Introduction
Today, we are releasing Ampersand Tutor (&Tutor), a model trained to teach, available via API.
&Tutor learns to teach by training against Persimmon, our large-scale user model, which plays the learner. &Tutor achieves state-of-the-art performance on TutorMoments [1], a benchmark built from real tutoring moments.
A core belief at humans& is that models can better understand us and optimize for our long-term goals by learning from long human interactions. Teaching brings this belief into sharp focus. It requires understanding what a learner knows [2] and building on that understanding to help them learn [3]. Our evaluations examine how this training changes &Tutor’s teaching behavior and how those changes are driven by Persimmon.
In this blog post, we describe how we built and evaluated &Tutor. Then, we describe the evaluations we designed to validate Persimmon’s ability to model human learners. We are making &Tutor available to the public today via API.
Yes n = 10
Let's check that! If n = 10, what is 10 × 10 × 10? Compute it out. Does it equal 10?
ohhhhhhh
Ha, I love that reaction! Tell me what you just realized. And then give my earlier question another try — is there any whole number n where n × n × n = 10? Why does that matter for taking cube roots?
Because cubes of numbers get big pretty fast. 1 to the third is 1, 2 to the third is 8, 3 to the third is 27. 10 is between 8 and 27, so there isn’t a whole number n such that n times n times n equals 10.
&Tutor
A good tutor helps learners develop lasting understanding. We built &Tutor’s training environment around this principle, rewarding the tutor by how much the student learns. We train against Persimmon as a learner model. Our evaluations below show that this training environment encourages many downstream teaching behaviors associated with good pedagogy.
Next, we go into detail on our evaluations of &Tutor.
TutorMoments
TutorMoments [1] is built from real tutoring sessions with real students. Experienced math teachers pick out key moments where a tutor has to decide whether to offer support or push the learner to think harder, and note what each moment calls for. Models continue each conversation from that moment, with an AI model playing the student, and an LLM judge scores their responses against the teachers’ annotations. We evaluate tutor models under both a plain prompt (which just asks the model to tutor well) and an evaluation-aware prompt (which explicitly describes the balance between support and rigor).
Results. Under the plain prompt, &Tutor scores 0.754 overall on TutorMoments, compared with 0.500 for the GLM-5.3 model it was trained from. It outperforms leading frontier models, including Claude Sonnet 5.5 (0.667) and GPT-6 Astra (0.634). With the evaluation-aware prompt, which raises every model’s score, &Tutor scores 0.919, just behind GPT-6 Astra (0.926), and has the highest score of any model on appropriate rigor (0.931).
Importantly, none of these behaviors are trained for directly. &Tutor is rewarded based on how much the student learns. The pedagogical skills of knowing when to scaffold, when to push for rigor, and when to hold back all emerge implicitly from training.
Our results also suggest that the learner model plays a key role in the training environment and determines the quality of the resulting tutor model. We ran an ablation training a tutor model with the same recipe but with GLM-5.3 as a learner instead of using Persimmon. This ablation significantly underperforms &Tutor trained against Persimmon. This result is explained by the fact that, as we later show in the learner evaluations, Persimmon is a much stronger model of a learner than GLM-5.3.
The real test of a model like &Tutor is how well it helps people learn over time in real-world settings. We look forward to collaborating with teams working in education to evaluate it in those settings.
Persimmon as a learner
Simulated learners have long been a research topic in education [4, 5], used to give teachers practice, evaluate teaching methods, and study how students learn, and language models have recently made them far more capable [6, 7, 8, 3, 9, 10, 11, 12]. Good simulated learners can in turn be used to train tutors [13, 14]. Scaling this approach depends on how well the simulated learners model real learning behavior. If they behave differently from real people, &Tutor may learn to exploit those differences, producing gains in simulation that fail to translate to real-world teaching. The quality of the learner model is therefore central to training &Tutor.
We used Persimmon as our model of learners after extensive validation of its ability to faithfully model the behavior of real human learners. We specifically show that Persimmon is better able to model the distribution of real learners’ behavior over many dimensions relevant to learning, including conversational flows, affect, and error patterns.
Datasets
To evaluate how closely simulated learners resemble real people, we generate learner responses using real tutoring conversations as a starting point. We then compare the simulated and real learners across four datasets.
Two datasets capture learners working with AI tutors. These are SimulatorArena Math [15], where people work through competition math problems, and StudentBench [16], one-hour GRE tutoring sessions. For these datasets, the simulated learner continues from the real learner’s first message over multiple learner turns, interacting with the same AI tutor prompt and model, or the closest available version where the original model has been retired.
The other two, Eedi [17] and TutorMoments [1], capture learners working with human tutors. Since we can’t replay a human tutor, we generate a single learner turn at a time, using the real conversation up to that point. We repeat this at each learner turn in the original conversation. In both settings, we match the original number of learner turns.
