English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Cultural Awareness Is Represented but Not Decoded in LLMs: A Four-Knife Dissection of the Mythology Gap

Forum topic · ✨步子哥 · 2026-08-04

Summary

This article explains a counterintuitive finding from arXiv:2608.02486: when 18 open-source LLMs across 8 architecture families (Llama, Qwen, Mistral, Gemma, Phi, Yi, OLMo, Falcon) are asked about thunder gods from Finnish, Slavic, Egyptian, or Chinese mythologies, they collapse answers to Zeus, Jupiter, or Thor. Yet the failure is not in representation. Using linear probing, logit lens, activation patching, and output extraction on the folkloristic Thompson Motif Index, the authors show the residual stream cleanly separates ten cultures well above a name-string baseline. The breakdown occurs only at readout: the decoder collapses culturally-specific tokens onto dominant-tradition defaults. Prompting in the target language does not help, because the decoder is gated on prompt language and fails independently per language. The piece generalizes this pattern, citing SOPHIA, SWE-Pruner Pro, EvoThink, and Token Budget as evidence that LLMs often internally encode richer, more correct information than they verbalize. Implications for AI safety and evaluation are discussed.

Key points

  • The phenomenon. Across 18 open-source LLMs (8 architecture families: Llama, Qwen, Mistral, Gemma, Phi, Yi, OLMo, Falcon), queries about thunder gods from dominant mythologies (Greek, Roman, Norse) succeed, while queries about Finnish, Slavic, Egyptian, or Chinese equivalents collapse to "Zeus, Jupiter, or Thor."
  • The method. Authors Iaroslav Chelombitko et al. (University of Nicosia, Cyprus) apply four "knives" to localize the failure: (1) linear probing of every residual-stream layer, (2) logit lens projections, (3) activation patching that transplants Finnish-context activations into Greek prompts, and (4) direct output extraction.
  • Representation is healthy. Linear probes classify cultures from residual-stream states far above a name-string baseline. The model internally distinguishes ten cultures.
  • Readout is broken. Starting at the logit-lens stage and worsening at final decoding, culturally-specific tokens are collapsed onto dominant-tradition defaults. The authors phrase it: *"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."*
  • The library analogy. Knowledge is correctly shelved and labeled in a basement "library" (residual stream), but a single front-desk clerk (decoder) only shouts three names regardless of which section is queried.
  • Language gating. Asking in the target language (e.g., Finnish) does not recover the answer. In-language failures correlate with each other, but cross-language failures decouple, implying the decoder is gated on prompt language and collapses independently per language.
  • Why the comparison is fair. The authors use the Thompson Motif Index, a folkloristic taxonomy of cross-cultural story elements, so "Greek thunder god" is matched structurally with "Finnish thunder god," not with a minor figure.
  • Mechanism, not just data. Adding more non-dominant training data may sharpen the residual stream further without fixing the readout collapse; remediation should target the decoder (unembedding, final LayerNorm, or sampling).
  • A broader pattern. The article links this to SOPHIA (2025), SWE-Pruner Pro (2025), EvoThink (2025), and Token Budget (2025), arguing a recurring law: internal representations in LLMs tend to be richer, more correct, and more fine-grained than their outputs.
  • Implications. For AI safety, output-only evaluation underestimates what an LLM "knows," including withheld hazardous knowledge. For benchmarking, conflating readout failures with representational deficits leads to misdiagnosis.
  • Limitations. Ten cultures is a modest set; linear separability does not guarantee the model uses the distinction at inference time; activation patching deserves deeper treatment; and the paper stops at "phenomenon" rather than offering a mechanistic account (unembedding norms vs. attention vs. LayerNorm).
  • Resources

  • Paper: https://arxiv.org/abs/2608.02486
  • Code & data: https://github.com/AragonerUA/folkmotif
  • Dataset: https://huggingface.co/datasets/Aragoner/folkmotif

Tags

#llm-evaluation#cultural-bias#representation-vs-readout#mechanistic-interpretability#linear-probing#activation-patching#thompson-motif-index#open-source-llms

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178585122