PaperlayerView exact revision on arXiv

Source-linked AI summary

Think in English, Answer in Korean: Efficient Adaptation of Multilingual Tool-Using Agents

Utsav Garg, Sungjin Hong, Jason Jung, Justin Lee, Shaan Desai, Joon Hee Kim, Anirudh Shrinivason, Edmond Wen, Susie Park

arXiv:2606.31648v1cs.AIcs.LG

TL;DR

Korean-English enterprise agents must combine private-data reasoning, tool use, language-consistent answers, and constrained memory and serving budgets. The model is adapted from fully post-trained Command A using preamble-conditioned hybrid SFT, RLVR, language-consistency rewards, preference alignment, and 4-bit quantization. Verifiable NL2SQL rewards, language-consistency penalties, and preference alignment improve tool use while preserving Korean and English instruction-following behavior; 4-bit quantization enables single-H100 deployment with little observed quality loss.

  • Problem

    Korean-English enterprise agents must combine private-data reasoning, tool use, language-consistent answers, and constrained memory and serving budgets.

    Source: §1 · p. 1

  • Method

    The model is adapted from fully post-trained Command A using preamble-conditioned hybrid SFT, RLVR, language-consistency rewards, preference alignment, and 4-bit quantization.

    Source: §1 · p. 1

  • Results

    Verifiable NL2SQL rewards, language-consistency penalties, and preference alignment improve tool use while preserving Korean and English instruction-following behavior; 4-bit quantization enables single-H100 deployment with little observed quality loss.

    Source: §1 · p. 1, §5 · p. 5

  • Takeaways & Limitations

    Adapting an existing post-trained model toward verifiable enterprise workflows can combine focused agent capabilities with preserved general behavior and lower serving requirements.

    Source: §4 · p. 5, §5 · p. 5

    Some evaluations are internal and proprietary data and evaluation details are limited, constraining reproducibility and requiring public benchmarks to complement deployment-specific tests.

    Source: §4 · p. 5

Abstract

from arXiv · show

We present LuckyStar 111B, a 111B-parameter hybrid reasoning model developed through a collaboration between Cohere and LG CNS for Korean-English enterprise agents under practical memory and serving constraints. The model trains from Cohere's fully post-trained Command A model rather than a new pretraining run, and uses preamble conditioning to switch between concise non-reasoning behavior and longer tool-oriented reasoning. We study four choices for scaling tool-using agents efficiently: multilingual supervised fine-tuning, reinforcement learning with verifiable rewards for multi-step tool-use tasks, language-consistency rewards for Korean user-facing responses, and 4-bit quantization for single-GPU serving. The adapted model improves mathematical reasoning, function calling, and agentic natural-language-to-SQL (NL2SQL) performance while preserving general Korean and English instruction-following quality. These results provide a practical recipe and failure-mode analysis for adapting post-trained multilingual models to verifiable agentic workflows under memory-constrained deployment.

1. Introduction

LuckyStar 111B adapts a fully post-trained multilingual model for Korean-English enterprise agents that must combine reasoning, tool use, and constrained deployment. Its hybrid design adds specialized capabilities while preserving general instruction following and supporting language-consistent, memory-efficient operation.

  • LuckyStar 111B adapts Cohere’s fully post-trained Command A for Korean-English enterprise agents instead of training a new model from scratch.The approach targets agent quality, verifiable tool use, and memory footprint together.Source: §1 · p. 1
  • Preamble conditioning lets one set of weights switch between longer reasoning for tools, mathematics, and NL2SQL and concise ordinary responses.Downstream systems select reasoning depth at inference time without separate models.Source: §1 · p. 1
  • The three-stage adaptation pipeline combines hybrid SFT, RLVR for verifiable reasoning and tool use, and preference alignment for concise user-facing behavior.This adds reasoning and tool-use behavior without full retraining.Source: §1 · p. 1, Figure 1 · p. 2
  • A language-consistency reward penalty addresses Korean prompts drifting toward English final answers during RLVR.The intervention targets multilingual reasoning failures while retaining English-heavy intermediate reasoning.Source: §1 · p. 1
  • 4-bit quantization preserves benchmark quality in the reported evaluations while enabling single-H100 serving for the 111B-parameter model.This reduces deployment requirements for enterprise settings.Source: §1 · p. 1

2. Method

The method adapts a post-trained multilingual model through hybrid SFT, verifiable-reward reinforcement learning, and preference alignment for reasoning and tool use. It addresses multilingual reasoning imbalance, constructs executable NL2SQL supervision, and corrects RLVR-induced language drift and verbosity.

