Korean-First Value Policy in AI Systems

• Korean-First Value Policy is a set of model-behavior constraints that prioritizes Korean language usage by default while permitting controlled translation when schema conditions demand.
• It is applied in tool-use systems by enforcing schema-level rules that retain Korean input to prevent API execution errors, improving accuracy and semantic integrity.
• In alignment frameworks, the policy guides culturally coherent responses by grounding safe completions in Korean statutes, social norms, and locally meaningful contexts.

Searching arXiv for the cited papers to ground the article in the specified sources. The Korean-First Value Policy denotes a family of model-behavior constraints in which Korean is treated as the default representational and normative reference point, with deviations permitted only under explicit constraints. In "SLM-Based Agentic AI with P-C-G: Optimized for Korean Tool Use" (Jeon et al., 19 Sep 2025), the policy is defined operationally for tool-call argument construction: all “value” fields are emitted in Korean by default, and only parameters whose schema explicitly forbids Korean are converted into English or another code through a controlled translation step. In "Korean Culture into LLM Alignment: Toward Cultural Coherence" (Jung et al., 5 Jun 2026), the same label is recast at the alignment layer as a constructive policy for culturally coherent response generation, requiring refusals and redirections to be grounded in Korean statutes, social norms, and locally meaningful alternatives. Taken together, these uses describe a Korean-first constraint on both symbolic interface behavior and semantic alignment.

1. Definition and scope

In the tool-use setting, the Korean-First Value Policy is a rule for constructing tool-call arguments such that all “value” fields are emitted in Korean by default (Jeon et al., 19 Sep 2025). Only when a parameter’s schema explicitly forbids Korean—such as English-only enumerations, regex or format constraints, or fields on a small whitelist—is the value converted into English or another code via a controlled translation step. The stated motivation is that, in Korean-centric services, unintended Korean→English code-switching in tool arguments often leads to back-end execution failures; examples include DB lookups for Korean place names failing if transliterated. The policy therefore aims to preserve semantics between user input, model output, and downstream APIs.

In the alignment setting, the term is extended from interface fidelity to cultural coherence. The policy demands not only suppression of harmful outputs but a constructive counterpart: models must refuse or redirect in a way that reflects how informed Koreans would themselves respond (Jung et al., 5 Jun 2026). The paper defines a culturally coherent response through three properties: P1 Sociolegal Anchoring, P2 Demographic Specificity, and P3 Grounded Refusal without Over-Refusal. This framing shifts the policy from a narrow language-retention rule to a broader normative program in which Korean law, social norms, and interpretive conventions shape the content of safe responses.

A plausible implication is that “Korean-first” names two related but non-identical design commitments. In one usage, it is a language-preservation policy for tool arguments; in the other, it is a cultural-coherence policy for aligned generation.

2. Formal rule in tool-use systems

The tool-use formulation is specified over a JSON schema $S$ for a tool’s parameters, where each parameter $p$ has an admissible-language set $L_p \subseteq \{\text{“kr”}, \text{“en”}\}$ and an optional whitelist flag $W_p \in \{0,1\}$ (Jeon et al., 19 Sep 2025). Let $v \in V_{kr}$ be the raw value extracted from the user in Korean, and let $T:V_{kr}\to V_{en}$ be a deterministic translation function. The paper defines indicator functions $I_{kr}(p)=1$ if “kr” $\in L_p$, else $0$, and $I_w(p)=W_p$, then sets the Korean-first output value $p$0 as

$p$1

The accompanying verbal rule is explicit. If the schema allows Korean, the system keeps the original Korean. Else if the parameter is on the English-whitelist, it translates. Otherwise, it flags a schema-violation error. This specification makes language admissibility a first-class property of parameter validation rather than an incidental consequence of prompting.

This formalization is significant because it constrains value realization at the schema level. Rather than allowing a model to decide opportunistically whether to transliterate or translate, the policy ties the decision to parameter metadata. This suggests a design in which language form is treated as part of API correctness.

3. Integration into the Planner–Caller–Generator architecture

The policy is embedded in the P-C-G architecture, which separates planning, calling, and generation by role (Jeon et al., 19 Sep 2025). The Planner receives a prompt template whose system instruction states: “When you design a call chain, assume all parameter values remain in Korean unless the schema explicitly forbids it.” The paper notes that the Planner’s reasoning remains unaffected beyond ensuring the Caller will apply the Korean-first rule, and that no pseudocode change is needed at this stage.

The Caller is identified as the core enforcer of Korean-first. For each tool in the tool_chain, it retrieves the schema, extracts raw values from context, applies the two indicator functions $p$2 and $p$3, validates the resulting arguments, constructs the call object, invokes the API, normalizes the response, and triggers replanning if the normalized response contains an error (Jeon et al., 19 Sep 2025). The workflow is defined by the following simplified pseudocode:

$T:V_{kr}\to V_{en}$1

Three implementation points are emphasized. Extraction from dialogue always yields a Korean string. The indicator functions drive the policy decision. Schema validation rejects any value violating pattern or language constraints. The Generator then receives the sequence of {tool_calls, tool_results}, is instructed to produce the final answer in Korean, and applies no further translation inside the Generator; it simply incorporates Korean tool outputs verbatim (Jeon et al., 19 Sep 2025).

