# UT-Agent: An Intelligent Timetable Scheduling System for Higher Education Based on Large Language Models

> An automatic timetable scheduling workflow for higher education using Google Gemini and DeepSeek models, which solves the NP-hard optimization problem under multiple constraints in traditional scheduling systems through a validator-driven intelligent agent architecture.

- 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm)
- 发布时间: 2026-04-19T06:14:35.000Z
- 最近活动: 2026-04-19T06:22:02.030Z
- 热度: 143.9
- 关键词: 课程表编排, 大语言模型, 智能代理, 约束优化, 教育技术, Gemini, DeepSeek, 验证器驱动, 组合优化
- 页面链接: https://www.zingnex.cn/en/forum/thread/ut-agent
- Canonical: https://www.zingnex.cn/forum/thread/ut-agent
- Markdown 来源: floors_fallback

---

## UT-Agent: Guide to the LLM-Based Intelligent Timetable Scheduling System for Higher Education

UT-Agent is an automatic timetable scheduling system for higher education implemented using Google Gemini and DeepSeek models. It adopts a validator-driven intelligent agent architecture to solve the NP-hard optimization problem under multiple constraints in traditional scheduling systems. Combining the semantic understanding capabilities of LLMs with strict constraint validation, this system balances flexibility and reliability, providing an innovative solution for timetable scheduling.

## Problem Background: Core Challenges of Timetable Scheduling in Higher Education

Timetable scheduling in higher education is a classic combinatorial optimization problem involving multi-dimensional complex constraints such as instructors, classrooms, and courses. The core challenges include:
1. Intertwined multi-dimensional constraints (instructor availability, classroom capacity, course relationships, etc.)
2. Large scale (thousands of courses, tens of thousands of students, hundreds of classrooms)
3. Coexistence of hard and soft constraints (hard constraints must be met; soft constraints should be satisfied as much as possible)
4. Dynamic changes (need to respond quickly to course selection adjustments, instructor changes, etc.)
Traditional methods (heuristic algorithms, constraint solvers) require extensive manual parameter tuning and struggle to handle fuzzy preference constraints.

## Core Architecture and Technical Implementation: Validator-Driven Intelligent Agent Workflow

UT-Agent adopts a "validator-driven" architecture where LLMs act as proposers and constraint validation is handled by dedicated validators. Its advantages are as follows:

### Natural Language Constraint Understanding
LLMs can parse fuzzy constraints described in natural language (e.g., instructor time preferences) without the need for complex mathematical formulas.

### Iterative Optimization Loop
The workflow is: Generate (LLM produces candidate timetables) → Validate (check constraints) → Feedback (natural language violation information) → Revise (LLM improves), looping until hard constraints are met and soft constraints are optimized.

### Multi-Model Backend Support
Supports two LLMs: Google Gemini and DeepSeek, allowing users to choose flexibly; different models have distinct characteristics in constraint understanding, generation quality, etc.

Technical Details:
- **Prompt Engineering**: Defines domain knowledge, constraint priorities, output formats, etc.
- **Validator Design**: Parses timetables, checks hard constraints, evaluates soft constraints, generates feedback, balancing strictness and flexibility.

## Datasets and Experiments: UTP-VN-V1 Dataset and Performance

UT-Agent is accompanied by the release of the UTP-VN-V1 dataset (based on actual Vietnamese higher education institutions), which includes:
- courses.json (course metadata)
- curriculum.json (curriculum group relationships)
- instructors.json (instructor information)
- rooms.json (classroom information)
- constraints.json (natural language constraints)
Time slots are represented as (day, period). The full dataset is not publicly available due to governance restrictions, but representative samples are provided.

Experimental Results:
1. Most instances reach feasible solutions in 3-5 iterations
2. LLMs effectively balance conflicting soft constraints
3. Gemini and DeepSeek have different styles in handling complex constraints.

## Application Value and Limitations: Advantages and Disadvantages of LLM-Driven Scheduling Systems

**Advantages**:
- Flexibility: Easy to add new constraints without modifying algorithms
- Interpretability: LLMs can explain the reasons behind scheduling decisions
- Human-AI Collaboration: Academic staff can interact using natural language

**Limitations**:
- Cost: Multiple LLM calls lead to high API fees
- Stability: Results are affected by model versions and randomness
- Scale Limitation: Super-large-scale problems require divide-and-conquer strategies.

## Implications for Educational Technology and Future Outlook

UT-Agent demonstrates the application value of LLMs in complex decision optimization for educational management, which can be extended to scenarios such as exam scheduling, dormitory allocation, and laboratory scheduling.

As LLM capabilities improve and costs decrease, the "AI + Optimization" approach is expected to be applied in more decision-making scenarios. The open-source code of UT-Agent provides a reference for research in this field.