Unverified50% confidenceBenchmarkExact time
2024年一项研究对GPT-4在Temperature=0下的代码生成任务进行100次测试，发现约8%的测试用例存在跨次运行的输出差异
1
Sources
50%
Confidence
Long-term
Relevance
7/9/2026
First Seen
Sources
Context Warp Drive: A Deep Dive into Deterministic Context Folding for AI Agents
hackernewshackernews7/7/2026
Related Claims
UnverifiedWhen testing LLM citation behavior, the same prompt should be run at least 20–50 times, ideally expanded to more than 100 times.69% similarUnverifiedA 2023 Stanford University study showed that multimodal models such as GPT-4V, when parsing flowcharts with nested conditional logic, had accuracy approximately 15–25 percentage points lower than equivalent textual descriptions.69% similarUnverified2024 年的多项基准测试表明，经过严格安全对齐的模型在代码生成任务上的拒绝率比基座模型高出 15-30%，在创意写作任务上的自我审查率高达 40%69% similarUnverified自2023年GPT-4发布以来，AI在HumanEval等主流编程基准测试上的通过率从不足50%跃升至90%以上67% similarUnverifiedThe generational capability gap between models is narrowing, with multiple models approaching or even surpassing GPT-4's reasoning advantage on certain benchmarks.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/319431API
curl https://kongchang.com/api/v1/knowledge/claims/319431MCP
get_claim(id=319431)