PACT: Protocolized Action-state Communication and Transmission

Paper: What Should Agents Say? Action-state Communication for Efficient Multi-Agent Systems

Run PACT on the split-evidence interaction setting — Qwen3-14B on HotpotQA.

Setup

pip install -r requirements.txt

A CUDA GPU is required to run Qwen3-14B. The demo uses vLLM by default; Qwen3-14B fits on a single 40 GB+ GPU with --tensor_parallel_size 1.

Get the data

bash scripts/download_hotpotqa.sh

This downloads hotpot_dev_distractor_v1.json (~50 MB) into data/. The data/ directory is git-ignored.

Run

Quick demo on a 50-question sample:

bash scripts/run_demo.sh

Full dev set (7,405 questions):

MAX_SAMPLES=-1 bash scripts/run_demo.sh

Or call run.py directly:

python run.py \
    --model_name Qwen/Qwen3-14B \
    --data_path data/hotpot_dev_distractor_v1.json \
    --use_vllm --tensor_parallel_size 1 \
    --seed 42 --generate_bs 64 --max_new_tokens 4096 --max_turns 4 \
    --max_samples -1 \
    --output_path results/pact_qwen3_14b_hotpot.jsonl

The run prints a summary (exact match, average token-F1, average total tokens per problem) and, with --output_path, writes per-sample results as JSON lines.

Citation

If you find PACT useful, please cite:

@article{huang2026what,
  title={What Should Agents Say? Action-state Communication for Efficient Multi-Agent Systems},
  author={Huang, Chen and Wu, Yuhao and Zhang, Wenxuan},
  journal={arXiv preprint arXiv:2606.05304},
  year={2026}
}