Du, Wenyu, Chung, Stephen · Oct 28, 2026
Du and Chung propose CopyLoop to speed up looped Transformer decoding
Markdown · JSON · Original source
Wenyu Du and Stephen Chung introduce CopyLoop, a training-free decoding algorithm for looped Transformers. These models repeatedly apply shared blocks, but their recurrent depth makes conventional full-depth decoding expensive. CopyLoop proposes multi-token suffixes from repeated context, screens them at shallow recurrent depth to filter incorrect branches, and schedules newer blocks while earlier ones continue toward full-depth verification. The authors report speedups over full-depth autoregressive decoding across six from the Ouro, Raven, and Huginn families, evaluated on code generation and multi-step mathematical reasoning. They report 2.98x–10.70x on Base models and 2.74x–3.52x on Thinking models, and say the method consistently outperforms or closely matches competing baselines across evaluation tasks. Their analyses also report greater speedups on longer generations with reusable contexts and a need for only lightweight . The proposed mechanism matters because it aims to reduce decoding by checking likely continuations early and overlapping work, rather than waiting for each full-depth step. These are reported evaluation results, not evidence here of performance in deployed systems. How well do the reported speedups hold on other models and workloads, especially when contexts are not reusable?
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Oct 28, 2026 (2026-10-28).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-13T07:16:57.489Z. First observed: 2026-10-30T06:45:36.991Z. Retrieved: 2026-10-31T07:46:51.217Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The evidence is an abstract; it does not provide detailed methods, per-task results, or evidence of performance in deployed systems. The reported evaluations cover six checkpoints from the Ouro, Raven, and Huginn families and code generation and multi-step mathematical reasoning; performance on other models or workloads is not established here. The abstract reports greater speedups with reusable contexts but does not quantify performance when contexts are not reusable.
Open question
How well do the reported speedups hold on other models and workloads, especially when contexts are not reusable?
Source-stated
Du and Chung introduce CopyLoop, a training-free decoding algorithm for looped Transformers that combines context-copy proposals, shallow screening, and continuous recurrent scheduling.
We introduce CopyLoop, a training-free decoding algorithm that combines low-cost context-copy proposals with shallow screening and continuous recurrent scheduling.
Source-stated
The authors evaluate CopyLoop on six checkpoints from the Ouro, Raven, and Huginn families, covering code generation and multi-step mathematical reasoning.
We evaluate CopyLoop across six model checkpoints from the Ouro, Raven, and Huginn families, covering code generation and multi-step mathematical reasoning.
Source-stated
The authors report decoding speedups of 2.98x–10.70x on Base models and 2.74x–3.52x on Thinking models, compared with full-depth autoregressive decoding.
CopyLoop achieves decoding speedups of 2.98x--10.70x over full-depth autoregressive decoding on Base models and 2.74x--3.52x on Thinking models
Source-stated
The authors say CopyLoop performs better or approximately as well as competing baselines across their evaluation tasks.
consistently outperforming or closely matching all competing baselines across evaluation tasks.
Source-stated
The authors report greater speedups on longer generations with reusable contexts and say CopyLoop needs only lightweight model calibration.
CopyLoop requires only lightweight model calibration and yields greater speedups on longer generations with reusable contexts.
Source and contributor context
Du, Wenyu, Chung, Stephen · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Model checkpoint
A saved snapshot of a model's learned numbers at one point in its training.
Model calibration
Whether a model's stated confidence matches how often it turns out to be right.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: 38963f09-3f32-43a2-8da8-15f1d79cc9c2.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-13T07:16:57.489Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: a32998592569d9d7a59579dd5b690c7165373835a392aea8da6fb5121011e919. Source text hash: 2abe6acb34c2c2a53fc93571128acbc9c10c66031ea7b6b262e4bf5d7e0eb020.
Revision and source-check history
Version 1 · 2026-11-13T07:16:57.489Z
Du and Chung propose CopyLoop to speed up looped Transformer decoding
Wenyu Du and Stephen Chung introduce CopyLoop, a training-free decoding algorithm for looped Transformers. These models repeatedly apply shared blocks, but their recurrent depth makes conventional full-depth decoding expensive. CopyLoop proposes multi-token suffixes from repeated context, screens them at shallow recurrent depth to filter incorrect branches, and schedules newer blocks while earlier ones continue toward full-depth verification. The authors report speedups over full-depth autoregressive decoding across six from the Ouro, Raven, and Huginn families, evaluated on code generation and multi-step mathematical reasoning. They report 2.98x–10.70x on Base models and 2.74x–3.52x on Thinking models, and say the method consistently outperforms or closely matches competing baselines across evaluation tasks. Their analyses also report greater speedups on longer generations with reusable contexts and a need for only lightweight . The proposed mechanism matters because it aims to reduce decoding by checking likely continuations early and overlapping work, rather than waiting for each full-depth step. These are reported evaluation results, not evidence here of performance in deployed systems. How well do the reported speedups hold on other models and workloads, especially when contexts are not reusable?