Within this decomposition, the policy is not distributed uniformly across modules. The Planner assumes it, the Caller enforces it, and the Generator preserves its outputs.

4. Validation, tokenization, and invocation mechanics

The implementation details in the tool-use paper are highly specific (Jeon et al., 19 Sep 2025). Tokenization uses a SentencePiece tokenizer with a bilingual vocabulary while ensuring that Korean Hangul syllables are split at the syllable-level, with no forced subword into Latin. The stated purpose is to preserve Korean words as compact tokens, thereby reducing the risk of inadvertent fragmentation or translation bias.

Before API invocation, the system runs JSON-schema checks on the arguments object. The standard JSON-schema validator is extended to include a language field: for each string parameter, the system verifies the Unicode block of $p$4 against $p$5. Violations—for example, Korean in an English-only enum—raise a SchemaLanguageError, triggering either user clarification or limited replanning. The runtime registry is also augmented with explicit allowed_languages tags so that the Caller can consult them during execution. In addition, a small hard-coded whitelist enumerates fields where translation is mandatory, with country-code parameters and currency codes given as examples.

Tool invocation itself is adapted to Korean settings by ensuring that all HTTP payloads carry JSON with UTF-8 encoding so as to preserve Korean characters (Jeon et al., 19 Sep 2025). This places the policy at multiple layers: tokenizer behavior, schema metadata, validator logic, runtime invocation, and error recovery.

A plausible implication is that the policy treats Korean preservation as an end-to-end systems property rather than solely a prompting heuristic. The paper’s mechanics indicate that robustness depends on synchronized constraints across representation, validation, and transport.

5. Cultural coherence as an alignment policy

The second paper generalizes the phrase “Korean-First Value Policy” from tool-use argument formatting to cultural alignment (Jung et al., 5 Jun 2026). It argues that cultural-aspect work on LLMs has been dominated by a negative target—outputs to suppress—and that a constructive counterpart is needed: a working definition of what a culturally coherent response is. The paper instantiates this for Korean and places a Korean-culturally-adapted safe-response policy at the center of an alignment-data pipeline.

The three defining properties are central. P1 Sociolegal Anchoring requires the response explicitly to name the applicable Korean statute or social norm. P2 Demographic Specificity requires that, when a protected group is involved, the response identify the group and the locally enforceable anti-discrimination framework. P3 Grounded Refusal without Over-Refusal requires that refusals be more than brittle templates, grounding themselves in locale-specific substance and, when safe, preserving the benign informational core of a query (Jung et al., 5 Jun 2026). The contrast is drawn sharply: rather than training a model only to refuse or safe-complete, the constructive framing requires each refusal or redirection to carry Korean cultural substance.

The policy is organized through a Korean harm taxonomy with five top-level domains and twelve subcategories grounded in Korean legal codes, social norms, and historical context. These domains are Harmful Content; Bias & Unfair Expression; Misinformation Harm; Privacy & Info Security; and Malicious Use. Each subcategory is tied to relevant Korean statutes and social norms, such as the Personal Information Protection Act for “Personal Sensitive Info Requests” and the Public Official Election Act for “Prohibited Advisory” during campaign periods (Jung et al., 5 Jun 2026).

At the per-category level, the safe-response policy specifies three fields for each subcategory: Core Principle, Judgment Criteria, and Response Strategy. For Privacy Violation, the core principle is that personal privacy is absolute and non-consensual collection or identification is prohibited under the Personal Information Protection Act; the response strategy is to give a firm Korean-legal refusal, educate on privacy importance, and offer constructive defensive advice (Jung et al., 5 Jun 2026). This design makes Korean statutes and institutional procedures part of the output policy itself.

6. Preference optimization and empirical results

The alignment pipeline fine-tunes six open-weight Korean-capable LLMs using Direct Preference Optimization (DPO) over triplets $p$6, where $p$7 is the preferred response for query $p$8 (Jung et al., 5 Jun 2026). The models listed are A.X-4.0-Light, EXAONE-3.5, Kanana-1.5, Qwen-2.5, Gemma-3, and Llama-3.1. The DPO objective is written as

$p$9

with $L_p \subseteq \{\text{“kr”}, \text{“en”}\}$0 and $L_p \subseteq \{\text{“kr”}, \text{“en”}\}$1 a KL-regularization weight. The reported training configuration uses LoRA adapters with rank $L_p \subseteq \{\text{“kr”}, \text{“en”}\}$2, $L_p \subseteq \{\text{“kr”}, \text{“en”}\}$3, dropout $L_p \subseteq \{\text{“kr”}, \text{“en”}\}$4 on all attention and MLP projections, 4-bit NF4 quantization, BF16 compute, batch size $L_p \subseteq \{\text{“kr”}, \text{“en”}\}$5, learning rate $L_p \subseteq \{\text{“kr”}, \text{“en”}\}$6, $L_p \subseteq \{\text{“kr”}, \text{“en”}\}$7, and $L_p \subseteq \{\text{“kr”}, \text{“en”}\}$8 balanced triplets across five top-level domains (Jung et al., 5 Jun 2026).

