--- type: "post" id: "humans-releases-tutor-a-tutoring-model-trained-against-02a024" slug: "humans-releases-tutor-a-tutoring-model-trained-against-02a024" title: "humans& releases &Tutor, a tutoring model trained against its Persimmon learner model, via API" description: "humans& has released Ampersand Tutor (&Tutor), a tutoring model available through an API." retrieval_nugget: "humans& has released Ampersand Tutor (&Tutor), a tutoring model available through an API." published_at: "2026-10-30" updated_at: "2026-10-30" record_date: "2026-10-30" date_kind: "published_at" topics: [] entities: ["humans&","Ampersand Tutor","Persimmon","Ampersand","GLM-5.3","TutorMoments"] related_patterns: [] novelty_label: "new_fact" what_changed: "Humans& has released a separate &Tutor model via API, trained against Persimmon as a learner to teach and reporting state-of-the-art TutorMoments results." prior_refs: ["memory:eab3c475-f148-826f-a172-dd93250e3cb4","memory:08169b2c-fe10-867d-a3ef-e76a3373ee49"] source_url: "https://tutor.humansand.ai/blog" source_urls: ["https://tutor.humansand.ai/blog"] schema_version: "newruntime-agent-readable-v0.2" stable_id: "post:humans-releases-tutor-a-tutoring-model-trained-against-02a024" status: "published" visuals: [] editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]} routes: {"html":"https://newruntime.com/posts/humans-releases-tutor-a-tutoring-model-trained-against-02a024/","markdown":"https://newruntime.com/posts/humans-releases-tutor-a-tutoring-model-trained-against-02a024.md","json":"https://newruntime.com/posts/humans-releases-tutor-a-tutoring-model-trained-against-02a024.json"} --- # humans& releases &Tutor, a tutoring model trained against its Persimmon learner model, via API ## Retrieval answer humans& has released Ampersand Tutor (&Tutor), a tutoring model available through an API. ## Summary humans& has released Ampersand Tutor (&Tutor), a tutoring model available through an API. Its central claim is not simply a higher tutoring benchmark score: the model was trained against Persimmon, a learned model of a student, with reward tied to student learning. For AI builders, it is a concrete example of optimizing an interaction agent against a user simulator rather than only prompting a general model. ## What happened humans& says it has released Ampersand Tutor, written as &Tutor, as a model trained to teach and available to the public through an API. The company frames teaching as a setting where a system must infer what a learner already knows and choose whether to support, challenge, or wait. Its stated training objective is how much a student learns, rather than direct reward for a list of individual tutoring behaviours. The training environment used Persimmon, humans&' large-scale user model, in the role of learner. &Tutor was trained from GLM-5.3, but humans& says it was not trained from Ampersand; the company plans to integrate it into future Ampersand training. The distinction matters because the release is presenting a specialized tutor as a separate result of a particular training environment, not as a capability already inherited by its collaborator model. The company evaluates the model on TutorMoments, a benchmark assembled from real tutoring sessions with real students. Experienced mathematics teachers identify points in those sessions at which a tutor must decide between offering support and asking the learner to reason further, and annotate what the moment calls for. A model continues the conversation, with an AI model playing the student, and an LLM judge scores the response against the teacher annotations. The evaluation is run both with a plain prompt asking a model to tutor well and with an evaluation-aware prompt that describes the desired balance of support and rigour. Under the plain prompt, humans& reports an overall TutorMoments score of 0.754 for &Tutor, compared with 0.500 for the GLM-5.3 model from which it was trained. The post lists 0.667 for Claude Sonnet 5.5 and 0.634 for GPT-6 Astra in the same condition. With the evaluation-aware prompt, &Tutor scores 0.919, while GPT-6 Astra scores 0.926. humans& also reports that &Tutor has the highest evaluation-aware score for appropriate rigour, at 0.931. The company says an ablation replacing Persimmon with GLM-5.3 as the learner performed materially worse, and attributes the difference to Persimmon matching learner moves, affect, and error rates more closely. These are vendor-reported results, including the account of the ablation. ## What the archive adds This is a new fact on a known humans& story. The earlier note on [Ampersand v0.1](https://newruntime.com/posts/humans-released-ampersand-v0-1-a-collaborator-model-2e438d/) described a collaborator model trained on long-horizon, multi-user simulations. &Tutor applies a related simulation idea to a narrower interaction: a tutor is trained against a modeled learner, with learning as the reward signal. ## Trend signal No supplied pattern genuinely fits this release closely enough. It is evidence about behaviour-specific training and user simulation, but the available pattern set does not make a specific claim about either tutoring agents or simulated-user training. ## Connections to other events The release connects directly to the earlier Ampersand v0.1 note because humans& says &Tutor is not trained from Ampersand but may be incorporated into future Ampersand training. The relationship is therefore prospective rather than evidence that Ampersand already has &Tutor's reported teaching performance. ## Why it matters Most teams building instructional, onboarding, support, or guided-workflow agents start with a capable general model and a well-written prompt. &Tutor makes a narrower engineering proposition: train for a user outcome inside an environment that can represent a user's evolving state. If that environment is realistic enough, it can reward decisions that are difficult to enumerate in a rubric, such as when to give a hint, when to ask for work, and when not to over-explain. The important implication is conditional. A simulated user can make specialized training more effective only if it preserves the behaviours that matter in real use. humans&' own ablation makes that simulator quality the pivotal component of its argument. Teams should treat the user model, its data and its evaluation as first-class parts of the product system, not as invisible scaffolding around a model fine-tune. ## What happens next - **Specialized interaction agents will be evaluated against held-out human outcomes, not only simulator-based scores.** (6 months, 55%). Watch for: a third-party study or a deployment report measuring learner progress against a comparison group. - **humans& will test whether the &Tutor training setup transfers into future Ampersand training.** (12 months, 45%). Watch for: an Ampersand release that identifies tutor-style or user-simulation training as a component. ## Limits and uncertainty The release does not establish that API users obtain better learning outcomes with real students. TutorMoments begins with real-session moments and teacher annotations, but the evaluated continuation uses an AI student and an LLM judge. The published comparisons are supplied by humans&, and the material does not provide independent replication, API pricing, rate limits, supported subjects, deployment details, or performance measurements from production learners. The evaluation-aware condition also changes the prompt for every model, making it a useful diagnostic but not a direct substitute for ordinary use. The Persimmon ablation supports the company's mechanism within its reported setup; it does not yet show that a learned user model will generalize across products or populations. ## Skill to build now For any agent meant to change a user's state, write down one observable outcome before tuning prompts: a concept correctly applied later, an onboarding task completed without help, or an error avoided on a new case. Then build a small test set of interaction moments where the system must choose between giving the answer and eliciting it. Review both the final answer and the user's next action. This will reveal whether the team is optimizing fluent assistance or actual progress. ## Reader takeaway &Tutor is a public API release, but its more useful signal is architectural: specialized interaction agents may be trained around a modeled user and a measurable user outcome. The reported benchmark lead is preliminary. The design question worth carrying forward is whether your agent's training loop contains a credible representation of the person whose progress it is meant to improve. ## Open question Can Persimmon's claimed advantage over a generic learner model survive evaluation with real students, across subjects and over repeated sessions? Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json ## Continue in the graph Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`. - collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis. - collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph. - whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)