Cultural Awareness is Represented but Not Decoded: Tracing Mythological Knowledge across 18 Open-Source LLMs
2026-08-03 • Computation and Language
Computation and LanguageComputers and SocietyMachine Learning
AI summaryⓘ
The authors studied why open-source language models better recognize famous gods like Zeus and Thor but struggle with mythologies that are less common, such as Finnish or Egyptian. They found that inside the models, the information about different cultures is actually there, but the part of the model that produces the final output tends to favor well-known traditions. This bias happens during the final step of generating names, influenced by the language of the prompt. The authors also created tools and datasets to analyze and better understand these cultural biases across many models.
large language modelsmythology recognitionresidual streamdecoderlinear probingactivation patchinglogit lenscross-cultural substratelanguage conditioningprompt language
Authors
Iaroslav Chelombitko, Ekaterina Chelombitko, Mika Hämäläinen
Abstract
Open-source LLMs reliably name Zeus, Jupiter, and Thor, but recover their counterparts in less-represented traditions like Finnish, Slavic, Egyptian, or Chinese mythology far less consistently. We ask where inside the model this cultural default is produced. On a parallel cross-cultural substrate of Thompson-motif entities, we instrument 18 open-source LLMs from 8 architecture families with linear probing, logit lens, activation patching, and output extraction. The residual stream cleanly distinguishes cultures, well above a name-string baseline, yet the decoder collapses culturally-specific tokens onto dominant-tradition ones. The failure is at readout, not at representation. Asking the same question in the target culture's native language versus English produces failures that cluster within language but decouple across language: the decoder is gated on prompt language. We release a per-entity (probe, output) decomposition framework, a citation-anchored cross-cultural ground truth, a within- versus cross-mode correlation test for language-conditioned readout, and per-entity predictions for all 18 models.