GenMentor is the official implementation of LLM-powered Multi-agent Framework for Goal-oriented Learning in Intelligent Tutoring System (WWW 2025, Industry Track, oral). You tell it where you want to be; a set of coordinated agents refines the goal, finds the gap between it and your background, schedules a path of sessions, writes each session's reading and quiz for you, and rebuilds your learner profile from what you actually did.
This repository is a single Next.js application. There is no account and no server-side database: the learner's archive lives in the browser and can be exported and imported as one JSON file. Agent calls run inside the app's API routes against any OpenAI-compatible endpoint or the Anthropic API.
| Paradigm | Typical characteristics | Primary focus |
|---|---|---|
| Traditional MOOC | Static syllabus; pre-recorded lectures; fragmented learning | Broad access, low personalization |
| Chatbot ITS | Reactive Q&A; rule/LLM-driven; session-based help | Instant support, limited long-term adaptation |
| Goal-oriented ITS | Proactive planning; personalized paths; goal-aligned assessments | Targeted skill acquisition, continual adaptation |
| Agent | Route | What it does |
|---|---|---|
| Goal Refiner | POST /api/refine-goal |
Turns a vague goal into an actionable one |
| Skill Mapper + Skill Gap Identifier | POST /api/identify-skill-gap |
Maps the goal to skills, infers the learner's level for each |
| Adaptive Learner Profiler | POST /api/profile |
Builds the profile, then rebuilds it after every session with quiz evidence |
| Learning Path Scheduler | POST /api/schedule-path |
Creates, refines and reschedules the session sequence (streamed) |
| Knowledge Explorer | POST /api/explore-knowledge |
Picks the knowledge points a session needs |
| Knowledge Drafter | POST /api/draft-knowledge |
Drafts each point, optionally grounded in web search (streamed) |
| Document Integrator | POST /api/integrate-document |
Synthesizes the drafts into one document (streamed) |
| Quiz Generator | POST /api/generate-quiz |
Writes questions from the document |
| Learner Feedback Simulator | POST /api/simulate-feedback |
Role-plays the learner to critique a path or content |
| AI Chatbot Tutor | POST /api/tutor |
Streams grounded answers during a session |
The prompts are the paper's prompts. The verbatim port is preserved at git tag
prompts-baseline; later edits are evaluated side by side against it
(scripts/eval-prompts.sh, wiki/reviews/).
Requirements: Node 24, pnpm 11, and one LLM credential.
make install # pnpm install in app/
cp app/.env.example app/.env.local
# set LLM_API_KEY and LLM_BASE_URL (or OPENAI_API_KEY / ANTHROPIC_API_KEY)
make dev # http://localhost:3000Learners can also bring their own key from the app (the key icon in the sidebar): it stays in their browser and is sent with each request, so a public deployment can run without a server key.
LLM_FAST_MODEL and LLM_SMART_MODEL split the work between a cheap model (tutor, quizzes,
goal refinement) and a stronger one (skill gap, scheduling, content). Defaults and the other
knobs are documented in app/.env.example.
No key at hand? GENMENTOR_LLM_MODE=replay make dev serves the recorded journeys in
e2e/fixtures/llm for the sample goal used by the tests.
| Target | How |
|---|---|
| Vercel | The public instance at genmentor.aurax.live ships without a server key: visitors add their own provider, key and models in Model settings. To run your own, import the repo with Root Directory app and optionally add the environment variables |
| Container | docker build -t genmentor app && docker run -p 3000:3000 --env-file app/.env.local genmentor |
app/ the Next.js application (routes, agents, prompts, schemas, UI)
e2e/ Playwright journeys, axe checks, screenshot evidence, recorded LLM fixtures
wiki/ product proposal, roadmap, architecture, design principles, API reference, pitfalls
scripts/ gate.sh and verify-ui.sh (acceptance gates), eval-prompts.sh (prompt evaluation)
docs/rebuild/ the plan this rebuild followed
make gate runs the static acceptance gate (layout, docs, size budgets, banned patterns,
types, lint, unit tests, build). make verify-ui builds the app, drives every journey in the
browser, checks accessibility in both themes and screenshots every page in three viewports.
AGENTS.md holds the constraints every contributor, human or agent, works under.
The Python implementation used for the paper (FastAPI backend, Streamlit frontend, LangChain
agents) lives on the v0.1-demo branch, also tagged v0.1.0 and paper-python-v1:
git checkout v0.1-demoIts README describes how to run it. That branch is frozen as the demo version; all further work
happens on main, which keeps the same nine agents and the same prompts (see prompts-baseline),
rewritten in TypeScript.
@inproceedings{wang2025llm,
title={LLM-powered Multi-agent Framework for Goal-oriented Learning in Intelligent Tutoring System},
author={Wang, Tianfu and Zhan, Yi and Lian, Jianxun and Hu, Zhengyu and Yuan, Nicholas Jing and Zhang, Qi and Xie, Xing and Xiong, Hui},
booktitle={Companion Proceedings of the ACM Web Conference},
year={2025}
}See LICENSE.