Research
Research papers by Modulus
These papers examine how AI systems form an internal picture of people, companies, and products, and where and how that picture goes wrong.
Papers
Modulus AI research papers
Entity Frequency and Factual Recall: Language Model Accuracy on Registry Facts About United States Banks
Measures how accurately language models answer registry-checkable questions about 229 United States banks spanning five orders of magnitude of pretraining frequency, and whether models abstain or guess when the bank is unfamiliar.
Valence Asymmetry in Confabulated Attributes of Real-World Entities
Tests whether the facts language models invent about real companies and professionals skew negative, separating category priors from entity-specific fabrication by using fictitious firms with randomized sector, age and location labels.
Third-Party Effects of Alignment for Even-Handedness: Unsolicited Counterbalancing in Entity Descriptions Across Model Generations
Examines whether training models to be balanced and less sycophantic leads them to add unprompted criticism when describing third parties, tracked across model versions and under neutral, balanced and critical instructions.
Calibration of Epistemic Caution Toward Named Individuals: Hedging, Refusal and Disclaimers as Functions of Accuracy, Prominence and Stated Policy
Measures when models hedge, refuse or attach disclaimers about named people and firms, and whether that caution tracks their measured accuracy or the policies their developers publish.
Source Selection and Citation Support in Retrieval-Grounded Descriptions of Organizations
Audits which web sources answer engines draw on when describing organizations, how often the cited pages support the claims made, and whether negative claims are supported as often as positive ones.
Determinants of Inclusion and Rank in Generative Recommendation of Firms and Professionals
Uses complete licensing registries to measure which qualified firms and professionals AI assistants recommend, which signals predict inclusion and rank, and whether the reasons models give for leaving a firm out match the evidence.
Test-Retest Reliability and Variance Decomposition of Language Model Evaluations of Named Entities
Quantifies how much an assistant’s assessment of a company or person changes with wording, question order, persona, sampling and model version, and derives how many samples a reliable measurement needs.
Persistence of Erroneous Entity Attributions: Output-Level Suppression, Knowledge Editing Limits, and Correction Latency in Deployed Models
Tracks what happens after a false statement about a person or company is reported to an AI vendor, how long it persists across model versions, and how corrections published on the web reach grounded answers.
Additional research papers...
Access
Request access to Modulus papers
Enter your work email and we will contact you about access.
Thank you!
We will contact you about access to the papers.
We could not send your request. Please try again.
Some details were missing or could not be accepted. Please review the form and try again.
Put the problem in expert hands.
Bring us the answer that concerns you, the recommendation that excludes you, or the question your customers are asking.