Caesar: Deep Agentic Web Exploration for Creative Answer Synthesis

Caesar is a novel agentic LLM architecture designed to bridge the gap between information gathering and the synthesis of new insights. Unlike existing agents that treat the web as a flat sequence of disconnected documents, Caesar leverages an extensive knowledge graph to foster associative reasoning, enabling the discovery of non-obvious connections between disparate concepts. It consists of two core components: (1) exploration driven by a dynamic context-aware policy that adapts its search strategy based on the current state of knowledge, and (2) synthesis controlled by an adversarial draft refinement loop that actively seeks novel perspectives rather than confirming established priors. This dual approach allows Caesar to generate artifacts and answers characterized by high novelty and structural coherence.

In benchmark evaluations, Caesar significantly outperformed state-of-the-art LLM research agents in tasks requiring creativity, demonstrating its ability to produce original and coherent outputs. The architecture represents a shift from convergent search—which often yields derivative summaries—to divergent, associative reasoning that can uncover new ideas. By integrating knowledge graphs and adversarial refinement, Caesar offers a pathway for autonomous agents to move beyond passive retrieval and engage in genuine creative discovery. The paper is available on arXiv and has been submitted to the Information Retrieval and Multiagent Systems categories.