Ampersand Releases &Tutor, a Tutoring Model Trained on Persimmon Learner Simulations

• Ampersand released &Tutor, an educational model available via API, which achieved state-of-the-art performance on the TutorMoments benchmark.
• &Tutor scored 0.754 on plain prompts, surpassing GLM-5.3 (0.500), GPT-6 Astra (0.634), and Claude Sonnet 5.5 (0.667).
• The model was trained using Persimmon as a simulated learner, a method that implicitly derived tutoring skills like scaffolding without direct instruction.
• Persimmon demonstrates high fidelity to human learners, with Jensen-Shannon distances of 0.06–0.24 for conversational moves and 0.03–0.05 for error rates, significantly lower than other models.

Persimmon Learner Fidelity

• Persimmon models human learner behavior more accurately than other models across four datasets.
• In Turing test methodologies, Persimmon generated conversations judged as human at rates of 20.2% (SimulatorArena Math) and 19.0% (StudentBench).
• Persimmon mimics human error frequency, being wrong on 19–33% of checkable claims compared to 25–30% for real learners and 15% or fewer for other models.

Training Methodology

• The training environment rewards the tutor based on the learner's progress, guiding behavior through the simulated student's responses.
• Ablation tests show &Tutor performance drops significantly when using GLM-5.3 instead of Persimmon as the learner (0.635 vs 0.754 at max reasoning effort).
• The approach is not derived from direct Ampersand model training but is planned for integration into future Ampersand learning processes.