Youtu-LLM: Unlocking the Native Agentic Potential for Lightweight Large Language Models
Youtu-LLM: Unlocking the Native Agentic Potential for Lightweight Large Language Models [78.7] ネイティブエージェントインテリジェンスと高い計算効率を調和させる軽量言語モデルであるYoutu-LLMを紹介する。 Youtu-LLMは、スクラッチから体系的に推論と計画能力の育成まで事前訓練されている。 論文参考訳(メタデータ) (Wed, 31 Dec 2025 04:25:11 GMT)
「Youtu-LLM significantly outperforms existing state-of-the-art models of similar scale across both general- purpose (Figure 2) and agentic benchmarks (Figure 1), and in several settings, rivals substantially larger models. Beyond performance gains, our analyses provide the first systematic evidence that agentic pre- training can unlock agent potential in lightweight LLMs, revealing phenomena such as scalable growth of agent capabilities.」と小規模、エージェント向けのモデルの提案。オンデバイスを狙うとエージェント関連の能力を保ったままの小型化が重要であり「We propose a principled training paradigm that enhances native agentic capabilities through innovations in tokenizer design, data allocation, and multi-stage learning, guided by an agent-centric philosophy.」とあるように狙って強化することもできるよう。