Large language models are among the most powerful tools ever built.
They can label, summarize, retrieve, reason, translate, and connect evidence at scales no human team could match. But they also can inherit the artifacts of their sources: bias, leakage, spurious correlation, brittle benchmarks, shifting meanings, and false confidence. Used naively, while analyzing the world, they can quietly alter the quantities we think we are measuring.
At Neuristemic, we study AI systems as instruments of knowledge. We build methods for determining when AI-generated outputs can be trusted, tested, and used as evidence at scale.
Our work builds the methodological, computational, and philosophical foundations needed to use language models responsibly in science and society. We combine deep learning with causal inference, computational linguistics, network science, statistics, multimodal AI, and epistemology to develop methods for measuring with models while accounting for their limits.
That work spans debiasing LLM-assisted parameter estimates, testing whether text can really be disentangled from sensitive or causal signals, preventing treatment leakage in text-based causal inference, benchmarking model capabilities across modalities, linking massive organizational datasets, improving automated content analysis, and using multimodal representations to study development and social conditions at scale.
The papers below are our working record. Wherever feasible, we release data, code, benchmarks, and documentation so that others can inspect, replicate, challenge, and extend the work. Our research is organized around three connected questions:
1. What can AI systems measure?
2. What do those measurements assume about language, knowledge, and action?
3. How can AI-based evidence be made scientifically trustworthy?
News
2026
- October 29 – Nicolas Audinet de Pieuchon presents new work at the 9th BlackboxNLP Workshop
- October 5 – New preprint: ExpertMuon-Compass: Alignment-Guided Step Sizes for Mixture-of-Experts Training
- September 2 – Platonic Representations for Poverty Mapping: Unified Vision-Language Codes or Agent-Induced Novelty? to appear at ICSC
- August 5 – Preprint posted: When Many Answers Are Valid, Voting Fails: Symbolic Verification for Best-of-K Causal Reasoning in LLMs
- July 18 – Ozzie Gerring’s Story Turing data released!
- July 14 – Correcting Bias When Using Latent Regressors forthcoming at Political Analysis
- July 2 – Multiplication in Multimodal LLMs: Computation with Text, Image, and Audio Inputs (with Sam Balter and Ethan Jerzak) to appear in ACL Findings
- May 24–26 – Rationality for What? A Workshop on the Nature and Value of Rationality at NUS organized by Ethan Jerzak
- May 11 – Preprint posted: Queryable LoRA: Instruction-Regularized Routing Over Shared Low-Rank Update Atoms
Large Language Models: Capabilities & Alignment
This work evaluates what large language models and multimodal AI systems can reliably measure, where they fail through bias, leakage, entanglement, or benchmark artifacts, and how LLM-assisted evidence can be debiased, tested, and aligned with trustworthy scientific use.
Omatharv Bharat Vaidya, Ashwin Vinod, Pedram Akbarian, Aditya Sai Ellendula, Connor T. Jerzak, Nhat Ho. ExpertMuon-Compass: Alignment-Guided Step Sizes for Mixture-of-Experts Training. ArXiv Preprint, 2026.
[PDF] [BibTeX]
Nicolas Audinet de Pieuchon, Adel Daoud, Connor T. Jerzak, Moa Johansson, Richard Johansson. Benchmarking Debiasing Methods for LLM-based Parameter Estimates. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2025.
[PDF] [BibTeX]
[Overview] [Video] [Data] [>]
Samuel Balter, Ethan Jerzak, Connor T. Jerzak. Multiplication in Multimodal LLMs: Computation with Text, Image, and Audio Inputs. Findings of the Association for Computational Linguistics (ACL), 2026.
[PDF] [BibTeX]
[Overview] [Video] [Data] [>]
Omatharv Bharat Vaidya, Connor T. Jerzak, Nhat Ho, Chandrajit Bajaj. Queryable LoRA: Instruction-Regularized Routing Over Shared Low-Rank Update Atoms. ArXiv Preprint, 2026.
[PDF] [BibTeX]
[Overview]
Nicolas Audinet de Pieuchon, Adel Daoud, Connor T. Jerzak, Moa Johansson, Richard Johansson. Can Large Language Models (or Humans) Disentangle Text? Proceedings of the Sixth Workshop on Natural Language Processing and Computational Social Science (NLP+ CSS 2024): 57-67, 2024.
