Hasan Hammoud

Just released "DiffCLIP", extending Differential Attention proposed by @ytz2024 to CLIP models - replacing both visual & text encoder attention with the differential attention mechanism! TL;DR: Consistent improvements across all tasks with only 0.003% extra parameters!

AraLingBench offers a vital diagnostic for developing Arabic LLMs with true linguistic mastery. Access the paper & dataset to empower your research: Paper: huggingface.co/papers/2511.14… Dataset: huggingface.co/datasets/hammh…

Unveiling AraLingBench: Deep Linguistic Evaluation for Arabic LLMs A new human-annotated benchmark of 150 expert-designed questions. It stress-tests grammar, morphology, spelling, comprehension & syntax, revealing LLMs often rely on memorization over true linguistic understanding.

🔥 Daily AI Paper (2025-11-19) 📄 AraLingBench A Human-Annotated Benchmark for Evaluating Arabic Linguistic Capabilities of Large Language Models 🔗 chatpaper.ai/dashboard/pape… #AI #ML #ChatPaper

🔥 Daily AI Paper (2025-09-18) 📄 Hala Technical Report: Building Arabic-Centric Instruction & Translation Models at Scale 🔗 chatpaper.ai/dashboard/pape… #AI #ML #ChatPaper

1. Hala Technical Report: Building Arabic-Centric Instruction & Translation Models at Scale 🔑 Keywords: Arabic-centric, Hala, translate-and-tune pipeline, lightweight language model, NLP 💡 Category: Natural Language Processing 🌟 Research Objective: - The primary goal is to develop Arabic-centric instruction and translation models that achieve state-of-the-art results using advanced methodologies. 🛠️ Research Methods: - Utilized a translate-and-tune pipeline, compression to FP8, and slerp merging, alongside fine-tuning a lightweight language model on bilingual supervision. 💬 Research Conclusions: - Hala models, trained with varying parameters from 350M to 9B, deliver state-of-the-art performance on Arabic-centric benchmarks, publishing resources to further Arabic NLP research. 👉 Paper link: huggingface.co/papers/2509.14…

Hala: New Arabic-centric models released on Hugging Face A family of state-of-the-art instruction and translation models, built with a novel translate-and-tune pipeline. Achieves SOTA performance in "nano" (≤2B) and "small" (7-9B) categories on Arabic benchmarks.

We just released Hala: open, state-of-the-art Arabic instruction & translation models! ✨ Includes: • 1.2B Translation model (very light-weight) • 4.6M Arabic Instruction Tuning Dataset • 4 models (350M–9B) 📄 Paper: huggingface.co/papers/2509.14… Don't forget to upvote :)!! 🤗 Models & Data: huggingface.co/collections/ha… 💻 GitHub: github.com/hammoudhasan/H… 🙏 Big thanks to my co-authors Mohammad Zbeeb and @BernardSGhanem !

We released 44B synthetic tokens from our CoT-guided rewriting, offering higher quality pretraining data than the average human-written web texts📈 🤗Data: huggingface.co/datasets/faceb… 📜Paper: arxiv.org/abs/2506.04689 (accepted at #COLM2025) Excited to see what the community builds!

AI, decoded in under a minute. Prof. Bernard Ghanem @BernardSGhanem from #KAUST, ranked #1 in the Middle East for producing #AItalent, breaks it into four pillars. The expertise driving Saudi Arabia’s bold #AI future.

New paper out ! Train Long, Think Less. We introduce Curriculum GRPO, start with long reasoning chains, then progressively tighten token budgets to train LLMs that think better with fewer tokens. 📈 +Accuracy, 🔻Token usage, across GSM8K, MATH500 & more. Special thanks to all co-authors! @KumailAlhamoud, Abed Hammoud, Eli Bou-Zeid, @MarzyehGhassemi, @BernardSGhanem. Amazing collaboration between @KAUST, @MIT, and @Princeton. Paper: arxiv.org/abs/2508.08940 Code: github.com/hammoudhasan/c…

🚀 Our MatchDiffusion was accepted to ICCV 2025 in Hawaii! 🌺 We generate two synchronized videos from text prompts—designed for match-cuts. Results: matchdiffusion.github.io Paper: arxiv.org/abs/2411.18677 #MatchDiffusion #ICCV2025 #DiffusionModels #TextToVideo #GenerativeAI

Web data, the “fossil fuel of AI”, is being exhausted. What’s next?🤔 We propose Recycling the Web to break the data wall of pretraining via grounded synthetic data. It is more effective than standard data filtering methods, even with multi-epoch repeats! arxiv.org/abs/2506.04689

📢I am attending #CVPR2025 (Jun 11 - 14). Come to our snap-research.github.io/Omni-ID/ poster to know more about how we achieved the highest ID preservation in personalization and further enables expression following in our follow ups. See you at Fri 4 - 6 pm, ExHall D Poster #326.

📢Excited to share our new paper "Motion-Aware Concept Alignment for Consistent Video Editing". A training-free framework for video semantic mixing: 🔁Blend new concepts into specific objects 🎯Maintain spatial stability & temporal coherence 📊Outperform basselines A thread🧵

If you work on long-context compression for LLMs, you've seen the Gisting approach: add a few "gist tokens" and adjust the attention mask so all context flows into them. Elegant and simple… But we found that it COMPLETELY BREAKS when compressing more than just a few tokens 🤯

@itanih0 @BernardSGhanem Arxiv Paper is Out :) arxiv.org/abs/2504.20708

Special thanks to the co-authors @itanih0 and @BernardSGhanem for this amazing collaboration ! 🌐 Project Page & Demo: hammoudhasan.github.io/SubthoughtReas… (Contains a very cool demo!!) 📘 GitHub (with paper copy until arXiv): github.com/hammoudhasan/S…

Excited to share our new paper "Beyond the Last Answer: Your Reasoning Trace Uncovers More than You Think"! We show that digging into an LLM’s intermediate “subthoughts” and aggregating their answers can significantly boost math reasoning performance. A thread 🧵