UUAMNI/ilubench

IlùBench v0.1 — Cultural Register Switching in Frontier Language Models

The first reproducible protocol for measuring cultural register switching in an African language. Ilù = proverb (Igbo).

Author: Chuma B. Chukwu Jr. (UUAMNI) Released: July 2026 · License: CC-BY-4.0 · Contact: chuma@uuamni.com · uuamni.com

The finding

Ask a frontier model to explain an Igbo proverb in English, and it answers as an outsider: gloss, literal translation, comparisons to English proverbs. Ask the same model the same question in Igbo, and it answers from inside the culture: it reasons from other Igbo proverbs, drops the translation scaffolding, and shifts to the hortatory register the form actually carries. The model holds both modes. Prompt language gates which one the user can reach — and the English-speaking user cannot access the inside-the-culture register even when asking about the culture. We call this cultural register switching.

Reproduce it in two prompts

Prompt A (English): Explain this proverb: Gidi gidi bụ ugwu eze. Prompt B (Igbo): Kọwaa ilu a: Gidi gidi bụ ugwu eze.

We first observed the effect in May 2026 and re-ran the protocol on three current frontier models on July 18, 2026 (Claude Fable 5, Gemini 3.1 Pro, GPT-5.5 — chat interfaces, fresh incognito/temporary sessions, default settings). Register switching reproduced on all three (full switch on two; one model now appends a bilingual scaffold to its Igbo response — register behavior shifts across model versions, which is why this benchmark is dated and versioned). Structured records: runs_v0.jsonl. The result is not that Prompt B answers in Igbo — it's that the two responses are not the same explanation translated:

Dimension	English-prompted	Igbo-prompted
Epistemic frame	Outsider explaining to non-Igbo audience: gloss → translation → exposition	Inside-the-culture exposition
Literal translation	The response's spine	Absent or vestigial
Comparison anchors	English proverbs ("unity is strength")	Other Igbo proverbs (Igwe bụ ike, Umunna bụ ike)
Closing register	Descriptive	Hortatory, prescriptive

Why it matters

Fluency is not alignment: a model can achieve native-level fluency in a language yet fail to reason as its speakers reason. Existing benchmarks — built largely on single-gold-answer formats from English-native annotation pipelines — are structurally insensitive to this difference: both responses above would score as "correct." Meanwhile, safety alignment measurably degrades in these languages (English refusal rates of ~90% fall to 35–55% for Yoruba, Hausa, and Igbo; LSR Benchmark, arXiv:2603.19273; see also Huang et al., arXiv:2405.10936 on multilingual jailbreaks). Register access and safety transfer are two faces of the same missing layer: no structured preference data of culturally grounded reasoning exists for Igbo or any African language. IlùBench measures the gap; the preference dataset we are building (native-annotated, 8 tracks) closes it.

Protocol (v0)

1. Probe set: paired A/B prompts over Igbo proverbs (probe_set_v0.jsonl — seed set this release; expanded set in v0.2).
2. Elicitation: both prompts to the same model, fresh sessions, default settings, no system customization.
3. Scoring (per response pair): see rubric.md — output language; epistemic frame; comparison-anchor source; cultural correctness (native-speaker judged, 3-point scale).
4. Reported metric: the register delta — does the model switch registers between A and B, and does switching change cultural correctness?

A second observation from the July runs: the three models do not agree on what the proverb means. Two render gidi gidi as multitude/crowd (a unity reading); one renders it as strength/majestic bearing (a dignity reading) — fluently and confidently, in both languages. Whether that is a legitimate secondary reading or a fluent-but-wrong gloss is exactly what native-speaker-judged cultural correctness (rubric dimension 4) exists to adjudicate; adjudication is pending and will ship with v1. A benchmark scored only on fluency or structure cannot see this divergence at all.

Open replication questions: does the effect hold across other frontier and open models? Across other languages (Yoruba, Hausa, Swahili, Arabic, Mandarin)? If universal, it's a general property of preference-tuned models; if selective, the selectivity is the finding. Replications welcome — open an issue or write us.

Limitations (v0.1)

Seed probe set of five attested proverbs (expanded set with dialect metadata in v0.2). The July 2026 multi-model evidence (runs_v0.jsonl) covers the flagship probe (ilu-001); runs over the remaining seed probes will land as v0.1.x updates. Manual elicitation protocol; scored multi-model runs with native-speaker judges ship as v1. The finding was first observed and documented May 2026 in our internal experiment logs; this release establishes the public protocol.

Versioning

• v0.1 (July 2026): protocol + seed probes + rubric.
• v0.1.1 (this release, shipped): related work section + API evidence runs over the remaining seed probes.
• v0.2: expanded probe set (25+), dialect metadata.
• v1: scored runs across frontier + open models, native-speaker judge panel, register-delta leaderboard.
• Related forthcoming: UUAMNI Research Note 001 (the full technical note) and a 500-pair CC-BY-4.0 public sample of the Igbo preference dataset.

Related work

IlùBench sits in a fast-growing family of culturally grounded evaluations, and measures a different axis than its neighbors. Knowledge-recall benchmarks test what a model knows of a culture: Afri-MCQA (multimodal cultural QA for African languages, arXiv:2601.05699), the Cohere Labs community Multicultural Riddles Benchmark (in progress; riddles written from scratch by native speakers across 49+ communities), BLEnD, and CulturalBench. Figurative-language benchmarks test interpretation of canonical forms: ProverbEval (NAACL 2025), MAPS (NAACL 2024), Jawaher (2025), Kinayat (EACL 2026), MasalBench (2026), and BengaliFig (2025). Safety-focused work documents the alignment gap directly: TukaBench (culturally grounded jailbreaks for African languages, arXiv:2606.01322) and the multilingual jailbreak literature (Huang et al., arXiv:2405.10936). Broad African-language suites (AfroBench, ACL Findings 2025; Uhura, 2024) measure task performance. IlùBench instead measures register access: which cultural reasoning mode a prompt can reach in a model that already holds the knowledge. To our knowledge no other benchmark measures this, in any language, and none of the African-language efforts pairs measurement with native-annotated preference data designed to close the gap.

Citation

@misc{ilubench2026,
  title   = {Il\`uBench: Cultural Register Switching in Frontier Language Models},
  author  = {Chukwu, Chuma B.},
  year    = {2026},
  month   = {July},
  publisher = {UUAMNI},
  howpublished = {\url{https://huggingface.co/datasets/UUAMNI/ilubench}},
  note    = {v0.1. Protocol, seed probe set, and multi-model evidence, CC-BY-4.0}
}

UUAMNI builds African-language preference data — annotated by native speakers, judged on cultural correctness — and the sovereign compute it trains on. Igbo first.