Unverified50% confidenceBenchmarkExact time
After installing the corpus, Llama-3.3-70B's harmful decisions out of 300 decisions dropped from 70 to 0, while its MMLU benchmark remained unchanged.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/28/2026
First Seen
Valid until: 1/26/2027
Sources
Related Entities
Related Claims
Unverified改变块大小时MMLU多选题几乎不受影响仍稳在78%，说明知识本身还在，崩掉的是生成过程62% similarUnverifiedTraditional secondary LLM review has a failure rate of over 80% in blocking attacks such as Base64 obfuscation, path traversal, and indirect prompt injection, while introducing more than 300ms of additional latency (as disclosed by the project team).61% similarUnverified哈达玛变换可用于LLM超低比特量化的预处理，缓解离群值导致的精度崩溃问题60% similarUnverified随着指令数量增加，LLM的遵循质量会整体下降，开始无差别忽略所有指令而非仅忽略靠后的指令60% similarUnverified研究表明LLM在处理过长、过复杂的指令集时整体遵循率会系统性下降，即指令遗忘现象60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1105906API
curl https://kongchang.com/api/v1/knowledge/claims/1105906MCP
get_claim(id=1105906)