💬 NLP

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

This paper investigates 18 open-source LLMs and finds that while they internally represent mythological knowledge from diverse cultures, their failure to consistently retrieve non-dominant traditions stems from a decoder readout bias that collapses these representations onto dominant-tradition tokens, a limitation further influenced by the prompt's language.

Original authors: Iaroslav Chelombitko, Ekaterina Chelombitko, Mika Hämäläinen

Published 2026-08-04

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Iaroslav Chelombitko, Ekaterina Chelombitko, Mika Hämäläinen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Great Cultural Glitch: Why AI Knows Zeus but Forgets Ukko

Imagine you are walking into a giant, high-tech library that contains almost every book ever written on the internet. This library is run by a super-smart robot librarian who has read every single page. You ask the librarian, "Who is the king of the gods in Greek stories?" The librarian instantly shouts, "Zeus!" You ask, "Who is the king of the gods in Roman stories?" "Jupiter!" You ask, "Who is the king of the gods in Norse stories?" "Thor!" The robot is perfect. But then, you ask, "Who is the supreme sky god in Finnish mythology?" The robot hesitates, stammers, and eventually gives you a wrong answer, or maybe even says, "I don't know."

This isn't just a random mistake; it's a pattern. The robot seems to know the "popular" stories (Greek, Roman, Norse) but struggles with the "less famous" ones (Finnish, Slavic, Egyptian, Chinese). This happens because the library the robot learned from was mostly filled with English books and Western stories. But here is the big mystery: Does the robot actually forget these lesser-known gods, or does it know them but just refuse to say the right name?

To solve this, scientists use special tools to look inside the robot's "brain" (which is actually a complex computer program called a Large Language Model). They want to see if the information is lost in the memory (the "encoder") or if the memory is fine, but the part that speaks the answer (the "decoder") is broken. Think of it like a student taking a test. If the student writes the wrong answer, did they not know the fact, or did they know it perfectly but get nervous and write the wrong thing? This paper is a detective story that goes inside the robot's brain to find out exactly where the cultural bias happens.

The Investigation: Inside the Robot's Brain

The researchers, Iaroslav and Ekaterina Chelombitko along with Mika Hämäläinen, decided to test 18 different open-source AI models. They didn't just ask random questions; they used a very clever, structured game based on old folklore books. They picked 27 specific roles from a famous catalog of myths (like "The Supreme Sky God," "The Sun God," or "The Trickster") and asked the robots to name the character for that role in 10 different cultures: Greek, Roman, Norse, Finnish, Ukrainian, Indian, Egyptian, Chinese, Japanese, and Mesopotamian.

To see what was happening inside the robot, they used four different "X-ray" tools:

Linear Probing: A test to see if the robot's internal memory contains the correct cultural label.
Logit Lens: A way to peek at the robot's thoughts layer by layer to see when it decides on an answer.
Activation Patching: A "what-if" experiment where they swap the robot's internal thoughts from one culture to another to see if the answer changes.
Output Extraction: Simply asking the robot the question and seeing what it actually says.

The Big Discovery: The Brain is Fine, the Mouth is Broken

The results were surprising and very clear. When the researchers looked inside the robot's memory (the residual stream), they found that the robot actually knew the culture perfectly.

Imagine the robot's brain is a massive filing cabinet. When you ask about a Finnish god, the robot's internal files clearly show a folder labeled "Finnish Mythology" with the correct name inside. The "Linear Probe" tool could read this folder with high accuracy (often between 60% and 88% correct). The robot knew the difference between a Greek god and a Finnish god.

However, when the robot was asked to speak, it messed up.

This is the main finding: The failure happens at the "readout," not the "representation." The robot knows the answer, but when it tries to say it, it collapses the specific, unique name (like "Ukko" for the Finnish sky god) into the most popular, default name (like "Zeus" or "Thor").

The researchers call this "Decoding Suppressed." In their study of 18 models, this was the biggest problem. For every single model, the majority of the time (between 51% and 76% of the time), the robot's internal memory was correct, but the final answer it gave was wrong. It's like a student who knows the answer is "Finnish" but, when writing it down, their hand automatically writes "Greek" because that's what they see on the teacher's desk every day.

