haihao Lu's cuPDLP.jl matched Gurobi on standard LP sets on a GPU in late 2023. Gurobi's support page records that in 2023 its team concluded GPUs were "not a good fit". Within two years COPT, HiGHS, NVIDIA cuOpt, Gurobi and Xpress had all absorbed it. The counter-current already exists: Rothberg's "Backing PDHG into a Corner" on PDHG's lower accuracy and costly crossover, and Gurobi's own count that only 72 of 2,633 internal models ran faster with PDHG.
Groups and people
| Group | Main people and background | What the background predicts |
|---|---|---|
| Montréal school, OR insiders | Andrea Lodi: Bologna, CPLEX R&D consultant 2006 to 2021, Canada Excellence Research Chair at Polytechnique Montréal, now at Cornell Tech. Co-wrote the manifesto, Gasse's branching paper, the CPLEX classifier and the critical VRP guidelines. Elias Khalil (Toronto) bridges both camps. | Someone who lived inside a commercial solver learns components that keep exactness, and demands fair benchmarks. That is his record. The one learned component that shipped in a commercial solver (CPLEX 12.10) came from him. |
| ML-first routing (Amsterdam, KAIST, InstaDeep) | Wouter Kool: OR MSc, ORTEC OR engineer, UvA PhD under Max Welling, wrote the reference attention model, then co-organised the competition that found deep learning does not beat OR on static routing. Jinkyoo Park (KAIST) leads AI4CO and founded Omelet (2023, Kakao Ventures). InstaDeep: Karim Beguir and Zohra Slim, bought by BioNTech in 2023 for about €500M for the remaining shares. | ML researchers optimise for speed on distributions of small instances and generality across variants, not the last percent on one large instance, which is what a dispatcher pays for. InstaDeep's COMPASS abstract concedes RL heuristics "have not yet replaced industrial solvers". |
| Google OR team | David Applegate; Haihao Lu (MIT OR PhD, Google, now MIT Sloan), who "co-developed the PDLP algorithm, which many industry solvers and tech companies have adopted"; Miles Lubin (co-created JuMP). | Classical LP solvers (simplex, barrier) spend their time on sparse matrix factorisation: hard to spread over thousands of GPU cores and hungry for memory (the page's own example: barrier hit 1 TB RAM on the biggest LPs). PDLP instead repeats one cheap operation, a matrix-vector product (multiply the constraint matrix by a vector), which is the operation GPUs are built for and needs only the matrix itself in memory. That operation is the "simple kernel", a kernel being the small inner routine that does nearly all the work. People who know both OR and hardware choose the algorithm that fits the machine, even though it converges less precisely, instead of porting the old algorithm. This is the background-predicts-bet reading of the Google team, not something they state. |
| NVIDIA cuOpt | Product team since 2021; Li & Lim records in 2023; Apache 2.0 now, MIP in beta ("proving feasible solutions optimal remains under active development"). | A hardware vendor needs every solver to run on its chips, so it builds a reference implementation and gives it away. |
| Incumbent solver vendors | Gurobi (Rothberg and the GPU team), FICO, HiGHS, COPT. | Robustness on every customer model, so GPU becomes one more algorithm in the portfolio. "There is no silver bullet." |
The bets, ordered by how much classical OR they keep
| School | What it keeps fixed | Bet | Evidence |
|---|---|---|---|
| End-to-end neural CO | Only the problem definition | A learned policy replaces hand heuristics; now "one foundation model for all routing variants" (RouteFinder, 48 variants). | Benchmarks; HGS-PyVRP stays ahead on quality, RouteFinder 1 to 3% above it but seconds instead of minutes. |
| Learned branching | The exact solver and its certificate | Learn the expensive decisions inside the solver. | Benchmarks; its 2019 tooling (Ecole) is "looking for a new home". One shipped classifier in CPLEX 12.10. |
| ML plus CO layer | The classical engine | Learn the inputs of a combinatorial layer and train through it. | Won the dynamic ORTEC competition on real data. |
| GPU first-order LP | The LP itself | Change the algorithm to fit memory and parallel hardware. | Google production since 2023; in every major solver; giant LPs (up to 6.3 billion nonzeros) where barrier hit 1 TB RAM. |
Ordered by how much classical OR each keeps, the order matches the order of real-world evidence. Language models are almost absent. They appear in 2026 in three thin places: as the agent calling a GPU solver (NVIDIA's AI Planner, internal, with what-if planning cut from about a day to under 10 minutes and a 2026 Edelman finalist), as code generators writing solver internals (LLM4Branch at ICML 2026, from eight training examples; LLM-guided cut management in SCIP at IJCAI 2026), and as generators of training instances (MILP-Evolve). The second is bucket B aimed at C's components.
