Research map / Vision-language and multimodal / Multimodal LLMs
CCCaption: Dual-Reward Reinforcement Learning for Complete and Correct Image Captioning
Zhijiang Tang, Linhua Wang, Jiaxin Qi, Weihao Jiang, Peng Hou, Anxiang Zeng, Jianqiang Huang
Read the paperCheck your idea against 137,217 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.
CCCaption trains captioners with GRPO using completeness and correctness rewards.
Problem
Captioning references are incomplete or incorrect.
Approach
Dual reward combining completeness and correctness under GRPO.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add RL training with dual rewards to Qwen3-VL-2B | Qwen3-VL-2B · CapArena (avg) | win-minus-loss score: 41.39 -> 46.39 | Improved |
| Add RL training with dual rewards to Qwen3-VL-2B | Qwen3-VL-2B · Hallucinations evaluation (avg) | correctness score: 70.19 -> 70.57 | Improved |
| Use RL with completeness reward only (no correctness reward, as in CapRL) | Qwen2.5-VL-3B · Hallucinations evaluation (avg) | correctness score: 67.77 -> 66.91 | Hurt |
| Add correctness reward to completeness reward (ablation) | CCCaption-2B · ChartQA/OCR Prism and Hallucination avg | average score: 67.56 -> 70.07 | Improved |
| Add completeness reward to correctness reward (ablation) | CCCaption-2B · ChartQA/OCR Prism and Hallucination avg | average score: 65.93 -> 70.07 | Improved |
| Add dynamic query sampling (ablation) | CCCaption-2B · ChartQA/OCR Prism and Hallucination avg | average score: 69.15 -> 70.07 | Improved |
| Use correctness-only reward instead of completeness-only on OCR (ablation) | CCCaption-2B · OCRBench Prism | accuracy: 50.16 -> 40.45 | Hurt |
| Use multi-MLLM CCaption-44k queries instead of single-MLLM Vanilla queries (ablation) | Qwen3-VL-2B · Prism evaluation (4-benchmark avg) | average accuracy: 44.07 -> 45.94 | Improved |
| Scale CCaption training data from 8k to 44k samples (ablation) | Qwen3-VL-2B · Prism evaluation (4-benchmark avg) | average accuracy: 44.30 -> 45.94 | Improved |
| Scale CCaption training data from 16k to 44k samples (ablation) | Qwen3-VL-2B · Prism evaluation (4-benchmark avg) | average accuracy: 44.94 -> 45.94 | Improved |
| Train on CCaption-44k with RL instead of no RL (ablation) | Qwen3-VL-2B · Prism evaluation (4-benchmark avg) | average accuracy: 43.25 -> 45.94 | Improved |
| Add CCCaption training to 2B model vs CapRL-3B | CCCaption-2B · Prism evaluation (avg) | accuracy: 51.07 -> 52.80 | Improved |
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.
- CapRL-3B — CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 2026) · better on 3 of 3 shared results
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
- CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 2026 · similarity 0.93)
- BalCapRL : A Balanced Framework for RL-Based MLLM Image Captioning (NeurIPS 2026 · similarity 0.92)
- Clarification as Supervision: Reinforcement Learning for Vision-Language Interfaces (NeurIPS 2026 · similarity 0.89)
- Painting with Words: Elevating Detailed Image Captioning with Benchmark and Alignment Learning (ICLR 2025 · similarity 0.88)
- ViCrit: A Verifiable Reinforcement Learning Proxy Task for Visual Perception in VLMs (NeurIPS 2025 · similarity 0.88)
- Learning by Correction: Efficient Tuning Task for Zero-Shot Generative Vision-Language Reasoning (CVPR 2024 · similarity 0.88)
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.
