Optimized Related Work Generation with Multi-Agent Systems

Minimal instructions to run the demo and benchmark.

Requirements

• Python 3.10+
• NVIDIA GPU (tested on ~12GB VRAM)
• vLLM installed and a Qwen3-8B (quantized) model available locally

Install

pip install -r requirements.txt

Start vLLM (example)

vllm serve ./Qwen3-8B-quantized.w4a16 \
  --port 8000 \
  --gpu-memory-utilization 0.85 \
  --max-model-len 16384 \
  --enforce-eager \
  --enable-reasoning \
  --reasoning-parser deepseek_r1

Run demos

# Single-agent demo (arXiv retrieval)
python single_agent.py

# Multi-agent demo (arXiv retrieval)
python run_mas.py

Run benchmark (MAS vs Single Agent on OARelatedWork)

# Compare MAS vs single agent on high-reference samples
python benchmark_optimized.py --samples 10 --min-refs 7 --compare --max-revisions 2