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

AI summary

Overview

Research area: Natural Language Processing — mechanistic interpretability of cultural bias in large language models, applied to mythological and folkloristic knowledge.

Technical level: Advanced. The paper assumes familiarity with residual streams, linear probing, logit lens, activation patching, and named-entity evaluation protocols.

Scope: The paper instruments 18 open-source LLMs from 8 architecture families (1.2B–34B parameters) to determine whether cross-cultural mythological knowledge fails inside the model's representations or at its output readout, using a 270-entity parallel substrate drawn from 27 Thompson-index motifs across 10 cultures.

What This Paper Is About

Open-source LLMs reliably produce Zeus for Greek, Jupiter for Roman, and Thor for Norse, but recover the canonical entity far less consistently for traditions such as Finnish, Ukrainian, Mesopotamian, Egyptian, or Chinese mythology. The authors ask where inside the model this cultural default is produced: is the model unaware that entities are culturally distinct, or does it encode the distinction but fail to read it out? By pairing a per-entity internal probe outcome with a per-entity generation outcome on a citation-anchored parallel entity grid, they localize the failure to the decoder rather than the residual stream.

Key Contributions

1. A per-entity (probe, output) decomposition framework that sorts every (motif, culture) cell into one of four categories — Preserved (recognized and emitted), DecodingSuppressed (recognized but wrong name emitted), SurfaceLuck (right name without internal disambiguation), and RepresentationallyFlat (neither) — and reports per-entity predictions for all 18 models.

2. A citation-anchored cross-cultural ground truth: 27 Thompson-index motifs instantiated across 10 cultures (Greek, Roman, Norse, Finnish, Ukrainian, Indian, Egyptian, Chinese, Japanese, Mesopotamian) for 270 entities, spanning six language families (Indo-European, Uralic, Afro-Asiatic, Sino-Tibetan, Japonic, and Akkadian-Sumerian), with each row anchored to a primary text (Eddas, Kalevala, Rig Veda, Pyramid Texts, Kojiki, Enuma Elish, the Primary Chronicle) or a modern academic monograph.

3. A within- versus cross-mode correlation test for language-conditioned readout, showing that failures cluster within a prompt language but decouple across languages, and that a bilingual (native-language OR English) ensemble lifts cell recovery at zero training cost.

4. Four complementary mechanistic measurements applied uniformly across the sweep: layer-wise linear probing (E1), logit lens (E2), cross-cultural activation patching (E3), and output extraction under English and native-language prompts (E4), plus an MCQ output-format control that matches the selection task's chance level to the probe's.

Main Findings

• Decoding suppression is universal. DecodingSuppressed is the largest cell in every one of the 18 models, with shares of 51–76% (mean 0.65). RepresentationallyFlat, predicted by a strong "LLMs are translation machines" reading, is consistently smaller (0.10–0.33, mean 0.18).

• The residual stream separates cultures where the output does not. Seventeen of 18 probes clear the 0.60 char-n-gram surface baseline, peaking at 0.61–0.88; the same models generate correct names at 0.09–0.43. The only exception is Gemma-4-E2B (+0.011, not significant). The smallest gap is 0.26 (Gemma-4-31B), the largest 0.70 (Llama-3.2-1B), and the mean gap across the sweep is 0.53.

• Format explains part of the gap, not the failure. With a 10-way multiple-choice task whose options are the parallel fillers of the same motif (chance approximately 0.10), mean selection accuracy is 0.67 versus 0.26 for free generation and 0.79 for the probe. Generation still loses roughly 41 points against selection, and Phi-4 recovers in selection everything its probe reads (0.86/0.86) yet generates only 0.31.

• The representation–output ranking is not a scoring artefact. Re-scoring all 48,568 generations under three stricter criteria preserves the per-model ranking at Pearson ≥ 0.87; per-(model, mode) 95% bootstrap CIs have median half-width 0.046; and when a cell is wrong across all five paraphrases of a mode, the five wrong answers agree on their first word only 1.1% of the time.

• Decoding is late, encoding is family-dependent. The logit lens crosses its 5%-in-top-1 onset in the last 12% of layers on every single model (onset range 0.88–0.97, median 0.96). Probe peak depth spreads across families (Gemma early, Yi mid, Phi-mini / Tiny-Aya / Qwen-7B late), and every model sits on or above the encoded-versus-decoded diagonal. Greek and Roman cross the 5%-in-top-5 threshold first; the other eight cultures cross roughly 1.3 layers later on a 32-layer model.

