Open scholarship
Scholarly profiles and repositories are the sources of papers and research software. Readers can follow and explore the assumptions on which the work was done, as well as contribute to it.
Papers and research resources →Decision theory · Agentic risk · Corrigibility · AI risk measurement
Mathematical research for the public good
AgentNorm is an independent research initiative studying the mathematical foundations of decisions under uncertainty. We examine utility representations, ambiguity, regret, and collective judgment through axioms, proofs, and counterexamples. Our applied questions include decision theory for agentic risk, corrigibility, and the measurement and assurance of AI risk.
Our public-interest mission brings together mathematical research, open scholarship, and student mentorship. We share papers and research software so that researchers and students can examine the assumptions behind consequential decisions.
Research in the public interest
AgentNorm aims to answer foundational questions and contribute to the larger research community. Shuo Li Liu coordinates the research programme and collaborates with student researchers on theory and computational evaluation.
Student research and mentorship
We engage in research on judgment aggregation, verification, and model collapse with student collaborators, bridging theoretical concepts with computational experiments.
Meet the research team →Public-benefit research
We study how individuals and institutions reason under uncertainty, and how researchers assess evidence about AI systems. Our aim is to contribute methods and knowledge that others can use.
Explore our research programme →Legal entity
AgentNorm is the public research identity of NORMATIVE AI INSTITUTE INC., a British Columbia corporation.
BC Corporate Registry notice →Grant acknowledgement
We gratefully acknowledge a $20,000 grant from the Corrigibility Research Fund supporting CAST Decision Theory by Xuanyi Wang and Shuo Li Liu. The fund is a program of Lightcone Infrastructure Inc.; applications are coordinated through Lightcone Commons.
Research directions
We start with explicit definitions and assumptions, prove what follows, and then give counterexamples that push a claim to its limits. In the applied work, we make a clear distinction between mathematically guaranteed results and computational evidence.
Axioms and representation
We study which assumptions allow preferences to admit a utility representation, and which additional conditions a characterization requires.
Expected uncertain utility →Convexity, ambiguity, and regret
We study the structure of preferences under uncertainty, and provide sufficient conditions and counterexamples for the analysis of diversification, transitivity, and regret.
Convexity under ambiguity →Decision theory for agentic risk
We apply decision theory and mathematical analysis to questions about dominance, judgment aggregation, verification, and the limits of deployment claims.
Savage Without Monotonicity →Corrigibility, risk measurement, and assurance
We study how corrigibility, oversight, and measurement can support cautious assessment of risks in agentic systems, distinguishing formal guarantees from empirical evidence.
Behavior-only deployment certification →Theorems, proofs, and counterexamples
Representation theorem
A Comment on: “Expected Uncertain Utility”
Econometrica 92(1):247–256 · Journal article
Constructs a counterexample to an expected uncertain utility characterization and proves a corrected representation theorem. The revised axioms establish the required structure of ideal events and state-independent interval utility.
Paper and source →Sufficient conditions
Convexity under ambiguity
Economic Theory Bulletin 13(2):353–373 · Journal article
Establishes sufficient conditions for convex preferences in portfolio choice under ambiguity. The result combines aversion to mean-preserving spreads with uncertainty aversion, providing a foundation for diversification without assuming concave utility.
Paper and source →Constructive counterexamples
Counterexamples to “Transitive Regret”
arXiv:2407.00055 · Preprint
Constructs counterexamples to a claimed implication from transitivity and continuity to indifference between identically distributed prospects. The analysis examines the assumptions underlying regret-based preference representations.
Paper and source →Dominance and finite additivity
Savage Without Monotonicity
Philosophy Meets Machine Learning (PhilML), ICML 2026 · Peer-reviewed workshop poster
Develops a conditional counterexample in which a finitely additive expected-utility representation assigns equal value to an act and its statewise improvement. It identifies additional additivity, dominance, or domain assumptions that prevent this conflict.
Paper and source →People
Research coordinator and collaborators
Shuo Li Liu coordinates AgentNorm’s research programme. Xuanyi Wang collaborates on CAST Decision Theory. Student mentorship is part of the initiative’s public-interest research mission; Hruday Sivakumar and Vivan Shankar collaborate on projects involving language-model judgments, model collapse, and verification.
Shuo Li Liu
Research programme coordinator, AgentNorm
Shuo coordinates AgentNorm’s research programme. He is a PhD student in Economics at Princeton University and works on mathematical decision theory, uncertainty, judgment aggregation, and the evaluation of agentic risk.
Xuanyi Wang
Research collaborator
Collaborates with Shuo Li Liu on CAST Decision Theory, supported by a $20,000 Corrigibility Research Fund grant.
