AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Escaping the Likelihood Trap: Geometric Diversity Optimization for Long-Form Image Captioning

Qingmei Tang, Shuai Hao, Rong Fu, Zirui Mo, Xiang Liu, Jiaxuan Lu, Wenyu Wang

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks

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.

Identifies a "Likelihood Trap" in which MLE and KL-regularized alignment push VLM captions toward generic high-probability templates, and proposes a geometric diversity optimization approach to recover fine-grained detail in long-form image captions.

Problem

VLMs produce generic, template-like long-form image captions because standard alignment objectives (MLE and KL regularization) favor high-probability outputs and suppress fine-grained visual details.

Approach

Frames the caption-generation collapse as a likelihood trap caused by MLE and KL-regularization, and counters it with a geometric diversity optimization objective rather than likelihood-based alignment alone.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add DPP volume reward (ablation)Geo-RL · DeCapBenchDistinct-2: 0.72 -> 0.84Improved
add DPP volume reward (ablation)Geo-RL · DeCapBenchSVS: 2.65 -> 3.55Improved
add DPP volume reward (ablation)Geo-RL · DeCapBenchCHAIRi: 7.7 -> 7.1Improved
add contrastive repulsion (ablation)Geo-RL · DeCapBenchDistinct-2: 0.77 -> 0.84Improved
add contrastive repulsion (ablation)Geo-RL · DeCapBenchSVS: 2.95 -> 3.55Improved
use RBF kernel instead of linear kernel (ablation)Geo-RL DPP reward · DeCapBenchSVS: 2.85 -> 3.55Improved
use RBF kernel instead of polynomial (deg=2) kernel (ablation)Geo-RL DPP reward · DeCapBenchDistinct-2: 0.68 -> 0.84Improved
use Sentence-BERT instead of SimCSE encoder (ablation)Geo-RL · DeCapBenchSVS: 3.21 -> 3.55Improved
use 8 samples per image instead of 4 (ablation)Geo-RL · DeCapBenchSVS: 3.05 -> 3.55Improved
use 8 samples per image instead of 32 (ablation)Geo-RL · DeCapBenchSVS: 3.55 -> 3.32Hurt
train on 50K images instead of 10K (ablation)Geo-RL · DeCapBenchSVS: 3.01 -> 3.55Improved
train on 100K images instead of 50K (ablation)Geo-RL · DeCapBenchSVS: 3.55 -> 3.50No clear effect

Surprising findings

Experimental setup

Builds on: MLE training KL-regularized alignment

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