"FlowBalance: Verifier-Grounded Self-Improvement from On-Policy Reasoning Experience" by Zixun Huang , Kishan Panaganti , Haitao Mi , Leowei Liang
TLDR:
The text introduces FlowBalance, a method that enhances reasoning models by using verifier-calibrated self-guidance with trajectory-level score reweighting and profile-based trajectory balance. This approach aims to improve reasoning models by learning a normalized distribution over complete responses, utilizing guidance from terminal verifiers, and avoiding overconcentration on narrow solution modes. FlowBalance calibrates trajectory-level self-guidance using verifier-derived group advantage and adjusts the energy to reweight a reference policy accordingly. The method demonstrates improvements in performance, training speed, stability, and correct-strategy diversity compared to existing models like FlowRL.
Read Paper / Blog
huggingface.co
Paper page - FlowBalance: Verifier-Grounded Self-Improvement from On-Policy Reasoning Experience
Join the discussion on this paper page
September 8, 2026 38
