Research lab @Princeton led by @adjiboussodieng. We develop theory and computational methods to tackle challenges in AI and the natural sciences.
Joined February 2022
- Tweets316
- Following98
- Followers1.8K
- Likes980
Pinned Tweet
New tool narrows the search for ideal material structures, via @EurekAlert! @eprinceton eurekalert.org/news-releases…
Vertaix® (AI & Science) retweeted
Congrats to @adjiboussodieng on being selected by @theNAEng to participate in the Grainger Foundation Frontiers of Engineering EU-US Symposium! 🎉
The symposium brings together engineers from the US and the EU to share ideas and build connections.
bit.ly/4Bzx1sN
ALT Adji Bousso Dieng
The Vendi Clustering paper is now published online. Check it out! aclanthology.org/2026.findin…
#NewPaper accepted in #EMNLP findings!
🔥 Clustering is a fundamental task in machine learning and science. In our new paper, we show you how you can do clustering using diversity as your mathematical object! We refer to this new algorithm as *Vendi Clustering*.
🧠 Instead of asking which two data points are closest, Vendi Clustering asks: which merge of datapoints causes the smallest loss in the diversity of the whole dataset? We optimize that global objective using the Vendi Score. When applied naively, the idea incurs an O(n^6) cost, which is prohibitive. Fortunately, we were able to derive an efficient version that reduces the naive intractable complexity of Vendi Clustering from O(n^6) to O(n^3).
📚 Vendi Clustering finds a natural application to the task of topic reduction. Embedding-based topic models like BERTopic often produce hundreds of fine-grained, overlapping topics. Topic reduction decides what structure survives in the final topic model. Poor topic reduction can merge genuinely distinct themes, preserve redundant ones, or discard useful data as noise. Existing topic reduction methods either make local pairwise merge decisions, require trial-and-error retuning of the clustering, or rely on expensive stochastic LLM prompting.
☀️ Vendi Clustering provides an efficient way to do topic reduction while preserving topic diversity. Some highlights from the results:
📌 0.65 -> 0.88 topic diversity (+35%) over topic reduction assisted by the Gemma-3-12b-it LLM on 20 Newsgroups.
📌 Over 40x faster than the LLM baselines in our experiments.
📌 Consistently improves coherence and diversity over standard agglomerative topic reduction across benchmarks
🖇️ Check the paper out! openreview.net/pdf?id=gpecgW…
💻 Open Source Code: github.com/vertaix/Vendi-Clu…
✒️ Authors: Sean Kim (an amazing undergrad student from Brown University) and @adjiboussodieng.
#VendiScore #VendiClustering #Diversity #Clustering #AI #Science
Check out our @adjiboussodieng’s Op-Ed published by @AJEnglish:
📝 “AI fears are being weaponized”
#AI #AIGovernance #AIRisk @PrincetonCS
Tech giants are fighting to protect their monopolies from cheap open-source models. Fear is their weapon. — #AJOpinion by Adji Bousso Dieng.
🔗: aje.news/16wj21
🚀 We are hiring! Join Vertaix at @PrincetonCS for Fall 2027! Full funding available.
🔬 Are you passionate about the intersection of machine learning and the natural sciences? I am looking for highly motivated PhD students to join my research group, Vertaix, to build the foundations of Investigative Intelligence.
☀️ We are moving AI beyond simple predictive pattern matching and building systems capable of autonomous scientific discovery. We will tackle critical challenges like exploring high-dimensional spaces, mitigating mode collapse, and leveraging spectral diversity to create trustworthy AI systems that actually discover. Full funding from an NSF CAREER Award and other sources available.
⁉️ How to Apply: Apply directly through the Princeton Computer Science department application process (see links below). In your Statement of Purpose, explicitly mention your interest in working with Prof. Adji Bousso Dieng (@adjiboussodieng) and the Vertaix lab. Deadline is December 15, 2026.
🐯 If you are a PhD student already at Princeton (in any department, not only in CS) and want to be advised by @adjiboussodieng, please feel free to email her with your interests and CV.
🖇️ Link 1: cs.princeton.edu/grad/admiss…
🖇️ Link 2: gradschool.princeton.edu/adm…
🖇️ Link 3: cs.princeton.edu/news/dieng-…
Let’s advance AI for scientific discovery together. Please share this with anyone who might be a great fit! #PhDposition #MachineLearning #AI #ScientificDiscovery #InvestigativeIntelligence #AI4Science #PrincetonUniversity #Vertaix
🤖 Made with AI
Vertaix® (AI & Science) retweeted
Yes. And Vendi Clustering is ~40 times faster than the LLM baseline!
#NewPaper accepted in #EMNLP findings!
🔥 Clustering is a fundamental task in machine learning and science. In our new paper, we show you how you can do clustering using diversity as your mathematical object! We refer to this new algorithm as *Vendi Clustering*.
