SD-GPS Framework Uses QwenVL3-2B for Solver-Driven Geometry Solving
June 29, 2026
The SD-GPS framework improves geometry problem solving by using QwenVL3-2B to unify supervised formal-language adaptation with solvability-guided reinforcement learning. This neuro-symbolic approach treats the symbolic solver as an execution oracle to enable autoformalization and propose local auxiliary lemmas when solvers hit deductive impasses.
HOW THIS AFFECTS YOU
●
researcherYou can leverage executability as a central training signal to bridge the gap between multimodal translation and symbolic reasoning.