Controllable Tool-Use Data Synthesis for Reinforcement Learning
Discover COVERT, a method for reliable, verifiable tool-use data synthesis that boosts reinforcement learning agent accuracy and robustness.
Pioneer Agent: Boosting Small Language Models in Production
Discover how Pioneer Agent automates continual improvement of small language models, enhancing performance and efficiency in real-world production.
Understanding Expert Specialization in MoEs: Geometry Over Domain
Explore why expert specialization in Mixture of Experts (MoEs) stems from hidden state geometry, not domain expertise, revealing new AI insights.
Tipiano: Realistic Piano Hand Motion Synthesis Using Fingertip Priors
Discover Tipiano, a novel framework for realistic piano hand motion synthesis that enhances accuracy and naturalness using fingertip priors and advanced mo...
Belief-Aware VLM for Enhanced Human-Like Reasoning
Discover how belief-aware VLMs improve human-like reasoning with retrieval memory and reinforcement learning for better AI adaptability.