Sentient has launched a new multi-agent open-source architecture called ROMA, which surprisingly achieves high inference and search performance. First, let's do a simple science popularization. What are the three question banks used by ROMA this time? Seal-0 question bank: It contains a carefully selected set of extremely difficult questions, each of which is iterated to the strength of "multiple cutting-edge models tried multiple times but almost all of them are wrong". It examines AI's ability to use web search/tool models for fact checking, reasoning, and denoising under conflicting, noisy, or useless search results. FRAMES question bank: RAG unified evaluation set proposed by Google/Harvard (Factuality+Retrieval+Reasoning), multi hop and multi constraint problems (824 questions+papers&datasets provided by the official). It examines whether AI retrieval is correct, whether citations are accurate, and whether reasoning is in place. SimpleQA question bank: OpenAI's benchmark set of short factual Q&A questions, with short questions, easy scoring, and wide coverage. It mainly tests whether AI answers correctly and has less hallucinations, and can also evaluate the model's self calibration (matching confidence with actual accuracy). Then let's talk about how it was implemented? ROMA mainly does it through task decomposition, which is a three-step process: 1. Judgment and disassembly: The parent node determines whether the task is simple or complex, and if it is a complex task, it is disassembled into several subtasks; 2. Targeted treatment: Sub nodes find the most suitable AI agents and tools to solve sub tasks; 3. Summary report: Summarize the results and report them layer by layer to form the final answer. In this way, ROMA achieved high scores in inference and search (although the disadvantage is that it also brings higher load and longer thinking time to the server). In addition, some multi-agent systems have also adopted similar architectures in the past, but many have encountered a problem called "error accumulation". For example, if the accuracy of a single AI is 90% and six layers are accumulated, the accuracy will only be around 50%. @The idea of SentientAGI ROMA architecture is to make the entire inference process transparent and open source, which facilitates developers to make targeted adjustments to the entire process to increase the accuracy of long tasks. This is the benefit of open source.