machine learning researcher • organizer @afciworkshop
Joined September 2018
- Tweets468
- Following340
- Followers798
- Likes1.3K
Awa Dieng retweeted
Avoiding common failures in AI for health and medicine
A reminder that benchmark performance is not deployment performance.
The paper focuses on erroneous outputs, unjustified performance gaps, deployment-time degradation, monitoring, and governance.
cell.com/cell/fulltext/S0092…
Awa Dieng retweeted
We’re coming out of stealth with some exciting AI-optimized compute results.
Some highlights:
1) Works on industrial-grade and consumer-grade GPUs.
2) It can optimize different compute layers, from the kernels to tokens, for higher tokens-per-watt on your GPU.
The entire industry is racing to make AI cheaper. Most people assume the answer is better chips or bigger GPU clusters. However, while the industry is buying its next 2× when there's a 1.78× speedup is sitting in the software.
We've been hard at work @unifiedsciences, building an autonomous optimization engine toward radically more efficient compute.
Today I'm very happy to share some early results where our unified engine:
1) wrote faster GEMM kernels for cuBLAS (up to 1.78x faster)
2) made an efficient AI model 6x smaller and 5x faster without any quality loss
3) optimized diverse everyday AI workloads on a consumer GPU!
Awa Dieng retweeted
Congratulations to AI2050 Early Career Fellow @adjiboussodieng on receiving an NSF CAREER Award for her research developing AI methods for materials discovery, with applications ranging from next-generation medicines to improved solar cells.
Read more: cs.princeton.edu/news/dieng-…
Awa Dieng retweeted
🚨 CFP is LIVE! 🚨
DAIH 2026 Workshop @ COLM — LLM/VLM Deployment Opportunities and Risks in Healthcare 📍 San Francisco | 📅 Oct 9, 2026
We want your research papers & real-world deployment case reports!
⏰ Deadline: June 23, 2026
🔗 daih2026.github.io/index.htm… (1/X)
Awa Dieng retweeted
That’s a wrap for #AFAA2026!!! 🎉 Thank you to all the authors, attendees, roundtable leads and speakers for the great presentations and insightful discussions!
1/4
Awa Dieng retweeted
Incredible work by @BediSuhana42170 and team on HealthAdminBench. Benchmarks that only measure steps will miss end-to-end performance. Great example of what it looks like to actually design for deployment conditions. arxiv.org/abs/2604.09937, Code: github.com/som-shahlab/healt…
Agent demos meet reality in healthcare.
On 𝐇𝐞𝐚𝐥𝐭𝐡𝐀𝐝𝐦𝐢𝐧𝐁𝐞𝐧𝐜𝐡, the best agent achieved only 𝟑𝟔.𝟑% end to end success. Reliability matters.
Paper
arxiv.org/abs/2604.09937
Leaderboard
healthadminbench.stanford.ed…
@stai_research @StanfordAILab
Awa Dieng retweeted
Hey world! I am looking for a Research Scientist (L4) team match at @GoogleDeepMind and @Google 🚀
I build things with RLHF, Model Evaluation, and Computer Using Agents — from agentic pipelines to RL systems.
If you're recruiting, let's talk! My DM is OPEN
Awa Dieng retweeted
We've been working on the Waxal dataset project since 2021, aiming to enhance the amount of data available for African languages. This public speech dataset initially covers 27 Sub-Saharan African languages spoken by over 100 million speakers across more than 26 countries. 🌍
This quoted post is unavailable.
Awa Dieng retweeted
Schmidt Sciences grantees @adjiboussodieng, an AI2050 Fellow, and @Yoshua_Bengio, a Science of Trustworthy AI grantee, were recently appointed by the @UN General Assembly to serve on the inaugural Independent International Scientific Panel on AI: un.org/independent-internati…
Awa Dieng retweeted
Faculty members @adjiboussodieng and @korolova have been named to the @UN Independent International Scientific Panel on AI.
Composed of 40 experts, the panel is the first global scientific body dedicated entirely to artificial intelligence.
bit.ly/3MuPGRY
ALT Adji Bousso Dieng and Aleksandra Korolova
Awa Dieng retweeted
Singular Bayesian Neural Networks
Mame Diarra Toure, David A. Stephens
arxiv.org/abs/2602.00387 [𝚜𝚝𝚊𝚝.𝙼𝙻 𝚌𝚜.𝙻𝙶 𝚜𝚝𝚊𝚝.𝙰𝙿]
Awa Dieng retweeted
🚨 Deadline Extended to Feb 5 (AoE)!
CFP still OPEN for the #AFAA2026 Workshop at @iclr_conf — on fairness across alignment & agentic AI systems.
Full & tiny papers welcome • Interdisciplinary work encouraged!
🔗 afciworkshop.org
#ICLR2026 #AFAA2026
🚨 CFP OPEN!
We’re launching the #AFAA2026 Workshop at @iclr_conf on fairness across alignment and agentic AI systems.
Submit full or tiny papers!
Interdisciplinary work especially welcome :D
🗓 Deadline: Jan 31 (AoE) | 🔗 afciworkshop.org/
#AFAA2026 #ICLR2026
Awa Dieng retweeted
Major release from our group today: “End-to-End Test-Time Training” (TTT-E2E). We show that continual learning enables strong long context performance without changing the underlying architecture!
