When Hidden States Drift: Can KV Caches Rescue Long-Range Speculative Decoding?

Fuente: arXiv

Gespeichert in:

Bibliographische Detailangaben
Hauptverfasser:	Liu, Tianyu, Shen, Yuhao, Hu, Xinyi, Zhang, Baolin, Zhang, Hengxin, Dai, Jun, Zhang, Jun, Ge, Shuang, Chen, Lei, Li, Yue, Wan, MingCheng
Format:	Preprint
Veröffentlicht:	2026
Schlagworte:	Computation and Language
Online-Zugang:	Acceder al recurso
Tags:	Tag hinzufügen Keine Tags, Fügen Sie den ersten Tag hinzu!

Ähnliche Einträge

ECHO: Elastic Speculative Decoding with Sparse Gating for High-Concurrency Scenarios
von: Hu, Xinyi, et al.
Veröffentlicht: (2026)

Draft Less, Retrieve More: Hybrid Tree Construction for Speculative Decoding
von: Shen, Yuhao, et al.
Veröffentlicht: (2026)

DeltaKV: Residual-Based KV Cache Compression via Long-Range Similarity
von: Hao, Jitai, et al.
Veröffentlicht: (2026)

G-KV: Decoding-Time KV Cache Eviction with Global Attention
von: Liao, Mengqi, et al.
Veröffentlicht: (2025)

QuantSpec: Self-Speculative Decoding with Hierarchical Quantized KV Cache
von: Tiwari, Rishabh, et al.
Veröffentlicht: (2025)

ParisKV: Fast and Drift-Robust KV-Cache Retrieval for Long-Context LLMs
von: Qi, Yanlin, et al.
Veröffentlicht: (2026)

Competitive Non-Clairvoyant KV-Cache Scheduling for LLM Inference
von: Feng, Yiding, et al.
Veröffentlicht: (2026)

Moment-KV: Momentum-Based Decode-Time KV Cache Compression for Long Generation
von: Jana, Soumyadeep, et al.
Veröffentlicht: (2026)

LongSpec: Long-Context Lossless Speculative Decoding with Efficient Drafting and Verification
von: Yang, Penghui, et al.
Veröffentlicht: (2025)

When, What, and How: Rethinking Retrieval-Enhanced Speculative Decoding
von: Fang, Min, et al.
Veröffentlicht: (2025)

MEDA: Dynamic KV Cache Allocation for Efficient Multimodal Long-Context Inference
von: Wan, Zhongwei, et al.
Veröffentlicht: (2025)

LouisKV: Efficient KV Cache Retrieval for Long Input-Output Sequences
von: Wu, Wenbo, et al.
Veröffentlicht: (2025)

Attention Drift: What Autoregressive Speculative Decoding Models Learn
von: Eldenk, Doğaç, et al.
Veröffentlicht: (2026)

PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling
von: Cai, Zefan, et al.
Veröffentlicht: (2024)

Cacheback: Speculative Decoding With Nothing But Cache
von: Ma, Zhiyao, et al.
Veröffentlicht: (2025)

StreamKV: Streaming Video Question-Answering with Segment-based KV Cache Retrieval and Compression
von: Chen, Yilong, et al.
Veröffentlicht: (2025)

TALON: Confidence-Aware Speculative Decoding with Adaptive Token Trees
von: Liu, Tianyu, et al.
Veröffentlicht: (2026)

When Speculation Spills Secrets: Side Channels via Speculative Decoding In LLMs
von: Wei, Jiankun, et al.
Veröffentlicht: (2024)

CXL-SpecKV: A Disaggregated FPGA Speculative KV-Cache for Datacenter LLM Serving
von: Liu, Dong, et al.
Veröffentlicht: (2025)

CacheClip: Accelerating RAG with Effective KV Cache Reuse
von: Yang, Bin, et al.
Veröffentlicht: (2025)

LagKV: Lag-Relative Information of the KV Cache Tells Which Tokens Are Important
von: Liang, Manlai, et al.
Veröffentlicht: (2025)

HybridKV: Hybrid KV Cache Compression for Efficient Multimodal Large Language Model Inference
von: Zeng, Bowen, et al.
Veröffentlicht: (2026)

DASH-KV: Accelerating Long-Context LLM Inference via Asymmetric KV Cache Hashing
von: Guo, Jinyu, et al.
Veröffentlicht: (2026)

DynamicKV: Task-Aware Adaptive KV Cache Compression for Long Context LLMs
von: Zhou, Xiabin, et al.
Veröffentlicht: (2024)

LongFlow: Efficient KV Cache Compression for Reasoning Models
von: Su, Yi, et al.
Veröffentlicht: (2026)

Scaling Speculative Decoding with Lookahead Reasoning
von: Fu, Yichao, et al.
Veröffentlicht: (2025)

Draft Model Knows When to Stop: Self-Verification Speculative Decoding for Long-Form Generation
von: Zhang, Ziyin, et al.
Veröffentlicht: (2024)

KVSwap: Disk-aware KV Cache Offloading for Long-Context On-device Inference
von: Zhang, Huawei, et al.
Veröffentlicht: (2025)

Make Every Draft Count: Hidden State based Speculative Decoding
von: Chen, Yuetao, et al.
Veröffentlicht: (2026)

Efficient Long-Context LLM Inference via KV Cache Clustering
von: Hu, Jie, et al.
Veröffentlicht: (2025)

CoKV: Optimizing KV Cache Allocation via Cooperative Game
von: Sun, Qiheng, et al.
Veröffentlicht: (2025)

XKV: Personalized KV Cache Memory Reduction for Long-Context LLM Inference
von: Li, Weizhuo, et al.
Veröffentlicht: (2024)

Model Tells You What to Discard: Adaptive KV Cache Compression for LLMs
von: Ge, Suyu, et al.
Veröffentlicht: (2023)

When Drafts Evolve: Speculative Decoding Meets Online Learning
von: Qian, Yu-Yang, et al.
Veröffentlicht: (2026)

RelayCaching: Accelerating LLM Collaboration via Decoding KV Cache Reuse
von: Geng, Yingsheng, et al.
Veröffentlicht: (2026)

Forcing-KV: Hybrid KV Cache Compression for Efficient Autoregressive Video Diffusion Models
von: Ji, Yicheng, et al.
Veröffentlicht: (2026)

SWIFT: On-the-Fly Self-Speculative Decoding for LLM Inference Acceleration
von: Xia, Heming, et al.
Veröffentlicht: (2024)

ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference
von: Liu, Xiang, et al.
Veröffentlicht: (2025)

FastKV: Decoupling of Context Reduction and KV Cache Compression for Prefill-Decoding Acceleration
von: Jo, Dongwon, et al.
Veröffentlicht: (2025)

BitDecoding: Unlocking Tensor Cores for Long-Context LLMs with Low-Bit KV Cache
von: Du, Dayou, et al.
Veröffentlicht: (2025)

if (filemtime($filename) + $maxlifetime < time()) { unlink($filename); $count++; } } return $count; } /** * A function that is called internally when session data is to be saved. * * @param string $sessId The current session ID * @param string $data The session data to write * * @return bool */ protected function saveSession($sessId, $data): bool { $sessFile = $this->path . '/sess_' . $sessId; if ($handle = fopen($sessFile, 'w')) { $return = false; // Lock the file for exclusive access to avoid issues with multiple // processes writing session simultaneously: if (flock($handle, LOCK_EX)) { $return = fwrite($handle, $data); // Make sure that there's no trailing data by truncating the file to // the correct length: ftruncate($handle, strlen($data)); fflush($handle); flock($handle, LOCK_UN); } fclose($handle); if ($return !== false) { return true; } } // If we got this far, something went wrong with the file output... // It is tempting to throw an exception here, but this code is called // outside of the context of exception handling, so all we can do is // echo a message.