The quantitative findings are reported as follows. Korset safe rate rises for all six models, with average $L_p \subseteq \{\text{“kr”}, \text{“en”}\}$9. KoBBQ bias compliance is preserved or modestly improved, with average $W_p \in \{0,1\}$0. KMMLU, Ko-MT, HRM8K, and HumanEval⁺ remain within noise levels, with $W_p \in \{0,1\}$1 (Jung et al., 5 Jun 2026). The paper therefore states that cultural-coherence gains come with no large general-capability trade-off.

The tool-use paper reports a different but complementary empirical profile for the P-C-G system built around the Korean-first policy (Jeon et al., 19 Sep 2025). Under Korean tool-use scenarios, the reported Call Accuracy is $W_p \in \{0,1\}$2 for “Ours (w/ policy),” compared with $W_p \in \{0,1\}$3 for GPT-4o-mini, $W_p \in \{0,1\}$4 for EXAONE-4.0-32B, $W_p \in \{0,1\}$5 for Qwen3-14B, and $W_p \in \{0,1\}$6 for Qwen3-8B. For Missing Functions Accuracy, “Ours (w/ policy)” reports $W_p \in \{0,1\}$7, compared with $W_p \in \{0,1\}$8, $W_p \in \{0,1\}$9, $v \in V_{kr}$0, and $v \in V_{kr}$1 for the listed baselines. For Task Success Rate, “Ours (w/ policy)” reports $v \in V_{kr}$2, compared with $v \in V_{kr}$3, $v \in V_{kr}$4, $v \in V_{kr}$5, and $v \in V_{kr}$6. On average tokens and latency for correct queries only, “Ours (w/ policy)” reports $v \in V_{kr}$7 tokens and $v \in V_{kr}$8 seconds (Jeon et al., 19 Sep 2025).

Setting	Metric	Ours (w/ policy)
Tool use	Call Accuracy	75.0
Tool use	Missing Functions	91.2
Tool use	TSR	79.7
Tool use	Tokens Avg	4360.3
Tool use	Latency (s)	9.1
Alignment	Average $v \in V_{kr}$9 Korset	+6.59
Alignment	Average $T:V_{kr}\to V_{en}$0 KoBBQ	+1.64

In the interpretation given by the tool-use paper, the superior Call Accuracy and high Missing-Function score illustrate robust, language-consistent API invocations thanks to Korean-first, while token usage and latency remain competitive (Jeon et al., 19 Sep 2025). In the alignment paper, the interpretation is that constructive, Korean-grounded refusals can improve cultural safety without large degradation on general-capability benchmarks (Jung et al., 5 Jun 2026).

7. Controversies, misconceptions, and open questions

A central misconception would be to treat the Korean-First Value Policy as merely “always output Korean.” The tool-use formulation explicitly rejects that simplification: English or another code is used when a parameter’s schema explicitly forbids Korean or when a field lies on a whitelist such as country-code parameters or currency codes (Jeon et al., 19 Sep 2025). The policy is therefore conditional and schema-governed rather than absolute.

A second misconception would be to equate the policy entirely with refusal behavior. The alignment paper argues against a purely negative harm-suppression approach and instead defines cultural coherence through P1–P3, which require named statutes, demographic specificity where relevant, and constructive alternatives when safe (Jung et al., 5 Jun 2026). In this view, the policy is not reducible to suppression; it is a content-structuring rule for locale-specific helpfulness under constraint.

The papers also delimit current evidence. The tool-use work states that it does not include a stand-alone ablation on the Korean-first policy, even though the overall P-C-G system is built around that policy (Jeon et al., 19 Sep 2025). Consequently, claims about the isolated causal contribution of the policy, as distinct from role specialization and schema-value validation more generally, should be read cautiously. The alignment paper likewise identifies limitations: portability requires new sociolegal artifacts per locale; temporal drift in law requires sustainable governance models for policy updates; the judge pipeline uses frontier LLMs under unanimous rules; and human validation across Korean subgroups remains an open research challenge (Jung et al., 5 Jun 2026).

These limitations indicate that the Korean-First Value Policy is best understood as an operational pattern rather than a closed doctrine. In one instantiation, it is a schema-conditioned language-retention mechanism for Korean tool use. In another, it is a culturally anchored alignment policy that requires refusals and redirections to be legally and socially meaningful in Korean settings. The convergence of these two instantiations suggests a broader design principle: Korean-first systems seek to preserve Korean linguistic and sociolegal semantics unless an explicit interface or policy constraint requires otherwise.