Skip to content

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

🕒 Published (v1): 2026-06-30 13:29 UTC · Source: Arxiv · link

Why this paper was selected

Cohere + LG CNS; multilingual tool-use agents at 111B scale — real enterprise harness

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

LuckyStar 111B is a 111B-parameter bilingual (Korean-English) enterprise agent model adapted from Cohere's Command A via a three-stage pipeline (hybrid SFT → RLVR → DPO) without full retraining. The key design insight is separating internal reasoning language (English) from user-facing response language (Korean), enforced by a language-consistency reward penalty during RLVR. 4-bit quantization enables single-H100 deployment with negligible quality loss.

Problem

Enterprise Korean-English agents must parse Korean requests, reason over English-heavy schemas/documentation, execute tools (SQL, retrieval), and return Korean-language answers with grounded numbers — all under tight GPU memory constraints. Existing post-trained multilingual models lack verifiable tool-use reasoning and exhibit language drift (Korean prompts producing English final answers) when RL-optimized for agentic tasks.

Method

The pipeline adapts Command A (111B, fully post-trained) through three sequential stages:

  1. Hybrid SFT: ~80% reasoning examples (math, code, NL2SQL) and ~20% non-reasoning instruction-following data, mixed under two preamble modes (reasoning preamble vs. concise preamble). For Korean, a mixed-language strategy is adopted: English reasoning traces paired with Korean final answers, generated via rejection sampling from 20,000 Korean prompts. NL2SQL SFT data is bootstrapped via best-of-N rejection sampling (8 candidates/prompt, verified by heuristic + LLM judge) from Spider, BIRD, and SynSQL, yielding >50,000 agentic prompts.

  2. RLVR: REINFORCE Leave-One-Out (RLOO) with binary rewards for NL2SQL (rule-based verifier + LLM judge fallback) and math. Prompts with pre-RLVR pass rate in \([12.5\%, 87.5\%]\) are selected. A language-consistency penalty is applied: $\(r = r_{\text{correctness}} - r_{\text{penalty}}, \quad r_{\text{penalty}} = \begin{cases} 0.5 & \text{if final answer language mismatches prompt language} \\ 0 & \text{otherwise} \end{cases}\)$ Rollouts exceeding 32k tokens are filtered (not zero-rewarded) to avoid discouraging long reasoning chains.

  3. Offline DPO: Low-learning-rate preference alignment to correct verbosity introduced by RLVR; concise outputs labeled as chosen, verbose RLVR outputs as rejected.

Inference-time preamble conditioning (no weight/architecture change) selects reasoning vs. non-reasoning mode.

Key Contributions

  • Three-stage adaptation pipeline (hybrid SFT → RLVR → DPO) for adding verifiable tool-use reasoning to a post-trained multilingual model without full retraining.
  • Language-consistency reward penalty in RLVR that reduces Korean-to-English answer drift from 12.2% to 0.8% over 200 training steps.
  • Empirical finding that reasoning in English while answering in the user's language outperforms Korean-only reasoning traces (AIME 2024: 69.3 vs. 50.9 pass@1 for Korean prompts).
  • 4-bit quantization reducing memory footprint ~50%, enabling single 80GB H100 deployment with near-parity benchmark performance.
  • Analysis of three agentic failure modes: NL2SQL cold-start, RLVR language drift, and post-RLVR verbosity — each with a targeted remedy.

Results

  • AIME 2024 (Korean prompt): LuckyStar 111B 69.3 vs. Command A 10.0; 4-bit version 66.0
  • MATH 500 (Korean prompt): 93.6 vs. 76.2 (Command A); 4-bit 95.6
  • AIME 2024 (English prompt): 73.7 vs. 13.3 (Command A); comparable to Claude 3.7 Sonnet (46.7) and below Qwen3 235B A22B (80.0)
  • BFCL v3 (function calling): 38.0 vs. 7.3 (Command A), 37.3 (GPT-4o), 38.0 (Claude 3.7 Sonnet)
  • Enterprise NL2SQL (internal, 150 questions): 70.6 vs. 64.4 (Command A), 71.7 (GPT-4o), 58.6 (Claude 3.7 Sonnet)
  • LG Agentic Eval (internal, 60 questions): 4.85 vs. 2.67 (Command A), 2.56 (GPT-4o), 5.04 (Claude 3.7 Sonnet)
  • KMMLU: 68.6 vs. 64.9 (Command A); Korean IFEval: 77.8 vs. 77.2; Korean MT-Bench: 8.09 vs. 8.18
  • 4-bit vs. FP8: within 0.7 points across all reported academic benchmarks

Limitations

  • Internal NL2SQL and LG Agentic benchmarks are not publicly reproducible; enterprise-specific schema/formatting conventions limit generalizability of results.
  • Proprietary training data details withheld due to legal/confidentiality constraints, limiting reproducibility.
  • Model still relies on English reasoning for Korean prompts; native Korean reasoning remains future work, especially for culturally or domain-specific knowledge.
  • No latency, throughput, or energy measurements for the 4-bit serving configuration — deployment feasibility is checked but not comprehensively characterized.
  • Tool-use benchmark results for the 4-bit model were not completed before the evaluation cutoff.
  • Baseline comparisons use a fixed evaluation snapshot; newer model releases were not re-evaluated.

Relevance to Harnesses / Meta-Harnesses

LuckyStar's three-stage pipeline (SFT bootstrapping → RLVR with verifiable rewards → DPO correction) is itself a training meta-harness: it orchestrates multiple training phases with different data curation pipelines, reward functions, and filtering heuristics to produce a deployable agent. The cold-start problem for NL2SQL — where the base model has <1% accuracy and RLVR alone cannot bootstrap — illustrates a general pattern in agentic training harnesses: verifiable RL requires a prior SFT stage to provide sufficient initial competence for reward signal to propagate. The language-consistency reward and long-rollout filtering are harness-level interventions that shape training dynamics rather than model architecture, directly analogous to meta-harness logic that governs when and how sub-agents are rewarded. The preamble-conditioning design (one model, two behavioral modes selected at inference time) is a lightweight harness-controlled dispatch mechanism that avoids maintaining separate model artifacts for different workflow types.