Every learned method that reached production kept a classical engine and its certificate. "A network that outputs the plan" was tested on real data twice and lost both times, and an LLM that emits a schedule directly is the same bet with a bigger network. The speed gains came from algorithm-hardware fit, which NVIDIA owns. A model lab has little to add on GPU kernels. The LLM4Branch pattern is the interesting exception: an LLM can tune or write a solver component for one customer's instance distribution from a handful of examples, where a GNN needed thousands of strong-branching labels.
3D. Richer models of reality
Lineage
Buckets A to C make the model easier to write or faster to solve. D asks whether the model is about the right world. OR has known for fifty years that often it is not, and that this, not solver speed, is why plans get overridden. Little (1970) wanted the manager's judgment as an input. Ackoff (1979) said OR had stopped being useful. The next decades built machinery that could only take judgment in formal form: constraint acquisition (learn the constraints by asking the user to classify examples), inverse optimisation (infer the objective from what planners did), stochastic and approximate dynamic programming (Powell's dispatch models). In 2020 Kesavan and Kushwaha measured the override problem in the field: overrides lost money on mature products and won on growth products. Both sides are partly right, and neither has a channel to give the other what it lacks.
2023 was the first time judgment in its native form, sentences, could be parsed. In twelve months: OptiGuide, Holy Grail 2.0 (Tsouros, Guns and others reviving "the user states the problem, the computer solves it"), OptiChat and MeetMate. Then the planning vendors arrived with agents: Blue Yonder, o9, Kinaxis, Pigment, Optimal Dynamics, SAP.
Groups and people
| Group | Main people and background | What the background predicts |
|---|---|---|
| Microsoft Research | Menache's group (OptiGuide); MeetMate from a separate group (Lawless, now Cornell; Suh, Rowan, Sarrafzadeh). | Operator-scientists with their own supply chain as testbed. Conservative and credible: solver as truth, LLM writes the query and the explanation. |
| Purdue, Can Li (OptiChat) | Chemical engineering; Tsinghua, CMU PhD in process systems optimisation, postdoc at Polytechnique Montréal. | Process systems engineering, where experts build large models and hand them to plant engineers who cannot read them. Aimed at that handover. |
| MIT, Chuchu Fan (LLMFP) | AeroAstro and LIDS; formal methods and verification for autonomous systems. Led by Yilun Hao. | Formalise, then let a solver guarantee. Robotics-style planning, not enterprise data. |
| KU Leuven, Tias Guns and Dimos Tsouros | From constraint acquisition (QuAcq2) to LLM modelling. | The only group whose pre-LLM agenda was literally getting the model out of the user's head. |
| IBM Haifa and Benelux OR (DOCP manifesto) | Wasserkrug, Boussioux, den Hertog, Mirzazadeh, Birbil, Kurtz, Maragno. | Practitioners who watched projects fail at "understand the business problem", which the manifesto names first. |
| Cambridge, Alexandra Brintrup | Seven-agent system reading news and mapping disruptions to supplier tiers. | Recommends; does not change or re-solve a model. |
| Planning-suite founders | o9 (Sanjiv Sidhu, i2 co-founder; Chakri Gottemukkala) and Blue Yonder (i2 to JDA to Blue Yonder, Panasonic) are two branches of the same 1988 family tree. Aera: Fred Laluyaux, ex-Anaplan CEO. Optimal Dynamics: Warren Powell's lineage. | They own the system of record and the planners' screens, so they build "agent that explains the plan" and "agent that fixes an exception". |
The bets
Research: the LLM is the interface, the solver is the truth. The research systems split by direction. OptiGuide and OptiChat translate outward (model to human). MeetMate and LLMFP translate inward (human to model). Only inward is bucket D proper. MeetMate has the most telling number: GPT-4 turned stated preferences into constraint code with about 95% precision and 94% recall, but the check for whether it had enough information was only about 80% accurate. Knowing when you do not know enough about reality is the bucket's core problem in miniature.
