When Cultural LLM Alignment Gets a Positive Target, Who Writes the Spec?

Cultural alignment work on LLMs has been dominated by subtractive targets: detect the harmful output, then suppress it. That dominance is the opening claim of a Korean cultural alignment paper from the ICML 2026 Workshop on Culture X AI, whose abstract characterizes the field as “dominated by a negative target.” The paper then calls that the easy half. The harder half is constructive: stating what a culturally coherent answer looks like. Once the target is a positive specification, someone has to write it, and whoever writes it owns the model’s working definition of a culture that is not one thing.

What’s the difference between subtractive and constructive alignment?

Cultural alignment work on LLMs has been dominated by subtractive targets: flag the harmful output, suppress it, refuse. The Korean paper’s contribution is to name what this leaves out. A model that only knows what not to say has no working definition of what to say, and the failures are visible in production: blanket refusals on queries a culturally competent model would answer, missing context where a user needed a specific institutional answer, and a quality ceiling that more list items will not raise.

The subtractive form is the one the field knows well. A suppression list enumerates triggers and refusals; the model learns to avoid. Each culture adds its own entries, the lists compose, and the engineering work reduces to detection and data labeling. The unit of work is the banned output, and the unit scales. The subtractive default is what the field’s broader alignment methods already produce: preference-optimization pipelines trained on human feedback to downweight undesired responses, in the lineage of InstructGPT, make suppression their default behavior and encode no positive cultural specification.

Constructive alignment inverts that unit. Instead of a list of banned outputs, the target becomes a specification of the good output: what counts as a culturally coherent response, what institutional knowledge it should carry, what register and conventions it should use. The paper instantiates this as a culturally adapted safe-response policy grounded in Korean legal frameworks, social norms, and interpretive conventions (arXiv:2606.06797). That artifact is a design document, not a dataset, and it changes who is doing the work.

The reason this matters for the rest of the argument is that the two forms have different operational properties. Suppression lists are cheap, parallelizable, and composition-friendly, which fits the paper’s account of why negative targets dominate the field. Constructive specs are none of those things, and pretending otherwise is where the trouble starts.

How does the Korean cultural coherence pipeline work?

The pipeline turns a hand-authored Korean cultural policy into preference signal by using three frontier models to write candidate responses, then fine-tuning six open-weight LLMs on the resulting triplets with DPO.

The seed is a prompt-based LLM generator that expands a Korean harm taxonomy. The safe-response policy it is measured against is grounded in Korean legal frameworks, social norms, and interpretive conventions, so the “correct” behavior is defined before any training runs. Three frontier models then produce candidate responses against that policy. The outputs become preference triplets, and DPO fine-tuning applies the resulting signal across six open-weight LLMs (arXiv:2606.06797). DPO, introduced by Rafailov et al. as direct preference optimization, fits a policy to preference pairs without a separately trained reward model, which is why the pipeline can apply it across six open-weight base models rather than a single frontier system.

The reported results are modest in scope and clear in direction: an improved Korean cultural safe rate across the six models, with no large degradation on Korean general-capability benchmarks. The capability-preservation check matters here, because a known risk in narrow-domain fine-tuning is catastrophic forgetting, where the model gets worse at the broader language while getting better at the narrow cultural target. The paper reports it avoided that tradeoff.

The qualitative payoff is the part worth reading. Fine-tuned models named Korean statutes and institutional procedures and supplied constructive Korean-context information alongside refusals rather than blanket suppression (arXiv:2606.06797). A query a suppression-only model would flatly refuse gets routed to a specific institutional answer. That is the constructive behavior the paper is arguing for, made concrete: the model declines and points somewhere. The pointing is the load-bearing part, because a pointer has to choose a destination, and the destination is where the politics enters.

Why does a positive spec change who owns the model’s cultural behavior?

A suppression list vetoes without choosing; a positive specification chooses, and that is the entire point of writing one. The ownership of the model’s cultural behavior moves with that difference.

Under subtractive alignment, a suppression list is legible and extensible. A Korean community group, a regulator, an internal trust-and-safety team can each propose additions, and the disagreements are about boundary cases rather than about which framing of the culture is correct. The list does not pick winners among cultural interpretations, because a veto is not an interpretation. The cultural behavior of the model is, in effect, owned by the data-ops function that maintains the list, and that function is replaceable: swap vendors, swap annotators, the behavior is preserved by the artifact, not the person.

