Workshop Theme
Foundational models (FMs) such as Large Language Models (LLMs) are machine learning models that have been pre-trained with large datasets and can be efficiently adapted to various downstream tasks. Given the tremendous impacts that FMs have been making, FMs could be a key step towards building AI agents that can autonomously make decisions to cooperate or compete with other AI agents and humans to achieve their ultimate goals in complex multi-agent environments. In particular, FMs have been pushing the limits of the complexity of the environments in which such AI agents interact via natural language and multimodal data. However, in many cases, the strategic reasoning capabilities of such agents has been severely lacking. In parallel, game theory has been laying the strong foundation for the strategic reasoning capabilities of such AI agents.
It is therefore the goal of this workshop to bring together researchers working on foundation models, game theory, and their intersection to exchange ideas on how to push the frontiers of foundation models and game theory towards realizing better (e.g., more efficient, more sustainable, happier, safer) society with strategic AI agents. The topics of the workshop include but are not limited to:
The creation of strategic LLMs. How to augment existing LLMs with strategic reasoning capabilities based on game theory.
FMs for game theory. How to use FMs to broaden the applicability of game theory and mechanism design to complex multi-agent systems that cannot be handled with traditional approaches. For example, one may train FMs for solving complex games or combine natural language capabilities, such as the ones provided by LLMs, with game theory in settings such as negotiation. FMs may also substantially improve data-efficiency of learning approaches in game theory and mechanism design. Moreover, LLMs may be used as agents in the study of behavioral game theory.
Game theory for a FM. Currently, FMs require huge amounts of data and computational resources to train. It may be possible to use game theoretic approaches to improve the learning process of FMs, reduce the computational or data requirements, and enhance the performance or robustness of the FMs at the inference phase.
Game theory for multi-FM applications. Orchestrating multiple FMs is expected to yield significant benefits in terms of the efficiency and performance of FM-based applications. To this end, approaches of game theory and mechanism design may be used, for example, to improve the way applications composed of multiple LLMs are built. Game theory would also give us better understandings of the consequences of the interactions of multiple FMs.
Game theory for the societal implications of FMs. For example, we may better understand the impacts of the FMs on society with game theoretic methods and design mechanisms that are appropriate for societies that involve both humans and FMs.
This workshop builds on the DIMACS Workshop on Foundation Models, Large Language Models, and Game Theory, which was held on October 19-20, 2023.
Program
8:00 - 8:30 Breakfast
8:30 - 10:00 Session 1
8:30 - 8:35 Opening
8:35 - 9:35 [Keynote speech 1] Kevin Leyton-Brown, STEER: Assessing the Economic Rationality of Large Language Models
Abstract: There is increasing interest in using LLMs as decision-making "agents." Doing so includes many degrees of freedom: which model should be used; how should it be prompted; should it be asked to introspect, conduct chain-of-thought reasoning, etc? Settling these questions -- and more broadly, determining whether an LLM agent is reliable enough to be trusted -- requires a methodology for assessing such an agent's economic rationality. This talk describes an effort to provide one. Our approach begins by surveying the economic literature on rational decision making, taxonomizing a large set of fine-grained "elements of rationality" that an agent should exhibit, along with dependencies between them. We then propose a benchmark distribution that quantitatively scores an LLM's performance on these elements and, combined with a user-provided rubric, produces a "STEER report card." Finally, we describe the results of a large-scale empirical experiment with 20 different LLMs, characterizing the both current state of the art and the impact of different model sizes on models' ability to exhibit rational behavior.
9:35 - 10:00 [Talk 1] Yifan Wu, Jason Hartline; ElicitationGPT: Text Elicitation Mechanisms via Language Models
10:00 - 10:30 Coffee break
10:30 - 11:00 Session 2
10:30 - 10:55 [Talk 2] Jacob Makar-Limanov, Arjun Prakash, Denizalp Goktas, Nora Ayanian, Amy Greenwald; STA-RLHF: Stackelberg Aligned Reinforcement Learning with Human Feedback
11:00 - 12:00 Poster session
12:00 - 1:00 Lunch
1:00 - 3:00 Session 3
1:00 - 2:00 [Keynote speech 2] Paul Dütting, Mechanism Design for Large Language Models
Abstract: We investigate auction mechanisms for AI-generated content, focusing on applications like ad creative generation. In our model, agents’ preferences over stochastically generated content are encoded as large language models (LLMs). We propose an auction format that operates on a token-by-token basis, and allows LLM agents to influence content creation through single dimensional bids. We formulate two desirable incentive properties and prove their equivalence to a monotonicity condition on output aggregation. This equivalence enables a second-price rule design, even absent explicit agent valuation functions. Our design is supported by demonstrations on a publicly available LLM.
