Improving Reasoning in Vision-Language Models via Perception Verified Self-Training

Achieving human-like reasoning in Vision-Language Models(VLMs) remains a long-standing challenge. Recent approaches leverageChain-of-Thought (CoT) rationales generated by human annotators orproprietary models to improve reasoning, which is costly and difficultto scale. Self-training offers a promising alternative by using model’sown outputs as supervision. However, existing methods often suffer fromvisual hallucinations - where rationales describe non-existent visual con-tent, and language shortcuts - where predictions rely on textual priorsrather than true visual grounding, as rationales are typically filtered onlyby answer correctness without verifying visual perception. To addressthis limitation, we propose a perception-verified self-training frameworkthat enforces visually grounded reasoning. First, our method employs aCoT template (caption-reasoning-conclusion) that disentangles percep-tion from reasoning, enabling independent verification of visual under-standing. To compensate for the absence of ground-truth captions, wepropose PerceptEval, an unsupervised method that evaluates captionquality based on its alignment with visual and textual elements presentin the image. Using caption verification together with answer correct-ness, we partition the data into three subsets: easy (correct caption andconclusion), medium (correct caption but incorrect conclusion), and hard(incorrect caption). Building on this partitioning, we design a two-stagecurriculum learning strategy. In Stage 1, the model is trained on easyexamples and subsequently in Stage 2, medium samples are incorporatedthrough a caption-guided reasoning enhancement procedure thatregenerates reasoning conditioned on verified captions. Only regeneratedsamples with the correct conclusions are retained. This ensures trainingin both stages is done exclusively on perceptually grounded reasoning,reducing hallucinations and language shortcuts. Extensive experimentsacross diverse domains and models demonstrate improvements of up to16% over standard self-training baselines. This demonstrates that theproposed framework is a scalable and cost-effective solution to advanc-ing multimodal reasoning without manually annotated CoT rationales.