Original source
cs.AI · Article
Source details
Original source date: Oct 28, 2026 (2026-10-28).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-13T07:16:57.489Z. First observed: 2026-10-30T06:45:36.991Z. Retrieved: 2026-10-31T07:46:51.217Z.
Limits
The evidence is an abstract; it does not provide detailed methods, per-task results, or evidence of performance in deployed systems. The reported evaluations cover six checkpoints from the Ouro, Raven, and Huginn families and code generation and multi-step mathematical reasoning; performance on other models or workloads is not established here. The abstract reports greater speedups with reusable contexts but does not quantify performance when contexts are not reusable.
Open question
How well do the reported speedups hold on other models and workloads, especially when contexts are not reusable?
Source-stated
Du and Chung introduce CopyLoop, a training-free decoding algorithm for looped Transformers that combines context-copy proposals, shallow screening, and continuous recurrent scheduling.
We introduce CopyLoop, a training-free decoding algorithm that combines low-cost context-copy proposals with shallow screening and continuous recurrent scheduling.
Source-stated
The authors evaluate CopyLoop on six checkpoints from the Ouro, Raven, and Huginn families, covering code generation and multi-step mathematical reasoning.
We evaluate CopyLoop across six model checkpoints from the Ouro, Raven, and Huginn families, covering code generation and multi-step mathematical reasoning.
Source-stated
The authors report decoding speedups of 2.98x–10.70x on Base models and 2.74x–3.52x on Thinking models, compared with full-depth autoregressive decoding.
CopyLoop achieves decoding speedups of 2.98x--10.70x over full-depth autoregressive decoding on Base models and 2.74x--3.52x on Thinking models
Source-stated
The authors say CopyLoop performs better or approximately as well as competing baselines across their evaluation tasks.
consistently outperforming or closely matching all competing baselines across evaluation tasks.
Source-stated
The authors report greater speedups on longer generations with reusable contexts and say CopyLoop needs only lightweight model calibration.
CopyLoop requires only lightweight model calibration and yields greater speedups on longer generations with reusable contexts.
Source and contributor context
Du, Wenyu, Chung, Stephen · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Model checkpoint
A saved snapshot of a model's learned numbers at one point in its training.
Model calibration
Whether a model's stated confidence matches how often it turns out to be right.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: 38963f09-3f32-43a2-8da8-15f1d79cc9c2.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-13T07:16:57.489Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: a32998592569d9d7a59579dd5b690c7165373835a392aea8da6fb5121011e919. Source text hash: 2abe6acb34c2c2a53fc93571128acbc9c10c66031ea7b6b262e4bf5d7e0eb020.
passed · 2026-11-13T07:16:57.489Z · anthropic/claude-opus-5.5
Publication: 38963f09-3f32-43a2-8da8-15f1d79cc9c2
attribution — pass: The reading correctly credits Wenyu Du and Stephen Chung as the authors of the arXiv preprint. Speedup and baseline results are framed as reported by the authors, not as independently established facts.
Authors:Wenyu Du, Stephen Chung
summary — pass: The reading leads with the actor and the method. It briefly explains looped Transformers and the cost problem they create, describes the three mechanisms in source order, and states the evidence scope and limits. Terms such as decoding, autoregressive, checkpoint, and Base/Thinking models stay technical; tooltip entries would help. The plain-language significance sentence (checking likely continuations early and overlapping work) keeps the core idea readable. The question follows from the source's reusable-context condition.
Rather than step-by-step generation, CopyLoop proposes multi-token suffixes from repeated context.
date — pass: The originalPublished value, 2026-10-28, matches the v1 submission date shown on arXiv. It is not the later feed date of Oct 31.
[Submitted on 28 Oct 2026]
support — pass: All the numbers, model families, tasks, and the baseline comparison match the abstract. The significance sentence restates the abstract's own mechanism of early filtering and concurrent processing. The reading explicitly declines to infer deployed performance. Its limitations accurately note that the abstract does not quantify results when contexts are not reusable.
evaluates them early at a shallow recurrent depth to efficiently filter out incorrect branches, processing newer blocks concurrently while earlier blocks are still progressing toward full-depth verification
media — pass: The selected media is the verified direct arXiv abstract page for this paper, which matches the source URL. No better candidate was supplied.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A saved snapshot of a model's learned numbers at one point in its training.
Definition referenceWhether a model's stated confidence matches how often it turns out to be right.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA saved snapshot of a model's learned numbers at one point in its training.
Definition referenceWhether a model's stated confidence matches how often it turns out to be right.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv