humans& Trains Ampersand Tutor by Teaching a Simulated Student

retained text

humans& Trains Ampersand Tutor by Teaching a Simulated Student
humans& trained a tutor model against a simulated learner called Persimmon, topping a real-classroom benchmark and beating Claude Sonnet 5.5 and GPT-6 Astra.
- humans& released &Tutor;, a tutoring model trained by teaching a simulated student until it understood.
- Scored 0.754 on TutorMoments plain prompt, beating Claude Sonnet 5.5 (0.667) and GPT-6 Astra (0.634).
- With evaluation-aware prompting, hit 0.919 overall and led all models on appropriate rigor at 0.931.
- Ablation: swapping Persimmon for GLM-5.3 as the simulated learner dropped tutor score from 0.659 to 0.477.
- Persimmon matches real learners on conversational moves, affect, and 25 to 30 percent error rates where frontier models stay under 15 percent.
- Available now via the Ampersand agent API; backed by NVIDIA, Bezos, GV, SV Angel.
humans&, a Redwood City research startup backed by NVIDIA, Jeff Bezos, GV, and SV Angel, has released Ampersand Tutor, also called &Tutor. According to the company, the model was trained to maximize how much a simulated student learns and is now available through its agent API.
The model starts from GLM-5.3 and trains against Persimmon, humans&'s separate model of user behavior. Each tutoring response receives a reward based on its effect on the simulated learner's understanding. The training objective contains no explicit labels for scaffolding, rigor, or when to withhold an answer.
The reward follows the learner
During training, &Tutor repeatedly teaches Persimmon, observes its responses, and adjusts its strategy according to the resulting learning gain. That loop gives the tutor practice with learners who guess, ask for help, make mistakes, become frustrated, or disengage.
- Persimmon adopts the role of a learner with a particular level of knowledge and behavior.
- &Tutor guides the learner through the material over multiple conversational turns.
- The training environment measures Persimmon's learning and uses that result as the reward.
- The tutor updates toward strategies associated with larger learning gains.
This design ties the reward to the learner's progress. Behaviors such as asking a useful follow-up question or allowing productive struggle can emerge when they improve that outcome.
A lead on TutorMoments
The main result comes from TutorMoments, a benchmark built from real math tutoring sessions. Experienced teachers identify moments when a tutor must decide whether to offer support, request more reasoning, or let the student continue working.
At each selected turn, a model continues the tutoring conversation while an AI model plays the student. An LLM judge then compares the response with the teachers' annotations. The benchmark therefore measures local pedagogical decisions in simulated continuations, rather than learning or retention among real students.
Under the evaluation-aware prompt, which targets the benchmark's criteria more directly, &Tutor also records a reported 0.931 for appropriate rigor. That measure captures whether the tutor asks the student to supply enough reasoning instead of giving away too much of the solution.
Because TutorMoments relies on an AI student and an LLM judge, the scores provide controlled evidence about conversational tutoring behavior. Human trials are still required to establish whether those behaviors improve understanding and retention in classrooms.
Persimmon changes the training result
The team's ablation keeps the tutor training recipe fixed while replacing Persimmon with GLM-5.3 in the learner role. This comparison isolates the effect of the simulated student more directly than the leaderboard does.
In the reported low-reasoning setting, the Persimmon-trained tutor scores 0.659, while the tutor trained against a GLM-5.3 learner scores 0.477. humans& attributes the gap to Persimmon's closer imitation of real learner behavior.
A learner simulator can create exploitable shortcuts when its behavior diverges from that of people. A tutor may discover strategies that raise the simulated student's score while failing with learners who misunderstand concepts, express uncertainty, or respond unpredictably. Researchers generally describe this failure mode as simulator overfitting.
How human is the simulator?
To test Persimmon's behavior, humans& used open-coding evaluations across four datasets of real tutoring conversations, including SimulatorArena Math and StudentBench. Open coding groups observed responses by behavior, such as guessing or requesting help, instead of forcing every response into a predetermined answer format.
Lower values indicate closer behavioral similarity for the first two dimensions. On checkable claims, real learners were wrong 25% to 30% of the time, placing Persimmon near the observed human range.
The tested assistant-tuned models made fewer errors than real learners, which limits their value as stand-ins for struggling students. A tutor trained against an unusually capable learner receives little practice diagnosing misconceptions or responding to persistent confusion.
One strategy across three models
Founded by Eric Zelikman, Georges Harik, Stanford professor Noah Goodman, and Andi Peng, humans& has organized its releases around models of people and collaboration. Zelikman is also known for his work on STaR, a method for self-improving reasoning models.
Persimmon models user behavior, Ampersand focuses on collaboration, and &Tutor applies the same approach to education. The common strategy treats user modeling as a separate technical problem from improving an assistant's general capabilities.
What developers can test now
Teams building educational software can access &Tutor through humans&'s agent API, with the model card and documentation linked from the product site. The release also gives developers several concrete points to examine before deployment:
- Access: The announced product is API-based, and the release does not describe downloadable model weights.
- Reproducibility: GLM-5.3 is identified as the starting model, but reproducing the training run also requires details about the reward calculation, training data, model versions, and compute.
- Evaluation: Product tests should include real learners, delayed retention checks, and subjects outside the benchmark's math conversations.
- Student data: Education teams should verify API policies for retention, training use, regional storage, and accounts involving minors.
- Robustness: Trials should cover different ages, knowledge levels, languages, accessibility needs, and patterns of disengagement.
Within humans&'s reported setup, a tutor trained against Persimmon outperformed the same training recipe with GLM-5.3 as the learner and ranked strongly on TutorMoments. Studies with real students over weeks will determine whether those gains transfer to durable learning.

document fetched 2026-10-31 02:55