DeepTutor: Towards Agentic Personalized Tutoring
Bingxi Zhao, Jiahao Zhang, Xubin Ren, Zirui Guo, Tianzhe Chu, Yi Ma, Chao Huang
cs.CY, cs.AI, cs.CL
2026-04-11
DeepTutor ties tutoring and question generation through a shared, evidence-backed learner model, scoring 10.8% above a naive tutor on the self-built TutorBench.
Today's LLM tutors lean on static pretrained knowledge and do not adapt to the individual; tutoring and question generation are built separately, with no shared learner model in between. The authors argue both flaws share a root: the system lacks a fine-grained, evolving model of the learner, keeping at best a coarse skill inventory that never shows where a student's reasoning actually breaks. Evaluation has a blind spot too, since most educational benchmarks take an instructor's view and treat students as generic receivers. DeepTutor's answer is a closed loop: weaknesses surfaced during tutoring directly shape the next question, and performance on those questions in turn refines the learner model.
The core is two pipelines sharing one context.
Static Knowledge Grounding (SKG) decomposes source documents into atomic content units, indexed through two structures, a knowledge graph and a dense embedding index, fused by reciprocal rank fusion into the retrieval corpus.
Dynamic Personal Memory (DPM) centers on a Trace Forest, where each tree records one complete tutoring interaction across three node levels: session-level input and global summary, intermediate planning units from task decomposition, and fine-grained execution records including tool outputs. It is exposed through TraceToolkit (SearchTrace, ListTraces, ReadNodes). Three memory agents build three profiles from it: session history, an evidence-backed inventory of recurring confusions and gaps, and pedagogical reflections guiding future interactions.
The tutoring pipeline has three stages: personalized investigation (the planner decomposes the query into meta-questions and gathers evidence across the knowledge base and trace forest), step-by-step guided solving (each sub-goal uses the shared tool suite, self-notes distill takeaways, completed sub-goals are hierarchically compressed, and inadequate ones trigger replanning), and evidence-driven iterative writing (every factual claim carries a traceable citation).
The question-generation pipeline has two stages: personalized idea generation (scanning the conceptual landscape from the learner's view to produce candidate ideas, filtered by an evaluator into templates) and critic-guided generation (each template yields a question/answer/explanation triple, validated by LLM checks plus sandboxed code execution). The key design is that the validator shares no reasoning chain with the generator and must independently verify correctness, cutting self-confirming errors. The loop property: after each interaction a new trace enters the forest and the three memory agents update profiles in parallel.
The authors built TutorBench: 30 knowledge bases across five disciplines, three learner levels each (beginner/intermediate/advanced), giving 90 profiles, and 270 interactive tasks after rejection sampling. Gemini-3-Flash serves as simulator and tutor; Claude Sonnet 4.6 is the judge.
| System | Overall quality |
| Naive tutor | 3.53 |
| CoT tutor | 3.52 |
| ReAct tutor | 3.52 |
| Self-refine tutor | 3.57 |
| DeepTutor | 3.91 (+10.76%) |
Overall quality varies by only 0.16 points across the five disciplines, so the gain is not propped up by one field. On a 45-session human-preference study, human and LLM win rates correlate strongly across ten metric-level pairs (Pearson r=0.82). Ablations show removing SKG mainly hurts grounding while removing DPM mainly hurts personalization: SKG anchors "what to say" and DPM decides "how to adapt to the person".
Pulled out on its own (personalization off) and run on five public reasoning benchmarks, the solver pipeline lifts five backbone families by 25.7% to 32.0% on average: Gemini-3-Flash on GAIA L1 goes 43.4 to 56.6 and LiveBench 70 to 96.
For people building educational AI, two things are worth borrowing. First, the closed loop: do not let tutoring and question generation work in isolation; tie them with a shared, evidence-backed learner memory so weaknesses feed directly into question design. Second, the validator/generator separation is useful in any correctness-sensitive setting, especially questions with computation. The large gains the solver pipeline posts on general agent benchmarks suggest this multi-stage design has transfer value beyond tutoring as a general agentic scaffold.
The authors' own caveats: the interactive evaluation relies on LLM student simulators and rubric judges, so it inherits the gap between controlled simulation and real learner behavior, and the multi-stage pipeline trades extra inference cost for controllability and personalization. A few more: TutorBench's 270 tasks are small, and +10.76% is a small absolute move (about 0.38 on a 1-5 scale) with no significance test given; an overall quality of 3.91/5 is still some way from "a good tutor". The solver pipeline's big gains on general benchmarks come in a setting where every backbone's base score is low, and there is no comparison against agent frameworks built specifically for those benchmarks, so take them with a discount.