Under constructive alignment, every clause of the spec is a commitment. “Grounded in Korean legal frameworks” commits the model to a specific jurisdiction’s statutes. “Social norms” commits it to a specific consensus, real or constructed. “Interpretive conventions” commits it to a reading practice. Each of those is a choice that a different Korean community might contest, and the model will enforce the chosen interpretation at inference time across every Korean-language query it serves. The cultural behavior is now owned by whoever authored the spec, because the spec is the working definition the model enforces.

That role is not replaceable the way the data-ops role is. Replacing the spec author means changing what the model believes, which is precisely the political act the suppression list was designed to avoid. The suppression list survived leadership changes and vendor swaps precisely because it carried no position. A constructive spec carries a position by definition, and the position travels with the author.

This is why the question in the title is not rhetorical. “Who writes the spec?” has a concrete operational answer under constructive alignment: the spec author does, and their draft becomes the model’s behavior for an entire language community until someone with equal or greater standing replaces it.

How does the cost of alignment change when the target is constructive?

Constructive alignment moves the cost from a currency that scales with labor to one that scales with legitimacy, and legitimacy does not scale with labor. That single change rewrites the economics of multicultural alignment.

Subtractive work is data labeling. It is cheap, distributable across annotation vendors, and it composes with majority-vote protocols. You can throw labor at it, and the marginal culture added to a global model costs roughly the marginal annotation budget. That budget amortizes across cultures, which fits the paper’s framing of negative-target alignment as the field’s dominant mode: it slots cleanly into a roadmap with a labeling line item.

Constructive work is representational politics. There is no majority-vote protocol that resolves whose social norms count as the canonical Korean social norms, because the disagreement is not about how to label a given example but about the standard the labels are measured against. Adding annotators does not resolve a question of standing; it produces more contested drafts. The cost has moved from a quantity that scales with headcount to a quantity that scales with authority, and authority does not respond to headcount.

The pricing consequence is durable. A lab that assumed cultural alignment cost would amortize across cultures may find that the constructive direction reintroduces per-culture cost in a form that resists amortization. The cheap direction scales horizontally across markets; the expensive direction scales politically within each market. These are not the same curve, and a model that looks affordable under the first becomes expensive under the second.

Whose Korean culture does the spec encode?

The spec reads as South Korean, and a particular variant of it. The paper grounds its safe-response policy in what it describes as Korean legal frameworks and social norms; the institutional context is South Korean, and the North Korean perspective is absent, even though Korean is the official language of both states and is spoken by roughly 81 million people across a fractured geography.

The fragmentation is not incidental to the argument; it is the argument. Korean culture has been divided between North Korea (Joseon, or Chosŏn) and South Korea (Hanguk), with vocabulary differences and dialectal diversity within each country (Korean language). The writing system itself splits along the border. Hangul, the alphabet developed during the 15th century, is shared, while Hanja, the Chinese-character script, is in supplementary use in South Korea and now less common in North Korea (Korean language). The Korean-speaking world also extends beyond the peninsula, with official recognition in Yanbian Korean Autonomous Prefecture and Changbai County in China.

A model fine-tuned against South Korean statutes will, by construction, produce answers that read as aligned with one state’s institutions and as foreign to speakers whose norms are mediated by a different state, or by the Chinese state, or by diaspora conventions. There is no neutral “Korean culture” available to align to. The choice to specify one at all, and to ground it in one jurisdiction’s law, is a political act dressed as a technical one. The paper’s title says “Korean culture.” The implementation says “South Korean, statutory, contemporary.”

This is not a complaint about the paper, which is upfront about its grounding. It is the mechanism by which a constructive spec becomes representational. The spec has to land somewhere, and wherever it lands, it takes sides. The same problem recurs for any culture large enough to contain a disagreement, which is to say, for any culture at all.

What does this mean for shipping a multicultural model?

For a lab shipping a multilingual model, the constructive direction implies that culturally coherent behavior now requires per-culture specifications that must be authored, maintained, and politically defended, a cost suppression lists never imposed. There are three plausible responses, each following from the cost structure above rather than from documented vendor practice, and none restores the old cost structure.

The first is to keep suppression-only and accept the quality ceiling. This is the cheap path and it still works for the bulk of harm reduction. The Korean paper is useful here precisely because it shows what suppression-only gives up: the constructive context, the institutional routing, the answer in place of the refusal. A lab that ships suppression-only ships a culturally competent floor with no ceiling, and the gap shows up wherever a user needed the model to know something rather than refuse something.

The second is to author the specs internally and absorb the representational cost. This is the expensive path and it does not parallelize. Each spec is a contested document requiring standing the vendor may not have, and each market adds another one. The cost line for this work does not look like a labeling budget; it looks like a policy and editorial function, complete with the legitimacy problems that implies. Few labs are staffed for it, and fewer still are accountable to the communities they would be specifying.

