Reasoning Research

At Menlo, we’re advancing AI systems with agentic reasoning abilities. Our research focuses on developing models that can autonomously analyze challenges, adapt their strategies, and augment human capabilities.

Research Projects

ReZero (Apr 2025)

🔄 ReZero: Enhancing LLM Search Ability by Trying One-More-Time  ReZero is a search-oriented language model that uses Guided Reinforcement Policy Optimization (GRPO) and a retry-reward mechanism to encourage persistent refinement of search queries until the desired result is found. Unlike typical approaches that avoid repetition, ReZero strategically employs it to improve search performance, achieving a significantly higher success rate over baseline models.

Links:

• Paper: https://arxiv.org/abs/2504.11001 
• Huggingface: https://huggingface.co/Menlo/ReZero-v0.1-llama-3.2-3b-it-grpo-250404 
• Github: https://github.com/menloresearch/ReZero