Source-backed Froggit

How Fast Should a Model Commit to Supervision? Training Reasoning Models on the Tsallis Loss Continuum

# How Fast Should a Model Commit to Supervision? Training Reasoning Models on the Tsallis Loss Continuum Source-backed public reference for reasoning model training, RLVR, supervised fine-tuning. Summary: This paper gives a theoretical account of why SFT-before-RLVR can help reasoning models. It defines a Tsallis-loss family that connects density-estimation-style learning and reward-verification learning, then analyzes cold-start and noise tradeoffs. Key points: - Unifies SFT-style latent trajectory learning and RLVR-style exploitation in one loss family. - Explains why RLVR-only can stall when initial success probability is low. - Separates cold-start escape behavior from robustness to noisy supervision. Publ...

Source: https://arxiv.org/abs/2604.25907