Jiwei Liu

131 posts

Jiwei Liu

Jiwei Liu

@JiweiLiu

Katılım Mart 2013
217 Takip Edilen542 Takipçiler
Jiwei Liu retweetledi
Raja Biswas
Raja Biswas@raja_biswas·
NVIDIA AI-Q just hit #1 on DeepResearch Bench I & II. Our team built a multi-agent architecture, fine-tuned Nemotron 3 Super, added middleware for long-horizon reliability, and applied time-tested Kaggle strategies to make this happen! Solution writeup: huggingface.co/blog/nvidia/ho…
English
3
5
16
1K
Jiwei Liu retweetledi
JFPuget 🇺🇦🇨🇦🇬🇱
JFPuget 🇺🇦🇨🇦🇬🇱@JFPuget·
NVIDIA won Deep Research Bench with NVIDIA AI Q and Nemotron 3 models. A blog is under way.
JFPuget 🇺🇦🇨🇦🇬🇱 tweet mediaJFPuget 🇺🇦🇨🇦🇬🇱 tweet media
JFPuget 🇺🇦🇨🇦🇬🇱@JFPuget

Very proud that a cross NVIDIA effort led to top positions in Deep Research Bench I and II. DRB-I evaluates overall report quality: comprehensiveness, insight, instruction-following, and readability whereas DRB-II is more difficult and tests for information recall, analysis, and presentation. This was led by David Austin and @raja_biswas from my team. I contributed a bit. We will share more ASAP. huggingface.co/spaces/muset-a… #leaderboard" target="_blank" rel="nofollow noopener">agentresearchlab.com/benchmarks/dee…

English
3
16
94
13.1K
Jiwei Liu
Jiwei Liu@JiweiLiu·
Our team finished 12th in the Kaggle Santa 2025 competition. As someone with little experience in optimization, I’m very grateful to my awesome teammates and Claude Code!
Jiwei Liu tweet media
English
1
0
1
138
Jiwei Liu retweetledi
Tony Mongkolsmai
Tony Mongkolsmai@tonymongkolsmai·
Today we are releasing our first public beta of Nsight Python! The goal is to simplify the life of a Python developer by proving a pythonic way to analyze your kernel code! Check it out, provide feedback! Nsight Python — nsight-python docs.nvidia.com/nsight-python/
English
10
48
341
29.4K
Jiwei Liu
Jiwei Liu@JiweiLiu·
@JFPuget It’s still a remarkable achievement! ARC needs another leaderboard for score/dollar
English
1
0
22
3.9K
JFPuget 🇺🇦🇨🇦🇬🇱
JFPuget 🇺🇦🇨🇦🇬🇱@JFPuget·
I have mixed feelings about Gemini 3 Pro results on arc agi 2. 1. They are impressive. 45% is impressive on that benchmark. No doubt. 2. They come at a much higher cost than our 29.72% at $0.2 per task. I wonder what we'd get with their compute budget. 3. We could not really enjoy our 15 mins fame as our top score did not make it officially to the ARC AGI2 leaderboard. It may be verified after Dec 5th, but this is not timely anymore. I don't mind to be beaten by a stronger system, but item 3 is somewhat frustrating. So be it. I hope we'll still win arc prize competition. That would be a nice positive conclusion.
English
33
10
561
122.2K
JFPuget 🇺🇦🇨🇦🇬🇱
JFPuget 🇺🇦🇨🇦🇬🇱@JFPuget·
I took the first place on @arcprize @kaggle competition with my colleague Ivan Sorokin. These are preliminary results only. Final results will be available on Dec 5th. We may drop from first place as our code will be evaluated on a different set of challenges. Nevertheless, this is a somewhat unexpected result for us. We climbed fast during the last week, but topping the leaderboard was just a wild dream. We will open source our solution and share what we did. Stay tuned. Note that our score can be compared to the scores from the arc-agi leaderboard. We crush SOTA LLMs. Best one, GPT5 Pro, has a score of 18.3% vs our 27.7%. And we get it at a tiny fraction of the cost because we run within Kaggle compute constrains (12 hours with 4 L4 GPUs) kaggle.com/competitions/a…
JFPuget 🇺🇦🇨🇦🇬🇱 tweet media
English
36
48
522
84.1K
Jiwei Liu retweetledi
Vibhu Jawa
Vibhu Jawa@JawaVibhu·
1 TB of data on a single Grace Hopper GPU 🔥
NVIDIA AI Developer@NVIDIAAIDev

🌟 XGBoost 3.0 now has new capabilities for GPUs: ✅ Train up to 1 TB of data on a single Grace Hopper GPU: Up to 8x faster than CPU with External-Memory Quantile DMatrix. ✅ Faster and lighter: GPU hist/approx methods an additional ~2x speed up and reduced memory use. ✅ Full feature support: External memory now supports categorical features, all objectives, and SHAP. ✅ Distributed mode: Experimental support for out-of-core training across a cluster. Technical deep dive ➡️ nvda.ws/4le4ZcX

English
0
1
2
271
Jiwei Liu retweetledi
Dieter
Dieter@kagglingdieter·
Why are there only 362 @kaggle Competition Grandmasters worldwide? Simply because it takes an insane amount of experience and effort to reach that tier. Even with a PhD in math, it still took me 4,000 hours to achieve Kaggle GM. I needed 2,000 more hours before winning my first competition, and over 10,000 hours before reaching #1 on the overall Kaggle ranking.
English
21
64
1.3K
144.2K

Keşfet

@MJeblick @jackyu815 @JFPuget @arcprize @kaggle @ChrisDeotte @The0Viel @0verfit