Incumbents: the agent lives where the plan lives. Blue Yonder bets on a knowledge graph as the editable model of reality. o9 on learning from plan versus execution ("post-game analysis"). Kinaxis on agents that "plan with you, not for you", built by customers in Agent Studio on GPT and Gemini. SAP on owning the transaction system, so the agent can act. Aera on the decision log ("decision memory"), which of the incumbents is closest to bucket D's thesis. Palantir on the Ontology as a live model of operations. Pigment (Paris, planning rather than OR) on an agent that edits the planning model with auditable change history.
New agent companies: optimisation core, agents at the messy edge. Optimal Dynamics states the thesis better than anyone: "optimization software tends to be brittle with respect to input data quality", and LLM agents alone cannot see how one decision ripples through a network. Its Scale product bids on loads across load boards, email and EDI. UiPath bought Peak, a decision layer. Airlines show the non-LLM version: Lufthansa Systems' aiOCC at Eurowings (reinforcement learning, controller decides, feedback evaluated) and SITA's purchase of Big Blue Analytics (OCCam, "up to 30%" lower disruption costs, LLMs planned on top).
Evidence, strongest first
The question was who runs agents that change the optimisation model in production, with named customers and measured outcomes. Strictly, there is no documented case.
| # | Who | What is documented | Caveat |
|---|---|---|---|
| 1 | Palantir at General Mills | Recommender that "consumes constraints, capacity, and network cost" in real time; over 70% accepted; about $14M a year; named customer executive. | Only part of the network. The document does not say an LLM is involved or that the model changes. |
| 2 | JD.com generative planning agents | Paper with Zuo-Jun Max Shen; about 22% better planning accuracy, about 2% higher in-stock in deployment. | Baselines and scope not in the abstract. |
| 3 | SITA OCCam, Lufthansa aiOCC | Production; aiOCC named at Eurowings. | Not LLMs. OCCam names no airline; aiOCC gives no numbers. |
| 4 | Microsoft OptiGuide | First-party production. | Outward only, no numbers beyond "days to minutes". |
| 5 | Kinaxis, o9, Blue Yonder, SAP, Aera | Named platform customers, productivity anecdotes (Kinaxis "up to 10x" at an unnamed pharma). | No agent-specific outcome on a planning KPI. SAP agents GA listed for Q4 2026. |
| 6 | Research | LLMFP 85% vs 39% on nine tasks; MeetMate n=10 user study; Brintrup on 30 synthetic scenarios. | Small and clean. |
Nobody publishes the closed loop: the planner says something, the agent turns it into a candidate constraint or objective change, tests it against history, and the model keeps it with a scope and an expiry. Aera's decision memory, Kinaxis' guardrails and aiOCC's feedback evaluation each hold one piece. The hard parts are telling knowledge from habit (the Kesavan result means an LLM that faithfully encodes every planner sentence encodes the bad overrides too), knowing what you do not know (MeetMate's 80%), governance (a rule from Tuesday's email must expire and be auditable), and evaluation, since no benchmark measures "the model became more faithful to reality".
China
Why China has its own solver scene
The domestic solvers came first, from 2017 to 2021: LEAVES (SUFE), CMIP (CAS), COPT (Cardinal Operations), MindOpt (Alibaba DAMO), OptVerse (Huawei, Huawei Cloud only). They were driven by Stanford-trained academics who started companies, by platform companies wanting the capability in-house, and by state utilities. China Southern Power Grid built its own solver, Tianquan, ran it in parallel with imported solvers from 2022, and now clears the southern regional market across five provinces with it. A 2026 review reports COPT and OptVerse first and second on Mittelmann's LP benchmark and COPT, OptVerse and TAYLOR the top three on MIPLIB2017, with the caveat that Gurobi left those benchmarks in August 2024 and MindOpt left in December 2024. In 2026 Gurobi ended free academic licences in China, Hong Kong and Macau without giving a reason. No central policy document naming solvers as a self-reliance target was found.
