aiqbench analysis
This research introduces the TSR framework, which applies hierarchical reinforcement learning to improve LLMs in complex social negotiations by separating strategic planning from linguistic execution. By dynamically routing rewards based on goal variance, the method enables models like Qwen2.5-7B to outperform GPT-4o benchmarks in multi-agent coordination tasks. The work addresses a critical gap where current uniform reward systems fail to capture the nuance required for long-term social adaptation.
Enhancing Social Intelligence in LLMs with Hierarchical Reasoning and Utterance-Level Goal Rewarding
arXiv:2608.05832v1 Announce Type: new Abstract: Large language models (LLMs) excel in structured tasks but struggle with dynamic social interactions, where success requires long-term goal coordination and rapid adaptation. Current methods often apply uniform goal-based rewards to every utterance, overlooking the specificity of objectives at each dialogue turn and failing to account for the rationale of potential strategies. Inspired by the Theory of Planned Behavior, we propose the Think-Strategy-Response (TSR) framework, which decomposes social dialogue into two hierarchical stages: high-level strategic planning and low-level linguistic execution. To optimize TSR, we introduce Linearized Hierarchical Reinforcement Learning with Variance-Gated Rewards (LHRL-VGR), a novel algorithm that dynamically routes rewards - balancing goal completion and strategy adherence - based on the variance of goal achievement scores. Experiments on the SOTOPIA benchmark show that our approach fine-tunes a Qwen2.5-7B agent to surpass the GPT-4o baseline by 7.32% in goal completion success, demonstrating state-of-the-art performance in multi-agent social negotiation tasks.
Abstract:Large language models (LLMs) excel in structured tasks but struggle with dynamic social interactions, where success requires long-term goal coordination and rapid adaptation. Current methods often apply uniform goal-based rewards to every utterance, overlooking the specificity of objectives at each dialogue turn and failing to account for the rationale of potential strategies. Inspired by the Theory of Planned Behavior, we propose the Think-Strategy-Response (TSR) framework, which decomposes social dialogue into two hierarchical stages: high-level strategic planning and low-level linguistic execution. To optimize TSR, we introduce Linearized Hierarchical Reinforcement Learning with Variance-Gated Rewards (LHRL-VGR), a novel algorithm that dynamically routes rewards - balancing goal completion and strategy adherence - based on the variance of goal achievement scores. Experiments on the SOTOPIA benchmark show that our approach fine-tunes a Qwen2.5-7B agent to surpass the GPT-4o baseline by 7.32% in goal completion success, demonstrating state-of-the-art performance in multi-agent social negotiation tasks.
Computation and Language (cs.CL)
arXiv:2608.05832 [cs.CL]
(or arXiv:2608.05832v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2608.05832 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 243 words
Related coverage
Handle with CARE: Can LLMs Reproduce How Online Communities React?
arXiv · Oct 8, 2026
Links #3: 2026/06 Part 1
LessWrong · Jun 15, 2026
Somewhat Contra Ted Chiang on AI Consciousness
LessWrong · Jun 13, 2026
The J-Space Debate, Agent Swarms, and Pacing Frontier AI - Digital Minds Newsletter #4
LessWrong · Sep 18, 2026