37. 论文阅读地图与 Benchmark Card
学习与参考 · 持续进行
学习目标
- 建立 Agent Evaluation 论文阅读顺序
- 学会区分方法贡献与实验结果
- 使用统一 Benchmark Card 记录知识
前置知识
- 全书任意章节
本章产物: 至少完成 8 篇核心论文阅读卡。
37.1 第一阶段:全景与方法论
- Evaluation and Benchmarking of LLM Agents: A Survey
- An Evaluation-Driven Approach to Designing LLM Agents
- Establishing Best Practices for Building Rigorous Agentic Benchmarks
- AlphaEval: Evaluating Agents in Production
读完回答:
- What to Evaluate?
- How to Evaluate?
- Production 与 Benchmark 差异?
- Eval 本身如何验证?
37.2 第二阶段:通用 Benchmark
- AgentBench
- AgentBoard
- GAIA
重点看:
- environment;
- task design;
- progress;
- generality。
37.3 第三阶段:交互/状态
- τ-bench
- τ²-bench
- WebArena
重点看:
- user simulator;
- database state;
- reproducible environment。
37.4 第四阶段:工具/代码
- ToolLLM / ToolBench
- StableToolBench
- BFCL
- SWE-bench
重点看:
- structured calls;
- execution;
- environment stability;
- tests。
37.5 第五阶段:安全
- AgentDojo
- Agent-SafetyBench
- OpenAgentSafety
重点看:
- threat model;
- benign utility;
- attack success;
- real tools。
37.6 第六阶段:Evaluator
- Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
- Agent-as-a-Judge
重点看:
- agreement;
- bias;
- process-level judge。
37.7 Benchmark Card 模板
Title:
Year:
Authors:
Problem:
Agent type:
Environment:
Task source:
Action space:
Observation:
Success signal:
Metrics:
Human validation:
Number of runs:
Strength:
Known limitation:
Contamination risk:
License:
What we should borrow:
What we should not copy:37.8 三遍阅读法
第一遍:问题、结构、主要结论。 第二遍:Task、Environment、Metric、实验设置。 第三遍:只复现一个最关键评测机制。
不要把论文中的模型排名当作最重要知识;真正值得复制的是评测设计。
本章依据
- Evaluation and Benchmarking of LLM Agents: A Survey
- Establishing Best Practices for Building Rigorous Agentic Benchmarks — Zhu et al., 2025
- AlphaEval: Evaluating Agents in Production — Lu et al., 2026
- AgentBench: Evaluating LLMs as Agents — Liu et al., 2023
- AgentBoard: An Analytical Evaluation Board of Multi-turn LLM Agents — Ma et al., 2024
- GAIA: a benchmark for General AI Assistants — Mialon et al., 2023
- $\tau$-bench — Yao et al., 2024
- WebArena — Zhou et al., 2023
- SWE-bench — Jimenez et al., 2023 / ICLR 2024
- AgentDojo — Debenedetti et al., 2024