Research#

Foundational concepts and methodology for designing, conducting, and interpreting research on agentic systems. Understand how experiments are classified along the controlled-to-natural spectrum to choose an appropriate design, then apply qualitative or quantitative analysis to draw conclusions from the resulting data.

Visit Qualitative Analysis for methods that make sense of non-numerical data and Quantitative Methods for the statistical toolkit used to test and quantify claims.

ablation study#

Definition: experimental method that systematically removes or disables components of a model to measure their individual contribution to overall performance; derived from ablative brain surgery in neuroscience where tissue removal revealed functional regions

Purpose: isolates the causal influence of specific model components; low-effort way to understand which modules matter and how much they contribute; commonly used to justify architectural choices in neural networks and complex systems

Example: in an object detection system with localization, feature extraction, and classification modules, an ablation study might train variants that keep two modules fixed while varying the third to measure each component’s impact on detection accuracy

Related Terms: edge case, empirical testing, experimental design

Source: Baeldung: “Machine Learning: What Is Ablation Study?” by Panagiotis Antoniadis

A/B test#

Definition: also known as split testing and/or randomized controlled trial; experimental method that randomly assigns participants to treatment or control groups to compare two versions of something and determine which performs better

Purpose: commonly used by tech companies to test features, interfaces, or algorithms; isolates the effect of a single change by holding all other conditions constant across groups

Bayesian experiment#

Definition: experimental approach using Bayesian statistics to incorporate prior knowledge and update beliefs with new data; contrasts with frequentist A/B testing by assigning probabilities to hypotheses rather than data

Purpose: particularly valuable when sample sizes are small, prior information is available, or continuous monitoring is needed; allows sequential analysis where results can be evaluated as data accumulates without inflating error rates; quantifies expected loss to support risk-aware decision-making

Example: after running a Bayesian A/B test on a new checkout flow, the analysis might conclude “there is a 92% probability the new flow increases conversion rate” rather than reporting a p-value; this directly answers the business question “should we ship the feature?” with a probabilistic statement about the hypothesis

Sources:

conceptual AI experiment#

Definition: test type in which AI exists as a label or framing device, but no AI is actually implemented; typically uses vignettes or scenarios to model the operational principles or consequences of AI

Purpose: offers high feasibility and is easy to scale and replicate, letting researchers study impractical or impossible scenarios; trades off lower naturalness since subjects don’t interact with actual AI

controlled vs natural#

Definition: experimental design distinction based on environment; controlled experiments are conducted in artificial settings - labs, online platforms - where researchers manipulate variables, while natural experiments are conducted in real-world settings - workplaces, platforms, markets - where AI is actually used

Purpose: frames the trade-off between control/replicability and external validity/ generalizability when choosing where to run a study

Related Terms: A/B test, experimental design, natural AI experiment

edge case#

Definition: scenario or condition that occurs at extreme operating parameters or unusual circumstances; falls outside normal operating conditions but within specified boundaries

Purpose: critical for testing AI reliability and robustness; surfaces failure modes that typical inputs don’t exercise

Example: unusual inputs, rare combinations of factors, boundary conditions

Related Terms: empirical testing, proxy test

empirical testing#

Definition: validation approach based on observation and experimentation rather than theory alone; applies algorithms with actual users, tasks, or environments to measure performance

Purpose: uses real data and measurable outcomes to evaluate hypotheses, grounding claims about AI performance in observed behavior rather than theoretical argument

Related Terms: edge case, evaluation, experimental design

experimental design#

Definition: systematic planning of how to conduct an experiment to answer a research question; defines variables, treatments, control conditions, randomization, and measurement approach

Purpose: isolates causal effects while minimizing confounding factors; guides decisions about sample size, data collection methods, and analysis approach before a study begins

Related Terms: A/B test, controlled vs natural, statistical inference

natural AI experiment#

Definition: test type that features AI in environments where it is actually used - platforms, workplaces, real services; often takes the form of A/B tests run by organizations to improve products or operations

Purpose: yields the highest naturalness and directly applicable findings; trades off low feasibility, difficulty replicating results, narrow scope, and limited control compared to controlled study designs

proxy test#

Definition: indirect measure used to evaluate something difficult to assess directly; substitutes an observable indicator for an unmeasurable or impractical characteristic

Purpose: easier to apply than a direct measure, at the cost of occasionally misclassifying cases where the indicator and the underlying characteristic diverge

Example: using “developed exclusively for research” as a proxy for AI sophistication

Related Terms: edge case, empirical testing

quasinatural AI experiment#

Definition: test type that combines the naturalness of real AI systems with the feasibility of lab experiments

Purpose: enables naturalistic AI and broad research scope with easier data collection than natural experiments; in exchange, researchers give up some control over algorithm construction

Example: testing commercial chatbots in controlled studies, pilot experiments before product launch

stylized AI experiment#

Definition: test type conducted in a controlled environment, since the AI typically doesn’t exist outside the study; AI is tailored to a research question using rule-based algorithms, historical data replication, or reinforcement learning

Purpose: gives tight control over algorithm features and is feasible, replicable, and broad in scope; trades off lower naturalness compared to real-world AI systems

taxonomy#

Definition: classification system that organizes concepts, objects, or phenomena into hierarchical categories; defines relationships between categories and provides structure to a domain

Purpose: helps unify fragmented literature and reveal underexplored questions; in AI research, provides frameworks for organizing types of experiments, algorithms, or agent behaviors

Example: the SAP Labs agent eval taxonomy organizes evaluation along Objectives and Process axes

Related Terms: experimental design, SAP Labs agent eval taxonomy

vignette study#

Definition: research method presenting hypothetical scenarios to elicit preferences or judgments; participants read descriptions of situations and state what they would do

Purpose: can model any situation without implementation constraints and is easy to scale; common in conceptual AI experiments studying ethical dilemmas or preference patterns, though responses may not reflect actual behavior and carry lower external validity

Related Terms: conceptual AI experiment