HiMCM 2024: Recruiting & the AI Modelling Course
2024 HiMCM 摇人——以及AI建模新课程
After several years running HiMCM teams, the consistent finding is that incoming student background correlates weakly with final results. The skills that matter — working fast, communicating ideas, directing AI tools toward a specific problem — are not the ones you typically arrive with. They are the ones the competition teaches under pressure. So we have stopped filtering on prior preparation.
The short version: HiMCM is worth doing. Win and you have a strong award. Lose and you leave with LaTeX, Python packages, MATLAB familiarity, and a modelling workflow that looks like two or three lines on any research CV. There is no bad outcome.
What the Course Is Actually About
We consulted with teachers who have been coaching HiMCM for years before designing this. The consistent view: stop drilling specific algorithms and start teaching students what the landscape of methods looks like, then how to direct AI to implement whatever they choose. HiMCM 2023 explicitly allowed AI assistance with disclosure. Nobody had built a curriculum around that permission yet. We did.
The goal is not to teach you to use a screwdriver. It is to teach you to recognise when a screw is there.
The analogy I keep using: traditional courses teach you how a screwdriver is manufactured and the historical development of the Phillips head. What we teach instead is: here are the jobs a screwdriver does, here is how you identify one when you need it, and here is how you get an AI to handle the turning while you hold the piece in place.
Five Modules
LaTeX proficiency is table stakes for a credible paper. With AI, it is genuinely fast: photograph clean handwritten working, upload it, instruct the model to output LaTeX with editorial polish. The module covers what LaTeX is doing and how to direct AI through it efficiently — not the historical development of TeX.
We will not spend time on syntax, data structures, or debugging loops. The focus is on understanding what AI can do to make a programmer redundant, and how students with no coding background can produce working, well-structured code for competition purposes. Counterintuitively, students without entrenched coding habits often outperform those who do — the latter think in terms of implementation difficulty; the former think in terms of what the problem needs.
The core of the course. Optimisation, prediction, classification, regression, genetic algorithms, neural networks — covered at the level of "what does this do, and when is it the right choice?" Implementation is handled by AI. The analogy: we teach you to recognise a screw, not manufacture a driver. This is the module where the most contest-valuable time is spent.
No formula derivations. No SPSS walkthroughs. The entire data-cleaning and analysis pipeline, handled with AI as fast as possible, so the team is not still processing spreadsheets at 2 a.m. on the last night. Efficiency here is sleep; sleep is coherent writing; coherent writing is a better paper.
Not AI generation — that we do not teach, because everyone already knows how to do it badly. What we teach: using AI to impose academic register on your own thinking, to structure a paper's argument, to catch missing units and mislabelled figures, and to translate a student's genuine reasoning into the form the judges actually read.
Everything above was used in last year's competition, explicitly disclosed, and produced strong results. The students who understand AI well enough to know what it can and cannot do are the ones who use it well. Treating AI as dangerous is neither accurate nor useful.