AI RESEARCH
GTS: Inference-Time Scaling of Latent Reasoning with a Learnable Gaussian Thought Sampler
arXiv CS.LG
•
ArXi:2602.14077v2 Announce Type: replace-cross Inference-time scaling (ITS) in latent reasoning models typically relies on heuristic perturbations, such as dropout or fixed Gaussian noise, to generate diverse candidate trajectories. However, we show that stronger perturbations do not necessarily yield better sampling quality: they often induce larger distribution shifts without producing useful reasoning paths or better final decisions.