Research map / Large language models / Agents, code and math
TRACE-CS: A Synergistic Approach to Explainable Course Scheduling Using LLMs and Logic
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Presents TRACE-CS, a hybrid system where an LLM (GPT-4) parses contrastive user queries into logic and refines SAT-based explanations into natural language for course scheduling.
Problem
Explaining contrastive scheduling decisions (e.g., "why course X instead of Y?") in a way that is both logically correct and understandable in natural language.
Approach
Combine a SAT-based symbolic contrastive explainer (encoding constraints with labels) with an LLM that parses user queries into logical clauses (with user verification) and rewrites symbolic explanations into natural language, so correctness comes from the solver rather than the LLM.
Key results
- TRACE-CS achieved 100% explanation correctness on 100 schedule-query pairs.
- Zero-shot GPT-4 achieved 44% and few-shot GPT-4 achieved 49% explanation correctness.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TRACE-CS's symbolic explainer and LLM interface instead of a zero-shot LLM-only approach | GPT-4 course-scheduling explanations · Course scheduling (100 schedule-query pairs) | Explanation correctness: 44% -> 100% | Helped |
| Use TRACE-CS's symbolic explainer and LLM interface instead of a few-shot LLM-only approach | GPT-4 course-scheduling explanations · Course scheduling (100 schedule-query pairs) | Explanation correctness: 49% -> 100% | Helped |
Limitations
- Evaluation is a small proof-of-concept on one undergraduate CS program with manual correctness checking by the authors.
- LLM query parsing may be error-prone, requiring user verification of extracted queries.
Experimental setup
Datasets: university computer science course catalog and degree requirements scraped from official website
Benchmarks: custom course scheduling queries (100 schedule-query pairs)
Metrics: explanation correctness
Baselines: zero-shot GPT-4 few-shot GPT-4
Builds on: Vasileiou, Previti, and Yeoh 2021 contrastive explainer
Closest papers by meaning
- Large Language Models Are Neurosymbolic Reasoners (AAAI 2024 · similarity 0.86)
- ZebraLogic: On the Scaling Limits of LLMs for Logical Reasoning (ICML 2025 · similarity 0.85)
- Scheduling Your LLM Reinforcement Learning with Reasoning Trees (ICLR 2026 · similarity 0.84)
- Enhancing Reasoning Capabilities of LLMs via Principled Synthetic Logic Corpus (NeurIPS 2024 · similarity 0.84)
- AbstRaL: Augmenting LLMs' Reasoning by Reinforcing Abstract Thinking (ICLR 2026 · similarity 0.84)
- Logic.py: Bridging the Gap between LLMs and Constraint Solvers (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.