Unverified50% confidenceReleaseExact time
The arXiv paper "Phase-HDC: Replacing Optimizer History with Gradient Thresholds in Discrete Phase Learning" proposes a memory-free training method that replaces optimizer history with gradient thresholds.
1
Sources
50%
Confidence
Long-term
Relevance
10/9/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLearned optimizers face the challenge of distribution shift, and the cold-start problem requires first accumulating execution samples with the traditional optimizer before switching to a learned approach.66% similarUnverifiedA cosine annealing learning rate schedule may help the model break through loss plateaus and avoid getting stuck in local optima.66% similarUnverified75倍嵌入学习率是一种保持矩阵不绑定、通过放大嵌入层学习率补偿稀疏更新的可行工程手段，但过高学习率可能导致嵌入空间不稳定震荡66% similarUnverifiedOut-of-sample design can substantially eliminate the interference of memorization retrieval on evaluation by using contemporary works created after the model's training cutoff date or from controlled sources.65% similarUnverified采用域随机化（在训练早期就让ping在整个范围内随机采样）而非单向退火，可以让每个版本模型从一开始就见过各种延迟条件，避免后期崩溃风险65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/990620API
curl https://kongchang.com/api/v1/knowledge/claims/990620MCP
get_claim(id=990620)