Maximov, Egor, Merkelov, Aleksei, Alekseev, Vasiliy, Krivorotov, Ivan, Beliaev, Valentin, Arsenin, Nikita, Bocharov, Dmitry, Goncharov, Aleksei · Oct 27, 2026
Authors present DCKV for compressing the prompt KV cache
Markdown · JSON · Original source
The authors present DCKV, a codec for compressing the KV cache—the part of a prompt’s state that hybrid language models store in their full-attention layers so it can be reused. They train its transform, quantizer and entropy models together, optimizing for both code length and changes to the frozen language model’s output. On Qwen3.5-2B, the authors report storing a compressed token in more than 55 times fewer bytes than bf16, with a loss of at most 1.8 points on three long-context . At more than 100 times fewer bytes, they report a loss of at most 3.6 points. These results make the tradeoff concrete: the reported compression levels are accompanied by losses on the benchmarks, rather than compression being presented as cost-free. The abstract does not establish how well the approach transfers beyond this model and these benchmarks.
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 27, 2026 (2026-10-27).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-14T04:04:32.361Z. First observed: 2026-10-28T07:10:38.633Z. Retrieved: 2026-10-30T21:03:10.749Z.
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 reported results in the abstract are for Qwen3.5-2B and three long-context ; it does not establish performance on other models or workloads.
Open question
Would DCKV show a similar compression-and--loss tradeoff on other models and workloads?
Source-stated
The authors present DCKV as a codec for the KV cache, trained end to end.
We present DCKV, a differentiable codec for the KV cache and the first KV-cache codec that is trained end to end.
Source-stated
The authors optimize the transform, quantizer and entropy models together for code length and changes to the frozen language model's output.
Its transform, its quantizer and its entropy models, which are conditioned on the prompt tokens and on already decoded neighbouring positions, are optimized together against the frozen language model, both for the code length that the entropy models assign and for the change of the model's output.
Source-stated
On Qwen3.5-2B, the authors report more than 55 times fewer bytes than bf16 at a loss of at most 1.8 points on three long-context benchmarks.
On Qwen3.5-2B, DCKV stores a compressed token in more than 55 times fewer bytes than bf16 at a loss of at most 1.8 points on three long-context benchmarks
Source-stated
The abstract reports a higher compression level with a larger stated loss bound.
in more than 100 times fewer bytes at a loss of at most 3.6 points.
Source and contributor context
Maximov, Egor, Merkelov, Aleksei, Alekseev, Vasiliy, Krivorotov, Ivan, Beliaev, Valentin, Arsenin, Nikita, Bocharov, Dmitry, Goncharov, Aleksei · 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.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 077b3d1e-da08-476e-816c-d106016b3b50.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T04:04:32.361Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 96889c01f40c04ac715a4a91c42b21d27b842fc1843da64e4dc3dddb6a21bba5. Source text hash: 06c798824dd416eb0a00e59b8a4fba2d8f6240d1bb50e7a5d9e7b857620665a0.
Revision and source-check history
Version 1 · 2026-11-14T04:04:32.361Z
Authors present DCKV for compressing the prompt KV cache
The authors present DCKV, a codec for compressing the KV cache—the part of a prompt’s state that hybrid language models store in their full-attention layers so it can be reused. They train its transform, quantizer and entropy models together, optimizing for both code length and changes to the frozen language model’s output. On Qwen3.5-2B, the authors report storing a compressed token in more than 55 times fewer bytes than bf16, with a loss of at most 1.8 points on three long-context . At more than 100 times fewer bytes, they report a loss of at most 3.6 points. These results make the tradeoff concrete: the reported compression levels are accompanied by losses on the benchmarks, rather than compression being presented as cost-free. The abstract does not establish how well the approach transfers beyond this model and these benchmarks.
Original source
cs.AI · Article
Source details
Original source date: Oct 27, 2026 (2026-10-27).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-14T04:04:32.361Z. First observed: 2026-10-28T07:10:38.633Z. Retrieved: 2026-10-30T21:03:10.749Z.
Limits
The reported results in the abstract are for Qwen3.5-2B and three long-context ; it does not establish performance on other models or workloads.
Open question
Would DCKV show a similar compression-and--loss tradeoff on other models and workloads?
Source-stated
The authors present DCKV as a codec for the KV cache, trained end to end.
We present DCKV, a differentiable codec for the KV cache and the first KV-cache codec that is trained end to end.
Source-stated
The authors optimize the transform, quantizer and entropy models together for code length and changes to the frozen language model's output.
Its transform, its quantizer and its entropy models, which are conditioned on the prompt tokens and on already decoded neighbouring positions, are optimized together against the frozen language model, both for the code length that the entropy models assign and for the change of the model's output.
Source-stated
On Qwen3.5-2B, the authors report more than 55 times fewer bytes than bf16 at a loss of at most 1.8 points on three long-context benchmarks.
On Qwen3.5-2B, DCKV stores a compressed token in more than 55 times fewer bytes than bf16 at a loss of at most 1.8 points on three long-context benchmarks
Source-stated
The abstract reports a higher compression level with a larger stated loss bound.
in more than 100 times fewer bytes at a loss of at most 3.6 points.
Source and contributor context
Maximov, Egor, Merkelov, Aleksei, Alekseev, Vasiliy, Krivorotov, Ivan, Beliaev, Valentin, Arsenin, Nikita, Bocharov, Dmitry, Goncharov, Aleksei · 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.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 077b3d1e-da08-476e-816c-d106016b3b50.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T04:04:32.361Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 96889c01f40c04ac715a4a91c42b21d27b842fc1843da64e4dc3dddb6a21bba5. Source text hash: 06c798824dd416eb0a00e59b8a4fba2d8f6240d1bb50e7a5d9e7b857620665a0.
uncertain · 2026-11-14T04:04:32.361Z · anthropic/claude-opus-5.5
Publication: 077b3d1e-da08-476e-816c-d106016b3b50
attribution — pass: DCKV and its results are consistently attributed to the paper's authors (Maximov et al., arXiv). Results are framed as what the authors report, not as established fact.
We present DCKV, a differentiable codec for the KV cache
summary — uncertain: The reading is faithful but drops the source's stated motivation: the cache grows with the prompt and must be stored or moved to reuse it. Without that, a general reader cannot see why a smaller cache matters. Smallest repair: add that clause. The summary also omits the authors' claim that DCKV is the first KV-cache codec trained end to end. Central terms (codec, bf16, quantizer, entropy models) need term-library tooltips, especially bf16, which is the baseline for the 55x/100x figures. The title and question are fair and non-leading.
this cache is still the part of a prompt's state that grows with the prompt and that must be stored or moved to reuse it
date — pass: originalPublished 2026-10-27 matches the v1 submission date on arXiv. It is not the later feed date.
[Submitted on 27 Oct 2026]
support — pass: All claims match the abstract verbatim in substance: joint optimization against the frozen model, 55x fewer bytes than bf16 at a loss of at most 1.8 points, and 100x at a loss of at most 3.6 points on three long-context benchmarks. The tradeoff framing and the limitation to a single model and these benchmarks are appropriately conditional. No benefits are invented.
in more than 100 times fewer bytes at a loss of at most 3.6 points
media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference