SemanticKV compresses KV cache for agentic video generation
Authors report a 3.35x KV-memory cut and 1.22x speedup on Causal Forcing at compression ratio 0.7.
Stories per day · 30 days
2 RSS
Relatedkv-cache5brotli3long-context3google2qwen32ruler2test-time-training2video-generation2attention1autoregressive1clickhouse1cloudflare1data-pipeline1denial-of-service1embeddings1hybrid-models1inference1infrastructure1
32 stories, newest first · one card per story
Stories tagged #compression
LowRAM compresses KV cache with low-rank keys and amortised values
Test-time value predictors plus cross-layer key compression give about 4× compression and up to 5× decode throughput.
Brotli vulnerability enables unauthenticated remote code execution on WordPress sites
CVE-2026-111497 corrupts memory in Brotli library used by WordPress for outbound responses; public proof-of-concept works against PHP 8.3.36.
Brotli vulnerability puts broad range of software at risk
The Brotli compression library flaw, CVE-2026-111497, reaches beyond Node.js and affects many other projects.
1
DCKV learns end-to-end compression of hybrid-model prompt KV caches
On Qwen3.5-2B, tokens use over 55 times fewer bytes than bf16 with at most 1.8 points lost on three long-context benchmarks.
Google Brotli before 1.0.10, 1.1.1 and 1.2.1 has a vulnerability
Canada's Cyber Centre urges administrators to review Google's advisory and apply updates, without naming impact or a CVE.
LEMS advances SVD compression for LLMs
LEMS and KFAC-SVD deliver up to 4.8 percentage point accuracy gains on Mistral, Qwen3 and Llama 3 models.
XZ Utils mishandles decoder memory failures, risking code execution
Ubuntu's USN-8945-1 says a bad write during decoder reinitialization could cause denial of service or arbitrary code execution.
OverSeer adds fixed auxiliary views to multi-agent world models
A delta-VAE encodes only changes from a static reference, cutting tokens for auxiliary views.
SpecMRL beats full-size MRL with half the embedding size
On MTEB, SpecMRL outperforms standard MRL at every tested reduced dimension, with larger gains under more aggressive compression.
Latent Compaction with Test-Time Training
Latent Compaction preserves long-context retrieval at 32x KV-cache compression via test-time training.
Latent Compaction with Test-Time Training
Latent Compaction keeps 96.7% RULER accuracy at 32x KV-cache compression via test-time training.
Rewinding the exponential: How we keep hardware usage down
Cloudflare details pipeline changes that deferred about 2.5 years of hardware growth.
4
H-DDCM: Hadamard Codebooks for Fast Generative Image and Video Compression
H-DDCM uses Hadamard codebooks for faster, lower-memory generative image and video compression.
Learning by Compression: Risk Estimates Attain the Law-of-Large-Numbers Rate
Compression-based learning risk bounds match the law-of-large-numbers rate without a logarithmic penalty.
OpenSSH 10.6 enables a post-quantum signature algorithm, so experimental keys need replacing
OpenSSH 10.6 disables risky compression, blocks shell-injecting usernames, and enables a post-quantum signature.
Learning to Learn a Language
A 300M-parameter model trained only on synthetic priors learns to predict real languages in context.
Compression Footprints as Security Signals for Model-Poisoning Defense in Federated Learning
CRAFT defends federated learning against model poisoning by scoring lossy-compression footprints of client updates.
Beyond Selection: Token Parameterization for Extreme Visual Token Compression
Braco compresses visual tokens 23–144× while preserving VLM accuracy and cutting prefill FLOPs.
Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression
Chunked KV-cache compression creates phase-dependent retrieval gaps of up to 40 points in long-context models.
UniCache: Task- and Type-Aware KV Cache Compression for Unified Multimodal Models
UniCache compresses multimodal KV caches up to 5× with little quality loss and higher throughput.
Weight Pair Encoding: Inducing a Smaller Grammar in Neural Network Weights
WeightPE finetunes int8 weights so Re-Pair grammars shrink to 0.38–0.43x of QAT, costing about 1–2 accuracy points.
Generalization behavior of OPTQ and the role of regularization
New bounds show OPTQ generalization depends on regularization lambda, and a revised lambda choice beats prior recommendations.
Beyond Compression: Training Latent Representations for Stable Long-Horizon Rollout in Neural Surrogate Solvers
Research on improving long-horizon rollout stability in neural surrogate solvers, reducing error by 40% while matching full-resolution models on physics benchmarks.