People
| Group | People and background |
|---|---|
| Cardinal Operations (杉数科技), Beijing, 2016 | Four Stanford PhDs. Luo Xiaoqu, CEO (Stanford GSB PhD, ex-BCG). Dongdong Ge, CSO (Stanford OR PhD, professor at SUFE, now Antai College at SJTU; leads COPT; consulted for Boeing, IBM, JD, SF-Express, Huawei, Didi). Zizhuo Wang, CTO (Stanford OR PhD, ex-Minnesota, now CUHK-Shenzhen). Chief scientific adviser Yinyu Ye (Dantzig's student, interior-point pioneer, 2009 von Neumann Prize). Investors include a national manufacturing fund, Ansteel and China Nonferrous Metals (2023 round of nearly RMB 500M). |
| Alibaba DAMO Decision Intelligence Lab | MindOpt and MindOpt Copilot; Wotao Yin co-authors both the solver review and OptLLM. |
| Huawei (Noah's Ark, Cloud, OptVerse) | 26-author ML-inside-OptVerse paper (Xijun Li, Jianye Hao, Junchi Yan, Defeng Sun and others). Mingxuan Yuan and Xialiang Tong connect Huawei to both main academic lines (EoH; the Zhejiang modelling survey). |
| CityU Hong Kong | Qingfu Zhang (PhD Xidian, MOEA/D at Essex), Fei Liu. LLM4AD platform. |
| Peking University | Zaiwen Wen (SJTU, Columbia PhD, UCLA postdoc; head of Industrial Engineering at PKU). MIPLIB-NL with Great Bay University and Huawei. |
The core people are mathematical programmers trained in the US OR tradition (Stanford MS&E, Columbia, the Dantzig and Ye lineage), often holding a professorship and a company role at once. That predicts exactness and solver verification at the centre, benchmarks and products tightly coupled (IndustryOR is a Cardinal artefact, MIPLIB-NL has Huawei co-authors), and growth through owning the stack.
Bets
Cardinal owns the full decision stack: domestic compute, COPT, an LLM modelling layer, sold as "decision AI". Its Qingtian platform (2026-07-22) turns language into models, then solves, simulates, backtests and does "knowledge retention", with an operator library for domestic chips. Wang: the key "is not just having more powerful AI models, but the ability to truly embed AI into key business systems". Its research arm fine-tunes open models: ORLM (in Operations Research, bases Mistral-7B, DeepSeek-Math-7B, LLaMA-3-8B) and SIRL (NeurIPS 2025, RL with the solver as reward on Qwen2.5, with COPT and Gurobi variants). Alibaba bet in 2023 on natural-language optimisation as a cloud service on Qwen (MindOpt Copilot, trial since June 2023), and nothing public on it followed after 2024. Huawei puts ML inside OptVerse (GCNs for the initial basis, RL for presolve and cut selection) and uses LLM-designed heuristics for the hardest routing. PKU bets on data and honest evaluation. Meituan, Cainiao, JD run classical OR plus ML at huge scale, with no documented LLM in dispatch or routing.
Evidence
| Tier | Case | What it shows |
|---|---|---|
| Peer-reviewed, no LLM | Meituan dispatch (Edelman finalist 2023) | Over 5 million couriers, 60 million orders a day; delivery time down 20.96%, distance per order down 23.77%, about USD 0.23 billion a year. |
| Peer-reviewed, no LLM | Alibaba / Cainiao routing | ALNS plus deep learning, over USD 50M a year. |
| Edelman finalist, no LLM | Huawei Cloud live-streaming scheduling (Edelman finalist 2023) | Network utilisation up more than 30%; Douyu and Huya named. |
| State media, no LLM | CSG Tianquan | Primary market-clearing solver for five provinces. |
| Vendor, no LLM | OptVerse at Tianjin Port | "Up to 144x" faster planning. |
| LLM, claims only | Cardinal ORLM, Qingtian; Alibaba MindOpt Copilot | Named users are universities. No industry numbers. |
| LLM, competition | Huawei-CityU, CVRPLib BKS challenge | Best-known solution on 51 of 100 instances (1,000 to 10,000 customers), with EoH-evolved operators inside AILS-II. The LLM was Gemini-2.5-pro, used offline, about 200 samples at one hour each per operator. |
The contrast with the West, and what it means for a European lab
Western bucket-A work mostly prompts frontier closed models and calls Gurobi. Chinese bucket-A work mostly fine-tunes open 7B to 32B models and grounds them in a solver the same group owns. Qwen2.5 is now the default base. Then Chinese groups took their own leaderboard apart. The Zhejiang/Huawei survey found error rates of at least 26.4% in NL4Opt, 23.7% in MAMO-ComplexLP and 54% in IndustryOR, Cardinal's own benchmark (Cardinal has since issued corrected sets). PKU's MIPLIB-NL showed fine-tuned systems mostly below 6% Pass@1 on real models, while frontier general models did several times better.
