Latent Compaction is a test‑time training method that learns a compact, document‑specific key–value cache for large‑context language models. It initializes cache entries using reconstruction importance, teacher attention, and key‑space diversity, then fits attention outputs and log‑normalizers to the full cache before jointly optimizing the entire cache via end‑to‑end distillation. Experiments on RULER, LongBench, and Phonebook Lookup show significant accuracy gains over budget‑matched compression baselines, achieving 96.7% accuracy at 32× compression on RULER and transferring well across model scales and families.
arXiv:2609.36835v1 Announce Type: new Abstract: Long-context large language model inference is bottlenecked by KV caches that grow linearly with sequence length. This burden is especially severe for...
EchoPress is a training‑free method for pruning key‑value caches in large language models. It approximates the reconstruction attention used by KVzip by leveraging queries and keys from standard prefill, reconstructing only the first chunk to calibrate importance scores for the rest of the context. Experiments on LongBench and RULER with Qwen3‑8B and Llama‑3.1‑8B‑Instruct show that EchoPress matches KVzip’s task accuracy across eviction ratios from 50% to 90%, while reducing compression overhead by 1.7–19.6× and total prefill time by up to 2.9×.
arXiv:2605. 25475v2 Announce Type: replace-cross Abstract: Large Language Models (LLMs) are increasingly expected to operate over long contexts, yet standard softmax attention incurs a KV cache that grows linearly with sequence length, quickly becoming the bottleneck for long context inference.
arXiv:2608.23843v1 Announce Type: new Abstract: Long-context inference in large language models (LLMs) is increasingly limited by the memory required for the key-value (KV) cache. KV cache compressio...
arXiv:2607. 16213v1 Announce Type: new Abstract: Large Language Models (LLMs) generate text autoregressively, relying on a key-value (KV) cache whose memory footprint grows linearly with context length, creating a major bottleneck.
Long-context large language model (LLM) inference is increasingly constrained by the memory footprint and decoding cost of key-value (KV) caches, limiting sustainable deployment on resource-constrained hardware. Existing KV cache eviction methods typically apply heuristic token scoring over all heads in GQA-based LLMs.
arXiv:2609.37988v1 Announce Type: new Abstract: As the context size of text processed with an LLM grows, the size of KV caches can outstrip the memory allocated for the original model weights. This i...
The paper introduces Stratified Sparse Attention (SSA), a framework that blends deterministic top‑k retrieval with score‑guided sampling to reduce the gap between sparse and dense attention in large language models. SSA attends to high‑scoring tokens deterministically while sampling the remaining attention mass using proxy scores, thereby recovering more of the attention distribution. Experiments on benchmarks such as RULER, LongBench, LOFT, and AIME show that SSA outperforms existing sparse methods across models ranging from 3B to 27B parameters, achieving significant accuracy gains and up to 2.96× decode speedup with optimized CUDA kernels.
arXiv:2606. 24467v1 Announce Type: new Abstract: Long-context large language model (LLM) inference is increasingly constrained by the memory footprint and decoding cost of key-value (KV) caches, limiting sustainable deployment on resource-constrained hardware.
arXiv:2609.17346v1 Announce Type: cross Abstract: To answer questions outside of their pre-training data, large language models (LLMs) need access to new information, which can be presented in the co...
arXiv:2610.07940v1 Announce Type: cross Abstract: Looped language models apply the same stack of layers T times to each token, which deepens the model without adding parameters but multiplies its key...
The paper critiques the common practice of rehearing all key‑value cache entries before compression, showing that under tight budgets this strategy actually harms performance. Experiments on RULER and LongBench reveal that full rereading retains far fewer useful entries than methods that rehearse nothing. The authors propose Attic‑KV, a training‑free rehearsal that quizzes the model with context‑based question‑answer pairs, which outperforms existing methods and scales better as compression budgets shrink.