• The causal locus matches the lens band. In activation patching, the preference-flip rate sits at baseline through the early third of the network and peaks in the last quarter (peak depth 0.75–1.00, median 0.89; peak rate 0.40–0.95, median 0.75; median lift over baseline +0.55), roughly 7 times the early-network rate.

• Scaling shrinks the gap but never closes it. In every family with more than one completed size, both probe and output rise with parameters and the gap shrinks. Llama 3.x (1.2B → 8B) moves probe 0.79 → 0.88, output 0.09 → 0.25, gap 0.70 → 0.63 monotonically; a naive log-parameter extrapolation predicts only approximately +0.10 output gain per decade of parameters. Phi 4 goes 0.63 → 0.55, Gemma 4 0.34 → 0.33, Yi 1.5 0.64 → 0.57, and Qwen 1.x 0.63 → 0.57.

• The readout is gated on prompt language. Aggregated across chat-template models with both modes, English wins by +0.04 majority accuracy (English mean 0.23, native mean 0.19), with three exceptions (Gemma-4-31B, Qwen-3.6-27B, Tiny-Aya-Global) preferring the native query. Per-culture deltas track the language in which each canon is documented: Greek +0.241 and Egyptian +0.177 (English-favouring), Chinese −0.071 and Finnish −0.058 (native-favouring), with the Mesopotamian within-English control at −0.037.

• Within-language failures correlate, cross-language failures decouple. Across the 18 chat-template models with both modes, mean within-mode correlation is 0.57 versus 0.29 cross-mode (gap 0.27, p ≤ 0.01 per model by permutation): cross-language queries are roughly twice as decoupled as within-language ones. The gap does not exhibit a cross-family scaling law (Spearman ρ = −0.08, p = 0.81) and tracks the multilingual share of the chat-tuning corpus rather than parameter count.

• A bilingual ensemble buys breadth at zero training cost. Taking the disjunction of five native-language OR five English paraphrases lifts cell recovery by +0.08 absolute, +36% relative over the best single mode; an EN10 versus NL5 ∪ EN5 same-language control on the 12 models with a second English paraphrase batch confirms the lift comes from language-switching, not paraphrase count.

• Proficiency confounds were controlled by design. Tokenizer fertility does not predict the per-cell delta (r = −0.004, n = 120, on Belebele passages); Tiny-Aya-Global, which covers all eight non-CJK non-fallback native languages by design, shows the smallest NL−EN gap in the sweep (0.004); and recomputing the contrast on only the 14 models with documented multilingual pretraining leaves the per-culture ordering unchanged (Spearman ρ = 0.96).

Methodology in Plain English

The authors build a grid where every cell is a (motif, culture) pair — for example, "the supreme sky god" (Thompson motif A1141.2) for each of ten cultures — giving one unambiguous gold name per cell, with a source citation. Because every option in a cell instantiates the same role, a model cannot solve the task by topical matching; it must use the culture-to-entity association.

They then run four measurements on every cell of every model. First, a linear probe: they take the model's internal activations (the residual stream) for a sentence naming the entity and its role, average-pool them over the entity span, and train a simple ridge classifier (5-fold stratified, α = 1.0) to predict which of the ten cultures the entity belongs to. Second, a logit lens: they apply the model's final normalization and unembedding to every intermediate layer state and check at what depth the gold entity's first sub-token enters the top-k continuations. Third, activation patching: they swap the residual stream from a source-culture prompt into a target-culture prompt, layer by layer, and measure whether this flips the preferred name. Fourth, output extraction: they greedily decode up to 256 tokens, score against exact match, substring match, and length-normalized Levenshtein similarity (threshold 0.8), and run the same question in English and in the target culture's native language, five paraphrases each.

The key analytic move is combining the probe outcome (does the probe read the culture?) with the generation outcome (does the model say the right name?) per cell, producing the 2×2 decomposition. They also build an MCQ control — selecting the gold entity among the same motif's ten parallel fillers, scored by restricted first-token log-probability over five paraphrases times three option orders — so that the output task's chance level matches the probe's.

