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

2026-06-30 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, quick

Summary

LuckyStar 111B, a 111B-parameter hybrid reasoning model, was developed by Cohere and LG CNS to create efficient Korean-English enterprise agents under practical memory and serving constraints. This model leverages Cohere's existing post-trained Command A model, avoiding a new pretraining run, and employs preamble conditioning to manage transitions between concise non-reasoning and tool-oriented reasoning behaviors. The research explored four key strategies for scaling tool-using agents efficiently: multilingual supervised fine-tuning, reinforcement learning with verifiable rewards for multi-step tool-use, language-consistency rewards for Korean responses, and 4-bit quantization for single-GPU serving. The adapted model demonstrates enhanced mathematical reasoning, function calling, and agentic natural-language-to-SQL (NL2SQL) capabilities, while maintaining general instruction-following quality in both Korean and English. These findings offer a practical recipe and failure-mode analysis for deploying post-trained multilingual models in memory-constrained agentic workflows.

Key takeaway

For MLOps Engineers deploying multilingual tool-using agents in resource-constrained enterprise environments, consider adapting post-trained models like LuckyStar 111B. You should implement a combination of multilingual supervised fine-tuning, verifiable reinforcement learning, and 4-bit quantization to achieve improved mathematical reasoning and NL2SQL performance on single GPUs. This approach offers a practical recipe to maintain instruction-following quality while optimizing for memory and serving efficiency.

Key insights

Efficiently adapting post-trained multilingual models for tool-using agents under memory constraints is achievable through specific fine-tuning and quantization strategies.

Principles

• Post-trained models adapt efficiently for agents.
• Preamble conditioning manages agent reasoning modes.
• Verifiable RL rewards enhance multi-step tool-use.

Method

The adaptation method combines multilingual supervised fine-tuning, reinforcement learning with verifiable rewards for multi-step tool-use, language-consistency rewards for Korean responses, and 4-bit quantization for single-GPU serving.

In practice

• Deploy Korean-English enterprise agents.
• Improve NL2SQL performance on constrained hardware.
• Enhance mathematical reasoning in multilingual contexts.

Topics

• LuckyStar 111B
• Multilingual Agents
• Tool-Using Agents
• 4-bit Quantization
• Model Adaptation
• Enterprise AI

Best for: AI Architect, NLP Engineer, AI Scientist, Machine Learning Engineer, AI Engineer, MLOps Engineer

Related on AIssential

• Translate-R1: Cost-Aware Translation Tool Use via Reinforcement Learning — A learned policy allows LLMs to introspect their comprehension and invoke translation tools only when necessary, balancing performance and cost.
• Towards Scalable Customization and Deployment of Multi-Agent Systems for Enterprise Applications — A two-stage framework customizes and efficiently deploys LLM-based multi-agent systems for enterprise use.
• Leveraging Routing Dynamics in Mixture-of-Experts Models for Efficient Language Adaptation — MoE language specialization emerges in final layers, enabling efficient adaptation by targeting those layers.
• Language Models Used in AI Agents: 8 Architectures Powering Intelligent Systems — Diverse language model architectures enable AI agents to perform specialized tasks from planning to multimodal understanding.
• Roberta Raileanu of Google DeepMind at RAAIS 2026 — Open-ended learning and tool use are critical for AI models to adapt and continuously improve in dynamic, real-world environments.
• Learning When to Translate for Multilingual Reasoning — RLMs can learn to selectively translate non-English inputs only when direct understanding is unreliable, improving multilingual reasoning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.