Zou, JianKun, Yuan, Tongtong · Nov 4, 2026
EchoKV uses iterative query feedback to repair KV cache eviction workspaces
Markdown · JSON · Original source
In the abstract for their paper, JianKun Zou and Tongtong Yuan argue that KV cache eviction is not just a matter of ranking tokens: the workspace retained to generate eviction queries can shape the quality of those queries. An early workspace error, they say, can distort later observations and reinforce a poor eviction decision, especially when evidence is dispersed across a long context. Their method, EchoKV, initializes a workspace using proposals at future decode positions, then uses early decode-generated queries to recover that workspace before the remaining lookahead steps. The authors report that EchoKV uses the same number of workspace forward passes as LAQ, with an added tensor re-scoring operation for workspace repair. On LongBench, across 16 tasks and two named models, they report improvements over LAQ on several tasks requiring dispersed evidence, including multi-hop cross-document QA and long single-document comprehension. They also report comparable performance on the remaining LongBench subsets. The abstract frames query–workspace coupling as a practical failure mode for query-driven KV eviction; it does not provide numerical gains.
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: Nov 4, 2026 (2026-11-04).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-09T09:18:06.828Z. First observed: 2026-11-05T05:06:42.879Z. Retrieved: 2026-11-09T09:17:17.474Z.
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 available evidence is the paper’s abstract, not its full text; it does not provide numerical gains or detailed evaluation results. The abstract reports LongBench results for Mistral-7B-Instruct-v0.2 and Qwen2.5-7B-Instruct; this evidence does not establish performance beyond the stated evaluation.
Open question
How large are EchoKV’s reported improvements over LAQ on the LongBench tasks requiring dispersed evidence?
Source-stated
The authors argue that the workspace retained for generating eviction queries affects query quality, and that early workspace errors can distort later observations and reinforce an incorrect eviction decision.
the workspace retained for generating eviction queries also determines the quality of those queries. Consequently, an early workspace error can distort later query observations
Source-stated
EchoKV initializes a feedback-compatible bootstrap workspace using proposals at future decode positions, then uses early decode-generated queries to recover the workspace before the remaining lookahead steps.
It initializes a feedback-compatible bootstrap workspace using proposals at future decode positions, then uses early decode-generated queries to recover the workspace
Source-stated
The authors say EchoKV uses the same number of workspace forward passes as LAQ, adding a tensor re-scoring operation for workspace repair.
It uses the same number of workspace forward passes as LAQ while adding a tensor re-scoring operation for workspace repair.
Source-stated
On LongBench, the authors report notable improvements over LAQ on several tasks requiring dispersed evidence, and comparable performance on the remaining subsets.
EchoKV shows notable improvements over LAQ on several tasks requiring dispersed evidence
Escape Meridian interpretation
The abstract does not quantify the reported improvements, so their magnitude cannot be assessed from this evidence.
EchoKV shows notable improvements over LAQ on several tasks requiring dispersed evidence
Source and contributor context
Zou, JianKun, Yuan, Tongtong · 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.
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: 3f7dd564-1825-4dcd-8487-80f5a9200e22.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-09T09:18:06.828Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: ca1214021e0ea90c0b080ec0e403aaf32f721d482d101d1d82d8955846b54b77. Source text hash: ae6ddeb2ff7950d16a6eab00eb4f766b8a908c65e3088ab1b79a2bfebff5dfb1.
Revision and source-check history
Version 1 · 2026-11-09T09:18:06.828Z
EchoKV uses iterative query feedback to repair KV cache eviction workspaces
In the abstract for their paper, JianKun Zou and Tongtong Yuan argue that KV cache eviction is not just a matter of ranking tokens: the workspace retained to generate eviction queries can shape the quality of those queries. An early workspace error, they say, can distort later observations and reinforce a poor eviction decision, especially when evidence is dispersed across a long context. Their method, EchoKV, initializes a workspace using proposals at future decode positions, then uses early decode-generated queries to recover that workspace before the remaining lookahead steps. The authors report that EchoKV uses the same number of workspace forward passes as LAQ, with an added tensor re-scoring operation for workspace repair. On LongBench, across 16 tasks and two named models, they report improvements over LAQ on several tasks requiring dispersed evidence, including multi-hop cross-document QA and long single-document comprehension. They also report comparable performance on the remaining LongBench subsets. The abstract frames query–workspace coupling as a practical failure mode for query-driven KV eviction; it does not provide numerical gains.
