Spaces:

---

huggingface

/

HuggingDiscussions

like 84

Running

https://arxiv.org/abs/2406.01954

We are excited to share our recent work on MLLM architecture design titled "Ovis: Structural Embedding Alignment for Multimodal Large Language Model".

Paper: https://arxiv.org/abs/2405.20797
Github: https://github.com/AIDC-AI/Ovis
Model: https://huggingface.co/AIDC-AI/Ovis-Clip-Llama3-8B
Data: https://huggingface.co/datasets/AIDC-AI/Ovis-dataset

@Yiwen-ntu for now we support only videos as paper covers in the Daily.

we are excited to share our work titled "Hierarchical Prompting Taxonomy: A Universal Evaluation Framework for Large Language Models" : https://arxiv.org/abs/2406.12644

Hi

Hi @AgPerry - I can remove the paper so you can upload an updated version. Let me know if that works for you.

@AdinaY So there is no way to edit the media, right?

@AdinaY I do not want to remove the paper, as this will lose the upvotes. So there is no way to edit the media, right?

It won't change the upvotes, you'll still have them iirc.

Thanks @AdinaY ! That's great to know. Please go ahead and remove it, and I'll re-upload with the correct gallery images.

Thanks @AdinaY ! That's great to know. Please go ahead and remove it, and I'll re-upload with the correct gallery images.

Done✅

Hi @meganariley @AdinaY , requesting a manual exception for Daily Papers submission — .

Paper: EMCompress: Video-LLMs with Endomorphic Multimodal Compression
arXiv: https://arxiv.org/abs/2508.21094 (v3, uploaded Apr 24, 2026)
Venue: ACL 2026 Findings (camera-ready)

The system blocks submission because v1 was uploaded Aug 2025. However, v3 is a completely rewritten paper, not a revision — the only thing shared with v1 is the arxiv ID:

• Everything fully rewritten: v1 was "Temporal Visual Screening for Video Question Answering (TVS)"; v3 is "EMCompress: Video-LLMs with Endomorphic Multimodal Compression (EMC)" — different title, different motivation / introduction / framing / formalization.
• New mathematical formulation: v3 introduces sufficient-statistic / Information Bottleneck framing (Markov chain A→(V,Q)→(v,q)A \to (V,Q) \to (v,q) A→(V,Q)→(v,q), classical sufficiency condition in VideoQA-natural form) — entirely absent in v1/v2.
• New experiments: v3 adds long-video benchmark evaluation (Video-MME, MLVU, LVBench, EgoSchema) and a full cost-efficiency analysis — neither was in v1.
• Benchmark restructured: YouCookII-TVS → EMCompress, repositioned as a standalone dual-task benchmark.

All five figures redrawn to reflect the new framework.

Peer-reviewed: this is the ACL 2026 Findings camera-ready, ~7 months after v2.

We kept the arxiv ID for citation continuity with the ACL submission metadata, but the work itself is new. Could you manually enable submission, or alternatively remove and re-index the paper page so we can resubmit during the v3 window? Happy to provide a v1↔v3 diff if helpful.

Thanks!