Hybrid SFT

  • Hybrid SFT teaches reasoning and non-reasoning modes, using roughly 80% reasoning examples and 20% instruction-following examples.Preambles select longer intermediate reasoning or concise final responses without changing weights or architecture.Source: §2.1 · p. 2

Multilingual Reasoning Strategy

  • English reasoning outperformed Korean reasoning on Korean prompts in early experiments, motivating a mixed-language strategy with English traces and Korean final answers.Machine-translating 30% of traces into Korean produced little improvement, so 20,000 Korean prompts were used for verified mixed-language examples.Source: §2.2 · p. 2

Agentic NL2SQL Data

  • Executable NL2SQL examples are created by retaining successful gold queries with outputs between one and 500 tokens, yielding 100,000 verifiable prompt-query pairs.Executable outputs support automated checking for RLVR and rejection sampling.Source: §2.3 · pp. 2–3, Figure 2 · p. 3
  • Because the base policy achieved less than 1% accuracy on the curated NL2SQL set, best-of-N rejection-sampled SFT bootstrapped the policy before RLVR.The RLVR mixture contains over 50,000 agentic NL2SQL prompts and 4,000 mathematics prompts, with binary verifier or judge rewards.Source: §2.3 · p. 3, §2.4 · p. 3

RLVR and Language Consistency

  • A language-consistency penalty reduces Korean-to-English answer drift by checking whether final responses match prompt language while reasoning remains in English.Prompt languages are pre-classified and ambiguous prompts are excluded.Source: §2.4 · p. 3, Table 2 · p. 3
  • RLVR improves reasoning and agentic NL2SQL while reducing repetition failures, but it also increases verbosity and harms brief-response instruction following.Rollouts over 32k tokens are filtered rather than assigned zero reward to avoid teaching premature termination.Source: §2.4 · p. 3

Offline Preference Alignment

  • Low-learning-rate DPO uses alignment and Korean-specific preference pairs to restore conciseness without compromising reasoning and tool-use capabilities acquired during RLVR.Direct answers are preferred over verbose post-RLVR outputs.Source: §2.5 · p. 3

3. Results

LuckyStar 111B는 수학 추론, 함수 호출, 기업용 NL2SQL에서 Command A보다 향상되었으며, 일반적인 한국어·영어 성능과 4비트 배포 효율을 함께 유지했다.

3.1. Reasoning and Tool Use

  • LuckyStar 111B는 Command A보다 한국어·영어 수학 추론, BFCL v3 함수 호출, 내부 기업용 도구 사용 평가에서 향상되었다.수학 평가는 AIME 2024와 MATH 500의 영어 및 한국어 번역 버전을 포함한다.Source: §3.1 · p. 4

3.3. Efficient Deployment

  • 내부 평가는 배포별 워크플로와 스키마를 반영하지만, 도구 사용 결과는 4비트 모델에 대해 보고되지 않았다.4비트 도구 사용 평가가 최종 평가 시점까지 완료되지 않았기 때문이다.Source: §3.1 · p. 4

3.1. Reasoning and Tool Use

  • Enterprise NL2SQL 점수는 7.3에서 38.0으로, LG Agentic Evaluation 점수는 2.67에서 4.85로 향상되었다.절대 점수는 현실적인 기업용 도구 사용이 여전히 어렵다는 점도 보여준다.Source: §3.1 · p. 4

3.2. General Instruction Following

  • 한국어 성능은 KMMLU 68.6, ARC Challenge 89.2, IFEval 77.8, MT-Bench 8.09에 도달했고, 영어 성능은 각각 82.7, 93.8, 89.4, 8.45였다.한국어 ARC, IFEval, MT-Bench는 영어 평가의 번역 버전이다.Source: §3.2 · p. 4

