humans& Releases &Tutor, an AI Model Trained on a Simulated Student's Learning Gains
The startup says its tutor beat frontier models on the TutorMoments benchmark under a plain prompt. The test uses an AI student and an AI judge, not real learning outcomes.
Published
Last updated / Fact-checked on
The AI startup humans& released Ampersand Tutor, or &Tutor, on October 30, a model it trained by rewarding how much a simulated student learned. The company says &Tutor outscored leading frontier models on the TutorMoments tutoring benchmark under a plain prompt. It is available now through the company's API.
Image: Illustrative photo of a child concentrating on schoolwork. Les Anderson / Unsplash via Wikimedia Commons — CC0 1.0
Who is humans&
humans& raised a $480 million seed round at a $4.48 billion valuation in January, Bloomberg reported. The round was led by SV Angel and co-founder Georges Harik, with Jeff Bezos, Nvidia and GV among the other investors.
Training against a learner model
&Tutor's training is built around one reward: how much the student learns. The student is Persimmon, humans&'s own large-scale model of user behavior, which plays the learner. The company released a research preview of Persimmon on September 10.
According to the blog post, &Tutor was not trained from Ampersand, the company's general model, though humans& plans to fold it into future Ampersand training. The tutor starts from the GLM-5.3 model.
humans& says behaviors usually associated with good teaching, such as knowing when to scaffold, when to push for rigor and when to hold back, were never trained directly. "The pedagogical skills ... all emerge implicitly from training," the company writes.
The learner matters, the company argues. In an ablation, humans& trained a tutor with the same recipe but used GLM-5.3 as the student instead of Persimmon. That version "significantly underperforms," the post says. humans& attributes the gap to Persimmon being a more faithful model of real learners.
The benchmark results
TutorMoments was developed by the Allen Institute for AI (Ai2). It is built from real one-on-one math tutoring sessions. Experienced math teachers flag moments where a tutor must choose between offering support and pushing the student to think harder. A model takes over the tutor role from that point, another AI model plays the student, and an LLM judge scores the replies against the teachers' annotations.
humans& reports these overall scores:
| Model | Plain prompt | Evaluation-aware prompt |
|---|---|---|
| &Tutor (max reasoning) | 0.754 | 0.919 |
| Claude Sonnet 5.5 | 0.667 | 0.902 |
| GPT-6 Astra | 0.634 | 0.926 |
| GLM-5.3 foundation (untrained, max reasoning) | 0.500 | 0.877 |
The plain prompt simply asks the model to tutor well. The evaluation-aware prompt spells out the balance between support and rigor, and it raised every model's score. Under that prompt, &Tutor finished just behind GPT-6 Astra overall but had the highest score of any model on appropriate rigor, at 0.931.
The limits of the test
These numbers measure decisions in a simulation, not learning. Ai2 says TutorMoments uses a simulated student, so its scores reflect "how a model acts at a decision point—not whether a real student learned." Its dataset is also narrow: US-based and mostly elementary and middle-school math.
humans& acknowledges a related risk in its own approach. If simulated learners behave differently from real people, it writes, &Tutor "may learn to exploit those differences," producing gains that don't carry over to real teaching. "The real test of a model like &Tutor is how well it helps people learn over time in real-world settings," the company says, adding that it wants to work with education teams on such evaluations.
Why it matters
humans& is betting that a tutor trained on measured learning, rather than on labeled teaching behaviors, will teach better. For now, the evidence comes from the company's own benchmark runs with a simulated student.
What developers and educators can do now:
- Try the model through the Ampersand API using the model ID
ampersand-tutor-v0.1. - Read the company's full results and Ai2's TutorMoments caveats before relying on the scores.
- Test with real students, including delayed checks on what they retain, before deploying it widely.
More coverage is on our Artificial Intelligence hub.
How we reported this
We checked the method, scores and quotes against humans&'s &Tutor blog post, and the release date against the company's blog index. Ai2's TutorMoments post supplied the benchmark design and its caveats. The Ampersand docs confirmed API access and the model ID. Funding details come from Bloomberg's January report. We removed an unsourced claim about discussion on X.
Published and fact-checked October 30, 2026.
Sources
- humans&, "Introducing &Tutor," Oct. 30, 2026: https://tutor.humansand.ai/blog/
- humans&, Blog index: https://humansand.ai/blogs/
- humans&, "Build with Ampersand" (API docs): https://ampersand.humansand.ai/docs/
- Ai2 on Hugging Face, "TutorMoments: Do AI tutors know when to help and when to hold back?", Aug. 7, 2026: https://huggingface.co/blog/allenai/tutormoments
- Bloomberg, "Nvidia, SV Angel Set to Back Humans& at $4.48 Billion Valuation," Natasha Mascarenhas, Jan. 20, 2026: https://www.bloomberg.com/news/articles/2026-01-20/nvidia-sv-angel-set-to-back-humans-at-4-48-billion-valuation
DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.