<?xml version='1.0' encoding='utf-8'?>
<rss version="2.0"><channel><title>LLM agent 評測基準與工具使用：連結排序</title><link>https://github.com/Eliot5566/JEV-Paper-Radar</link><description>Retrieved public metadata abstracts; not full-text review. Retrieval: 0 records without complete title/abstract, 3 errors. These were not scored; see report.json for details. Scores were made in the current Claude conversation and are not calibrated probabilities. Chat usage and cost are unavailable to this report. This renderer made no API requests. Interest: LLM agent 的評測基準（benchmarks）與工具使用能力（tool use）</description><item><title>AgentBench: Evaluating LLMs as Agents</title><guid isPermaLink="false">arxiv:2308.03688</guid><link>https://arxiv.org/abs/2308.03688</link><description>直接命中「LLM agent 評測基準」：提出 8 個環境的 agent benchmark，比較商用與開源模型並分析失敗原因；工具使用僅隱含於環境互動，未專門評測 API 呼叫。</description></item><item><title>Toolformer: Language Models Can Teach Themselves to Use Tools</title><guid isPermaLink="false">arxiv:2302.04761</guid><link>https://arxiv.org/abs/2302.04761</link><description>高度相關於「工具使用能力」：自監督學習何時、如何呼叫計算機、搜尋、翻譯等 API；但它是訓練方法而非 agent 評測基準，也非多步 agent 設定，故列為 maybe。</description></item></channel></rss>
