A 4B model mimics real students better than GPT-5.4
Two stories about where models are being pointed today: a student simulator that beats frontier models at being one particular learner, and Meituan's shopkeeper agent crossing a million restaurants.
Microsoft researchers trained a 4B model to imitate individual students — and it beats GPT-5.4 and GPT-4o at the job. The system, called StudentSim, attacks a real bottleneck in AI tutoring: knowing which explanation works for which student requires watching real learners, and that feedback is slow, expensive and sparse. Prior approaches each covered half the problem — knowledge-tracing models reproduce how a student behaves but cannot digest a tutor's explanation, while a prompted frontier model follows guidance fluently but does not actually play that student. StudentSim trains in two stages instead: a pooled domain model learns what students share (common mistake patterns, how answers change after a hint), then a second stage specializes it to one learner from a very thin record. In the English-writing dataset the median student had written three essays; some learners had as few as 73 usable records.
The comparison is the interesting part, because the small model wins on both axes the paper measures. Across 60 students in chess, second-language English writing and mathematics, the fine-tuned 4B simulator beat Maia2 — a model purpose-built to predict human chess moves from rating — on behavioral fidelity, and it beat GPT-5.4 and GPT-4o on responsiveness to guidance. On a single chess position where three real players chose three different moves, StudentSim reproduced each player's choice while the rating-based model predicted the same move for all three and GPT-5.4 got all three wrong. That is the point about per-player resolution: a description of someone's past games does not constrain a model the way training on those games does.
Why it matters beyond tutoring: the researchers then used the simulator as the reward signal to train a chess tutor with reinforcement learning, and professional players rated that tutor highest on factual accuracy, explanation quality and adaptation to the individual. The tutor trained with GPT-5.4 playing the student scored worse on factual accuracy than a tutor with no extra training at all. If a simulated student can serve as the reward model, education labs get a way to iterate on tutors without recruiting classrooms — which is exactly the constraint that has kept adaptive tutoring behind the underlying models. The authors are candid that this is a proof of concept: chess works because an engine can objectively judge a move, while essays and open-ended math lack a reliable scoring function, so the same loop does not transfer for free.
Meituan says its AI shopkeeper now covers about 1.3 million restaurants nationwide, and it is built on the company's own LongCat models. At its in-store dining conference in China today, the company put numbers on 智能掌柜 — "Smart Shopkeeper" — for the first time at this scale: roughly 1.3 million restaurant merchants, more than 130,000 brand stores, and about 8.6 million merchant questions resolved. Meituan vice president Wei Wei says the tool now spans business analysis, reputation management, marketing plans, ad placement, store management, industry research and new-store operations, and can flag food-safety and serious-complaint risks in advance. Meituan has been showing this stack since the September services trade fair, where it also displayed the trillion-parameter LongCat-2.0 and its CatPaw business workbench — so the number that matters is not the feature list but the distribution: a domestic model family running the daily operations of over a million small businesses.
What to watch: whether simulator-trained tutors move past chess into domains with workable scoring functions, and whether Meituan exports its LongCat-based merchant tools outside China.
Would you trust a model trained on your own mistakes to judge the tutor teaching you? Tell us in the comments.
Sources: Microsoft StudentSim (GitHub)
