Unverified50% confidenceFactExact time
编码 Agent 评估经历了从 HumanEval（164 道函数级编程题）到 SWE-bench（真实 GitHub Issue 修复）再到 Terminal Bench 的演进
1
Sources
50%
Confidence
Long-term
Relevance
8/27/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSWE-bench requires models to fix issues in real GitHub repositories, while HumanEval/MBPP are function-level code generation benchmarks.71% similarUnverifiedSWE-bench评估模型解决真实GitHub Issue的能力，被认为比HumanEval更贴近Agent实际工作场景69% similarUnverifiedSWE-bench requires models to resolve real software defects on GitHub and is the authoritative benchmark for evaluating agent-level programming capabilities.69% similarUnverifiedSWE-bench (Software Engineering Benchmark) is currently the core benchmark for measuring agent coding capability, requiring the model to start from real GitHub Issues, autonomously locate problems in the codebase, and submit fix patches.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/808666API
curl https://kongchang.com/api/v1/knowledge/claims/808666MCP
get_claim(id=808666)