China is ahead on owning the stack, not on LLM models of reality. Nobody there, or anywhere, has shown an LLM taking messy enterprise context into a validated industrial model in production. Small fine-tuned translators hit a ceiling on industrial models, which is good news for a frontier lab. The best Chinese discovery result used a US frontier model, so China's edge in B is people and problems, not models.
For a European lab: Europe has no domestic commercial solver at COPT or OptVerse scale, so it would rely on Gurobi or FICO (US), HiGHS or SCIP (open), or a partner. The Chinese experience says owning or deeply partnering a solver matters, because the solver is both the product engine and the RL reward. The sovereignty argument that built Tianquan (an engine you control at the centre of critical infrastructure) is available in Europe for grids, rail and defence, but only paired with an exact solver. The people to read first: Ge, Wang, Ye, Wen, Zhang and Liu, Yuan and Tong.
4. Across the buckets
Who has really tested any of this in business
Ranked across all buckets, separating "LLM involved" from "OR or ML without an LLM", because the strongest evidence is mostly the second kind.
| # | Case | Bucket | LLM? | Strength |
|---|---|---|---|---|
| 1 | Meituan dispatch; Alibaba/Cainiao routing | ChinaC | No | Peer-reviewed, Edelman-grade numbers, years in production. |
| 2 | Google: PDLP in traffic engineering; AlphaEvolve's Borg heuristic (0.7% of compute) | CB | AlphaEvolve yes | Production, self-reported by the operator. |
| 3 | FM Logistic with AlphaEvolve | B | Yes | The only named external classical-OR deployment of an LLM method with a number (10.4%). One site. |
| 4 | Palantir at General Mills | D | Not stated | Named customer, $14M a year, partial network. |
| 5 | cuOpt at AT&T, clicOH | C | No | Named, vendor-reported numbers, no baseline for clicOH. |
| 6 | Microsoft OptiGuide; NVIDIA AI Planner | AC | Yes | Internal production. AI Planner an Edelman finalist; day to under 10 minutes. |
| 7 | JD.com generative planning agents | DChina | Yes | Paper claims deployment, about 22% accuracy gain; baselines not in abstract. |
| 8 | Gurobi, AIMMS, Nextmv, AMPL, Kinaxis, o9, Blue Yonder, Aera, Pigment | AD | Yes | Shipped; customers for platforms, not for the LLM part. |
| 9 | Everything else | all | mostly | Benchmarks and contests. |
Pattern across all buckets: the OR is in production, the LLM layer is in pilots, platforms and press. Where an LLM is in production, it is either at a company using it on its own models with its own OR team (Microsoft, NVIDIA, Google), or it writes a heuristic against an evaluator someone built from real operations (FM Logistic).
