Research map / Reinforcement learning / RL from feedback
DIVA-GRPO: Enhancing Multimodal Reasoning through Difficulty-Adaptive Variant Advantage
Haowen Gao, zhenyu zhang, Liang Pang, Fangda Guo, douhongjian, Guannan Lv, ShaoGuo Liu, Tingting Gao, Huawei Shen, Xueqi Cheng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add adaptive variant generation (ablation) | DIVA-GRPO · Ablation average (MathVista/MathVerse/MMK12test) | accuracy (%): 61.6 -> 66.1 | Improved |
| add difficulty-weighted scaling (ablation) | DIVA-GRPO · Ablation average | accuracy (%): 64.0 -> 66.1 | Improved |
| add RRB-Rescaling (ablation) | DIVA-GRPO · Ablation average | accuracy (%): 63.8 -> 66.1 | Improved |
| add global-local balance (ablation) | DIVA-GRPO · Ablation average | accuracy (%): 64.1 -> 66.1 | Improved |
| add RRB-Rescaling to standard GRPO | GRPO · MathVista/MathVerse/MMK12test average | accuracy (%): 60.23 -> 62.17 | Improved |
| train with DIVA-GRPO instead of GRPO | Qwen3-VL-4B-Instruct · Math/Physics/Chemistry/Biology average | accuracy (%): 70.9 -> 73.4 | Improved |
| train with DIVA-GRPO instead of GRPO | Qwen3-VL-8B-Instruct · Math/Physics/Chemistry/Biology average | accuracy (%): 74.3 -> 78.5 | Improved |
| fine-tune with DIVA-GRPO | Qwen-2.5-VL-7B · six multimodal math benchmarks average | accuracy (%): 46.23 -> 54.58 | Improved |
| direct SFT on GPT-o3 reasoning traces | Qwen-2.5-VL-7B · six multimodal math benchmarks average | accuracy (%): 46.23 -> 44.23 | Hurt |
| generate variants with GPT-o3 instead of Qwen-Plus (ablation) | DIVA-GRPO · Math/Physics/Chemistry/Biology average | accuracy (%): 69.8 -> 70.2 | No clear effect |
| use relative instead of absolute difficulty weighting (ablation) | DIVA-GRPO · validation | accuracy | No clear effect |
| set sensitivity k to 0.1 rather than 0 or 1 (ablation) | DIVA-GRPO · validation | accuracy | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- R1-ShareVL-7B — R1-ShareVL: Incentivizing Reasoning Capabilities of Multimodal Large Language Models via Share-GRPO (NeurIPS 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- R1-ShareVL: Incentivizing Reasoning Capabilities of Multimodal Large Language Models via Share-GRPO (NeurIPS 2025 · similarity 0.92)
- Enhancing Multi-Modal LLMs Reasoning via Difficulty-Aware Group Normalization (ICML 2026 · similarity 0.90)
- SRPO: Enhancing Multimodal LLM Reasoning via Reflection-Aware Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- Multi-Task GRPO: Reliable LLM Reasoning Across Tasks (ICML 2026 · similarity 0.90)
- DeepVideo-R1: Video Reinforcement Fine-Tuning via Difficulty-aware Regressive GRPO (NeurIPS 2025 · similarity 0.90)
- DisCO: Reinforcing Large Reasoning Models with Discriminative Constrained Optimization (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