The third, and the most likely in practice, is to delegate spec authorship to a local partner. This externalizes the representational problem onto an entity that presumably has more standing in the culture, but it hands that partner ownership of the model’s working definition of the culture. Governance follows the delegation: the partner now shapes behavior at inference time for every query in that language, and the lab has ceded the editorial function that constructive alignment was supposed to internalize.

The practical read for builders: the constructive direction produces better outputs, but it does not inherit the operational properties of suppression lists. A cultural spec behaves more like an externally authored content policy than like a dataset. It is binding at inference time, and unlike a suppression list it will need a governance process behind it, because the question of whether the spec is right is exactly the question a labeling pipeline was never asked to answer.

What are the honest limits of this approach?

Three caveats temper how far to take this read: the spec is single-country, no vendor has deployed it, and the evaluation may be partly circular.

The spec is single-country. The paper grounds its policy in what it calls Korean norms and law; in its institutional context this reads as South Korean, and North Korean perspectives are not represented (arXiv:2606.06797). For a paper titled around Korean culture, that is a real scoping decision, and it is the same decision that makes the spec politically legible: grounding in one jurisdiction’s statutes is what makes “constructive” tractable in the first place. The generality claim is weaker than the title suggests, and a reader should treat “Korean” in the results as “South Korean, present-day.”

There is no vendor adoption. This is an ICML 2026 Workshop paper, and there is no evidence in the brief of commercial deployment or vendor uptake. The constructive direction is a research contribution and an argument, not a shipping practice. Treating it as the new alignment playbook would be ahead of the evidence, and it would repeat the familiar error of reading a workshop result as a product spec.

The evaluation may be partially circular. The Korean cultural safe rate appears to be measured against a notion of safe behavior derived from the policy at the center of the pipeline, so part of what the metric captures may be how well the model conforms to its own training signal. The abstract does not spell out the evaluation rubric in full, so this should be read as an inference about the methodology rather than a confirmed feature of it. Layered on top of that, the training triplets were generated by three frontier models, whose existing cultural biases may be baked into the constructive baseline the paper treats as ground truth (arXiv:2606.06797). The constructive spec is only as neutral as the models that instantiated it, and those models arrived with their own embedded assumptions about what a coherent answer looks like.

The durable takeaway is not the Korean result, which may age as alignment methods move past DPO and past frontier-model-generated training signal. It is the conceptual shift the paper names. Cultural alignment has a subtractive form that scales and a constructive form that does not, and the constructive form relocates the cost from data labeling into representational politics. Whoever writes that spec owns the model’s working definition of a culture, and Korean culture, like any culture, will not sit still long enough to be specified once. The question the title asks, who writes the spec, is also the answer: whoever does, the model will enforce their draft until someone with equal standing rewrites it.

Frequently Asked Questions

Does this constructive approach require a unified legal framework?

The Korean instantiation worked because South Korea provides a single statutory system to ground the policy. For cultures split across jurisdictions, like Kurdish speakers spanning Turkey, Iraq, Iran, and Syria, the approach would force a choice about which state’s law represents the culture. That choice is the political act the suppression list was designed to avoid.

How does DPO differ from RLAIF-style cultural alignment?

DPO fits a policy directly to preference pairs without a separately trained reward model. RLAIF pipelines, such as those used by Anthropic for Claude, rely on AI-generated reward models to score responses. The DPO advantage is computational efficiency, which is why the Korean paper could apply it across six open-weight base models rather than a single frontier system.

What’s the minimum staffing needed to maintain constructive specs across 20 languages?

Korean used three frontier models to generate candidate responses. Scaling to 20 languages means 60 model calls per prompt, plus subject-matter experts for each culture to vet outputs. Suppression lists scale with annotators who can label across cultures. Constructive specs scale with domain experts whose expertise does not transfer, and headcount does not solve the standing problem.

Does the frontier-model generation step introduce bias?

The training triplets came from three frontier models, which arrived with their own embedded assumptions about what a coherent answer looks like. The constructive baseline may inherit those biases even as it tries to replace them. The evaluation circularity risk is that the Korean cultural safe rate may partially measure how well the model conforms to its own training signal rather than to an independent standard.

How would vendors handle competing constructive specs for the same language?

Korean has 81 million speakers across both Koreas plus recognized communities in China’s Yanbian Prefecture and Changbai County. If North Korea later develops its own spec grounded in Joseon legal norms, vendors would need to route users by detected dialect or self-reported location. That routing challenge does not exist under suppression lists, where the same veto applies to all Korean-language queries regardless of origin.