Unverified50% confidenceOpinionExact time
Extending AsymLoop to MoE backbones means the layers in the recurrent decoder are themselves sparsely activated, further amplifying parameter efficiency.
1
Sources
50%
Confidence
Long-term
Relevance
10/21/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAsymLoop's separated design brings computational efficiency: the global KV needs to be computed only once, and only the local KV and state need updating during recurrent iterations.69% similarUnverifiedUnder matched effective depth and training token budget, AsymLoop exhibits lower validation perplexity and higher average downstream accuracy than Full Loop.62% similarUnverifiedThe MoE sparse architecture activates only a small fraction of all experts during each forward pass, dramatically scaling up model parameter count without a proportional increase in computation.62% similarUnverifiedStreaming ASR adopts a chunked decoding strategy and often cannot use the full future context in order to reduce latency, so local acoustic ambiguity more easily leads to early errors.61% similarUnverifiedAsync scheduling splits rollout and training onto different GPU groups to run in parallel, at the cost of policy lag, since the training data comes from a slightly earlier version of the model, which may harm convergence stability and final performance.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1078928API
curl https://kongchang.com/api/v1/knowledge/claims/1078928MCP
get_claim(id=1078928)