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

Yukun Chen; Jiaming Li; Longze Chen; Ze Gong; Jingpeng Li; Zhen Qin; Hengyu Chang; Ancheng Xu

cs.CLFeb 25, 2026

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

Authors: Yukun Chen, Jiaming Li, Longze Chen, Ze Gong, Jingpeng Li, Zhen Qin, Hengyu Chang, Ancheng Xu, +5 more

Organizations: Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences · University of Chinese Academy of Sciences · Alibaba Group · School of Biomedical Engineering, UNSW Sydney

Abstract

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 the model's competence. By dynamically adjusting rubric weights during training, RuCL guides the model from mastering foundational perception to tackling 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%.

Figures & tables

Figure 1: Comparison of reward paradigms. We move beyond (A) outcome-only signals and (B) unstructured dense feedback. (C) Our RuCL framework organizes rubrics into a stratified curriculum, aligning reward complexity with the model’s progressive learning stages.

Figure 2: Overview of Stratified Rubric-based Curriculum Learning (RuCL). The framework proceeds in two stages: (Top) Generalized Rubric Construction and Stratification , where evaluation rubrics are generated and categorized into Foundational ( $\mathcal{R}_{\text{easy}}$ ) and Advanced ( $\mathcal{R}_{\text{hard}}$ ) tiers based on empirical difficulty. (Bottom) Dynamic Curriculum Learning , where the rubric-based reward is synthesized via a dynamic weighting mechanism controlled by a scheduler. By adjusting the weight $\lambda$ based on real-time performance, RuCL progressively shifts the optimization focus from mastering basic skills to tackling complex reasoning.

Entity Extraction	Isolates the specific Region of Interest (ROI).
Rubric Criterion	Evaluation Focus	Difficulty Stratum	Evaluator
Visual Presence	Penalizes object and attribute hallucinations.	Foundational ( $\mathcal{R}_{\text{easy}}$ )	LLM Judge
Intent Alignment	Checks compliance with scope and constraints.
Conclusion Match	Ensures the answer logically follows the reasoning.
Step Coherence	Detects logical gaps and internal contradictions.	Advanced ( $\mathcal{R}_{\text{hard}}$ )
Evidence Grounding	Validates inferences using specific visual cues.	Advanced ( $\mathcal{R}_{\text{hard}}$ )

Table 1: The stratified reward system. The evaluation rubrics are categorized by difficulty and implemented via either a generative LLM Judge or a deterministic Answer Verifier. Detailed rubric definitions and scoring criteria are provided in Appendix D .

Proprietary Models
Model	Mathematical Reasoning				General			Avg.
Model	MathVerse	MathVision	MathVista	WeMATH	MMMU	LogicVista	Counting	Avg.
GPT-4o	50.20	30.30	63.80	68.80	69.10	45.90	-	-
Claude-3.5-Sonnet	57.64	46.48	67.70	73.05	68.30	43.97	-	-
Open-Source General-Purpose Models
Qwen2.5-VL-7B	48.98	24.18	70.20	58.52	51.00	39.26	73.50	52.23

Table 2: Performance comparison on Mathematical Reasoning and General benchmarks. The “Avg.” column reports the average score across all seven evaluated benchmarks. The best results among open-source reasoning models are highlighted in bold , while the second-best are underlined .

Figure 3: Left: Training dynamics of Foundational (blue) and Advanced (red) rubric rewards. Middle: Ablation study results on rubric aggregation and scheduling strategies. Right: Sensitivity analysis of the reward balancing hyperparameter.

Method	MMMU	LogicVista	Counting	Avg.
Vanilla GRPO	54.89	46.53	76.00	59.14
Uniform Averaging	55.44	50.11	77.00	59.29
Linear Stratification	55.44	47.43	79.50	60.79
RuCL	56.67	49.66	85.50	63.94

Table 3: Ablation results on General benchmarks.

