Nouha Dziri
I’m a research scientist at Allen Institute for AI (AI2) working with Yejin Choi. Currently, I co-lead the safety and post-training effort at Ai2 to build (OLMo): a highly capable and truly open LLM to advance AI.
Besides this, I work on understanding the limits of Transformers and their inner workings. Check out “Faith and Fate” to learn about the limits of Transformers on reasoning tasks. I also work on studying alignment in LLMs; checkout “Roadmap to Pluralistic Alignment”, “Finegrained RLHF”, and “RewardBench” the first evaluation tool for reward models..
My work has been featured in TechCrunch, LeMonde, The Economist, Science News and QuantaMagazine.
I was fortunate to work with brilliant researchers in the field. I have worked with Siva Reddy at Mila/McGill, with Hannah Rashkin, Tal Linzen, David Reitter, Diyi Yang, and Tom Kwiatkowski at Google Research NYC and have worked with Alessandro Sordoni, and Goeff Gordon at Microsoft Research Montreal.
News
| Jan 2025 |  Check out the new blog post: DeepSeek R1: Innovative Research and Engineering Can Rival Brute-Force Scaling |
|---|---|
| Dec 2024 | System 2 Reasoning at Scale workshop (NeurIPS 2024) was a success  including keynote talks, panel, posters and lightening talks. Check out details! |
| Dec 2024 | Invited talk “In-Context Learning in LLMs: Potential and Limits” at the Language Gamification Workshop @ NeurIPS 2024  |
| Dec 2024 | Invited as a panelist at the Meta-Generation Algorithms for Large Language Models Tutorial at NeurIPS 2024 |
| Oct 2024 | WildTeaming and WildGuard got accepted at NeurIPS 2024. See you in Vancouver |
| Sep 2024 |   New blogpost about AI safety “Current Paradigms of LLMs Safety Alignment are superficial” |
| Sep 2024 |   New blogpost about o1 models and LLMs reasoning “Have o1 Models Cracked Human Reasoning?” |
| Aug 2024 | Super excited that our workshop “System 2 Reasoning At Scale” was accepted to NeurIPS24, Vancouver! Mark your calendar for Dec 15, 2024! |
| Jul 2024 | Check out our  new safety moderation tool WildGuard: a state-of-the-art open tool for assessing safety risks, jailbreaks, and refusals in LLMs. |
| Jul 2024 | New red-teaming method  WildTeaming: an automatic red-teaming framework that discovers novel jailbreaks based on in-the-wild user-LLMs interactions. |
| Jul 2024 | Check out my interview with Science News Magazine about “LLMs reasoning skills” featuring “Faith and Fate” and “Generative AI Paradox”. |
| Jul 2024 | I will serve as a Demo Chair for NAACL 2025. |