Overall distribution matching
To produce a model that can teach many kinds of human learners, we need to train it against a learner model that can reproduce the distribution of human learning behavior. For this reason, we evaluate Persimmon’s ability to match the overall distribution of real learning behavior. We report results using the multi-user Turing test methodology we introduced in our previous blog post.
Open coding
For the evaluations below, we use open-coding methodology [18], automated with language models as in recent work [19], to build taxonomies of learning behavior across different dimensions.
Method How we build and validate the taxonomies
For each dataset and dimension, we sample 3,000 real learner turns uniformly from conversations that are not in the evaluation set, and hold out every turn from 20% of those conversations. A model (Gemini 3 Flash) reads the preceding conversation and describes, in a few words, what the learner is doing along that dimension. A second model (Claude Sonnet 4.6) groups these descriptions into a small set of mutually exclusive categories. For conversational moves, these describe actions such as answering incorrectly or tentatively, requesting an explanation, proposing a strategy, or challenging the tutor.
Three independent annotator models (Gemini 3 Flash, Claude Sonnet 4.6 and Claude Haiku 4.5) then label every sampled turn, and the two-of-three majority vote is the label. Categories with fewer than 3% of the majority labels are folded into “other”, repeated until none remain. We then refine for up to three rounds: categories the annotators often confuse with each other are merged or have their definitions sharpened, stopping once agreement on the building split reaches κ = 0.72. We report agreement on the held-out conversations, which played no part in building the taxonomy. Held-out Fleiss’ κ is shown below.
Does Persimmon talk the way real learners do?
Learners exhibit different conversational “moves” [20, 21, 22, 23, 24]. The moves learners make in a conversation, such as answering, guessing, explaining their reasoning, asking for help or disengaging, reflect how they are engaging with the material and are associated with how much they learn [25].
MovesExplore learner moves by dataset
Percentages show each move's share of the real learners' turns in the taxonomy-building sample. Categories are built per dataset, so names differ across rows.
Results. Persimmon's mix of conversational moves is the closest to real learners' on all datasets, at a distance of 0.06–0.24 from the real learners, against 0.10–0.65 for the other models.
Does Persimmon show emotions the way real learners do?
Real learners move between confusion, frustration, engagement and boredom [26]. We open-code each learner's turn for the affect the learner expresses, build a dataset-specific taxonomy of affective states, and compare the real and simulated distributions.
AffectExplore learner affect by dataset
Percentages show each category's share of the real learners' turns in the taxonomy-building sample. Categories are built per dataset, so names differ across rows.
Results. Persimmon also comes closest to the affect real learners express on all datasets, at a distance of 0.06–0.22, against 0.12–0.73 for the other models.
Does Persimmon make mistakes the way real learners do?
A realistic learner should be wrong about as often as real learners. We label every learner turn on all four datasets as a correct claim, an incorrect claim, or no checkable claim. Figure 6 compares each model's overall mix of these three labels with the real learners'.
Results. Persimmon's mix of correct, incorrect and non-claim turns is the closest to real learners' on all datasets, at a distance of 0.03–0.05, against 0.16–0.33 for the other models. Real learners are wrong on 25–30% of their checkable claims, Persimmon on 19–33%, and the other models on 15% or fewer.
Conclusion
We built &Tutor by training it to help learners understand, with Persimmon playing the learner. This idea extends beyond tutoring. Helping people understand is central to good communication and to giving them more agency. Both are at the heart of the AI we want to build. We’re excited about how far this can go. As our models of people improve, so will the AI we can train to help them.
References
- Albert Zhang, Alexis Ross, Kajal Patel, Julian Bernado, Rebecca Bowie, Ana Trindade Ribeiro, Daniel Halper, Haripriya Valayaputtur, Jacob Andreas, Susanna Loeb, Li Lucy, Kyle Lo, and Ryan Knight. 2026. When Help is Unhelpful: Evaluating AI Tutors for Productive Struggle.
- Chris Piech, Jonathan Bassen, Jonathan Huang, Surya Ganguli, Mehran Sahami, Leonidas J. Guibas, and Jascha Sohl-Dickstein. 2015. Deep Knowledge Tracing. In Advances in Neural Information Processing Systems 28.
- Alexis Ross and Jacob Andreas. 2024. Toward In-Context Teaching: Adapting Examples to Students' Misconceptions. In Proceedings of ACL 2024.
- Kurt VanLehn, Stellan Ohlsson, and Rod Nason. 1996. Applications of simulated students: An exploration. International Journal of Artificial Intelligence in Education 5, 135–175.
- Noboru Matsuda, Evelyn Yarzebinski, Victoria Keiser, Rohan Raizada, William W. Cohen, Gabriel J. Stylianides, and Kenneth R. Koedinger. 2013. Cognitive anatomy of tutor learning: Lessons learned with SimStudent. Journal of Educational Psychology 105(4), 1152–1163.