[PDF] [BibTeX]
[Overview] [Data] [Code] [>]
Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R. Brown, et al. Beyond the Imitation Game: Quantifying and Extrapolating the Capabilities of Language Models. Transactions on Machine Learning Research, 2023.
[PDF] [BibTeX]
[Overview] [Data] [>]
Causal Inference & Scientific Applications
This research develops causal inference, data-linkage, automated content analysis, and multimodal representation methods for using AI to study organizations, poverty, language, and social conditions without mistaking prediction, proxy signals, or treatment leakage for valid scientific evidence.
Satiyabooshan Murugaboopathy, Connor T. Jerzak, Adel Daoud. Platonic Representations for Poverty Mapping: Unified Vision-Language Codes or Agent-Induced Novelty? Proceedings of the 7th International Conference on Social Computing (ICSC 2026), of Communications in Computer and Information Science, Springer Nature, 2026.
[PDF] [BibTeX]
[Overview] [Video] [Data]
Connor T. Jerzak, Stephen Jessee. Correcting Bias When Using Latent Regressors. Political Analysis, 2026.
[PDF] [BibTex]
[Overview] [Video] [Code]
Omatharv Bharat Vaidya, Connor Thomas Jerzak, Zayne Rea Sprague, Fangcong Yin, Nhat Ho. When Many Answers Are Valid, Voting Fails: Symbolic Verification for Best-of-K Causal Reasoning in LLMs. ArXiv Preprint, 2026.
[PDF] [BibTeX]
Brian Libgober, Connor T. Jerzak. Linking Datasets on Organizations Using Half a Billion Open-Collaborated Records. Political Science Research and Methods: 1-20, 2024.
[PDF] [BibTeX]
[Overview] [Data] [Code] [>]
Connor T. Jerzak, Gary King, Anton Strezhnev. An Improved Method of Automated Nonparametric Content Analysis for Social Science. Political Analysis, 31(1): 42-58, 2023.
[PDF] [BibTeX]
[Overview] [Data] [Code] [>]
Adel Daoud, Connor T. Jerzak, Richard Johansson. Conceptualizing Treatment Leakage in Text-based Causal Inference. Proceedings of the Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL): 5638-5645, 2022.
[PDF] [BibTeX]
[Overview] [Video] [>]
Adel Daoud, Richard Johansson, and Connor T. Jerzak. Detecting and Mitigating Treatment Leakage in Text-Based Causal Inference: Distillation and Sensitivity Analysis. Arxiv Preprint, 2026.
[PDF] [BibTeX]
[Overview]
Epistemological & Conceptual Foundations
These papers examine the philosophical foundations of knowledge, rational action, meaning, logic, and convention to clarify what it means for AI-generated measurements, explanations, and decisions to count as evidence.
Ethan Jerzak, Alexander W. Kocurek. Knowing What to Do. Noûs, 59(1): 160-190, 2025.
[PDF] [BibTeX]
[Overview]
Ethan Jerzak. Narrow Content. The Philosophical Review, 130(3): 475-480, 2021.
[PDF] [BibTeX]
[Overview]
Alexander W. Kocurek, Ethan J. Jerzak. Counterlogicals as Counterconventionals. Journal of Philosophical Logic, 50(4): 673-704, 2021.
[PDF] [BibTeX]
[Overview]
Alexander W. Kocurek, Ethan Jerzak, Rachel Etta Rudolph. Against Conventional Wisdom. Philosophers’ Imprint, 20(22): 1-27, 2020.
[PDF] [BibTeX]
[Overview]
Ethan Jerzak. Two Ways to Want. Journal of Philosophy, 116(2): 65–98. February 2019.
[PDF] [BibTeX]
[Overview]
Ethan Jerzak. Paradoxical Desires. Proceedings of the Aristotelian Society, 119(3): 335–355.
[PDF] [BibTeX]
[Overview]
Ethan Jerzak. Non-Classical Knowledge. Philosophy and Phenomenological Research, 98(1): 190-220, 2019.
[PDF] [BibTeX]
[Overview]
See also [Bibliography].