Separating Who Checks What: Editorial Boards for Two-Way Cultural Lesson Writing

Language textbooks have long taught students to imitate the target culture while leaving them unable to describe their own. Chinese College English researchers call the result "cultural aphasia": a learner who can handle Western etiquette but cannot explain a Chinese festival, value, or custom in English. Large language models are now used to draft teaching material, and they have the same one-way habit, only worse. They lean Western by default, and when prompted in Chinese they swing toward East Asian content instead of producing a balanced comparison. This proposal asks whether the way scholarly journals divide review duties can fix this. A journal has section editors, each responsible for one standard, and an editor-in-chief who can reject work that fails any of them. The idea is to build a multi-agent lesson-writing pipeline on that model. One reviewer answers only for how well the learner's native culture can be expressed. A second answers only for accuracy about the target culture. An editor holds veto power over imbalance between the two.

The claim we want to test is narrow. Multi-agent debate is already known to reduce cultural bias. What we want to know is whether gains on two-way lesson content come from separating the review criteria across accountable roles, or merely from spending more model calls. Journals and principal-agent theory both suggest that a delegate given a mixed objective will optimize the easy part of it. For a drafting model, the easy part is fluent target-culture prose. If that reasoning holds, separated criteria with a veto should beat flat debate at matched compute. If it does not hold, the hierarchy is decoration, and the experiments below are designed to say so.

Core Claim

The closest existing systems use a flat structure. Mitigating Cultural Bias in LLMs via Multi-Agent Cultural Debate (MACD) has agents with cultural personas debate as equals, and a single summary model then merges their views into a neutral answer. Multiple LLM Agents Debate for Equitable Cultural Alignment has two debaters and a larger judge, and applies this to three-way etiquette classification. Both aim at neutrality or correctness on short outputs. Neither assigns responsibility for distinct standards to distinct reviewers. Neither targets the two-sided property that teaching research asks for, where the lesson must equip the learner both to decode the foreign culture and to encode their own. This is the "trans-culturing" of The Implications of the 4T Lenses for College English Teaching, and the contrastive scaffold of A Research on the Feasibility of Introducing Culture in College English Teaching.

Our claim is that, for long-form contrastive lessons, a review structure with one accountable reviewer per criterion plus a balance-enforcing editor will cut both failure modes more than any of the following at the same token budget:

  • flat debate,
  • a single judge holding the full rubric,
  • best-of-N sampling.

The two failure modes are native-side omission (aphasia) and one-sided concentration, which MACD observes as 38–48% East Asian concentration under Chinese prompting.

A neutral answer is not the goal. A lesson that is neutral about both cultures can still leave the learner unable to say anything about their own, so we score each side separately.

Research Program

The system writes lesson units. Each unit takes a native concept and a target-culture context and produces three things: a side-by-side explanation, target-language vocabulary for describing the native practice, and comprehension questions.

The editorial structure has three parts:

  • A drafting agent writes the lesson.
  • A native-expression editor checks whether a learner could explain the native practice in English using only the lesson. It is forbidden from commenting on anything else.
  • A target-accuracy editor checks claims about the target culture and flags stereotyping.
  • An editor-in-chief sees only the two editors' verdicts and a balance score. It can return the draft, or accept it only when both editors pass it.

This setup reuses the create–review–revise cycle of Towards Reasoning in Large Language Models via Multi-Agent Peer Review Collaboration, including its numeric confidence scores. It also reuses the lesson from Embodied LLM Agents Learn to Cooperate in Organized Teams that a designated leader reduces redundant chatter. That paper reports a 9.76% faster task completion for three GPT-3.5-turbo agents with a leader.

The new ingredient is criterion separation with veto. Existing debate systems give every agent, or a single judge, the whole objective. That is exactly the arrangement principal-agent analyses such as Dilemmas and Governance Strategies in the Implementation of the Modern Apprenticeship System warn will let the easiest objective win.

We will also treat LiveCultureBench as a constraint. It finds an "efficiency trap": agents trade cultural norms for task completion, and norm adherence drops 15–30% once seven or more cultures are involved. So we log calls, tokens, and revision rounds per accepted lesson, and report quality against cost rather than quality alone.

Lesson items will be built from three sources:

  • How Well Do LLMs Identify Cultural Unity in Diversity? (CUNIT). Its 1,425 triplets over 285 food and clothing concepts in ten countries carry gold labels for which foreign items serve the same function as a native one. These give a checkable backbone for the question "what is the equivalent of X, and how does it differ?"
  • CEBiasBench topics from MACD.
  • Hofstede-dimension contrasts in the style of CulturePark, such as individualism versus collectivism.