🧠 Instead of asking which two data points are closest, Vendi Clustering asks: which merge of datapoints causes the smallest loss in the diversity of the whole dataset? We optimize that global objective using the Vendi Score. When applied naively, the idea incurs an O(n^6) cost, which is prohibitive. Fortunately, we were able to derive an efficient version that reduces the naive intractable complexity of Vendi Clustering from O(n^6) to O(n^3).
📚 Vendi Clustering finds a natural application to the task of topic reduction. Embedding-based topic models like BERTopic often produce hundreds of fine-grained, overlapping topics. Topic reduction decides what structure survives in the final topic model. Poor topic reduction can merge genuinely distinct themes, preserve redundant ones, or discard useful data as noise. Existing topic reduction methods either make local pairwise merge decisions, require trial-and-error retuning of the clustering, or rely on expensive stochastic LLM prompting.
☀️ Vendi Clustering provides an efficient way to do topic reduction while preserving topic diversity. Some highlights from the results:
📌 0.65 -> 0.88 topic diversity (+35%) over topic reduction assisted by the Gemma-3-12b-it LLM on 20 Newsgroups.
📌 Over 40x faster than the LLM baselines in our experiments.
📌 Consistently improves coherence and diversity over standard agglomerative topic reduction across benchmarks
🖇️ Check the paper out! openreview.net/pdf?id=gpecgW…
💻 Open Source Code: github.com/vertaix/Vendi-Clu…
✒️ Authors: Sean Kim (an amazing undergrad student from Brown University) and @adjiboussodieng.
#VendiScore #VendiClustering #Diversity #Clustering #AI #Science
Vertaix® (AI & Science) retweeted
Experte de l’IA auprès de l’ONU et professeure à la prestigieuse université américaine de Princeton, la Sénégalaise Adji Bousso Dieng promeut une IA pour tous·tes les Africain·es.
#NewPaper accepted at #EMNLP findings: Vendi-RAG!
Retrieval-augmented generation (RAG) with large language models (#LLMs) has a redundancy problem -- and it's costing you multi-hop accuracy.
⁉️ The problem? RAG retrievers rank each document by similarity to the query, independently. On any real corpus that means the top-k comes back as several restatements of the same fact. But answering a multi-hop question requires chaining evidence across different documents. If the retrieved set is redundant, the chain is broken before the LLM ever sees it.
📌 Why does this matter? No amount of reasoning recovers evidence that was never retrieved. Most work on multi-hop RAG has gone into better loops, better prompting, better readers, while the retrieval objective itself stayed relevance-only.
🧠 Our approach: Vendi-RAG. We make semantic diversity a first-class retrieval objective. The Vendi Score (VS) measures the effective number of distinct documents in a set. Unlike the popular Maximum Marginal Relevance (MMR) or Determinantal Point Processes, VS scores the set as a whole and is directly interpretable. We combine it with query relevance into a single Vendi Retrieval Score (VRS). VRS intuitively implements the following idea: when answers are less accurate, diversify and explore the space; when they're accurate, consolidate.
☀️ Results:
+ 61.2% --> 71.4% (+10.2) supporting-facts recall over MMR on a matched single retrieval round. This means the performance gain comes from the scoring function, not extra iterations.
+ Retrieval diversity goes up from 5.21 to 6.82 Vendi Score on HotpotQA
+ Results hold across five instruction-tuned backbones and two dense retrievers
Paper: arxiv.org/abs/2502.11228
Code: github.com/vertaix/Vendi-RAG (pip install-able.)
Authors: the awesome Mohammadreza Rezaei and @adjiboussodieng.
@PrincetonCS @PrincetonEngineering
#RAG #LLM #NLP #InformationRetrieval #MachineLearning #Diversity #VendiScore #ScientificLiterature
Vertaix® (AI & Science) retweeted
Adji Bousso Dieng, la pépite sénégalaise en passe de révolutionner l'IA
l.jeuneafrique.com/cDx
AI is very good at learning what is common. But science often advances by finding what is rare.
Our @adjiboussodieng spoke with @laxmevy of @Sequencermag about why she believes diversity may be a fundamental technical principle for building #AI systems that go beyond reproducing what we already know, and help us discover what we don’t.
“Diversity is a technical solution to the problems facing AI in science.” #InvestigativeIntelligence @PrincetonCS @Princeton
sequencermag.com/diversity-i…
Are you at ICML (@icmlconf)? Meet Vertaix undergrad Shirley Yu this Saturday July 11 at the Workshop on AI for Physics. She'll deliver an oral presentation on the paper titled:
"Quality-Weighted Determinantal Point Processes for Sample Efficient DFT Surrogate Models"
The paper describes a new coreset selection method for materials property prediction. #AI4Science
Link to workshop: ai4physics-workshop.github.i…
