Unverified50% confidenceFactExact time
The results of multidimensional item response theory (MIRT) modeling show that HarmBench's scores strongly do not support the "single attribute" assumption; it simultaneously captures multiple distinct dimensions.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/13/2026
First Seen
Valid until: 1/11/2027
Sources
Related Entities
Related Claims
UnverifiedTraditional alignment methods typically rely on holistic scalar rewards, summarizing complex multi-dimensional quality with a single score, which produces confusing signals when a model performs poorly on one dimension but well on others57% similarVerifiedLocal small models still struggle to match GPT-4-level cloud services on complex reasoning and long-chain tasks.57% similarUnverifiedAdaptive sampling algorithms such as Min-P and Mirostat better avoid the probability collapse problem and have been adopted as defaults by tools such as Ollama.57% similarUnverifiedIf multiple models share similar training biases, multi-model adversarial debate may only amount to echo-chamber agreement and may fail to genuinely expose errors (model homogenization bias)57% similarUnverified偏差-方差权衡原理表明单一模型难以同时最小化偏差和方差，通过聚合多个具有差异性的模型可以突破这一限制56% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1013425API
curl https://kongchang.com/api/v1/knowledge/claims/1013425MCP
get_claim(id=1013425)