Unverified50% confidenceFactTime unknown
Haiku 4.5在安全评估中的错位行为（misalignment behavior）发生率在所有模型中最低。
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/2/2026
First Seen
Valid until: 9/30/2026
Sources
Related Claims
UnverifiedB站UP主通过5项可视化编程任务对Claude Haiku 4.5进行了系统性测试，结果全部表现不佳65% similarUnverified低比特量化（Q4）会导致权重矩阵的秩降低，影响模型对低频但关键token（如JSON花括号、冒号）的预测置信度65% similarUnverified测试者认为Claude Haiku 4.5在3D建模理解、代码性能优化、库引用准确性、物理逻辑推理、指令遵循能力五个维度均表现为差或极差65% similarUnverifiedClaude Opus 4.8 has notably lower rates of misaligned behavior compared to Claude Opus 4.7 on safety alignment evaluations64% similarUnverified在GPU的SIMT执行模型下，Karatsuba算法的递归分解会引发线程分歧，有效计算效率可能降至峰值的1/4甚至更低62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/54473API
curl https://kongchang.com/api/v1/knowledge/claims/54473MCP
get_claim(id=54473)