// Anecdotal testing Today and Yesterday seems to indicate destroy() // is called by the garbage collector and everything is good. // Something to keep in mind though. return 0; } /** * Write function that is called when session data is to be saved. * * @param string $sessId The current session ID * @param string $data The session data to write * * @return bool */ public function write($sessId, $data): bool { if ($this->writesDisabled) { return true; } return $this->saveSession($sessId, $data); } /** * A function that is called internally when session data is to be saved. * * @param string $sessId The current session ID * @param string $data The session data to write * * @return bool */ abstract protected function saveSession($sessId, $data): bool; }

* * @return void */ public function writeClose() { // The assumption is that we're using PHP's ext/session. // session_write_close() will actually overwrite $_SESSION with an // empty array on completion -- which leads to a mismatch between what // is in the storage object and $_SESSION. To get around this, we // temporarily reset $_SESSION to an array, and then re-link it to // the storage object. // // Additionally, while you _can_ write to $_SESSION following a // session_write_close() operation, no changes made to it will be // flushed to the session handler. As such, we now mark the storage // object isImmutable. $storage = $this->getStorage(); if (! $storage->isImmutable()) { $_SESSION = $storage->toArray(true); session_write_close(); $storage->fromArray($_SESSION); $storage->markImmutable(); } } /** * Attempt to set the session name * * If the session has already been started, or if the name provided fails * validation, an exception will be raised. * * @param string $name * @return SessionManager * @throws Exception\InvalidArgumentException */ public function setName($name) { if ($this->sessionExists()) { throw new Exception\InvalidArgumentException( 'Cannot set session name after a session has already started'

Environment & details:

GET Data empty

POST Data empty

Files empty

Key	Value
VUFIND_SESSION	"fd6b46637e60a3678af459e7da1c9fd7"
language	"de"
ui	"standard"

Session

Key	Value
__Laminas	array:2 [`"_REQUEST_ACCESS_TIME" => 1790990089.4316 "_VALID" => array:1 [ "Laminas\Session\Validator\Id" => "fd6b46637e60a3678af459e7da1c9fd7" ]`]
SessionState	Laminas\Stdlib\ArrayObject {#938}
FlashMessenger	Laminas\Stdlib\ArrayObject {#1433}
SessionHelper	Laminas\Stdlib\ArrayObject {#1759}

Server/Request Data

Key	Value
REDIRECT_VUFIND_LOCAL_DIR	"/usr/local/vufind/local"
REDIRECT_HTTPS	"on"
REDIRECT_SSL_TLS_SNI	"info.umar.mx"
REDIRECT_STATUS	"200"
VUFIND_LOCAL_DIR	"/usr/local/vufind/local"
HTTPS	"on"
SSL_TLS_SNI	"info.umar.mx"
HTTP_ACCEPT	"/"
HTTP_USER_AGENT	"Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; ClaudeBot/1.0; +claudebot@anthropic.com)"
HTTP_COOKIE	"VUFIND_SESSION=fd6b46637e60a3678af459e7da1c9fd7; language=de; ui=standard"
HTTP_ACCEPT_ENCODING	"gzip, br, zstd, deflate"
HTTP_HOST	"info.umar.mx"
PATH	"/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/snap/bin"
SERVER_SIGNATURE	"<address>Apache/2.4.66 (Ubuntu) Server at info.umar.mx Port 443</address>\n"
SERVER_SOFTWARE	"Apache/2.4.66 (Ubuntu)"
SERVER_NAME	"info.umar.mx"
SERVER_ADDR	"200.23.222.19"
SERVER_PORT	"443"
REMOTE_ADDR	"216.73.217.142"
DOCUMENT_ROOT	"/usr/local/vufind/public"
REQUEST_SCHEME	"https"
CONTEXT_PREFIX	""
CONTEXT_DOCUMENT_ROOT	"/usr/local/vufind/public"
SERVER_ADMIN	"admin@localhost"
SCRIPT_FILENAME	"/usr/local/vufind/public/index.php"
REMOTE_PORT	"10452"
REDIRECT_URL	""" /Record/\n arxiv_https___arxiv_org_abs_2604_26412/Similar """
GATEWAY_INTERFACE	"CGI/1.1"
SERVER_PROTOCOL	"HTTP/1.1"
REQUEST_METHOD	"GET"
QUERY_STRING	""
REQUEST_URI	"/Record/%0Aarxiv_https___arxiv_org_abs_2604_26412/Similar"
SCRIPT_NAME	"/index.php"
PHP_SELF	"/index.php"
REQUEST_TIME_FLOAT	1790990089.425
REQUEST_TIME	1790990089

Environment Variables empty

Registered Handlers

0. Whoops\Handler\PrettyPageHandler