20 Nov 2026
AI tutors matched expert humans, with a teacher checking every message. Plus: AI coaching for tutors, and this week's open models
LDPro teamSTAFF
Welcome
Welcome to edition 2 of the LDPro Newsletter. At LDPro we think learning should be AI-driven, procedural, and always evolving, with people still in charge of quality. This week's research makes that last part concrete: what human oversight of AI tutoring actually looks like, and what it costs. Every claim below links to its source, so you can check the detail yourself.
Research: AI tutoring and feedback
Supervised AI tutors matched expert human tutors in a 10-school trial
Google and Eedi published a technical report (late Oct 2026) on a preregistered randomised trial in 10 UK schools. Over 10 weeks, 1,179 students were randomly assigned to standard remediation, expert human tutors, or one of two Gemini-based tutor arms. An educator reviewed every AI draft before a student saw it.
The AI arms matched expert tutors on mistake correction and near transfer (about 57%, vs about 43% under standard remediation), and did slightly better on a mid-trial quiz.
On the externally scored Star Maths test, no arm, human or AI, beat standard remediation.
Tutors approved 94.7% of AI drafts unchanged. But when they did edit, they rewrote a median 69% of the text.
A "personalised context" arm added almost nothing, partly because a platform bug left out earlier chat summaries.
Tutoring conversations also slowed coverage: students got through 99 units with AI tutoring vs 132 under standard remediation.
Why it matters: human review isn't a temporary scaffold. Most drafts pass, but the ones that don't need heavy rewriting, so supervision is an ongoing cost to plan for. Tutors reported it saved time but was harder to manage. And richer tutoring can mean less content covered, which is a real pacing trade-off.
Caveats: the report was written by the developers (Google and Eedi) and hasn't been peer reviewed. It covers 10 weeks of K-12 maths, not adult or workplace learning. Some analyses (gender and coverage) were exploratory. There's a readable summary of the trial here.
Try this: measure AI learning at two levels, the immediate task and a broader outcome measured later. And treat reviewer edit rates and reviewer workload as product metrics, not afterthoughts.
Automated feedback changed how tutors teach, but not how much learners talked
A new preprint (Nov 2026) reports a randomised trial with Stanford and Schoolhouse.world covering 1,030 volunteer tutors, 7,443 learners and 6,508 sessions. Automated feedback on tutors' discourse increased how often they drew out learner ideas, built on learner contributions, and prompted learner reasoning, by 8–15%. The effect replicated across two bootcamps. Tutor talk time (91.5% of the session) didn't change, and feedback sent to learners added nothing.
Why it matters: it's evidence that automated, AI-generated coaching can shift facilitator behaviour at scale, which is directly relevant to live and cohort-based programmes. It also suggests that group-level statistics shown to learners don't change what they do.
Caveats: this is a preprint and hasn't been peer reviewed. It's SAT maths with volunteer tutors, and the effects are on discourse measures, not learning outcomes.
Try this: if you coach facilitators, give each one individual feedback on specific talk moves rather than relying on group-level dashboards.
Models
Qwen4: a hosted preview only, no open weights yet
According to secondary reporting on the announcement, Qwen announced Qwen4-Max on 11 Nov as a preview on its Token Plan subscription and in Qwen apps, not on the pay-as-you-go API, running until 25 Nov. Coverage reports a 1M-token context, but there's no model card, licence or per-token price yet. As of 13 Nov, Qwen's Hugging Face organisation shows no Qwen4 repository. The open-weight Qwen4-27B tier still has no release date.
Why it matters: there's nothing to fine-tune yet. Max-tier Qwen models have so far been API-only, so for teams building their own learning models, the 27B is the one to watch. Until then, Qwen3.8-27B, covered in edition 1, remains a sensible open baseline.
Thinking Machines releases Inkling 2 Small under Apache 2.0
Thinking Machines released Inkling 2 Small on 12 Nov 2026, with the repo tagged Apache 2.0 (announcement). It's a 594B-parameter mixture-of-experts model with 17B active parameters, and it accepts image input. It supports vLLM, SGLang and Unsloth, and can be fine-tuned through Thinking Machines' hosted Tinker service. Tinker's current list price is $0.30 prefill, $1.00 sampling and $1.50 training per million tokens at 64K context.
Why it matters: a genuinely permissive licence on a strong model, plus hosted fine-tuning, means you can run supervised fine-tuning experiments without renting GPUs.
Caveats: it's far too large to self-host on a single GPU. The benchmarks are the vendor's own and focus on coding and agents, and Thinking Machines itself says it's "less optimized for the chat experience" and tends to over-react to safety-related queries. Check the pricing page before you spend.
Try this: if you want a second baseline, run a small instructional-design evaluation through a hosted fine-tuning service and compare cost and quality against your current model.
The takeaway for L&D teams
This week's evidence sharpens the picture from edition 1: AI tutoring can match skilled humans, but only with people reviewing the work, and that review is a lasting cost worth measuring. The same AI can also coach the humans, one facilitator at a time. That's the kind of always-evolving, human-overseen learning we're building at LDPro.
Sources are linked inline. Preprints, developer-written reports and vendor-reported figures are labelled as such. The LDPro Newsletter is a digest, not an endorsement of any model or vendor.
Questions about this?
Comments are off while the community is read-only. Ask us and we reply within 48 hours.
Ask us