Unverified50% confidenceFactExact time
The "single-loop tax" phenomenon was reproduced in a small recurrent Transformer setting, namely that the single-loop version has a systematic performance penalty relative to the multi-loop version.
1
Sources
50%
Confidence
Long-term
Relevance
11/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedFull-stack recurrent Transformers place context construction and iterative state refinement within the same recurrent computation, without clearly distinguishing their roles.72% similarUnverifiedA recurrent Transformer needs a separate KV cache stored at each recurrent step, causing the KV cache to grow linearly with the number of recurrent steps.71% similarUnverifiedIn small recurrent Transformers, re-selecting experts on each loop brings observable performance differences.71% similarUnverifiedThe recurrent architecture is analogous to the unrolling of a classical RNN across time steps, except that the object being recurred is the Transformer block itself.69% similarUnverifiedThe Looped Transformer repeatedly calls the same set of layers in a loop, thereby inserting far more serial computation steps between two text outputs than in a conventional forward pass.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1167456API
curl https://kongchang.com/api/v1/knowledge/claims/1167456MCP
get_claim(id=1167456)