Kernekoncepter
결정 인식과 일반화된 도구 사용을 강조하는 프레임워크 제안
Resumé
이 논문은 대형 언어 모델의 도구 사용에 대한 새로운 방향을 제시하고, 다양한 실험을 통해 제안된 방법의 효과를 입증합니다. 논문은 결정 인식과 도구 사용의 일반화에 초점을 맞추고 있습니다. 또한, 다양한 실험을 통해 제안된 방법의 효과를 입증합니다.
Abstract
- Tool-augmented large language models (LLMs) are gaining attention for accessing knowledge and reducing hallucination issues.
- Current efforts focus on enhancing open-source LLMs' tool-usage capabilities.
- Proposed DEER framework improves decision-making awareness and generalizability of LLMs.
Introduction
- Despite advancements in LLMs, issues like hallucination and knowledge limitation persist.
- Tool-augmented LLMs offer solutions for accessing domain-specific knowledge.
- Proposed DEER framework addresses limitations in tool-usage of LLMs.
Tool-Usage Paradigms
- Template-driven and token-triggered tool-usage methods are compared.
- Template-driven methods constrain interactions, while token-triggered methods lack generalizability.
- DEER framework introduces decision-aware tool-usage for diverse user queries.
Methodology
- DEER framework constructs tool-usage samples with decision branches.
- Supervised fine-tuning enhances decision-making prowess.
- Mixture of sampling strategies boosts generalization on unseen tools.
Experiments
- DEER outperforms baselines in decision-making awareness and generalization.
- Extensive experiments demonstrate the effectiveness of DEER.
- Proposed method achieves state-of-the-art performance in various scenarios.
Statistik
ChatGPT와 GPT-4는 각각 78.1%와 87.6%의 정확도를 달성함
DEER-13B는 98.6%의 정확도를 달성함
Citater
"For general queries, LLMs should give an answer with their own knowledge rather than resorting to external tools."
"Our proposed DEER is effective and significantly outperforms baselines across various datasets."