AgreeMate: Teaching LLMs to Haggle

A framework for training Large Language Models (LLMs) to perform strategic price negotiations through natural language. AgreeMate combines role-specialized fine-tuning with systematic model comparison across scales and architectures.

Overview

AgreeMate presents a comprehensive finetuning/testing framework and negotiation system that explores LLMs' capabilities in strategic price bargaining. Our approach combines:

• Role-specialized fine-tuning (buyer, seller, generalist)
• Systematic model comparison across scales (3B to 70B parameters)
• Chain of thought prompting
• Personality-driven behavior modeling
• Attention mechanism analysis

For more details, read our full paper on arXiv or refer to the copy included in this repository (paper.pdf).

You can also listen to a summary of our research paper on NotebookLM.

Related Work

Our work builds on "Decoupling Strategy and Generation in Negotiation Dialogues" by He et al. (2018), which introduced a modular approach to negotiation using coarse dialogue acts for strategy control.

AgreeMate advances this by leveraging LLMs to unify strategy and generation without requiring explicit intermediate representations, enabling more nuanced and scalable negotiation behaviors.

Key Features

• Model Variants: Specialized fine-tuned models for buyer, seller, and generalist roles
• Multiple Scales: Support for LLaMA models from 3B to 70B parameters
• Personality Types: Configurable negotiation personalities (aggressive, fair, passive)
• Chain of Thought: Enhanced reasoning capabilities through CoT prompting
• Comprehensive Metrics: Novel measurements for fairness, bias, efficiency, and more

Key Technologies

• Parameter-Efficient Fine-Tuning (PEFT) with LoRA
• 4-bit quantization for memory efficiency
• Gradient checkpointing
• Mixed precision training
• Advanced learning rate scheduling

Results & Findings

• Larger models demonstrate improved agreement rates and fairness
• Chain of Thought prompting enhances exploratory behavior
• Personality traits significantly influence negotiation dynamics
• Attention analysis reveals semantic understanding of negotiation concepts

Critical Project Structure

Data

• data/craigslist_bargains/:

  • Dataset: 6.6k human-human negotiation dialogues from the Craigslist Bargains dataset.
  • Files:
    • train.csv, test.csv, validation.csv: Split, reformatted datasets.
  • Source: Stanford NLP's CraigslistBargains dataset.

• data/deal_or_no_deal/:

  • Dataset: 12k multi-issue bargaining dialogues.
  • Files:
    • buyer_training.csv, seller_training.csv, generalist_training.csv: Split, reformatted datasets for buyer, seller, and generalist fine-tuning.
  • Source: Facebook's Deal or No Deal dataset.

Finetuning

• finetuning/:
  • Contains all fine-tuning and training scripts.
  • Key Components:
    • data_loader.py: Utility for loading and preparing datasets.
    • finetuner.ipynb: Fine-tuning workflow for LLaMA models.
    • model_loader.py: Script to load pre-trained models.
    • models-<role>-finetuned-<scale>: Checkpoints for fine-tuned models at various scales (3B–70B).

Results

• results/:
  • Repository for analysis outputs and key findings.
  • Key Files:
    • analysis.ipynb and analysis_interactive.ipynb: Jupyter notebooks for model evaluation and visualizations.
    • Performance CSVs: Extensive evaluation results across buyer-seller roles, model scales, and negotiation styles:
      • Baseline comparisons (e.g., baseline-3B-buyer+baseline-3B-seller.csv)
      • Chain of Thought prompting results (e.g., baseline-COT-3B-buyer+baseline-3B-seller.csv)
      • Updated datasets for improved model iterations.
  • Visualization Outputs: Use these notebooks to analyze performance trends, fairness, and strategy success metrics.

Supporting Files

• paper.pdf: PDF of our research paper detailing AgreeMate’s methodology, experiments, and findings.
• poster.png: Visual summary from our halfway-point.
• LICENSE: MIT License information.
• requirements.txt: Dependency list for easy setup.

Paper

• paper/:
  • Contains main paper files such as .tex, .bib, etc.
  • figures/:
    • Stores visualizations generated during analysis for the paper.

Basic Usage

Installation

git clone https://github.com/GeneralCoder365/agreemate
cd agreemate
pip install -r requirements.txt

Basic Negotiation

from agreemate import Negotiator

# initialize buyer and seller agents
buyer = Negotiator(role="buyer", model="llama-3.2-3B-buyer")
seller = Negotiator(role="seller", model="llama-3.2-3B-seller")

# run negotiation
dialogue = negotiate(buyer, seller, scenario="car_charger")

Personality Configuration

# create agents with specific personalities
aggressive_buyer = Negotiator(
    role="buyer",
    personality="aggressive",
    model="llama-3.2-3B-buyer"
)

fair_seller = Negotiator(
    role="seller",
    personality="fair",
    model="llama-3.2-3B-seller"
)

Contributors

• Ainesh Chatterjee - Dataset preparation, Fine-tuning, and personality analysis
• Samuel Miller - Probing analysis and visualization
• Nithin Parepally - Architecture design and evaluation framework

Citation

@misc{chatterjee2024agreemateteachingllmshaggle,
      title={AgreeMate: Teaching LLMs to Haggle}, 
      author={Ainesh Chatterjee and Samuel Miller and Nithin Parepally},
      year={2024},
      eprint={2412.18690},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2412.18690}, 
}

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• He et al. (2018): For their foundational work, "Decoupling Strategy and Generation in Negotiation Dialogues", introducing a modular approach to negotiation by leveraging coarse dialogue acts for strategy control, inspiring our efforts to unify strategy and generation in negotiation through LLMs. Additionally, for providing the CraigslistBargains dataset, which comprises over 6,600 human-human negotiation dialogues across various product categories, facilitating research in negotiation strategies and language generation.

• Facebook Research: For the Deal or No Deal Negotiator dataset, a substantial collection of human-human negotiations on a multi-issue bargaining task, enabling the development of end-to-end models for negotiation.

• Meta AI: For developing the LLaMA model family, which served as the foundation for our large language model experiments.

Academic Context

This project was conducted as part of the CMSC723: Graduate Natural Language Processing (Fall 2024) course at the University of Maryland, College Park.

The content and findings reflect our independent research and do not imply any endorsement or ownership by the university.