What They Ruled Out

The paper also tested some other ideas to make sure they were right:

It's not that the robot is "translation-only": Some people thought robots just translate everything into English inside their heads and then translate back. If that were true, the robot wouldn't even know the cultural difference inside its brain. But the study showed the brain does know the difference clearly.
It's not just about model size: The researchers checked if bigger models (with more "brain power") fixed the problem. They found that while bigger models got slightly better, the gap between "knowing the answer" and "saying the answer" never closed. Even the biggest models still had this bias.
It's not just a language problem: They tested if the robot failed because it couldn't read the native language (like Finnish or Ukrainian). They found that even when the robot could read the language, it still preferred to give the "English/Western" default answer. The bias is in the decision-making part of the robot, not just its ability to read.

The Language Twist

One of the most interesting parts of the study was asking the same question in two ways: in English and in the native language of the culture (e.g., asking about Finnish gods in English vs. asking in Finnish).

They found that the robot's brain acts like a switch. When you ask in English, it uses one set of pathways to decide the answer. When you ask in Finnish, it uses a slightly different set. Sometimes, asking in the native language helped the robot give the right answer, but often, the robot still defaulted to the famous Western names.

The researchers discovered that if you ask the robot the same question in both English and the native language, and then take the best answer from either, you can fix about 36% more errors without teaching the robot anything new. It's like having two friends check the answer; if one says "Zeus" and the other says "Ukko," you can pick the right one.

The Conclusion: Fixing the Mouth, Not the Brain

The paper concludes that the problem isn't that the AI doesn't know about Finnish, Slavic, or Egyptian myths. The AI knows them. The problem is that the part of the AI that generates text is "gated" by cultural defaults. It has a habit of flattening unique, specific cultural details into the most common, dominant ones it sees in its training data.

This means that to fix cultural bias in AI, we shouldn't just try to feed the robot more books about Finnish myths (which would help the memory, but that memory is already there). Instead, we need to fix the part of the robot that speaks. We need to teach the robot that it's okay to say "Ukko" instead of "Zeus," even when it feels like "Zeus" is the safer, more common answer.

The study suggests that the solution lies in "readout" interventions—changing how the robot selects its final words—rather than just trying to retrain its entire brain. It's a reminder that just because a machine has all the information doesn't mean it will tell you the truth; sometimes, it just tells you what it thinks you want to hear.

Technical Summary: Cultural Awareness is Represented but Not Decoded

Problem Statement

Large Language Models (LLMs) exhibit a well-documented cultural bias, defaulting to Anglo-American or Western frames of reference across tasks ranging from moral reasoning to named-entity recognition. While behavioral studies have established that models fail to retrieve culturally specific knowledge (e.g., failing to name the supreme sky god in Finnish or Slavic mythology while reliably naming Zeus or Thor), they have not identified where inside the model this failure occurs.

The central research question is whether this bias stems from a representational failure (the model's internal state does not distinguish between cultures) or a decoding failure (the model distinguishes cultures internally but collapses them onto dominant-tradition tokens during output generation). Resolving this distinction is critical because the two regimes require opposite interventions: representational failures require new information in pretraining, while decoding failures require changes to the readout mechanism.

Methodology

The authors instrument 18 open-source LLMs (ranging from 1.2B to 34B parameters across 8 architecture families, including Llama, Gemma, Phi, Qwen, Yi, OLMo, and Aya) using a parallel cross-cultural substrate and four mechanistic tools.

1. Substrate: Parallel Cross-Cultural Entity Set

To ensure a fair comparison across cultures, the authors constructed a dataset based on the Thompson Motif-Index of Folk-Literature.

Structure: 27 specific mythological motifs (e.g., "supreme sky god," "theft of fire") evaluated across 10 cultures (Greek, Roman, Norse, Finnish, Ukrainian, Indian, Egyptian, Chinese, Japanese, Mesopotamian).
Ground Truth: Each (motif, culture) pair has a single canonical "gold" entity name, anchored to primary texts (e.g., Kalevala, Rig Veda) or modern academic monographs.
Prompting: Two prompt modes were used:
- Contextual: " ${name}` embodies the role of `$ {description}." (Used for hidden-state extraction; culture word is absent to prevent trivial copying).
- Chat: "In ${culture}` mythology, what is the name of the `$ {description}?" (Used for output generation).

