Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Assistive Large Language Model Agents for Socially-Aware Negotiation Dialogues

Yuncheng Hua, Lizhen Qu, Gholamreza Haffari

TL;DR

This work introduces assistive large language model agents for socially-aware negotiation dialogues, featuring a remediator that rewrites norm-violating utterances to improve both task outcomes and relational goals. A tuning-free, value-impact based ICL method selects high-quality exemplars to guide remediation without fine-tuning, using a formalized value function and hierarchical exemplar search. Across three negotiation topics, the ValueImpact ICL remediator consistently outperforms baselines in success rate, deal value, and social metrics, with human evaluation favoring its remediations. The approach enables efficient, scalable intervention in real-time negotiations and provides a public dataset and code for reproducibility and further study.

Abstract

We develop assistive agents based on Large Language Models (LLMs) that aid interlocutors in business negotiations. Specifically, we simulate business negotiations by letting two LLM-based agents engage in role play. A third LLM acts as a remediator agent to rewrite utterances violating norms for improving negotiation outcomes. We introduce a simple tuning-free and label-free In-Context Learning (ICL) method to identify high-quality ICL exemplars for the remediator, where we propose a novel select criteria, called value impact, to measure the quality of the negotiation outcomes. We provide rich empirical evidence to demonstrate its effectiveness in negotiations across three different negotiation topics. We have released our source code and the generated dataset at: https://github.com/tk1363704/SADAS.

Assistive Large Language Model Agents for Socially-Aware Negotiation Dialogues

TL;DR

This work introduces assistive large language model agents for socially-aware negotiation dialogues, featuring a remediator that rewrites norm-violating utterances to improve both task outcomes and relational goals. A tuning-free, value-impact based ICL method selects high-quality exemplars to guide remediation without fine-tuning, using a formalized value function and hierarchical exemplar search. Across three negotiation topics, the ValueImpact ICL remediator consistently outperforms baselines in success rate, deal value, and social metrics, with human evaluation favoring its remediations. The approach enables efficient, scalable intervention in real-time negotiations and provides a public dataset and code for reproducibility and further study.

Abstract

We develop assistive agents based on Large Language Models (LLMs) that aid interlocutors in business negotiations. Specifically, we simulate business negotiations by letting two LLM-based agents engage in role play. A third LLM acts as a remediator agent to rewrite utterances violating norms for improving negotiation outcomes. We introduce a simple tuning-free and label-free In-Context Learning (ICL) method to identify high-quality ICL exemplars for the remediator, where we propose a novel select criteria, called value impact, to measure the quality of the negotiation outcomes. We provide rich empirical evidence to demonstrate its effectiveness in negotiations across three different negotiation topics. We have released our source code and the generated dataset at: https://github.com/tk1363704/SADAS.
Paper Structure (54 sections, 4 equations, 3 figures, 14 tables, 2 algorithms)

This paper contains 54 sections, 4 equations, 3 figures, 14 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Assistive Systems for Negotiations
  3. Definition of Social Norm Violation
  4. Problem Setting
  5. LLM-based Assistive Agents
  6. Exemplars with High Value Impact
  7. Exemplar Filtering using Value Impact
  8. Value of an remediation
  9. Value Impact of exemplar(s)
  10. Individual exemplar Filtering
  11. Selection of a Set of ICL Demonstrations
  12. Search for Optimal exemplars.
  13. Experiments
  14. Experimental Settings
  15. Dialogue generation.
  16. ...and 39 more sections

Figures (3)

  • Figure 1: An illustration of our social agent(s). We expect to apply the remediator to real-world negotiations where the remediator can rectify potential social norm violations in the dialogue, thereby reducing conflicts and misunderstandings caused by cultural differences.
  • Figure 2: A running example: In this conversational exchange between two interlocutors, an utterance from the seller breaches a social norm. Our remediator intervenes to rectify it by generating a remedial response.
  • Figure 3: An illustration of using Hierarchical traversal with early pruning to search for the optimal exemplars.