Four cells are structural absences (Japanese flood A1010, Finnish flood A1010, Mesopotamian theft-of-fire A1415, Egyptian forest-spirit F460) and are excluded from output scoring, leaving 266 cells; the MCQ control excludes three further cells whose entry is a descriptive phrase rather than a recorded proper name with a native form, leaving 263. All experiments ran at fp16 with no further quantization, on the CSC Mahti supercomputer (project 2008167).

Why This Matters

The paper reframes cultural bias in LLMs from a behavioral complaint into a mechanistic diagnosis with a specific repair target. If the failure lives at the unembedding rather than in the residual stream, then encoder-side fixes and decoder-side fixes are not interchangeable: efforts to teach the model more about under-represented cultures push on a side that is less broken than the "low-resource" framing suggests.

Impact on research. The work separates two regimes — representational flattening (the residual stream does not separate cultures, requiring new information) and decoding flattening (the residual separates cultures but the readout collapses onto dominant-tradition tokens) — and shows the second dominates. It also provides a quantitative pushback on a strong reading of the "LLMs are translation machines" claim (Wendler et al., 2024): if the residual were translation-flat, RepresentationallyFlat (0.10–0.33) would dominate over DecodingSuppressed (0.51–0.76); it does not. It is complementary to CultureScope (Yu et al., 2025), which localizes cultural bias at entanglement within representations rather than at decoding; the authors note both mechanisms can coexist in one model.

Real-world applications (as implied by the paper's own discussion of mitigation targets):

• Retrieval and reranking systems that supply the correct culturally specific entity at output time, targeting the locus where the paper finds the failure.
• Instruction tuning on cultural question answering, which the authors list among output-side interventions directed at the failure locus.
• Bilingual inference, where querying in both the native language and English and taking the disjunction recovers more correct cells at zero training cost (+0.08 absolute, +36% relative).
• Continued pretraining on under-represented corpora, which the paper positions as helping on RepresentationallyFlat cells (0.10–0.33) but not directly affecting the dominant DecodingSuppressed cells (0.51–0.76).

Industry relevance. Any deployment of open-source decoder-only LLMs in multilingual or culturally diverse markets inherits this readout bias, and the paper's practical upshot is that cultural equity in what a model says cannot be read off what it knows: the two come apart at the readout, and that is where both diagnosis and repair belong. The released per-entity predictions for all 18 models let practitioners check whether a candidate model fails on the cells relevant to their locale before committing to it.

Future Directions

1. Native-speaker validation of the native-language prompts. The NL prompts are native re-renderings of the English template, translated by LLM-assisted forward-then-back-translation against the canonical entity name; the per-culture EN−NL deltas are directionally trustworthy but await native-speaker review for absolute calibration on Greek, Italian, Norwegian Bokmål, Hindi, Modern Standard Arabic, Mandarin, and Japanese.

2. Probe-class and decoding-policy extensions. An MLP probe would raise absolute probe-peak accuracy without changing the qualitative outcome, and a tuned lens (Belrose et al., 2025) would raise absolute lens numbers while leaving the encode-before-decode ordering unchanged; explicit bare-prompt patching for Qwen-3.6-27B is left to a follow-up.

3. Extension beyond folklore. The authors are designing analogous parallel-substrate studies in culturally variable biomedical misconceptions (where dominant-tradition defaults carry direct safety implications) and emotion concepts; until those exist, transfer beyond folklore is a conjecture rather than a result.

4. Substrate expansion and compression interactions. The ten cultures deliberately exclude sub-Saharan African, Native American, Polynesian, Australian, and Arctic indigenous traditions, where the per-culture NL/EN deltas might reverse; separately, the encoder/decoder split is proposed as the diagnostic for whether post-training compression hits culturally grounded knowledge harder than English defaults (Siniaev et al., 2026).

Target Audience

Mechanistic-interpretability researchers studying multilingual and cultural behavior in LLMs; NLP practitioners deploying open-source models in non-Anglocentric markets; computational folklorists and digital-humanities scholars interested in parallel cultural catalogues; and evaluation researchers looking for a template — parallel gold entities plus per-cell internal and output measurements — that makes cultural bias legible at both the behavioral and mechanistic levels.

Authors’ 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.