---
title: "Cultural Awareness is Represented but Not Decoded: Tracing Mythological Knowledge across 18 Open-Source LLMs"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.02486v1"
permalink: "https://feed.astralor.com/items/itm_7347d842a6ee605ebcdf6f6c"
published: "2026-08-03T16:53:15.000Z"
score: 78
channel: "ai"
tags: ["Interpretability","Open-source LLM","Cultural Bias","Activation Patching","Linear Probing","Cross-Cultural Mythology"]
---

# Cultural Awareness is Represented but Not Decoded: Tracing Mythological Knowledge across 18 Open-Source LLMs

## AI commentary

> This paper turns the observation that 'models know but fail to answer' into a localizable computational fact. For teams deploying multilingual models, it shows the decoder readout is the real bottleneck for cultural bias; for the research community, it offers a reusable analysis framework to run per-entity probes on your own model before deciding whether to fix training or decoding.

## At a glance

This paper probes mythological knowledge across 18 open-source LLMs from 8 architecture families, finding that residual streams cleanly distinguish cultures, yet decoders collapse culturally-specific tokens onto dominant Western traditions. Using linear probing, logit lens, activation patching, and output extraction, the authors localize the failure to the readout stage rather than representation. They also show that prompt language gates the decoder, with failures clustering within language but decoupling across languages. The release includes a per-entity decomposition framework, citation-anchored cross-cultural ground truth, and per-entity predictions for all 18 models.

## Key points

- Models distinguish mythological entities across cultures internally, but outputs collapse less-represented traditions (Finnish, Slavic, Egyptian, Chinese) onto dominant Greco-Roman or Norse counterparts.
- Linear probing and activation patching show cultural information exists in residual streams, localizing the failure to decoder readout rather than representation.
- Prompt language gates decoding: failures cluster within language and decouple across languages.
- Releases include a per-entity probe/ouput decomposition framework, citation-anchored cross-cultural ground truth, and per-entity predictions for all 18 models.

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_7347d842a6ee605ebcdf6f6c