Large Content And Behavior Models To Understand, Simulate, And Optimize Content And Behavior
Published in Submitted to ICLR 2024 — Pre-rebuttal reviews: 10,8,6,5, 2024
Recommended citation: Ashmit Khandelwal, Aditya Agrawal, Aanisha Bhattacharyya, Yaman K Singla, Somesh Singh, Uttaran Bhattacharya, Ishita Dasgupta, Stefano Petrangeli, Rajiv Ratn Shah, Changyou Chen, Balaji Krishnamurthy, 2023. "Large Content And Behavior Models To Understand, Simulate, And Optimize Content And Behavior", arXiv preprint doi.org/10.48550/arXiv.2309.00359
Shannon’s communication theory comprises three levels: technical, semantic, and effectiveness. While the technical and semantic levels have made substantial progress, the effectiveness levels, involving receiver behavior, has been largely unaddressed.
The paper introduces Large Content and Behavior Models (LCBMs) to bridge this gap by reintroducing behavior tokens into LLM training data. These models demonstrate generalization capabilities in simulating and explaining receiver behavior, understanding content, and adapting to various behavior domains using the Content Behavior Corpus (CBC).