Research map / Vision-language and multimodal / Visual question answering
Beyond Multiple Choice: Verifiable OpenQA for Robust Vision-Language RFT
Yesheng Liu, Hao Li, Haiyu Xu, Baoqi Pei, Jiahao Wang, Mingxuan Zhao, Jing-Shu Zheng, Zheqi He, JG Yao, Xi Yang, Bowen Qin, Jiajun Zhang
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 |
|---|---|---|---|
| RFT on MCQA data instead of base model (ablation) | Qwen2.5-VL-7B · MMMU | MCQA-OpenQA gap: 30.2 -> 39.3 gap; open 21.4 -> 17.1 | Hurt |
| RFT on MCQA data instead of base model (ablation) | Qwen2.5-VL-3B · MMLU-Pro | MCQA-OpenQA gap: 18.4 -> 27.0 | Hurt |
| train on ReVeL-rewritten OpenQA instead of MCQA (ablation) | Qwen2.5-VL-7B · Overall open-ended accuracy | OpenQA accuracy: 24.7 -> 34.0 | Improved |
| train on ReVeL-rewritten OpenQA instead of MCQA (ablation) | Qwen2.5-VL-7B · Overall MCQA accuracy | MCQA accuracy: 47.8 -> 46.8 | No clear effect |
| train on ReVeL-rewritten OpenQA instead of MCQA (ablation) | Qwen2.5-VL-3B · Overall score | overall score: 30.1 -> 34.3 | Improved |
| train on ReVeL-rewritten OpenQA instead of original ViRL OpenQA (ablation) | Qwen2.5-VL-3B · Overall open-ended accuracy | OpenQA accuracy: 21.8 -> 28.0 | Improved |
| add original ViRL OpenQA data on top of ReVeL OpenQA (ablation) | Qwen2.5-VL-7B · Overall score | overall score: 40.4 -> 39.8 | No clear effect |
| train InternVL3-4B on ReVeL OpenQA | InternVL3-4B · Overall score | overall score: 42.4 -> 42.8 | No clear effect |
| use rule-based hybrid judging instead of pure LLM judge | ReVeL evaluation · 600 sampled responses | judge accuracy: 97.3 -> 98.5 | Improved |
| add answer options to open-ended questions | frontier LLMs and VLMs · SimpleQA, VisualSimpleQA | accuracy | Improved |
| replace ground-truth option with 'None of the above' | VLMs · MMLU-Pro, MMMU | reasoning-answer mismatch rate: 18% -> 50% | Hurt |
| switch evaluation from MCQA to OpenQA | GPT-5 · MMMU | accuracy: 79.2 -> 59.5 | Hurt |
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-OneVision-7B — R1-Onevision: Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization (ICCV 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
- Automated Generation of Challenging Multiple-Choice Questions for Vision Language Model Evaluation (CVPR 2025 · similarity 0.90)
- FRAMES-VQA: Benchmarking Fine-Tuning Robustness across Multi-Modal Shifts in Visual Question Answering (CVPR 2025 · similarity 0.89)
- Answer, Assemble, Ace: Understanding How LMs Answer Multiple Choice Questions (ICLR 2025 · similarity 0.88)
- Rethinking Generative Large Language Model Evaluation for Semantic Comprehension (ICML 2024 · similarity 0.88)
- MoHoBench: Assessing Honesty of Multimodal Large Language Models via Unanswerable Visual Questions (AAAI 2026 · similarity 0.88)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.87)
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.