- Megha Srivastava and Noah Goodman. 2021. Question Generation for Adaptive Education. In Proceedings of ACL-IJCNLP 2021.
- Julia M. Markel, Steven G. Opferman, James A. Landay, and Chris Piech. 2023. GPTeach: Interactive TA Training with GPT-based Students. In Proceedings of Learning @ Scale 2023.
- Eric Zelikman, Wanjing Ma, Jasmine Tran, Diyi Yang, Jason Yeatman, and Nick Haber. 2023. Generating and Evaluating Tests for K-12 Students with Language Model Simulations: A Case Study on Sentence Reading Efficiency. In Proceedings of EMNLP 2023, 2190–2205.
- Alexis Ross and Jacob Andreas. 2025. Learning to Make MISTAKEs: Modeling Incorrect Student Thinking And Key Errors.
- Alexis Ross, Megha Srivastava, Jacob Blanchard, and Jacob Andreas. 2026. Modeling Student Learning with 3.8 Million Program Traces. In Proceedings of AIED 2026.
- Rose Niousha, Minwoo Kang, and Narges Norouzi. 2026. INSIDE the Student's Mind: Jointly Modeling Latent Reasoning and Action in LLM Student Simulators. In Proceedings of COLM 2026.
- Rose Niousha, Mihran Miroyan, Abigail O'Neill, Joseph E. Gonzalez, Gireeja Ranade, John DeNero, and Narges Norouzi. 2026. ParaStudent: Closing the Sim2Real Gap in User Simulators for AI Tutor Evaluation.
- David Dinucu-Jianu, Jakub Macina, Nico Daheim, Ido Hakimi, Iryna Gurevych, and Mrinmaya Sachan. 2025. From Problem-Solving to Teaching Problem-Solving: Aligning LLMs with Pedagogy using Reinforcement Learning. In Proceedings of EMNLP 2025, 272–292.
- Alexander Scarlatos, Naiming Liu, Jaewook Lee, Richard Baraniuk, and Andrew Lan. 2025. Training LLM-based Tutors to Improve Student Learning Outcomes in Dialogues.
- Yao Dou, Michel Galley, Baolin Peng, Chris Kedzie, Weixin Cai, Alan Ritter, Chris Quirk, Wei Xu, and Jianfeng Gao. 2025. SimulatorArena: Are User Simulators Reliable Proxies for Multi-Turn Evaluation of AI Assistants? In Proceedings of EMNLP 2025.
- Curtis Northcutt, Inaara Hasmani, Kevin Feng, Trevor Khangi, Andreas Plesner, and Jonas Mueller. 2026. StudentBench: AI and Human Tutoring Yield Equivalent GRE Learning Gains.
- Matthew Zent, Digory Smith, and Simon Woodhead. 2025. PIIvot: A Lightweight NLP Anonymization Framework for Question-Anchored Tutoring Dialogues. In Proceedings of EMNLP 2025.
- Anselm Strauss and Juliet Corbin. 1990. Basics of Qualitative Research: Grounded Theory Procedures and Techniques. Sage Publications.
- Michelle S. Lam, Janice Teoh, James Landay, Jeffrey Heer, and Michael S. Bernstein. 2024. Concept Induction: Analyzing Unstructured Text with High-Level Concepts Using LLooM. In Proceedings of CHI 2024.
- John McH. Sinclair and R. Malcolm Coulthard. 1975. Towards an Analysis of Discourse: The English Used by Teachers and Pupils. Oxford University Press.
- Arthur C. Graesser and Natalie K. Person. 1994. Question asking during tutoring. American Educational Research Journal 31(1), 104–137.
- Abhijit Suresh, Jennifer Jacobs, Charis Harty, Margaret Perkoff, James H. Martin, and Tamara Sumner. 2022. The TalkMoves Dataset: K-12 Mathematics Lesson Transcripts Annotated for Teacher and Student Discursive Moves. In Proceedings of LREC 2022.
- Dorottya Demszky, Jing Liu, Zid Mancenido, Julie Cohen, Heather Hill, Dan Jurafsky, and Tatsunori Hashimoto. 2021. Measuring Conversational Uptake: A Case Study on Student-Teacher Interactions. In Proceedings of ACL-IJCNLP 2021, 1638–1653.
- Dorottya Demszky and Heather Hill. 2023. The NCTE Transcripts: A Dataset of Elementary Math Classroom Transcripts. In Proceedings of the 18th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2023), 528–538.
- Michelene T. H. Chi and Ruth Wylie. 2014. The ICAP framework: Linking cognitive engagement to active learning outcomes. Educational Psychologist 49(4), 219–243.
- Sidney D’Mello and Art Graesser. 2012. Dynamics of affective states during complex learning. Learning and Instruction 22(2), 145–157.