Introduction

The field of human-AI collaboration and language models is rapidly evolving, with a focus on developing more pragmatic and socially aware language agents. Researchers are exploring innovative approaches to enhance human-AI collaboration, including the development of dual-process theory of mind frameworks and long-chain-of-thought critic models.

General Direction

The current trend in this field is towards creating more sophisticated language models that can reason about shared goals and beliefs, and adapt to diverse human behaviors in dynamic scenarios. This involves developing frameworks that can decouple knowledge and reasoning, and facilitate slow and deliberate System-2-like thinking.

Noteworthy Papers

• Collaborative Rational Speech Act (CRSA) is a notable paper that introduces an information-theoretic extension of the Rational Speech Act framework, demonstrating more consistent and collaborative behavior in referential games and doctor-patient dialogs.
• RefCritic is another significant paper that proposes a long-chain-of-thought critic module based on reinforcement learning, generating high-quality evaluations with actionable feedback that effectively guides model refinement.
• R4ec is a reasoning, reflection, and refinement framework that evolves recommendation systems into weak System-2 models, demonstrating superiority on Amazon-Book and MovieLens-1M datasets and showing a 2.2% increase in revenue on a large-scale online advertising platform.