The benchmark reality
| Benchmark | What it shows |
|---|---|
| NL4Opt (2022) | 1,101 LP word problems. Best competition formulation accuracy 0.899; untuned gpt-3.5-turbo 0.927 on the combined task. Textbook level, near saturated. |
| ORLM on IndustryOR (2024) | Best ORLM 38.0% vs GPT-4 prompting 28.0%. SIRL-Qwen2.5-32B later 61.9% on the cleaned set, 87.4% on MAMO-Complex. |
| The cleanup (2025 to 2026) | Xiao et al.: errors in at least 26.4% of NL4Opt, 23.7% of MAMO-ComplexLP, 54% of IndustryOR. OptiMind team: 30 to 50% of public test data flawed. |
| MIPLIB-NL (2026) | 223 real MIPLIB 2017 models in language. Average Pass@1 17.85%. GPT-5.1 39.05%, Gemini-3-Pro 36.19%, DeepSeek-V3.2 23.81%. Fine-tuned systems mostly below 6%. "The main bottleneck is verification rather than generation." |
| FrontierOR (2026) | 180 tasks from OR papers. The best one-shot model beats Gurobi on quality and speed in 31%. |
| ORAgentBench (2026) | Best agent passes 35.51% overall, 20.59% of hard tasks. |
| Gurobi user survey (2025) | 6% already combine optimisation with LLMs, 24% experiment, 30% interested; 51% see model generation or code assistance as a use case. |
Formulation is far from solved on realistic sets, and the old leaderboards were partly measuring their own errors.
People movements
Talent flows from solver companies and OR groups toward model and hardware companies, and between incumbents. Oliver Bastert went from FICO to Gurobi CTO (2025). Dirk Zechiel went from co-founding Gurobi GmbH to Quantagonia to Strangeworks CTO. Pushmeet Kohli spent a decade at Microsoft Research before DeepMind. Haihao Lu went from Google's OR team to MIT Sloan, and his algorithm went into every solver. Andrea Lodi went from CPLEX consulting to Montréal to Cornell Tech. Robert Lange interned on DeepMind's Discovery team before ShinkaEvolve at Sakana. Lawless did MeetMate at Microsoft and is now at Cornell. Cardinal's Wang Xi came from Google. The Chinese founders went Stanford to Shanghai, Beijing and Shenzhen.
5. Where the white space is
A is crowded and commoditising. Every solver vendor has a modeller assistant and an MCP server, and general coding agents already write gurobipy. A pure words-to-model product has no moat. What is left is the checker: spec plus acceptance tests, infeasibility diagnosis, "why did it do that", and the corpus of real industrial formulations with known-correct answers. Solver companies sit on that, which is a reason a lab would want people from them.
B is where the labs are, and still mostly scored on puzzles. The loop itself is free (OpenEvolve, ShinkaEvolve). The open gap is building evaluators from a company's own operations quickly, and earning trust in the result. Evolve the heuristic for this warehouse, this fleet, this week. That is the "engines start from a blank page" claim in a new form.
C is a hardware story and NVIDIA is winning it. A lab adds little except as a good caller of solvers, and maybe as the writer of solver components tuned to one customer's instances.
D is the emptiest and the most valuable. Nobody publishes the closed loop: read the contracts and the planner's notes, propose a typed, testable model edit (constraint, parameter, objective weight, with scope and expiry), abstain when information is missing, replay history to decide if it is knowledge or habit, and keep it versioned. That needs an exact solver so correctness never depends on the LLM, a design partner with years of plan, override and outcome logs, and a model trained to produce edits rather than whole models. European on-premise deployment matters because the inputs are exactly the sensitive documents that OptiGuide was designed never to send to an LLM.
6. Sources
All pages were opened by the research lanes or for this page. Claims that only appeared in search snippets, or that a lane could not confirm, were left off.
A. Front end
- GAMS company history
- AMPL about · AMPL AI
- AIMMS AI capabilities
- Gurobi, Wikipedia
- NL4Opt report · NL4Opt site
- OptiGuide paper · OptiGuide blog · repo
- OptiMUS v1 · OptiMUS scalable · Udell homepage
- Chain-of-Experts, ICLR 2024
- ORLM · ORLM repo · LLMOPT
- Cosmo Tech copilot
- NVIDIA cuOpt agent, 2024 · cuOpt agent skills, 2026
- HBR article record at MSR · HBR · MIT Exec Ed summary
- Gurobi AI Modeling · Gurobot · Gurobi MCP docs · Intelligence Hub launch
- Gurobi 2025 survey
- Bastert appointment
- Strangeworks acquires Quantagonia · Aura release
- OptiMind · Menache profile
- MIPLIB-NL · FrontierOR · ORAgentBench · Xiao et al. survey
- Nextmv MCP · Nextmv about
- IBM Modeling Assistant
- Dongdong Ge bio · Yinyu Ye profile
B. Algorithm discovery
- Burke et al. 2013, hyper-heuristics survey
- FunSearch, Nature 2023
- Evolution through Large Models
- AlphaTensor
- EoH · EoH repo · LLM4AD
- ReEvo · LLaMEA
- AlphaEvolve blog · paper · Cloud preview · GA
- FM Logistic case · FM Logistic on trans.info
- ALE-Bench · ShinkaEvolve · ShinkaEvolve paper · AHC058 · Sakana home
- AWTF 2025 · AWTF 2026 Heuristic · OfficeChai
- VRPAgent · PyVRP+
- Kohli profile · MIT Technology Review · T3F
- Qingfu Zhang · van Stein · Lange bio
- OpenEvolve
- ORTEC science · ORTEC competition news · Perron profile
C. Better solvers
- Pointer Networks · Bello et al. · Dai et al. · Khalil et al.
