Search

Vikram Sharma
Vikram Sharma @v4vix ·
Replying to @v4vix
Multimodal Large Language Models (MLLMs) like MMLU and SwinBert are pushing the boundaries of language and vision integration. But, as Loc3R-VLM demonstrates, they still struggle with spatial understanding & viewpoint-aware reasoning. #VisionLanguage
14
AI Hot Sheets
AI Hot Sheets @aiHotSheets ·
🔥 VLMs struggle with multi-round visual reasoning, failing to iteratively refine understanding across visual contexts. 🌊 RegionReasoner enables iterative visual understanding via region-grounded multi-round reasoning. #AI #VisionLanguage #Reasoning arxiv.org/abs/2602.03733
arXiv logo
RegionReasoner: Region-Grounded Multi-Round Visual Reasoning

Large vision-language models have achieved remarkable progress in visual reasoning, yet most existing systems rely on single-step or text-only reasoning, limiting their ability to iteratively...

From arxiv.org
17
AI Hot Sheets
AI Hot Sheets @aiHotSheets ·
🔥 Multimodal models memorize visuals but fail to describe them in text. This "modal aphasia" challenges unified AI. 🌊 We reveal this dissociation: models recall images but can't articulate their content. @josh_swanson_ #AI #Multimodal #VisionLanguage arxiv.org/abs/2510.21842
arXiv logo
Modal Aphasia: Can Unified Multimodal Models Describe Images From Memory?

We present modal aphasia, a systematic dissociation in which current unified multimodal models accurately memorize concepts visually but fail to articulate them in writing, despite being trained...

From arxiv.org
57
LocalAI
LocalAI @LocalAI_API ·
🚨 New model alert! 🚨 We've got Qwen3-VL-8B-Instruct & Qwen3-VL-8B-Thinking added to LocalAI! 🎉 These are 8B parameter vision-language models. Try it out: `local-ai run qwen3-vl-8b-instruct` or `local-ai run qwen3-vl-8b-thinking` 🚀 #LocalAI #Qwen3 #VisionLanguage
221
Long Lian
Long Lian @LongTonyLian ·
👀 🏋️‍♂️ Train smarter, not just larger. VisGym’s scalable visual tasks reveal where VLMs still struggle and how to push them further. Try it out! #MachineLearning #VisionLanguage #VisGym
Zirui "Colin" Wang Zirui "Colin" Wang @zwcolin ·
🎮 We release VisGym: Diverse, Customizable, Scalable Environments for Multimodal Agents (w/ @junyi42 @aomaru_21490) 🌐 With 17 environments across multiple domains, we show systematically the brittleness of VLMs in visual interaction, and what training leads to. 🧵[1/ZPL
1
2.7K
Waseem M Ansari
Waseem M Ansari @wsmaisys ·
VL‑JEPA predicts the meaning of an image in a brain‑like sketch, using 50% fewer parameters and 2.85× less decoding work. Faster, leaner, and already beating CLIP on video tasks. #AI #VisionLanguage
17
Darshan Jain
Darshan Jain @i_darshanjain ·
Serving Qwen2-VL 7B with vLLM V1 on VisionArena benchmarks. At high QPS the V1 engine significantly outperforms V0. If you're still on the old architecture for multimodal workloads you're leaving perf on the table. #vLLM #VisionLanguage #Benchmark
8