I’m doing it with LLMs or I’m not doing it at all.
📜 google scholar | ✉️ aholtzman@uchicago.edu | 💡ideahub
- Head of Conceptualization Lab
- Second ever Communication & Intelligence researcher
- Check out our C&I blog
- Assistant Professor at
- UChicago CS
- UChicago DSI
- TTIC (by courtesy)
- Office: DSI 315
- I am actively looking for PhD students to apply through both UChicago CS and the Data Science Institute. Experience with LLMs/generative models is a plus, but not required! I no longer have time to reply to all emails from prospective PhD students but I highly encourage you to check out my lab website and apply!
- Previously:
- Post-Doc @ FAIR @ Meta
- PhD from UW advised by Luke Zettlemoyer
- Communication & Intelligence
- Machine Communication
- Generative Models as a Complex Systems Science
I work on pragmatic narrative—stories about models and models that produce stories.
Events
- 2025/01—Gave a talk at Northwestern’s CS seminar on Seeing Like a Language Model
- 2025/12—Gave a talk at CogInterp @ NeurIPS
- 2025/11—Gave a talk at Université du Québec à Montréal’s Séminaire en Informatique Cognitive and UChicago’s Theoretical Philosophy Workshop
- 2025/10—Gave a talk at CMU’s LTI Colloquium seminar on Halloween 🎃👻
- 2025/09—Gave a talk at the COLM Visions of Language Modeling Workshop!
- 2025/07—the Economist published my letter
- 2025/06—gave a talk about Articulating the Ineffable at MMLS!
- 2024/10—The UChicago Communication and Intelligence Symposium was a great success!
Some things I’m interested in right now:
- I want to figure out how Transformer LLMs communicate with themselves in the residual stream. In my opinion, both the alignment and MechInterp communities have become somewhat less ambitious. I’ve often said that I think MechInterp is overrated. I think MechInterp is very cool—just the vast majority of students want to do MechInterp, when I think behavioral work is where much of our insight comes from. But, I’m slowly becoming convinced that Transformer LLMs are simpler than I thought, they just don’t line-up with the kind of explanations people were looking for, so I’m throwing my hat back into MechInterp after being briefly excited and then abandoning it in 2021.
- I think we should build communication games, games where the main mechanic is communication. There are some (Disco Elysium, Chants of Sennaar, Keep Talking and Nobody Explodes, etc.), but I want ones that are NPC-driven (not MMORPGs or friend-group co-ops) and use LLMs to build a novel social ecosystem you have to navigate. This will happen, let’s be a part of it!
- The jury is still out on whether passive-learning based AI can ever produce truly interesting media. My guess is yes, but I’m excited to look at this either way. Let’s see if LLMs have a story to tell or two!
- For other ideas, see my twitter or IdeaHub
Selected Publications
Prompting as Scientific Inquiry Ari Holtzman & Chenhao Tan
Predicting vs. Acting: A Trade-off Between World Modeling & Agent Modeling
Margaret Li, Weijia Shi, Artidoro Pagnoni, Peter West, Ari Holtzman
Generative Models as a Complex Systems Science:
How can we make sense of large language model behavior?
Ari Holtzman, Peter West, Luke Zettlemoyer
Surface Form Competition: Why the Highest Probability Answer Isn’t Always Right
EMNLP 2021
=Ari Holtzman, =Peter West, Vered Shwartz, Yejin Choi, and Luke Zettlemoyer
= equal contribution
The Curious Case of Neural Text Degeneration
ICLR 2019
Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, Yejin Choi