Research map / Vision-language and multimodal / Multimodal LLMs
Boosting MLLM Reasoning with Text-Debiased Hint-GRPO
Qihan Huang, Weilong Dai, Jinlong Liu, Wanggui He, Hao Jiang, Mingli Song, Jingyuan Chen, Chang Yao, Jie Song
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 |
|---|---|---|---|
| convert multi-choice questions to fill-in-the-blank in training data (ablation) | GRPO on Qwen2-VL-7B · Geo170K | accuracy (%): 35.81 -> 38.46 | Improved |
| inject hint in the answer rather than in the query (ablation) | Hint-GRPO on Qwen2-VL-7B · Geo170K | accuracy (%): 41.64 -> 45.62 | Improved |
| add adaptive hints (Hint-GRPO) to GRPO | Qwen2-VL-7B · Geo170K | accuracy (%): 38.46 -> 45.62 | Improved |
| add adaptive hints (Hint-GRPO) to GRPO | Qwen2.5-VL-3B · Geo170K | accuracy (%): 45.49 -> 53.32 | Improved |
| add text-bias calibration at test time (ablation) | Hint-GRPO on Qwen2-VL-7B · Geo170K | accuracy (%): 45.62 -> 46.68 | Improved |
| add text-bias calibration at test time (ablation) | Hint-GRPO on Qwen2.5-VL-3B · Geo170K | accuracy (%): 53.32 -> 55.31 | Improved |
| raise calibration strength gamma from 0.8 to 1.6 (ablation) | Debiased Hint-GRPO on Qwen2-VL-7B · Geo170K | accuracy (%): 46.68 -> 44.69 | Hurt |
| increase number of hint groups M from 1 to 3 (ablation) | Hint-GRPO on Qwen2-VL-7B · Geo170K | accuracy (%): 38.46 -> 45.62 | Improved |
| increase number of hint groups M from 3 to 4 (ablation) | Hint-GRPO on Qwen2-VL-7B · Geo170K | accuracy (%): 45.62 -> 45.89 | No clear effect |
| increase number of hint groups M from 3 to 4 (ablation) | Hint-GRPO on Qwen2.5-VL-3B · Geo170K | accuracy (%): 53.32 -> 52.65 | Hurt |
| use adaptive M=3 hint groups in place of M=1 (no hints) (ablation) | Hint-GRPO on Qwen2.5-VL-3B · Geo170K | accuracy (%): 45.49 -> 53.32 | Improved |
| use SFT instead of the original model | Qwen2-VL-7B · Geometry average (7 datasets) | accuracy (%): 30.40 -> 25.50 | Hurt |
Closest papers by meaning
- Multi-Task GRPO: Reliable LLM Reasoning Across Tasks (ICML 2026 · similarity 0.87)
- Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models (ICLR 2026 · similarity 0.86)
- R1-ShareVL: Incentivizing Reasoning Capabilities of Multimodal Large Language Models via Share-GRPO (NeurIPS 2025 · similarity 0.85)
- GRIT: Teaching MLLMs to Think with Images (NeurIPS 2025 · similarity 0.85)
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.85)
- DreamPRM: Domain-reweighted Process Reward Model for Multimodal Reasoning (NeurIPS 2025 · similarity 0.85)
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.
