Explore Professor Banghua Zhu's presentation at the Open AGI Summit in Brussels, where he introduces Nexus Flow's unique approach to creating language software agents using customized language models. Professor Zhu, co-founder of Nexus Flow and incoming Assistant Professor at the University of Washington, describes the advantages of separated models where separate models are used for extractive and abstractive reasoning to yield impressive results in capabilities.
Key Highlights:
Complex Instruction Following and Tool Use: Addressing tasks like AI-driven navigation and multi-step instructions with enhanced accuracy, reduced hallucination, and lower latency compared to GPT-4.
Extractive vs. Abstractive Reasoning: Nexus Flow's approach separates these reasoning types in smaller models, resulting in significant success.
Nexus Flow Models: NexusRaven-V2-13B and Starling-7B, two open-sourced models that excel in tool use and chatbot interactions. NexusRaven-V2-13B, based on CodeLlama-13B, surpasses GPT-4 in complex tool use with minimal hallucination. Starling-7B, built on Mistral-7B, leads in chatbot performance.
NexusRaven-V2 LLM: This model enhances extractive reasoning and tool API handling, achieving 7% better accuracy on parallel and nested APIs with a significantly smaller model size. On RoTBench, it boasts a 71% win rate over GPT-4.
Starling-7B: Demonstrates state-of-the-art data efficiency in human-preference alignment, ranking 13th on Chatbot Arena, outperforming models like Llama-2-Chat 70B and Gemini-Pro-V1.0.
Future Developments: An open model stronger in chat and function calling is coming soon from Nexus Flow, promising further advancements in language agent capabilities.
Join us to understand how Nexus Flow's innovative models are shaping the future of AI agents, making complex instruction following and tool use more accurate and reliable. Don't miss this detailed presentation by Professor Zhu on the forefront of AI research and development.
Тэги:
##AI ##NexusFlow ##LanguageModels ##AIResearch ##OpenAGISummit ##BanghuaZhu ##UniversityOfWashington ##ExtractiveReasoning ##Sentient ##OpenAGI ##AbstractiveReasoning ##NexusRavenV2 ##Starling7B ##ToolUse ##ChatbotPerformance ##RoTBench ##FutureAI