Research map / Label-efficient and robust learning / Test-time adaptation
DISCODE: Distribution-Aware Score Decoder for Robust Automatic Evaluation of Image Captioning
Nakamasa Inoue, Kanoko Goto, Masanari Oi, Martyna Gruszka, Mahiro Ukai, Takumi Hirose, Yusuke Sekikawa
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 |
|---|---|---|---|
| replace token-probability score smoothing with ATT-loss Gaussian-prior decoding | LLaVA-Next-72B evaluator · MCEval mean | accuracy: 82.5 -> 83.6 | Improved |
| add cross-entropy term to the ATT loss (ablation) | DISCODE · MCEval | accuracy: 81.8 -> 83.6 | Improved |
| add divergence term to the ATT loss (ablation) | DISCODE · Flickr8k-Expert | Kendall tau-c: 49.9 -> 56.1 | Improved |
| use adaptive alpha based on raw score instead of fixed (ablation) | DISCODE · MCEval | accuracy: 83.0 -> 83.6 | Improved |
| use continuous 0.0 to 1.0 scale instead of 0 to 9 (ablation) | DISCODE · Flickr8k-Expert | Kendall tau-c: 55.4 -> 56.1 | Improved |
| use numeric scale instead of letter scale A to E (ablation) | DISCODE · Flickr8k-Expert | Kendall tau-c: 54.4 -> 56.1 | Improved |
| use 1 to 5 scale instead of continuous scale (ablation) | DISCODE · MCEval | accuracy: 83.6 -> 83.6 | No clear effect |
| use weighted KL instead of standard KL divergence (ablation) | DISCODE · MCEval | accuracy: 83.0 -> 83.6 | Improved |
| use weighted KL instead of Jensen-Shannon divergence (ablation) | DISCODE · MCEval | accuracy: 81.6 -> 83.6 | Improved |
| apply DISCODE to InternVL-2.5-78B instead of FLEUR smoothing | InternVL-2.5-78B · MCEval | accuracy: 74.0 -> 78.1 | Improved |
| apply DISCODE to Qwen2-VL-7B instead of FLEUR smoothing | Qwen2-VL-7B-Instruct · MCEval | accuracy: 75.2 -> 75.2 | No clear effect |
| apply DISCODE to MiniCPM-V-2.6 instead of FLEUR smoothing | MiniCPM-V-2.6 · Composite | Kendall tau-c: 66.2 -> 66.2 | No clear effect |
Closest papers by meaning
- Painting with Words: Elevating Detailed Image Captioning with Benchmark and Alignment Learning (ICLR 2025 · similarity 0.87)
- CAPability: A Comprehensive Visual Caption Benchmark for Evaluating Both Correctness and Thoroughness (NeurIPS 2025 · similarity 0.86)
- BalCapRL : A Balanced Framework for RL-Based MLLM Image Captioning (NeurIPS 2026 · similarity 0.86)
- G-VEval: A Versatile Metric for Evaluating Image and Video Captions Using GPT-4o (AAAI 2025 · similarity 0.86)
- CaptionQA: Is Your Caption as Useful as the Image Itself? (CVPR 2026 · similarity 0.86)
- CapsFusion: Rethinking Image-Text Data at Scale (CVPR 2024 · 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.