2. Mechanistic Instruments

The study applies four complementary measurements to every (entity, model) pair:

E1 Linear Probing: A ridge classifier is trained on the residual stream to predict the 10-way culture label. This tests if cultural information is linearly decodable from the internal representation.
E2 Logit Lens: The model's final norm and unembedding are applied to intermediate hidden states to determine the depth at which the gold token enters the top- $k$ continuations.
E3 Activation Patching: Residual streams from a source-culture prompt are swapped into a target-culture prompt to identify the causal layer where the model's preference flips to the source culture's name.
E4 Output Extraction: Greedy decoding of the chat template to measure actual generation accuracy.

3. Decomposition Framework

By combining the probe outcome (E1) and the generation outcome (E4), the authors define a $2 \times 2$ decomposition:

Preserved: Probe correct, Output correct.
Decoding-Suppressed: Probe correct, Output wrong (the core hypothesis).
Representationally Flat: Probe wrong, Output wrong.
Surface Luck: Probe wrong, Output correct.

Key Results

1. Decoding Suppression is Universal (RQ1)

Across all 18 models, the Decoding-Suppressed cell is the dominant failure mode, comprising 51–76% of cases (mean 65%).

Evidence: The linear probe consistently achieves high accuracy (0.61–0.88), significantly above the character-n-gram surface baseline (0.60), indicating that the residual stream cleanly distinguishes cultures.
Contrast: However, the generation accuracy is drastically lower (0.09–0.43).
Conclusion: The cultural distinction is present in the model's internal state but is not emitted by the decoder. The failure is at readout, not representation.

2. Late Decoding and Causal Localization (RQ2 & RQ3)

Layer-wise Emergence: The logit lens shows that the gold token enters the top- $k$ predictions only in the last 12% of layers (median depth 0.96).
Causal Bound: Activation patching confirms that swapping residual streams only flips the model's preference in the last quarter of the network (median depth 0.89).
Implication: The network encodes culture early (probe peaks in the first half), but the causal link to the output token is established very late, coinciding with the unembedding layer.

3. Scaling Effects (RQ4)

Increasing model size within a family (e.g., Llama 1.2B $\to$ 8B) improves both probe accuracy and generation accuracy, but does not close the gap.

The gap shrinks slightly (e.g., from 0.70 to 0.63 in Llama 3.x) but remains substantial even at larger scales.
This suggests that simply scaling parameters does not solve the decoding suppression issue.

4. Language-Conditioned Readout (RQ5)

The failure is gated by the prompt language.

Decoupling: Cross-language paraphrases (English query vs. Native Language query) are roughly 2x more decoupled than within-language paraphrases ( $r_{cross} = 0.29$ vs $r_{within} = 0.57$ ).
Asymmetry: The direction of the bias (English vs. Native advantage) tracks the language in which the cultural canon is documented today (e.g., Greek/Egyptian favor English queries; Finnish/Chinese favor native queries), rather than tokenizer efficiency or model capability.
Bilingual Ensemble: Taking the disjunction of English and Native Language outputs (without training) recovers an additional +8% absolute accuracy (+36% relative), confirming that the two modes access partially disjoint subsets of the same representation.

Significance and Claims

The paper claims a mechanistic re-framing of cultural bias in LLMs:

Representation vs. Readout: The primary locus of cultural flattening is the unembedding layer, not the residual stream. The models "know" the cultural distinctions (they are linearly separable) but fail to "say" them.
Implications for Mitigation: Interventions targeting the encoder (e.g., continued pretraining on under-represented corpora) may help the small minority of "Representationally Flat" cases but will not directly fix the dominant "Decoding-Suppressed" cases. Effective mitigation must target the readout (e.g., retrieval-augmented generation, reranking, instruction tuning on cultural QA, or bilingual ensembling).
Limitations: The findings are established on open-source models with white-box access. The authors note that the "Decoding-Suppressed" dominance is a conjecture for closed-source models, though the language-conditioned readout prediction (RQ5) generalizes to them as it relies only on output behavior. The substrate is limited to folklore (Thompson motifs) where parallel gold labels are constructible; transfer to other domains (e.g., biomedical misconceptions) is proposed as a future direction but not claimed as a result of this study.

In summary, the paper argues that cultural equity in what a model says cannot be read off what it knows; the two diverge at the readout stage, and that is where both diagnosis and repair must be focused.