Original source
cs.AI · Article
Source details
Original source date: Nov 4, 2026 (2026-11-04).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-09T09:18:06.828Z. First observed: 2026-11-05T05:06:42.879Z. Retrieved: 2026-11-09T09:17:17.474Z.
Limits
The available evidence is the paper’s abstract, not its full text; it does not provide numerical gains or detailed evaluation results. The abstract reports LongBench results for Mistral-7B-Instruct-v0.2 and Qwen2.5-7B-Instruct; this evidence does not establish performance beyond the stated evaluation.
Open question
How large are EchoKV’s reported improvements over LAQ on the LongBench tasks requiring dispersed evidence?
Source-stated
The authors argue that the workspace retained for generating eviction queries affects query quality, and that early workspace errors can distort later observations and reinforce an incorrect eviction decision.
the workspace retained for generating eviction queries also determines the quality of those queries. Consequently, an early workspace error can distort later query observations
Source-stated
EchoKV initializes a feedback-compatible bootstrap workspace using proposals at future decode positions, then uses early decode-generated queries to recover the workspace before the remaining lookahead steps.
It initializes a feedback-compatible bootstrap workspace using proposals at future decode positions, then uses early decode-generated queries to recover the workspace
Source-stated
The authors say EchoKV uses the same number of workspace forward passes as LAQ, adding a tensor re-scoring operation for workspace repair.
It uses the same number of workspace forward passes as LAQ while adding a tensor re-scoring operation for workspace repair.
Source-stated
On LongBench, the authors report notable improvements over LAQ on several tasks requiring dispersed evidence, and comparable performance on the remaining subsets.
EchoKV shows notable improvements over LAQ on several tasks requiring dispersed evidence
Escape Meridian interpretation
The abstract does not quantify the reported improvements, so their magnitude cannot be assessed from this evidence.
EchoKV shows notable improvements over LAQ on several tasks requiring dispersed evidence
Source and contributor context
Zou, JianKun, Yuan, Tongtong · 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.
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: 3f7dd564-1825-4dcd-8487-80f5a9200e22.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-09T09:18:06.828Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: ca1214021e0ea90c0b080ec0e403aaf32f721d482d101d1d82d8955846b54b77. Source text hash: ae6ddeb2ff7950d16a6eab00eb4f766b8a908c65e3088ab1b79a2bfebff5dfb1.
uncertain · 2026-11-09T09:18:06.828Z · anthropic/claude-opus-5.5
Publication: 3f7dd564-1825-4dcd-8487-80f5a9200e22
attribution — pass: The reading correctly attributes the argument, method and results to authors JianKun Zou and Tongtong Yuan in their arXiv abstract. It keeps the results framed as their reported findings rather than established facts.
Authors:JianKun Zou, Tongtong Yuan
summary — uncertain: The reading is faithful but not understandable on its own to a non-specialist. It never explains in plain language what a KV cache is or why models evict from it (to save memory on long inputs), so the stakes stay unclear. LAQ is never introduced, and workspace, decode positions, forward passes and tensor re-scoring go unexplained. Smallest repair: add one plain sentence on what KV cache eviction does and why it matters. Identify LAQ as the prior query-driven method being compared, using tooltips for the remaining terms. The question is substantive and grounded.
KV cache eviction is not merely a token-ranking problem
date — pass: The originalPublished date of 2026-11-04 matches the v1 submission date on arXiv. The feed date of Nov 5 is a later announcement and was correctly not used.
[Submitted on 4 Nov 2026]
support — pass: All claims closely track the abstract: the coupling argument, the bootstrap workspace, equal forward passes to LAQ plus re-scoring, and selective gains with comparable results elsewhere. The interpretation that the magnitude of gains is unquantified accurately reflects the absence of numbers in the abstract. No benefits or costs are invented.
Across budgets and backbones, EchoKV selectively improves evidence-intensive tasks while maintaining comparable performance on the remaining LongBench subsets.
media — pass: The selected media is the verified direct link to the original arXiv abstract page, which matches the source URL and 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.
arXiv 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