Research
Reasoning capability in AI systems has progressed substantially, but it has been measured almost entirely on problems with a single checkable answer. The questions that decide things in the world (in law, medicine, policy, ethics, scientific discovery, and philosophy) are not like that. Utumno exists to evaluate, train, and secure reasoning on the questions that have no ground-truth answer.
Research Programs
Foundations
The formal theory behind machine reasoning and cognition.
Research advancing the formal theory behind machine reasoning and learning. Topics such as optimization, learning theory, dynamical systems, game theory and mechanism design, and mathematical models of argumentation and belief revision may fall under this direction, as may any work that is primarily theorem-driven and concerned with convergence, incentives, guarantees, or the formal properties of reasoning processes.
Cognition & Architectures
The structures of reasoning and new kinds of intelligent systems.
Research on the structures that reasoning runs on and the design of new kinds of intelligent systems. Work in this direction may include topics such as novel AI architectures, multi-agent systems, computational models of cognition, representations for holding and revising beliefs, and the study of human cognitive processes as both a source of design insight and a standard of comparison.
Evaluation & Measurement
Instruments for understanding and assessing reasoning systems.
Development of the instruments by which reasoning systems are understood and assessed. Research here may involve topics such as benchmarks, datasets, metrics, and evaluation software, particularly for domains without verifiable answers, along with auditing methods that recover the standards a system actually applies, human studies and corpora, and the validation of measurement tools themselves.
Alignment & Validation
The safety, security, and trustworthiness of reasoning systems.
Research on the safety, security, and trustworthiness of reasoning systems. Topics that may fall under this direction include how values and evaluative standards enter models through training, methods for making those standards explicit and auditable, adversarial robustness of belief and memory systems, interpretability of reasoning processes, and the validation of system outputs against real-world outcomes.
Philosophy & Epistemology
Knowledge, justification, truth, and normative reasoning.
Fundamental inquiry into knowledge, justification, truth, and normative reasoning, pursued both on its own terms and through the new methods that machine reasoning makes available. Research in this direction may engage topics such as the nature of defeasible knowledge, the status of machine belief, the structure of ethical and evaluative frameworks, and the epistemology of reasoning under irreducible uncertainty.
Scientific Discovery
Automating and augmenting scientific reasoning.
Research on the automation and augmentation of scientific reasoning. Topics such as hypothesis generation, evidence synthesis, the evaluation of competing explanations, and AI systems that participate in inquiry as rigorous, self-correcting reasoners may fall under this direction, reflecting a view of science as a defeasible enterprise in which progress depends on structured criticism and revision.