cad-memotrap-construction

IN premise — summaries/2026/08/24/shi-2024-context-aware-decoding-s3-experimental-setup.md

Created 2026-08-25T02:58:35+00:00

MemoTrap presents well-known proverbs with an instruction to end them with a different word (e.g., 'Better late than ___' → 'early' instead of the memorized 'never'), testing instruction-following over memorized completions.

Summary

MemoTrap is a diagnostic test that checks whether a model can override its strong urge to autocomplete a famous phrase and instead follow whatever specific ending the user actually requested. It matters because a system that defaults to pattern-recall over instruction-following will silently produce wrong answers whenever the expected output isn't the most common one.