- Kool et al. attention model · POMO · Bengio, Lodi, Prouvost · Gasse et al. · Nair et al.
- CPLEX classifier (Bonami, Lodi, Zarpellon)
- Amazon Last Mile winner · Accorsi, Lodi, Vigo · ML4CO
- EURO meets NeurIPS 2022 · Kleopatra, Transportation Science
- PDLP · Google PDLP blog · PDLP revision
- cuPDLP.jl · cuPDLPx · Backing PDHG into a Corner
- Gurobi GPU support page · Gurobi GPU blog
- cuOpt LP · cuOpt open source · cuOpt repo · cuOpt records, AT&T · clicOH · NVIDIA AI Planner
- RouteFinder · RL4CO · Ecole
- Lodi · Khalil · Kool profile · Haihao Lu · Omelet
- InstaDeep · BioNTech acquisition · Poppy · COMPASS · Jumanji · DB Netz CTMS
- LLM4Branch · LLM cuts in SCIP · MILP-Evolve
D. Richer models of reality
- Little 1970 · MIT DSpace
- Zeleny on Ackoff · Punk Rock OR
- QuAcq2 · Inverse optimisation survey · CASTLE
- Kesavan and Kushwaha summary
- Holy Grail 2.0 · OptiChat · MeetMate · DOCP manifesto
- Can Li · Chuchu Fan · LLMFP, MIT News · Brintrup agents
- JD generative planning
- Blue Yonder agents · o9 · Kinaxis agents · Agent Studio · SAP agents
- Aera platform · Aera customers
- Palantir General Mills impact study · Palantir and NVIDIA
- Pigment Modeler Agent
- Optimal Dynamics · Scale · Series C · UiPath buys Peak
- Lufthansa aiOCC · SITA buys Big Blue Analytics
- Sanjiv Sidhu · Blue Yonder · Laluyaux at Aera · Kinaxis
China
- Huang, Xiao, Ge, Yin, Liang, solver overview, 2026 · Mittelmann benchmarks
- Tianquan, Science and Technology Daily
- Gurobi China academic licences · 中华网 on the licence change
- 钛媒体 interview with Ge · 36Kr PitchHub, Cardinal · 36Kr 2016 · Zizhuo Wang · Yinyu Ye bio
- Qingtian launch · SIRL · SIRL repo · Cardinal at WAIC 2025
- OptLLM · MindOpt Copilot
- ML inside OptVerse · OptVerse product page · Tianjin Port · Huawei Cloud Edelman
- OptVerse-CityU CVRPLib report · CityU Optima · Qingfu Zhang, Xidian
- Zaiwen Wen · MAMO
- Meituan dispatch, IJAA · Alibaba routing, IJAA · JD assortment and inventory
History, one timeline
The goals are old. Each generation found a new tool for the same three wishes: let non-experts write models (A), let machines design heuristics (B), and get the manager's judgment into the model (D). The new thing since 2022 is a mutation operator and a parser that both understand language and code.
