Unverified50% confidenceFactExact time
The GMR training objective records the model's length-normalized log-likelihood of the correct answer when a fact is first introduced as an anchor, and after subsequent memory updates applies a penalty if the score falls below the anchor by more than a tolerance margin, with the retention loss formula being L_ret = (1/B) * sum_f max(0, anchor_f - margin - current_score_f).
1
Sources
50%
Confidence
Long-term
Relevance
10/21/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedAfter geometric transformation, the transformation parameters must be recorded and the predicted boxes must be inverse-transformed to align them; ignoring this step will cause semantic misalignment in the consistency loss and harm training effectiveness70% similarUnverifiedA cosine annealing learning rate schedule may help the model break through loss plateaus and avoid getting stuck in local optima.69% similarUnverifiedWhen measuring prediction error after fractional integration, there is an exact 1/L leading-order error for forecasting from a finite context of length L.69% similarUnverifiedThe main result of the paper is that Q-Learning with zero initialization and constant step size can, with high probability, reduce the sup-norm error to within ε using Õ(H^4/(μ_min·ε^2)+h_star) state transitions (for 0<ε≤1).68% similarVerifiedGPTQ is a post-training quantization scheme based on layer-wise error compensation, with minimal precision loss.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1080347API
curl https://kongchang.com/api/v1/knowledge/claims/1080347MCP
get_claim(id=1080347)