Research map / Multi-agent systems and games / LLM agents
MASTER: Enhancing Large Language Model via Multi-Agent Simulated Teaching
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.
Proposes MASTER, a multi-agent simulated teaching method (error correction, debate, analogical reasoning) that augments instruction data into BOOST-QA, which improves fine-tuned Qwen/LLaMA/Mistral 7B models on math, coding, and multiple-choice benchmarks.
Problem
Obtaining high-quality instruction fine-tuning data for LLMs is costly, and existing synthetic data methods rely on handcrafted prompts and lack realistic interaction mechanisms.
Approach
Simulate teacher-student classroom interactions with multiple LLM agents under three pedagogical scenarios (error correction, debate, analogical retrieval) to turn existing QA data into multi-turn ShareGPT-format training data.
Key results
- LLaMA3-8B average across benchmarks rises from 45.98 (Ori) to 51.26 (BOOST-QA) and Qwen2.5-7B from 37.27 to 60.64.
- HumanEval for Qwen2.5-7B improves from 22.56 to 42.07 with BOOST-QA.
- Ablation: using only one augmentation scenario underperforms the original data on all test sets; full three-scenario MASTER gives average 43.45 vs 37.52 for Ori on the ablation subset.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply MASTER augmentation instead of Ori-Data | LLaMA3-8B · MATH | accuracy: 21.58 -> 23.90 | Helped |
| Apply MASTER augmentation instead of Ori-Data | LLaMA3-8B · MMLU-PRO-MATH | accuracy: 13.55 -> 27.39 | Helped |
| Apply MASTER augmentation instead of Ori-Data | LLaMA3-8B · MBPP | Pass@1: 65.30 -> 67.20 | Helped |
| Apply MASTER augmentation instead of Ori-Data | LLaMA3-8B · HumanEval | Pass@1: 39.02 -> 50.61 | Helped |
| Apply MASTER augmentation instead of Ori-Data | LLaMA3-8B · MMLU | accuracy: 48.13 -> 48.13 | No clear effect |
| Apply MASTER augmentation instead of Ori-Data | LLaMA3-8B · ARC | accuracy: 57.76 -> 61.52 | Helped |
| Apply MASTER augmentation instead of Ori-Data | LLaMA3-8B · SCI-Q | accuracy: 76.50 -> 80.10 | Helped |
| Apply MASTER augmentation instead of Ori-Data | LLaMA3-8B · Average | score: 45.98 -> 51.26 | Helped |
| Use all three teaching scenarios rather than only Make Error and Debate (ablation) | LLaMA3-8B fine-tuning data · Average | score: 32.78 -> 43.45 | Helped |
| Use all three teaching scenarios rather than only Make Error and Expand (ablation) | LLaMA3-8B fine-tuning data · Average | score: 32.40 -> 43.45 | Helped |
| Use all three teaching scenarios rather than only Debate and Expand (ablation) | LLaMA3-8B fine-tuning data · Average | score: 37.11 -> 43.45 | Helped |
| Add Debate and Expand to Make Error-only augmentation (ablation) | LLaMA3-8B fine-tuning data · Average | score: 29.72 -> 43.45 | Helped |
Surprising findings
- Single-scenario augmentation (ME, DB, or EP alone) underperforms training on original data across all test sets.
- Qwen2.5-7B on MATH and Mistral-7B on MBPP/SCI-Q show slightly lower scores with BOOST-QA than with Ori-Data.
Limitations
- Gradient-based justifications for the augmentation are heuristic approximations rather than rigorously derived results.
- Evaluation relies on a single LLM judge (Qwen2.5-14B-Instruct) for subjective questions, and the MMLU and table numbers show inconsistencies (e.g., Ori and BOOST-QA identical for LLaMA3-8B on MMLU).
Experimental setup
Datasets: Orca-Math-200k ProcQA OpenHermes 2.5 HumanEval MBPP MATH MMLU-PRO MMLU ARC SCI-Q
Benchmarks: MATH MMLU-PRO-MATH MBPP HumanEval MMLU ARC SCI-Q CODEMMLU
Metrics: accuracy pass@1
Baselines: RandomAug SpellingAug TAGCOS CoT Collection Ori-Data
Builds on: EDA TAGCOS CoT Collection Orca-Math ProcQA
Closest papers by meaning
- Star-Agents: Automatic Data Optimization with LLM Agents for Instruction Tuning (NeurIPS 2024 · similarity 0.88)
- AgentRefine: Enhancing Agent Generalization through Refinement Tuning (ICLR 2025 · similarity 0.87)
- Boosting Multi-Domain Fine-Tuning of Large Language Models through Evolving Interactions between Samples (ICML 2025 · similarity 0.87)
- MAGDi: Structured Distillation of Multi-Agent Interaction Graphs Improves Reasoning in Smaller Language Models (ICML 2024 · similarity 0.87)
- Recursive Introspection: Teaching Language Model Agents How to Self-Improve (NeurIPS 2024 · similarity 0.87)
- TAIA: Large Language Models are Out-of-Distribution Data Learners (NeurIPS 2024 · similarity 0.87)
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.