New LLM framework enhances social intelligence and negotiation skills

Researchers have developed a new framework called Think-Strategy-Response (TSR) to improve the social intelligence of large language models (LLMs). This framework, inspired by the Theory of Planned Behavior, breaks down social dialogue into strategic planning and linguistic execution stages. To optimize TSR, they introduced Linearized Hierarchical Reinforcement Learning with Variance-Gated Rewards (LHRL-VGR), which dynamically adjusts rewards based on goal achievement variance. Experiments on the SOTOPIA benchmark showed that a Qwen2.5-7B agent fine-tuned with this method outperformed GPT-4o by 7.32% in social negotiation tasks. AI

IMPACT This research could lead to LLMs that are more adept at complex social interactions and multi-agent negotiations.

RANK_REASON Academic paper detailing a new framework and algorithm for LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• GPT-4o
• Linearized Hierarchical Reinforcement Learning with Variance-Gated Rewards (LHRL-VGR)
• qwen2.5:7b
• SOTOPIA benchmark
• theory of planned behavior
• Think-Strategy-Response (TSR) framework

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →