RuCL: Stratified Rubric-Based Curriculum Learning for Multimodal Large Language Model Reasoning

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a prevailing paradigm for enhancing reasoning in Multimodal Large Language Models (MLLMs). However, relying solely on outcome supervision risks reward hacking, where models learn spurious reasoning patterns to satisfy final answer checks. While recent rubric-based approaches offer fine-grained supervision signals, they suffer from high computational costs of instance-level generation and inefficient training dynamics caused by treating all rubrics as equally learnable. In this paper, we propose Stratified Rubric-based Curriculum Learning (RuCL), a novel framework that reformulates curriculum learning by shifting the focus from data selection to reward design. RuCL generates generalized rubrics for broad applicability and stratifies them based on model competence, dynamically adjusting their weights to guide the model from foundational perception to advanced logical reasoning. Extensive experiments on various visual reasoning benchmarks show that RuCL yields a remarkable +7.83% average improvement over the Qwen2.5-VL-7B model, achieving a state-of-the-art accuracy of 60.06%.

Multimodal Large Language Models (MLLMs) that understand both images and text are typically trained with a single binary reward: correct answer or not. Like grading a student solely on the final exam score, this encourages the model to chase answers by any means—often through shallow shortcuts that happen to work, producing reasoning that is brittle and uninterpretable. We introduce **RuCL** (Stratified Rubric-Based Curriculum Learning), a training method that evaluates and teaches reasoning more like an attentive teacher. Instead of a single score, RuCL decomposes feedback into a **hierarchy of rubric dimensions** spanning both foundational skills (e.g., "did the model attend to the relevant visual region?") and advanced ones (e.g., "does its logical chain remain self-consistent?"). The key innovation is a **stability-aware curriculum over these rubric signals**: early in training, the model receives substantive feedback almost exclusively from foundational rubrics, effectively masking out harder dimensions whose gradients are too noisy to be useful. As learning progresses and these harder signals stabilize, the method **gradually shifts the reward mix**—mathematically a convex combination controlled by a schedule parameter $\lambda_t$—admitting increasingly challenging objectives. We provide a theoretical analysis showing that this schedule directly controls gradient variance in policy optimization, explaining why the curriculum improves training stability compared to naïvely combining all rubrics from the start. Applied to a state-of-the-art vision-language model, RuCL boosts average accuracy by nearly 8% across seven visual reasoning benchmarks, setting a new performance record among models of comparable scale. Beyond accuracy, by rewarding coherent reasoning chains rather than lucky guesses, RuCL produces models whose decision processes are more transparent and auditable.