AI RESEARCH
MetaDNS: Enhancing Exploration in Discrete Neural Samplers via Well-Tempered Metadynamics
arXiv CS.LG
•
ArXi:2605.21722v1 Announce Type: cross Sampling from discrete distributions with multiple modes and energy barriers is fundamental to machine learning and computational physics. Recent discrete neural samplers like MDNS suffer from mode collapse and fail to sample high-energy barrier regions between modes, which is critical for free energy estimation and understanding phase transitions. We propose Metadynamics Discrete Neural Sampler (MetaDNS), a general framework integrating well-tempered metadynamics into discrete diffusion or autoregressive samplers.