Hruday Sivakumar
Student researcher
Research collaboration on Bayesian Truth Serum and the evaluation of language-model panels.
Vivan Shankar
Research mentee
Research assistance on model collapse, verification, and the evaluation of learning systems.
Publications & research
Publications and reports
↑ ContentsJournal articles, a letter, a proceedings abstract, and an edited report.
- 2024
A Comment on: “Expected Uncertain Utility”
Econometrica 92(1):247–256
Journal article
Constructs a counterexample to an expected uncertain utility characterization and proves a corrected representation theorem. The revised axioms establish the required structure of ideal events and state-independent interval utility.
- 2025
Convexity under ambiguity
Economic Theory Bulletin 13(2):353–373
Journal article
Establishes sufficient conditions for convex preferences in portfolio choice under ambiguity. The result combines aversion to mean-preserving spreads with uncertainty aversion, providing a foundation for diversification without assuming concave utility.
- 2020
Patent Collateralization: The Effect of Policy and Impact on Patenting
Academy of Management Proceedings 2020(1):20539
Proceedings abstract
- 2019
Assessing cell-based animal proteins
Science 363(6429):826
Letter
- 2018
Agricultural Value Chains Analysis in Oe-Cusse, Timor-Leste
UNDP Timor-Leste
Report · Editorial contribution
Workshop papers
↑ ContentsPeer-reviewed posters accepted at ICML 2026 workshops.
- 2026
Savage Without Monotonicity
Philosophy Meets Machine Learning (PhilML), ICML 2026
Peer-reviewed workshop poster
Summary
Develops a conditional counterexample in which a finitely additive expected-utility representation assigns equal value to an act and its statewise improvement. It identifies additional additivity, dominance, or domain assumptions that prevent this conflict.
- 2026
Model Collapse Despite Ground-Truth Verification
Hypothesis Testing, ICML 2026
Peer-reviewed workshop poster
Summary
Separates exact verification of generated artifacts from coverage of the target task population, showing how missing task types limit risk identification and developing an audit bound for target-risk claims.
- 2026
Behavior-Only Deployment Certification Requires Bridge Assumptions
Failure Modes in Agentic AI (FAGEN), ICML 2026
Peer-reviewed workshop poster
Summary
Shows why behavioral evaluation alone cannot certify deployment safety without assumptions connecting evaluation and deployment histories, and quantifies the sample cost of detecting rarely reached failures.
- 2026
Bayesian Truth Serum for LLMs?
Statistical Frameworks for Uncertainty in Agentic Systems (AgenticUQ), ICML 2026
Peer-reviewed workshop poster
Summary
Tests Bayesian Truth Serum aggregation for language-model panels, finding lower accuracy than majority voting but improved calibration on the evaluated review and benchmark tasks.
Working papers and preprints
↑ ContentsPreprints, manuscripts, and related research in progress.
- 2024
Counterexamples to “Transitive Regret”
arXiv:2407.00055
Preprint
Constructs counterexamples to a claimed implication from transitivity and continuity to indifference between identically distributed prospects. The analysis examines the assumptions underlying regret-based preference representations.
- 2025
Clarifications and Extensions to “Regret Theory: A New Foundation.”
SSRN 5356990
Working paper
- 2026
AI Labor Rubric: Evaluating Multi-Agent Research Systems by Verified Marginal Value
Working paper
- 2026
Constitutional AI as Judgment Aggregation
Working paper
- 2026
Public Knowledge Labor after Generative AI
Manuscript in preparation
Research on how generative AI changes public technical knowledge and participation in online question-and-answer communities.
- 2026
What Is Left for Human Researchers? Coverage Profiling Against Support Collapse in Agentic Drug Discovery
Working paper
- 2024
Asset Demand Monotonicity
Working paper
- 2022–2024
Regret theory
Working-paper series
Research projects and software
↑ ContentsResearch tools and evaluation infrastructure.
- 2026
Agent Memory and Feedback Architecture Benchmark
Research project
Evaluation of memory and feedback for language model based systems.
- 2026
Distributed AI Research Evaluation Pipeline
Research software
Infrastructure for research workflows that include drafting, adversarial review, and independent checking.
- 2026
Multi-Agent Research Organization Simulator
Research software
A simulation framework to explore the organisation of research involving multiple agents.
- 2026
PRF-Bench — RLHF Preference-Aggregation Audit
Research project
Auditing how pooled preference models represent heterogeneous judgments.
Contact
For research collaboration, open scholarship, and student research:
admin@agentnorm.tech
You can also find Shuo Li Liu on OpenReview and Google Scholar.