Unverified50% confidenceFactExact time
The above comparison is conducted under the same LoRA budget, ruling out the explanation that more parameters make it stronger.
1
Sources
50%
Confidence
Long-term
Relevance
11/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLoRA fine-tuning weights can be merged after fine-tuning to eliminate inference overhead, which differs from MaRN's inference cost characteristics.68% similarUnverifiedLoRA's parameter count, memory usage and optimization difficulty grow linearly with the rank r, rather than quadratically with the model width.66% similarUnverifiedIn evaluation, if the thinking budget is not fixed, scores of the same model are not comparable, and the budget effectively becomes a hidden hyperparameter.64% similarUnverifiedLoRA uses a scaling factor alpha/r to scale the update, so adjusting r does not require extensively re-tuning the learning rate, reducing the cost of hyperparameter search.63% similarUnverifiedThese theoretical results suggest that the choice of LoRA rank should refer to the bottleneck rank, and that beyond the bottleneck rank threshold, extra rank may not bring observable changes in output.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1125567API
curl https://kongchang.com/api/v1/knowledge/claims/1125567MCP
get_claim(id=1125567)