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

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Utsav Garg, Sungjin Hong, Jason Jung, Justin Lee, Shaan Desai, Joon Hee Kim, Anirudh Shrinivason, Edmond Wen, Susie Park
• Venue: arXiv: cs.AI
• Published: 2026-06-30
• Collected: 2026-07-01
• URL: http://arxiv.org/abs/2606.31648v1
• PDF: https://arxiv.org/pdf/2606.31648v1
• DOI: UNKNOWN

Plain-Language Brief

Researchers present LuckyStar 111B, a hybrid reasoning model developed by Cohere and LG CNS for enterprise agents that can switch between concise non-reasoning behavior and longer tool-oriented reasoning in Korean and English. The model was adapted for efficient tool use through four methods: multilingual supervised fine-tuning, reinforcement learning with verifiable rewards, language-consistency rewards for Korean responses, and 4-bit quantization for single-GPU serving. The adapted model improves mathematical reasoning, function calling, and agentic NL2SQL performance while maintaining general Korean and English instruction-following quality.

What the Document Says

• The study introduces LuckyStar 111B, a hybrid reasoning model for enterprise agents that can switch between concise non-reasoning behavior and longer tool-oriented reasoning in Korean and English.
• The model was developed through collaboration between Cohere and LG CNS and was adapted from Cohere's Command A model.
• Four methods were used to scale the model for efficient tool use: multilingual supervised fine-tuning, reinforcement learning with verifiable rewards, language-consistency rewards for Korean responses, and 4-bit quantization for single-GPU serving.
• The adapted model showed improvements in mathematical reasoning, function calling, and agentic NL2SQL performance while maintaining general Korean and English instruction-following quality.

Key Concepts Mentioned

• LuckyStar 111B
• hybrid reasoning model
• Cohere
• LG CNS
• Korean-English enterprise agents
• practical memory and serving constraints
• mathematical reasoning
• function calling
• agentic natural-language-to-SQL (NL2SQL) performance
• multilingual supervised fine-tuning
• reinforcement learning
• language-consistency rewards
• 4-bit quantization

Claims Made by the Document

• LuckyStar 111B can switch between concise non-reasoning behavior and longer tool-oriented reasoning in Korean and English.
• The adapted model improves mathematical reasoning, function calling, and agentic NL2SQL performance while maintaining general Korean and English instruction-following quality.

Evidence Used

• The text describes the development of LuckyStar 111B as a hybrid reasoning model that can switch between concise non-reasoning behavior and longer tool-oriented reasoning.
• Improvements in mathematical reasoning, function calling, and agentic NL2SQL performance are mentioned in the results section.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

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.