3.3. Efficient Deployment

  • 4비트 양자화는 메모리 사용량을 약 50% 줄이고 단일 80GB H100에서 111B 모델을 실행하게 하면서 보고된 벤치마크 품질을 FP8과 가깝게 유지했다.이는 완전한 서빙 연구가 아니라 배포 가능성 점검이며, 지연시간·처리량·에너지 측정은 남은 과제다.Source: §3.3 · p. 4

4. Discussion

논의는 영어 내부 추론과 한국어 사용자 응답을 분리하는 설계, NL2SQL의 콜드스타트 필요성, 그리고 배포 중심 적응의 실용성을 강조한다.

4. Discussion

  • 영어 추론 흔적은 검증 가능한 추론에 더 강한 학습 신호를 제공하고, 언어 일관성 보상은 한국어 프롬프트의 영어 최종 답변을 줄인다.그러나 한국어 고유 추론은 문화적·도메인 특화 지식이 중요한 작업에서 여전히 미래 과제다.Source: §4 · pp. 4–5
  • NL2SQL에서는 거부 샘플링 SFT가 RL이 정책을 개선할 만큼의 초기 능력을 제공하므로, 콜드스타트 데이터가 필요하다.이는 보상 품질과 프롬프트 선택이 검증 가능한 에이전트 RL의 핵심이라는 관찰과 일치한다.Source: §4 · p. 5
  • 개발 중 확인된 NL2SQL 콜드스타트, 한국어-영어 언어 이동, 추론 최적화 후 장황함은 각각 SFT, 언어 보상, 선호 정렬로 다뤄졌다.저자들은 이를 검증 가능한 보상, 다국어 행동, 배포 지향 적응의 접점에서 발생하는 문제로 본다.Source: §4 · p. 5
  • 배포 중심 관점에서는 모든 능력에 최적화된 단일 모델보다, 기존 사후학습 모델을 검증 가능한 워크플로에 맞추고 양자화하는 방식이 유용할 수 있다.이 방식은 일반 지시 따르기를 보존하면서 서빙 요구량을 줄이는 것을 목표로 한다.Source: §4 · p. 5
  • 일부 평가는 내부 워크플로와 제한된 독점 데이터에 의존하므로, 공개 벤치마크를 대체하기보다 보완하며 재현성도 제한된다.법적·개인정보·기밀성 의무가 독점 데이터와 내부 평가 세부사항 공개를 제한한다.Source: §4 · p. 5

5. Conclusion

LuckyStar 111B combines hybrid reasoning, tool-use improvements, and language-consistency training while preserving bilingual instruction following. Four-bit quantization enables deployment on a single H100 with little observed quality loss.

  • 4-bit quantization makes LuckyStar 111B deployable on a single H100 with little observed quality loss in the reported evaluations.Source: §5 · p. 5
  • Preamble conditioning lets one model support both reasoning and non-reasoning modes.Source: §5 · p. 5
  • Verifiable NL2SQL rewards, language-consistency penalties, and preference alignment improve tool use while preserving Korean and English instruction-following behavior.Source: §5 · p. 5
  • Future work targets native-language reasoning data, Korean domain adaptation, comprehensive serving-efficiency measurement, and closer integration with enterprise knowledge bases and tools.Source: §5 · p. 5

LLM/Agent Usage Disclosure

The paper used LLMs throughout model development and also for proofreading, while retaining author responsibility for technical claims, experiments, and final text.

  • LLMs generated candidate reasoning traces, judged answer equivalence when deterministic verification was insufficient, checked final-answer language consistency, and produced candidate SQL traces with a read-only execution tool.Source: LLM/Agent Usage Disclosure · pp. 5–6
  • LLMs also helped proofread the paper for clarity and flow, while the authors remained responsible for technical claims, experiments, and the final text.Source: LLM/Agent Usage Disclosure · p. 6

Related papers

Loading 2606.31648v1…