AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

The Telephone Game: Evaluating Semantic Drift in Unified Models

Sabbir Mollah, Rohit Gupta, Swetha Sirnam, Qingyang Liu, Ahnaf Munir, Shah Mubarak

ECCV 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs

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
move from simple to complex compositional prompts (ablation)BAGEL · GenEval-R MGGMGG: 92.4 -> 75.9Hurt
move from simple to complex compositional prompts (ablation)Janus-7B · GenEval-R MGGMGG: 55.0 -> 14.9Hurt
move from simple to complex compositional prompts (ablation)VILA-U · GenEval-R MGGMGG: 49.5 -> 8.6Hurt
use BAGEL (shared-weight, large-scale interleaved pretraining) instead of VILA-Uunified model · Nocaps+Docci400MCD_avg: 0.2829 -> 0.4414Improved
replace CLIP cross-modal scorer with SigLIP2 (ablation)SDP MCD · Nocaps+Docci400MCD_avg: 0.4414 -> 0.3592No clear effect
use MAGViT2 tokenizer features instead of CLIP features (ablation)Show-o · Nocaps+Docci400 text-to-text MPNetMCD: 0.5749 -> 0.5957Improved
use Detailed 1 recaption prompt instead of Short (ablation)VILA-U · GenEval-R MGGMGG: 19.66 -> 24.65Improved
use Detailed 1 recaption prompt instead of Short (ablation)BAGEL · GenEval-R MGGMGG: 78.14 -> 76.62No clear effect
use OWLv2 detector instead of GroundingDINO (ablation)MGG · GenEval human auditagreement: 24.9 -> 81.0Improved
evaluate multiple generations (MGG) instead of single-pass GenEvalevaluation protocol · human T2I rankingPearson correlation: -0.753 -> -0.821Improved
use proprietary ChatGPT instead of BAGELunified model · small subsetMGG: 83.3 -> 93.0Improved

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