| 1960s to 2014: the old wishes | ||
| 1960s, 2000 | B | Idea of machines designing heuristics; the term "hyper-heuristics" from 2000 (2013 JORS survey by Burke, Gendreau, Hyde, Kendall, Ochoa, Özcan, Qu). |
| 1969 to 1970 | D | John Little's "Models and Managers": a decision calculus processes "data and judgments". Judgment was meant to be an input. |
| 1970s, 1987 | A | GAMS starts in a World Bank economic modelling group; commercial product in November 1987. |
| 1979 | D | Ackoff, "The Future of Operational Research is Past". The charge is usefulness, not mathematics. |
| 1985 | A | AMPL designed at Bell Labs by Fourer, Gay and Kernighan "to let people express optimization models in familiar algebraic form". |
| 1990 | D | Warren Powell founds CASTLE Lab at Princeton; dispatch models used at Yellow Freight, Schneider, Norfolk Southern. |
| 2008 | A | Gurobi founded by Gu, Rothberg and Bixby. |
| 2015 to 2021: learning the solver, and the first reality checks | ||
| 2015 to 2016 | C | Pointer Networks output TSP tours; Khalil et al. learn to imitate strong branching in a MIP solver. Replace the solver versus improve one decision inside it, from day one. |
| 2017 to 2020 | China | LEAVES (SUFE, 2017), CMIP (CAS, 2018), COPT 1.0 (Cardinal, 2019-05-28), MindOpt 0.9 (Alibaba, 2020-08). |
| 2018 | C | Kool, van Hoof, Welling's attention model for routing; Bengio, Lodi, Prouvost's "tour d'horizon" sets the ML-for-solver programme; Google starts the work that becomes PDLP. |
| 2019 | C | Gasse et al. put a graph neural network on the MIP and beat expert branching rules in a solver. |
| 2020 | D | Kesavan and Kushwaha's field experiment: merchant overrides of a data tool lose money on mature products and win on growth products. |
| 2021 | CChina | Amazon Last Mile challenge won by classical combinatorial optimisation (Cook, Held, Helsgaun), 42% ahead of second. PDLP at NeurIPS. Huawei announces OptVerse. |
| 2022 to 2023: the language turn | ||
| 2022 | BAC | OpenAI's "Evolution through Large Models" (June): a code model as the mutation operator. AlphaTensor (October). NL4Opt competition at NeurIPS. EURO meets NeurIPS on real ORTEC data: deep learning "fails to outperform OR baselines" on static routing. |
| 2023 | ABCD | PDLP in Google traffic engineering (May). OptiGuide (July). Holy Grail 2.0, OptiChat and MindOpt Copilot (August). OptiMUS (October). cuPDLP.jl matches Gurobi on standard LP sets (November). FunSearch in Nature and MeetMate (December). |
| 2024 to 2025: split into prompting versus fine-tuning, academics versus clouds | ||
| 2024 | BACD | EoH (January), ReEvo (February), LLaMEA (May). DOCP manifesto (February). ORLM fine-tunes 7B models (May). NVIDIA's cuOpt agent uses Codestral-22B (July). COPT and HiGHS add PDLP. Gurobi leaves the Mittelmann benchmarks (August). |
| 2025 | ABCD | HBR piece on Microsoft's LLM supply planning (January). Gurobi AI Modeling (February). cuOpt goes open source (March). AlphaEvolve (May) and Blue Yonder agents. OpenAI 2nd at AtCoder World Tour Finals Heuristic (July). Strangeworks buys Quantagonia (August). Benchmark error survey (August). Gurobi survey: 6% use LLMs with optimisation (September). Gurobi GPU alpha (September). AlphaEvolve Cloud preview and Sakana's AHC058 win (December). |
| 2026: the reckoning and the MCP turn | ||
| Jan to Mar | AChinaBD | OptiMind; 30 to 50% of benchmark test data flawed (January). Huawei-CityU wins the CVRPLib BKS challenge (January to February). MIPLIB-NL (February). FM Logistic AlphaEvolve case and Pigment Modeler Agent (March). Nextmv MCP server (preview March, GA April). |
| Apr to Jul | ABCChina | cuOpt agent skills (May). FrontierOR (May), ORAgentBench (June). SITA buys Big Blue Analytics (June). Gurobi Intelligence Hub with MCP; AlphaEvolve GA on Google Cloud; OpenAI wins the AtCoder Heuristic final; Strangeworks Aura; Cardinal's Qingtian platform (all July). |
| Aug to Oct | ChinaD | Gurobi ends free academic licences in China, Hong Kong and Macau (help page updated 2026-08-31). SAP announces planning agents for Q4 GA (October). |