EXAONE 4.0: Unified Large Language Models Integrating Non-reasoning and Reasoning Modes
EXAONE 4.0: Unified Large Language Models Integrating Non-reasoning and Reasoning Modes [42.3] EXAONE 4.0は、EXAONE 3.5の優れた使いやすさとEXAONE Deepの高度な推論能力の両方を達成するために、非推論モードと推論モードを統合している。 EXAONE 4.0シリーズは、高性能に最適化された中型32Bモデルと、オンデバイスアプリケーション用に設計された小型1.2Bモデルである。 論文参考訳(メタデータ) (Tue, 15 Jul 2025 15:24:51 GMT)
LLM/LRMハイブリッドなLGのモデル。「Unified Mode Training In the combined dataset, the NON-REASONING data primarily consists of diverse tasks, while the REASONING data is centered on Math and Code domains. Rather than fine-tuning the two modes sequentially, we combine both modes and train them together.」とのこと。構築過程の「After unified NON-REASONING/REASONING mode fine-tuning, to address domain imbalance, we perform a second round of training using high-quality REASONING data from the Code and Tool Use domains, reusing these samples to further enhance the performance.」が興味深い。