AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

ReflectCAP: Detailed Image Captioning with Reflective Memory

Kyungmin Min, Minbeom Kim, Kang-il Lee, Seunghyun Yoon, Kyomin Jung

ECCV 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Training-free token reduction for MLLMs · 1 citations (OpenAlex)

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 changedWhereMeasuredEffect
add Structured Reflection Notes (ReflectCAP)InternVL3.5-4B zero-shot · IIW-400F1: 59.1 -> 63.0Improved
add Structured Reflection Notes (ReflectCAP)Qwen3-VL-32B zero-shot · IIW-400F1: 75.7 -> 76.4Improved
add Structured Reflection Notes (ReflectCAP)GPT-4.1-mini zero-shot · CapArena-Autoscore: 57.7 -> 90.0Improved
add few-shot human exemplarsInternVL3.5-4B · IIW-400F1: 59.1 -> 51.1Hurt
add CapMAS multi-agent verificationGPT-4.1-nano · CapArena-Autoscore: 21.2 -> -14.7Hurt
add self-correctionQwen2.5-7B · IIW-400F1: 62.7 -> 60.4Hurt
add SFT on human-authored DOCCI captionsQwen2.5-VL-7B · IIW-400F1: 62.7 -> 57.6Hurt
add SFT on ReflectCAP-generated captionsQwen2.5-VL-7B · IIW-400F1: 62.7 -> 66.5Improved
add separate passes for Avoid and Include notes (vs combined prompt) (ablation)ReflectCAP on InternVL3.5-4B · IIW-400 (100 images)F1: 57.3 -> 64.4Improved
add separate passes for Avoid and Include notes (vs combined prompt) (ablation)ReflectCAP on GPT-4.1-mini · IIW-400 (100 images)F1: 77.1 -> 78.3Improved
add Grounded Base Caption (Avoid notes only) (ablation)Qwen2.5-VL family · IIW-400factuality: marginal improvement or degradationMixed
increase exemplar images N from 30 to 100 (ablation)ReflectCAP · IIW-400 (100 images)F1: plateau around N=30, slight decline at N=100No clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy