In-context-learning contamination

Classic questionnaires and vignettes are all over the public internet, which is to say all over LLM training corpora. A model that has effectively memorized the Cognitive Reflection Test is not being measured by it; it is reciting. Contamination therefore threatens the validity of any study that administers a well-known instrument to a model without checking for prior exposure.

A second, subtler channel operates inside the prompt itself: in-context learning. If the prompt's structure, examples, or earlier items telegraph the expected response pattern, the model can adapt to the test within the session — again producing scores that measure adaptation rather than disposition.

Detection strategies include memorization probes (can the model complete an item verbatim from a fragment?), comparing performance on original versus carefully modified items that preserve the construct but break surface memorization, and flagging suspicious familiarity patterns. Instruments that fail these checks need modified or alternate forms before their scores can be interpreted.

AIScholar includes ICL-detection and contamination-assessment tools: memorization probes, generation of modified items for flagged instruments, and alternate-forms reliability checks to verify the modified items still measure the same construct.

• Walkthrough: psychometric assessment of an LLM

• Psychometric validity for LLMsRe-establishing reliability and validity when human instruments are administered to models.
• Machine psychologyStudying LLMs as behavioral subjects with experimental methods from psychology.
• Method Specification Prompt (MSP)A self-contained method specification precise enough for independent re-implementation.

Browse all 11 glossary terms