Teaching LLMs to haggle like humans through strategic price negotiations and natural dialogue.
AgreeMate, a framework enabling LLMs to conduct strategic price negotiations through natural language, combining prompt engineering, fine-tuning, and chain-of-thought prompting to enhance performance in automated bargaining.
-----
https://arxiv.org/abs/2412.18690
Original Problem 🤖:
Traditional automated negotiation systems relied on complex modular architectures separating strategy planning from language generation. This made them inflexible and limited in handling multi-step reasoning and natural dialogue.
-----
Solution in this Paper 🛠️:
→ AgreeMate introduces a unified framework where LLMs handle both strategy and language generation simultaneously.
→ The system uses role-specialized fine-tuning to create buyer, seller and generalist negotiation agents.
→ Implements parameter-efficient training using LoRA and 4-bit quantization to enable training on limited hardware.
→ Leverages chain-of-thought prompting to improve strategic reasoning and decision-making during negotiations.
-----
Key Insights 🔍:
→ Larger models (70B parameters) demonstrate better fairness and agreement rates
→ Chain-of-thought prompting improves exploratory behavior but introduces bias in smaller models
→ Aggressive negotiation strategies dominate in efficiency but compromise fairness
→ Passive strategies enable smoother negotiations but show higher buyer bias
-----
Results 📊:
→ Achieved stable GPU memory usage (peak: 22GB, stable: 3.56GB)
→ Final validation losses converged around 4.05
→ Generalist model achieved slightly lower EMA loss of 4.49
→ Higher probing ratios and increased aggressiveness with CoT
