Unverified50% confidenceFactExact time
The researchers also conducted experiments on diffusion Transformers trained on ImageNet and obtained the same two findings: block orthogonality brings no additional benefit, and spherical optimization is already sufficient.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedTransformer、残差连接、批归一化等突破性架构与训练技巧均先在实践中被证明有效，理论解释往往滞后数年71% similarUnverifiedThe work of REPA and RECON shows that the semantic information learned by the early layers of a diffusion Transformer may not automatically flow to the later layers, and that designing supervision signals hierarchically and in stages may be an underestimated direction for improving training efficiency.68% similarUnverifiedUnsloth通过内核级优化实现了相比原生Hugging Face Transformers训练流程的训练速度大幅提升，同时降低了显存占用68% similarUnverifiedIn the AD training stage, the wav2vec 2.0 backbone is frozen, the two task vectors are injected into each Transformer layer with per-layer learnable merging coefficients, and the entire training process optimizes only these merging coefficients and one linear classification head, with trainable parameters under 0.1% of full fine-tuning.67% similarUnverified一项研究提出仅训练单个Transformer层，其效果就能与全参数RL训练相媲美67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1129391API
curl https://kongchang.com/api/v1/knowledge/claims/1129391MCP
get_claim(id=1129391)