Introducing &Tutor

Open original

humans& has released &Tutor, a model trained to teach by getting rewarded for how much a simulated learner (Persimmon) picks up. It scores 0.754 on the TutorMoments benchmark against 0.500 for its base model and beats several frontier models, and the team says the learner model's realism is what makes the tutor work.

  • On TutorMoments under a plain prompt, &Tutor scores 0.754, versus 0.500 for the GLM-5.3 base model and 0.667 for Claude Sonnet 5.5.
  • Teaching behaviors like knowing when to scaffold or push for rigor are never trained directly; they emerge from rewarding learning gains.
  • Swapping in GLM-5.3 as the learner instead of Persimmon produced a noticeably weaker tutor, which the authors attribute to Persimmon being a more realistic learner.

Ampersand Tutor, or &Tutor, is a model trained to teach and released today via API by humans&. Its training setup is the interesting part. Rather than training on human-written teaching transcripts, &Tutor learns by teaching Persimmon, the company's large-scale simulated learner, and gets rewarded according to how much the simulated student learns. The team says behaviors like knowing when to scaffold, when to push for rigor, and when to hold back emerge from that reward signal without being trained directly. Notably, &Tutor was trained from GLM-5.3, and an ablation that swapped Persimmon for GLM-5.3 as the learner performed significantly worse, which the authors attribute to Persimmon being a more faithful model of real learners. On the TutorMoments benchmark, built from real math tutoring moments annotated by experienced teachers, &Tutor scores 0.754 under a plain prompt versus 0.500 for its GLM-5.3 base. It beats Claude Sonnet 5.5 (0.667) and GPT-6 Astra (0.634) in that setting. Under an evaluation-aware prompt that spells out the support-versus-rigor balance, scores rise across the board: &Tutor reaches 0.919, just behind GPT-6 Astra at 0.926, and posts the top score on appropriate rigor at 0.931. Those are the headline numbers, and they're worth reading with the caveat that the benchmark is judged by an LLM against teacher annotations, and the authors themselves say the real test is longitudinal learning outcomes in classrooms or tutoring products, which they have not yet shown. The bulk of the post defends Persimmon as a learner model, since a simulated student that behaves differently from real people would let a tutor learn to exploit quirks instead of teaching. The validation uses four datasets: SimulatorArena Math and StudentBench, where learners interact with AI tutors, and Eedi and TutorMoments, where learners interact with human tutors. For the human-tutor sets, the simulator generates one learner turn at a time from the real conversation history. Behavior is compared through an open-coding pipeline: about 3,000 real learner turns per dataset are labeled by LLM annotators into taxonomies of conversational moves such as answering incorrectly, requesting explanation, proposing a strategy, or challenging the tutor, with inter-annotator agreement reported as Fleiss' κ on held-out conversations. The core question the section tries to answer is whether Persimmon's distribution of moves, affect, and error patterns matches what real learners produce, and the article says it does better than the alternatives on those dimensions.

Ask about this article