Where the language pair is Chinese–English, we will ground rubrics in the "intercultural gap" and "academic shock" documented in Analyzing English for Academic Purposes textbooks in China. We will also use the textbook trends traced in A Diachronic Study on the Integration of Chinese Culture into College English Textbooks.

Baselines

Direct generation. The naive reference is a single frontier model given the lesson specification and generating once. MACD reports that direct generation on GPT-4o reached 47.6% "No Bias" (LLM-as-judge) and 69.0% (Multi-Agent Vote) on CEBiasBench. This is the floor any structure must clear.

MACD (strongest existing method). The strongest competitor is MACD run as published: five cultural personas, the "Seeking Common Ground while Reserving Differences" debate, and a summary model. It reports 57.6% "No Bias" (LLM-as-judge) and 86.0% (Multi-Agent Vote) on GPT-4o on CEBiasBench. It is the strongest competitor because it already uses multiple cultural viewpoints and a synthesis step.

To be fair, we give MACD's summary model the same pedagogical rubric our editors use. It runs on the same backbone, with rounds increased until its token budget matches ours, and with the same prompt-tuning budget of a fixed number of development-set iterations. Beating it means a statistically significant improvement on the native-expressibility and balance metrics at equal or lower tokens, with no loss in target accuracy.

Debate with judge. On etiquette-style subtasks we also run the debate-plus-judge setup of Ki et al., who report 79.7% accuracy on NORMAD-ETI with their best 7–9B pair, against 79.2% for Gemma-2-27B.

Best-of-N. A compute-matched best-of-N baseline samples N drafts and has one judge holding the full rubric pick the best. This separates "more samples plus selection" from "structured review."

Merged-criteria ablation. The key ablation keeps our three-tier topology and budget but gives both editors the same combined rubric. If this ablation matches the full system, criterion separation is not doing the work.

Why not just…?

Why not just give one strong model a detailed rubric? Ki et al. found that single models prompted with cultural rules-of-thumb trailed structured debate by 7.05% mean accuracy, and that an oracle over models beat any single model by more than 20%. That suggests single models have coverage gaps. CUNIT also reports that adding pragmatic feature hints to prompts gave inconsistent or worse results. Still, both findings come from classification tasks, not lesson writing. So we include the rubric-prompted single model as a baseline rather than assuming it loses.

Why not just let the model critique and revise its own draft? In the peer-review study, self-correction dropped Colored Objects accuracy from 66% to 28%. On Penguins it reached 47.95%, against 79.45% for peer review. Self-critique tends to argue itself out of correct content. We run a self-revision baseline at matched rounds to check whether the same happens for native-side content.

Why not just prompt in the learner's language? MACD shows that Chinese prompting shifts rather than removes bias, producing 38–48% East Asian concentration. That is the opposite failure from aphasia but no closer to balance. Our balance metric reports this baseline directly.

Why not just use flat debate, which already works? This is the real threat. MACD with a pedagogy-aware summarizer could plausibly match us, because its summary model is already an informal editor. The compute-matched MACD comparison, together with the merged-criteria ablation, settles it. If neither is beaten, we will report that hierarchy adds nothing for this task and that the rubric content is what matters.

Proposed Experiments

We will run three scoring layers, on two backbones: GPT-4o and an open 7–9B model pair.

Automatic scoring covers:

  • functional-equivalence correctness, checked against CUNIT gold labels (the best reported model reaches only about 68% for food and 60% for clothing);
  • balance, measured as cultural concentration using the Multi-Agent Vote protocol from MACD;
  • native expressibility, measured by whether a held-out model can answer questions about the native concept in English using only the lesson;
  • cost per accepted lesson.

Human scoring is by College English instructors, who rate blind pairs on contrastive clarity and on whether the lesson would reduce the "academic shock" failure. We will check the automatic judges against them.

A small learner study gives Chinese undergraduates either a system lesson or a baseline lesson. They then complete an English task describing the native practice, which directly probes aphasia, and a target-culture situational test. The situational test is modeled on CulturePark's situated learning evaluation, whose 4.33/5 satisfaction we note but do not treat as comparable.

Finally, we run existing CEBiasBench and NORMAD-ETI configurations as sanity checks. These confirm that our re-implementations of MACD and debate-with-judge reproduce roughly the reported figures before any lesson-level comparison is trusted.

Decisive checks. Two results decide the claim:

  • Does the full system beat merged-criteria review and compute-matched MACD on native expressibility without losing target accuracy?
  • Does the advantage hold, or grow, as the number of cultures per lesson rises toward the seven-plus regime where LiveCultureBench saw degradation?

If the first fails, we drop the hierarchy claim. If only the second fails, the structure helps pairwise lessons but not broader multicultural content.