Research map / Vision-language and multimodal / Multimodal LLMs

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

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RuCL, which generates generalized rubrics, stratifies them by pass rate into foundational and advanced tiers, and schedules their weights during GRPO training on Qwen2.5-VL-7B to reach 60.06% average over seven visual reasoning benchmarks.

Problem

Outcome-only RLVR for MLLMs invites reward hacking, while instance-level rubric rewards are costly to generate and treat all rubrics as equally learnable, producing noisy gradients.

Approach

Applies curriculum learning to reward design rather than data: a reusable pool of generalized rubrics is filtered by applicability, stratified by pass rate into foundational and advanced tiers, and their weights are shifted over training by a performance-triggered sigmoid/linear schedule, combined with a rule-based answer reward in GRPO.

Key results

• RuCL reaches 60.06% average over seven benchmarks, a +7.83 point gain over Qwen2.5-VL-7B (52.23%).
• WeMATH improves from 58.52% to 71.49% and LogicVista from 39.26% to 49.66% over the base model.
• In the ablation on general benchmarks, RuCL (63.94 avg) beats Vanilla GRPO (59.14) and Linear Stratification (60.79).

Surprising findings

• Uniform averaging of rubrics gave LogicVista 50.11%, higher than RuCL's 49.66%, a case where the uniform baseline is better on one benchmark.

Limitations

• Rubric pool is fixed and generated offline; online rubric construction is left for future work.
• Results focus on 7B-scale models; scaling to larger architectures is not tested.
• Ablation and sensitivity results are reported with inconsistent numbers across tables (e.g., averages of 57.13/58.41 in text vs 59.14/60.79 in table).

Experimental setup

Datasets: ViRL-39K

Benchmarks: MathVista MathVerse MathVision WeMATH MMMU LogicVista Super-CLEVR Counting

Metrics: accuracy average accuracy

Baselines: Qwen2.5-VL-7B Qwen2.5-VL-32B InternVL2.5-8B InternVL2.5-38B GPT-4o Claude-3.5-Sonnet MM-Eureka-7B OpenVLThinker-7B Perception-R1-7B Vision-R1-7B R1-Onevision-7B ThinkLite-VL-7B

Builds on: GRPO RLVR curriculum learning verl

Closest papers by meaning

• Step-wise Rubric Rewards for LLM Reasoning (NeurIPS 2026 · similarity 0.91)
• Breaking the Exploration Bottleneck: Rubric-Scaffolded Reinforcement Learning for Open-Ended LLM Reasoning (ICML 2026 · similarity 0.89)
• Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.88)
• Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models (ICLR 2026 · similarity 0.88)
• Training Large Language Models for Reasoning through Reverse Curriculum Reinforcement Learning (ICML 2024 · similarity 0.87)
• Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.87)