一作ACL 2026Poster· Natural Language Processing
Does RLVR Extend Reasoning Boundaries? Investigating Capability Expansion in Vision-Language Models
Minghe Shen · Zhuo Zhi · Chonghan Liu · Shuo Xing · Zhengzhong Tu · Che Liu
摘要
Recent studies posit that Reinforcement Learning with Verifiable Rewards (RLVR) primarily amplifies behaviors inherent to the pre-training distribution rather than inducing new capabilities, but these insights are predominantly limited to language-only domains, leaving the dynamics of visual-cent...