Model Summary
| Field | Description |
|---|---|
| Model name | &Tutor |
| Model type | Tutor model |
| Developer | humans& |
| Base model | GLM 5.3 |
| Architecture | Mixture-of-experts transformer: 78 layers, 256 routed experts plus 1 shared expert, 8 routed experts active per token |
| Weight format | NVFP4 |
| Context length | 131,072 tokens in evaluations |
| Modalities | Text input, text output |
| Version & API identifier | v0.1 · ampersand-tutor-v0.1 |
Overview
&Tutor is a tutor model trained to maximize a student’s understanding. It learns from interactions with Persimmon acting as a simulated student. See the blog for the training approach and teaching evaluations.
Intended Uses
For educational applications that help learners reason through problems, identify misconceptions and build understanding. &Tutor supports learners and educators; consequential educational decisions require human judgment.
Inputs & Outputs
Input: text conversation history and tutoring instructions. Output: tutor replies. Evaluations use native GLM chat formatting without tools or the Ampersand harness.
Training
&Tutor is built on GLM 5.3 and trained with reinforcement learning in simulated tutoring conversations. Persimmon plays the learner, and a judge evaluates a student’s learning.
Evaluations
TutorMoments
TutorMoments tests whether a tutor provides appropriate support and productive challenge. The tutor continues selected moments from real tutoring sessions with a simulated student, and an LLM judge scores its behavior against expert-teacher annotations.
Plain prompt
| Dimension | &Tutor · max reasoning | &Tutor · low reasoning |
|---|---|---|
| Appropriate scaffolding | 93.08% (242/260) | 84.23% (219/260) |
| Appropriate rigor | 57.69% (150/260) | 47.69% (124/260) |
| Avoids over-scaffolding | 76.73% (399/520) | 67.31% (350/520) |
Both settings scored all 520 cases using a local five-turn protocol, temperature 0 and a 57,344-token generation limit.
Safety & Security
Safety evaluations cover hazardous knowledge, prompt-injection resistance and responses to harmful requests. Across these evaluations, 7,213 of 7,303 cases were scored; 90 remain unscored.
| Evaluation | Metric | GLM-5.3 | &Tutor | &Tutor scored / total |
|---|---|---|---|---|
| WMDP (bio) | Accuracy | 87.9% | 88.0% | 1,273 / 1,273 |
| WMDP (chemistry) | Accuracy | 81.4% | 82.4% | 408 / 408 |
| WMDP (cyber) | Accuracy | 80.6% | 79.0% | 1,954 / 1,987 |
| LAB-Bench (public) | Accuracy | 53.6% | 41.3% | 1,951 / 1,967 |
| SecFid | Injection execution rate ↓ | 0.668 | 0.657 | 1,168 / 1,168 |
| FORTRESS (adversarial) | Attack risk score ↓ | 21.4 | 19.0 | 459 / 500 |
↓ Lower is better.
Limitations
Results with simulated learners and automated judges may not generalize to human learners, other subjects or longer interactions. &Tutor may give incorrect explanations, miss misconceptions or provide too much or too little help. Short tutoring evaluations do not establish lasting learning gains. Hazardous-knowledge accuracy measures capabilities, not refusal behavior or deployment safety.