Jason-Huang/MCTS-guided-LLM-Negotiation

MCTS-guided LLM Negotiation — actor & analyzer adapters

Two LoRA adapters (r=32) over Qwen3-14B, used together by MCTS-guided-LLM-Negotiation.

A buyer agent negotiates the price of an item against an LLM-simulated seller. Rather than sampling utterances directly, it runs Monte Carlo Tree Search over dialogue acts — propose a counter, use comparatives, ask a question — and the two adapters fill different roles in that loop.

Subfolder	Objective	Role
actor/	DPO	Selects the dialogue act and writes the buyer's utterance. Trained on preference pairs harvested from MCTS rollouts, where the chosen act is the one with the highest search value at that turn.
analyzer/	SFT	Reads the transcript after each exchange and emits a structured estimate of the seller — target price, likely next action, and persona traits — which becomes the opponent model inside subsequent rollouts.

Both are served together through vLLM multi-LoRA, so a single model process runs the buyer, the analyzer and the simulated seller.

Usage

vllm.lora.request.LoRARequest takes a local directory, so download first:

from huggingface_hub import snapshot_download

snapshot_download(
    "Jason-Huang/MCTS-guided-LLM-Negotiation",
    local_dir="checkpoints",
)

That lays the adapters out exactly where the evaluation scripts expect them — checkpoints/actor and checkpoints/analyzer are the defaults for --actor-adapter and --analyzer-adapter:

python -m experiments.cb.run_negotiation_cb_mcts \
    --data-file CB/test.csv --end-index 300 --seed 42

To load one directly:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", device_map="cuda")
model = PeftModel.from_pretrained(
    base, "Jason-Huang/MCTS-guided-LLM-Negotiation", subfolder="actor"
)

Details

• Base model. Trained against unsloth/qwen3-14b-unsloth-bnb-4bit; evaluation runs use Qwen/Qwen3-14B.
• LoRA config. r=32, alpha=32, dropout 0.05, over q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj. 560 tensors each.
• Training data. Harvested from the agent's own MCTS rollouts on CraigslistBargain — the actor from ranked strategy preference pairs, the analyzer from the seller turns collected in the same rollouts.
• Domains. There is no A2A-specific checkpoint; A2A evaluation reuses these as cross-domain transfer.

Licence

The adapters follow the base model, Qwen3-14B, which is Apache-2.0. The surrounding code is MIT.

Citation

@software{mcts_guided_llm_negotiation,
  title  = {MCTS-guided LLM Negotiation: Strategy-level search with a learned opponent model},
  year   = {2026},
  url    = {https://github.com/AI-Application-and-Integration-Lab/MCTS-guided-LLM-Negotiation}
}

Evaluation datasets:

@misc{he2018decoupling,
      title={Decoupling Strategy and Generation in Negotiation Dialogues},
      author={He He and Derek Chen and Anusha Balakrishnan and Percy Liang},
      year={2018},
      eprint={1808.09637},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/1808.09637},
}

@misc{zhu2025automatedriskygamemodeling,
      title={The Automated but Risky Game: Modeling and Benchmarking Agent-to-Agent Negotiations and Transactions in Consumer Markets},
      author={Shenzhe Zhu and Jiao Sun and Yi Nian and Tobin South and Alex Pentland and Jiaxin Pei},
      year={2025},
      eprint={2506.00073},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2506.00073},
}