LLM memory is considered one of the hardest problems in AI.
All we have today are endless hacks and workarounds. But the root solution has always been right in front of us.
Next-token prediction is already an effective compressor. We don’t need a radical new architecture. The missing piece is to continue training the model at test-time, using context as training data.
Our full release of End-to-End Test-Time Training (TTT-E2E) with @NVIDIAAI, @AsteraInstitute, and @StanfordAILab is now available.
Blog: nvda.ws/4syfyMN
Arxiv: arxiv.org/abs/2512.23675
This has been over a year in the making with @arnuvtandon and an incredible team.
Awa Dieng retweeted
Princeton researchers have developed a new tool to speed the discovery of advanced materials known as metal organic frameworks potentially leading to better battery chemistry, more efficient carbon capture and improved access to clean water. bit.ly/4ptp3u6
Awa Dieng retweeted
With support from @StanfordHAI, a team of Stanford researchers investigates when and why current domain generalization benchmarks fall short and how they might be improved: stanford.io/4kTMQl3
Awa Dieng retweeted
Our paper introducing Alternators has been accepted at TMLR. Alternators are a new class of generative models that outperform state-space models, such as Mambas, and diffusion models such as ScoreGrad. Check it out!
GitHub: github.com/vertaix/Alternato…
Paper: openreview.net/forum?id=Q70C…
Awa Dieng retweeted
#NewPaper What does the concept of information have to do with diversity? Introducing the Vendi Information Gain, a new way of quantifying information in science and machine learning using diversity ☀️
🗝️Shannon introduced mutual information (MI) in 1948 in his seminal paper titled "A Mathematical Theory of Communication," which birthed information theory. MI has since pervaded science and engineering. However, MI has several limitations that become more pronounced in high dimensions and/or under limited data.
💡Our new paper proposes a flexible alternative to MI that scales nicely to high dimensions and works well under limited data settings. We call it the Vendi Information Gain (VIG). VIG is defined as the difference between the marginal Vendi entropy of a random variable and its conditional Vendi entropy given another variable. Here we use "Vendi entropy" to mean the logarithm of the Vendi Score.
🟩 VIG has many nice properties, all of which we describe in the paper. In particular, it generalizes MI. MI appears as a special case of VIG when you assume there are no similarities between samples. Importantly, VIG doesn't require specifying a probability distribution over samples like MI does: it works directly with samples! Finally, VIG is asymmetric, reflecting the natural asymmetry of information. We demonstrate other properties in the paper.
🆕VIG allows us to develop a novel information-theoretic framework for active data acquisition, a paradigm widely used in data-driven science. This framework outperforms STRADDLE, considered SOTA in level-set estimation.
🧾We demonstrate VIG empirically in many applications, including in cognitive science to model human response times to external stimuli, in epidemiology to model epidemic processes and detect disease hotspots in many countries via level-set estimation. #InformationTheory #VendiScoring #Vertaix #Princeton @PrincetonCS
Link to paper: arxiv.org/abs/2505.09007
Authors: Quan Nguyen (@the_subtrahend) and Adji Bousso Dieng (@adjiboussodieng)
Awa Dieng retweeted
📣 Proud to share HealthBench, an open-source benchmark from our Health AI team at OpenAI, measuring LLM performance and safety across 5000 realistic health conversations. 🧵
Unlike previous narrow benchmarks, HealthBench enables meaningful open-ended evaluation through 48,562 unique physician-written rubric criteria spanning several health contexts (e.g., emergencies, global health) and behavioral dimensions (e.g., accuracy, instruction following, communication).
Blog, paper, code: openai.com/index/healthbench…
Awa Dieng retweeted
Our LLM4Mat-Bench paper is now published @MLSTjournal ✨ Test your favorite LLM on the benchmark to predict material properties.
📖Paper: iopscience.iop.org/article/1…
💻Code: github.com/vertaix/LLM4Mat-B…
#NewPaper Have you been wondering how your favorite LLM, e.g. Llama, Mistral, or Gemma performs on materials property prediction? We have just released LLM4Mat-Bench, an extensive benchmark for materials property prediction with LLMs!
LLM4Mat-Bench has unique features:
☀️It spans 10 data collections, containing more than 2.6 Million data points.
☀️It covers 45 distinct material properties.
☀️It covers three different material representations: CIF, text description, and composition.
☀️It provides baseline results from different types and sizes of LLMs, e.g. Llama, Mistral, Gemma, MatBERT, and LLM-Prop.
With materials data scattered everywhere, we believe LLM4Mat-Bench represents a unified data source for driving research on leveraging LLMs for materials science. The benchmark will be maintained and we look forward to your task and data contributions.
Our @andre_niyongabo will present the paper at the AI4Mat #NeurIPS2024 workshop this December.
Paper: arxiv.org/abs/2411.00177
Code: github.com/vertaix/LLM4Mat-B…
Authors: Andre Niyongabo Rubungo (@andre_niyongabo), Kangming Li (@KangmingLi_), Jason Hattrick-Simpers, and Adji Bousso Dieng (@adjiboussodieng)
#AI4Materials #MatSci #NLP4Science #Benchmarks #LLMs #Vertaix