Window Size $w$	Mathematical	General	Avg.
$w=10$	55.57	60.8	57.81
$w=20$	57.15	63.94	60.06
$w=30$	56.64	63.74	59.67

Table 4: Sensitivity analysis of sliding window size

w

in curriculum triggering.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Optimization & Rollout
Hyperparameter	Value
Training Batch Size	256
Global Batch Size	128
Rollout Number ( $G$ )	8
Sampling Temperature	1.0
Learning Rate	$1\text{e-}6$

Appendix

Table 5: Detailed hyperparameters for DR-CL training.

ID	Applicability (%)	Accuracy (%)	ID	Applicability (%)	Accuracy (%)
Cand_01 (R01)	99.1	98.3	Cand_11 (R04)	100.0	69.4
Cand_02 (R02)	99.4	99.4	Cand_12	78.4	94.8
Cand_03	18.8	98.1	Cand_13 (R05)	99.2	68.6
Cand_04	78.1	99.3	Cand_14	34.0	90.4
Cand_05	71.0	96.7	Cand_15	27.3	98.2
Cand_06	22.9	87.3	Cand_16	33.5	85.8

Appendix

Table 6: Statistics for the 20 rubric candidates. The model selects Candidates 01, 02, 07, 11, 13, and 17 to form the core reward metrics R01–R06. Candidate 20 serves as the equivalent of the ground truth accuracy for the final assessment.

Method	EvadeBench Acc. (%)
Qwen2.5-VL-7B-Instruct	43.90
Vanilla GRPO	44.47
RuCL	45.86

Appendix

Table 7: Performance evaluation on EvadeBench

Explore similar work

Seeing with You: Perception-Reasoning Coevolution for Multimodal Reasoning
Mar 30, 2026Ziqi Miao, Haonan Jia, Lijun Li +4Multimodal Reasoning Multimodal Large Language Models
Step-wise Rubric Rewards for LLM Reasoning
May 17, 2026Weichu Xie, Haozhe Zhao, Wenpu Liu +15Reward Modeling Credit Assignment in RL
Peking University · JD Explore Academy · Shanghai Jiao Tong University +2
Offline-Online Curriculum RL for Multimodal Reasoning
Jul 26, 2026Wendi Deng, Hang Du, Guoshun Nan +11Reinforcement Learning RL for Language Model Reasoning
Beijing University of Posts and Telecommunications, Beijing, 100876, China. · BUPT Shenzhen Institute, ShenZhen, 518038, China. · Carnegie Mellon University, Pittsburgh, USA. +4

Mar 30, 2026J/K move · Enter open · S save

Seeing with You: Perception-Reasoning Coevolution for Multimodal Reasoning

Ziqi Miao, Haonan Jia, Lijun Li, Chen Qian+3

Multimodal Reasoning Multimodal Large Language Models RL for VLMs Reinforcement Learning with Verifiable Rewards

Open paper AskPDF arXiv

Reinforcement learning with verifiable rewards (RLVR) has substantially enhanced the reasoning capabilities of multimodal large language models (MLLMs). However, existing RLVR approaches typically rely on outcome-driven optimization that updates both perception and reasoning using a shared reward based solely on the final answer. This shared reward blurs credit assignment, frequently improving reasoning patterns while failing to reliably enhance the accuracy of upstream visual evidence extraction. To address this perception bottleneck, we introduce PRCO (Perception-Reasoning Coevolution), a dual-role RLVR framework with a shared policy. PRCO consists of two cooperative roles: an Observer that generates an evidence caption tailored to the question and a Solver that predicts the final answer based on this caption. Crucially, PRCO employs role-specific reward signals: the Solver is optimized using verifiable outcome rewards on the final answer, while the Observer receives a utility reward derived from the Solver's downstream success. Extensive experiments across eight challenging multimodal reasoning benchmarks demonstrate that PRCO yields consistent improvements across model scales by over 7 points on average accuracy compared to the base model, outperforming prior open-source RL-tuned baselines.

Abstract

Explore similar work

Seeing with You: Perception-Reasoning Coevolution for Multimodal Reasoning

Step-wise Rubric Rewards for LLM Reasoning

Offline-Online Curriculum RL for Multimodal Reasoning