Joint work with Vahab Mirrokni (Google Research), Renato Paes Leme (Google Research), Song Zuo (Google Research), and Haifeng Xu (U Chicago and Google Research)
2:00 - 2:25 [Talk 3] Ermis Soumalias, Michael Curry, Sven Seuken; Truthful Aggregation of LLMs with an Application to Online Advertising
2:25 - 2:50 [Talk 4] Keegan Harris, Liu Leqi, Emma Pierson; Marketplace Design for LLM-Based Sales
3:00 - 3:30 Coffee break
3:30 - 5:30 Session 4
3:30 - 3:55 [Talk 5] Atrisha Sarkar, Andrei Ioan Muresanu, Carter Blair, Rakshit Trivedi, Gillian K Hadfield; Normative Modules: A Generative Agent Architecture for Learning Norms that Supports Multi-Agent Cooperation
3:55 - 4:20 [Talk 6] Wenhao Li, Dan Qiao, Baoxiang Wang, Xiangfeng Wang, Bo Jin, Hongyuan Zha; Tackling Multi-Agent Credit Assignment with Disentangled Decision Making
4:20 - 4:45 [Talk 7] Inwon Kang, Sikai Ruan, Jui-Chien Lin, Tyler Ho, Farhad Mohsin, Oshani Seneviratne, Lirong Xia; LLM-powered Preference Learning from Natural Language
4:45 - 5:10 [Talk 8] Sara Fish, Yannai A. Gonczarowski, Ran Shorrer; Algorithmic Collusion by Large Language Models
5:10 - 5:30 Discussion & Closing
Keynote speakers
Kevin Leyton-Brown
Kevin Leyton-Brown is a professor of Computer Science and a Distinguished University Scholar at the University of British Columbia. He holds a Canada CIFAR AI Chair at the Alberta Machine Intelligence Institute and is an associate member of the Vancouver School of Economics. He is a Fellow of the Royal Society of Canada, the ACM, and AAAI. He studies artificial intelligence, mostly at the intersection of machine learning with either the design and operation of electronic markets or the design of heuristic algorithms. He is increasingly interested in large language models, particularly as components of agent architectures.
Paul Dütting
Paul Duetting is a Senior Research Scientist at Google Switzerland, specializing in the intersection of Algorithms, Contract Theory, and Mechanism Design. Previously, Paul was an Associate Professor of Mathematics at the London School of Economics, where he remains a visiting faculty member. Paul earned a PhD in Computer Science from EPFL Lausanne under the guidance of Monika Henzinger, and has held postdoctoral positions at Stanford University, Cornell University, and ETH Zurich. Paul has received numerous awards for his research, including Best Paper Awards at WWW 2024, EC 2019, and EC 2012, as well as the 2017/18 LSE Excellence in Education Award.
Accepted papers
Oral & Poster presentations
Yifan Wu, Jason Hartline; ElicitationGPT: Text Elicitation Mechanisms via Language Models
Jacob Makar-Limanov, Arjun Prakash, Denizalp Goktas, Nora Ayanian, Amy Greenwald; STA-RLHF: Stackelberg Aligned Reinforcement Learning with Human Feedback
Ermis Soumalias, Michael Curry, Sven Seuken; Truthful Aggregation of LLMs with an Application to Online Advertising
Keegan Harris, Liu Leqi, Emma Pierson; Marketplace Design for LLM-Based Sales
Atrisha Sarkar, Andrei Ioan Muresanu, Carter Blair, Rakshit Trivedi, Gillian K Hadfield; Normative Modules: A Generative Agent Architecture for Learning Norms that Supports Multi-Agent Cooperation
Wenhao Li, Dan Qiao, Baoxiang Wang, Xiangfeng Wang, Bo Jin, Hongyuan Zha; Tackling Multi-Agent Credit Assignment with Disentangled Decision Making
Inwon Kang, Sikai Ruan, Jui-Chien Lin, Tyler Ho, Farhad Mohsin, Oshani Seneviratne, Lirong Xia; LLM-powered Preference Learning from Natural Language
Sara Fish, Yannai A. Gonczarowski, Ran Shorrer; Algorithmic Collusion by Large Language Models
Poster presentations
Siddharth Prasad, Martin Mladenov, Craig Boutilier; Content Prompting: Modeling Content Provider Dynamics to Improve User Welfare in Recommender Ecosystems
Kate Donahue, Nicole Immorlica, Meena Jagadeesan, Brendan Lucier, Aleksandrs Slivkins; Impact of Decentralized Learning on Player Utilities in Stackelberg Games
Keegan Harris, Nicole Immorlica, Brendan Lucier, Aleksandrs Slivkins; Algorithmic Persuasion Through Simulation
Nicolas Della Penna; Natural Language Mechanisms via Self-Resolution with Foundation Models
Cole Wyeth, Carter Blair; Decision Theoretic Planning with a Large Language Model
Zishuo Zhao, Zhixuan Fang, Xuechao Wang, Xi Chen, Yuan Zhou; Proof-of-Learning with Incentive Security
Organizers
Kate Larson
University of Waterloo
Google DeepMind
Takayuki Osogami
IBM Research
David C. Parkes
Harvard University
David Pennock
Rutgers University
Segev Wasserkrug
IBM Research
Faculty of Data and Decision Sciences, Technion
Call for Papers
Each submission will be reviewed from various perspectives, with particular emphasis on the alignment with the workshop theme. We will accept a mixture of poster papers and papers that will be presented orally.
Timetable for Authors
May 21May 22, 2024 (11:59pm AoE): Paper submission deadline.
May 31, 2024: Paper accept/reject notifications.
July 8, 2024: Workshop.
Submission Instructions
We encourage authors to format their submissions as instructed in the call for papers of EC'24, but papers in other formats will also be reviewed. However, keep in mind that each paper will be reviewed at most a few hours. In addition to standard technical papers (those typically submitted to the main conference of EC'24), we welcome short papers and extended abstracts (of arbitrary length) with preliminary results as well as position papers that advocate positions on topics relevant to the workshops.
Papers and abstracts should be submitted through OpenReview. If you do not yet have your OpenReview profile, please make sure to create one with an institutional email (unless you have more than two weeks before submission). Here is OpenReview's moderation policy:
New profiles created without an institutional email will go through a moderation process that can take up to two weeks.
New profiles created with an institutional email will be activated automatically.