GX Xu

No training needed! While improve your SLM to o-1 level math performance with particle filtering based inference scaling 🎊 Paper, code, and blogpost are all released 🔥 If this interesting to you, please like share, comment to support open-sourcing cutting-edge AI at RedHat!

Random Samples, our weekly seminar series that bridges the gap between cutting-edge AI research and real-world application, continue this Friday, July 18! Title: Grounding Feedback is All You Need: Aligning Small Vision-Language Models Abstract: While recent vision-language models (VLMs) excel at integrating visual and linguistic information, their performance hinges on vast quantities of curated image-text pairs. This reliance makes the alignment process both time-consuming and resource-intensive. In this talk, we'll introduce Sampling-Based Vision Projection (SVP), a novel framework that improves vision-language alignment using automated feedback and minimal human supervision. Our results show that SVP significantly enhances image captioning, improves object recall, and reduces hallucination, enabling smaller models to match the performance of much larger systems. This approach offers a promising path toward developing powerful, efficient, and accessible multimodal AI. Watch live on YouTube: youtube.com/watch?v=ddnRm-… Or join us via Google Meet: meet.google.com/teh-ijwx-rbz

LLM inference is too slow, too expensive, and too hard to scale. 🚨 Introducing llm-d, a Kubernetes-native distributed inference framework, to change that—using vLLM (@vllm_project), smart scheduling, and disaggregated compute. Here’s how it works—and how you can use it today:

[1/x] 🚀 We're excited to share our latest work on improving inference-time efficiency for LLMs through KV cache quantization---a key step toward making long-context reasoning more scalable and memory-efficient.

@nataninvesting 1. Supplier are liable to risk of excessive inventory which temu asks for. 2. Suppliers profit take a huge dent with free returns and customer quality complaints. 3 top suppliers may dump temu for SHEIN or Amazon, and temu will be stuck in low quality products

@GX_NLP Why do you think it could be not sustainable?

$PDD price hasn't changed in last year. In the same time: - 87% Revenue growth y/y - 130% EPS growth y/y - 55% FCF growth y/y Stock is down following an insane multiple compression: - 23x PE 1 year ago - Only 11x PE today

@nataninvesting Give the risk, I still think it’s a strong buy.

.@RedHat AI Innovation team just dropped a new research paper on inference-time scaling! 🚨 All built on @vllm_project. Paper and code here: …abilistic-inference-scaling.github.io Cheers to paper authors @variational_i, @xukai92, @GX_NLP, Shivchander Sudalairaj, and @ishapuri101!

[1/x] can we scale small, open LMs to o1 level? Using classical probabilistic inference methods, YES! Joint @MIT_CSAIL / @RedHat AI Innovation Team work introduces a particle filtering approach to scaling inference w/o any training! check out …abilistic-inference-scaling.github.io

Congrats @GoogleDeepMind on the new Gemma-2 27B & 9B release! Gemma-2 was tested in the Arena under the codename "*late-june-chatbots" and now out of stealth. Its early result matches the best open models (Llama-3-70B, Nemotron-340B) with only 27B parameters! Impressively, Gemma-2-9B is ranked as high as Qwen-2-72B. The rate of improvement is fast. Great news to the open community! More links below.

@bindureddy To get rich, you need to do things that scale. Plumbing don’t scale…

TBH new college entrants should really consider acquiring some blue collar skills like electrical or plumbing! AI may not “replace” white collar jobs but it will most definitely reduce the need for so many writers, designers and programmers, decreasing wages in white collar professions. In the future, you won’t get rich with a CS degree, but you may by mastering plumbing! 😀

@sunjiao123sun_ Amazing 🤩

Honored to receive the 🥇BEST PAPER AWARD🥇 from CVPR 2024, please consider using our collected fine grained feedback! Huge shout out to our work DreamSync, the key method that we use for using the fine grained feedback to improve the model, detail in my pined tweet! 🚀

IBM’s Agent101 tops AWE Benchmark! Amazing work led by @aviaviavi__

A new RL alignment method, here’s Gaurav’s excellent blog that explains why BRAIn is more stable and gives better performance than PPO and DPO 🔥

TLDR: Looking for a RLHF method that combines the best of PPO and DPO, stable training, and gives amazing result? BRAIN theoretically unites DPO and PPO, and empirical shown to out-perform! An earlier pre-print of the ICML paper is available now🔥

@Sauers_ I tried again today, it seems to got fixed

@GX_NLP @RuiboLiu It did not fall for it for me

Even powerful LLM like Claude3 Opus breaks with the simplest attacks to start hallucinating about “non-existing” context about “steps”. The kind of mistake that a human 5 year old wouldnt make. 😉

A personal note: Unitxt originated within the @LChoshen fusing team, aiming to streamline the sharing of academic outputs, primarily through model weights but also data. In the process of training various models on numerous datasets, we encountered significant challenges related to data inconsistency and a lack of preprocessing transparency. The decision was made! Together with @LChoshen @Shachar_Don There’s a better way to share data preprocessing techniques.

Raw footage of Devin! Expect to see more soon 👀

I'd given up using ChatGPT for all but the most basic tasks -- I just wasn't getting answers that were good enough to be of practical use to me. But Claude 3 Opus is being genuinely useful, and it's making me use LLM chat again. Thanks @AnthropicAI!