Unverified50% confidenceBenchmarkExact time
Multi-position aggregation is highly effective against local perturbations (loss ≤ 0.5 percentage points), but can only attenuate distributed perturbations, with performance still dropping by approximately 3.8 percentage points.
1
Sources
50%
Confidence
Long-term
Relevance
9/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedMainstream quantization schemes typically cause only a 1–3% performance loss at 4-bit quantization precision, but performance degradation becomes significantly more pronounced at 2-bit quantization.74% similarUnverifiedPerturbation-augmented training can only narrow the probe performance gap to -3.7 percentage points, with limited effectiveness.70% similarUnverified当叠加超过3到5个LoRA时，效果出现明显的边际递减68% similarVerifiedGradient descent has three variants: batch gradient descent, stochastic gradient descent (SGD), and mini-batch gradient descent.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/924819API
curl https://kongchang.com/api/v1/knowledge/claims/924819MCP
get_claim(id=924819)