source_id,title,authors,venue_status,year,url,access_date,revision,repository_url,commit,checkpoint_availability,supported_backends,license,relevance,reproduction_status,verification
swat2026,Swarm Telepathy: Prefill-Free KV-Cache Transfer Across Families,Abdulrahman Alabdulkareem; Irene Tenison; Yang Zhang; Yada Zhu; Boris Katz; Dan Gutfreund,arXiv preprint (cs.LG); no venue stated,2026,https://arxiv.org/abs/2610.15095,2026-10-29,v1 (2026-10-10); latest v1,https://github.com/Ar-Kareem/SwaT,ae2de2285caf82f4d500526bf6e6992747b98eea,"none released; mapper (~300M params) must be trained locally (<1 GPU-day, 2 GPUs); default pair Granite-3.3-8B -> SmolLM3-3B",PyTorch 2.6.0 + HF Transformers 5.2.0; multi-GPU CUDA; GPU model not stated; no vLLM/MLX/MPS mention,"none (no LICENSE file, GitHub API license=null)",Primary: first prefill-free cross-family KV transfer; translate-then-AM-compact pipeline (Eq. 8); receiver needs token IDs of source text but no prefill,pending,"verified: arXiv abs+API+HTML, GitHub API, raw README/requirements; unverified: GPU type, MPS support"
compaction2026,Fast KV Compaction via Attention Matching,Adam Zweiger; Xinghong Fu; Han Guo; Yoon Kim,arXiv preprint (cs.LG); no venue stated,2026,https://arxiv.org/abs/2602.16284,2026-10-29,v1 (2026-02-18); latest v2 (2026-05-26),https://github.com/adamzweiger/compaction,2020f6d37940fe013e7e7f1ebf409e5d107c5670,not applicable (training-free); only per-model head-budget JSON files shipped; demo Qwen/Qwen3-4B,PyTorch 2.8.0 + HF Transformers 4.57.1 + vLLM 0.11.0 (query generation) + accelerate; Python 3.12; H100/H200 in paper; no MLX/MPS mention,MIT,Primary: same-model latent KV compaction used by SwaT+AM; closed-form keys/values/bias from reference queries; no cross-model claim,pending,"verified: arXiv abs+API+HTML, GitHub API, LICENSE, requirements; unverified: MPS support"
c2c2025,Cache-to-Cache: Direct Semantic Communication Between Large Language Models,Tianyu Fu; Zihan Min; Hanling Zhang; Jichao Yan; Guohao Dai; Wanli Ouyang; Yu Wang,Published in ICLR'26 (arXiv Comments); cs.CL cs.LG,2025,https://arxiv.org/abs/2510.03215,2026-10-29,v1 (2025-10-03); latest v2 (2026-03-02),https://github.com/thu-nics/C2C,5eac72b90faf7b3dc2638e874fe7cba1e74c8bcf,"released: https://huggingface.co/nics-efc/C2C_Fuser (apache-2.0; 7 fuser pairs: Qwen3-0.6B<-Qwen2.5-0.5B-Instruct, Qwen3-0.6B<-Llama-3.2-1B-Instruct, Qwen3-0.6B<-Qwen2.5-Math-1.5B, Qwen3-0.6B<-Qwen3-4B, Qwen3-0.6B<-Qwen3-4B-Base, Qwen3-1.7B<-Qwen2.5-1.5B-Instruct, Qwen3-8B<-Qwen2.5-7B-Instruct); demo huggingface.co/spaces/nics-efc/C2C_demo","PyTorch 2.6.0 + HF Transformers 4.52.4; CUDA (examples device=cuda, linux-64 conda env); A100 timing; no vLLM/MLX/MPS mention",Apache-2.0 (LICENSE; pyproject classifier says MIT - discrepancy),"Primary: cross-family KV fusion via trained per-layer fuser with gate; receiver still prefills source text; +3.1-5.4% vs T2T, 2.5x latency",pending,"verified: arXiv abs+API+HTML, GitHub API, raw README/pyproject, HF API model repo; unverified: training GPU type, MPS support"
latentmas2025,Latent Collaboration in Multi-Agent Systems,Jiaru Zou; Ruizhong Qiu; Gaotang Li; Xiyuan Yang; Katherine Tieu; Pan Lu; Ke Shen; Hanghang Tong; Yejin Choi; Jingrui He; James Zou; Mengdi Wang; Ling Yang,ICML 2026 Spotlight (arXiv Comments + repo); cs.CL cs.AI cs.LG,2025,https://arxiv.org/abs/2511.20639,2026-10-29,v1 (2025-11-25); v2 (2025-12-08); v3 (2026-06-01); latest v4 (2026-08-03),https://github.com/Gen-Verse/LatentMAS,9a9e4d331eb11430bd9e64754c6b252b06d73031,"none (training-free; stock HF models Qwen3-4B/8B/14B, Llama-3 3B/8B); experiment logs on Google Drive",PyTorch + HF Transformers (unpinned); optional patched vLLM backend; 8x A100-80G in paper; no MLX/MPS mention,Apache-2.0,Primary: training-free latent thoughts + shared KV working memory between same-architecture agents; lossless-transfer theorem; homogeneous agents only,pending,"verified: arXiv abs+API+HTML, GitHub API, LICENSE, requirements; unverified: dedicated limitations section not located, MPS support"
heo2026,Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse,Taekyung Heo; Rasoul Shafipour; Ritchie Zhao; Maximilian Golub; Mohammad Mahdi Kamani; Ritika Borkar; Makesh Tarun Chandran; Pantea Zardoshti; Bita Darvish Rouhani,arXiv preprint; no venue stated,2026,https://arxiv.org/abs/2608.03893,2026-10-29,v1 (2026-08-04); latest v1,unverified,unverified,unverified,unverified,unverified,Same-family closed-form ridge KV mapper (Qwen3 14B->32B); cited by SwaT as 'Closed Form' baseline,pending,arXiv export API metadata only (title/authors/dates/abstract); repo not checked
qu2026,CacheBridge: Efficient Cross-Model KV Cache Transfer,Xingyu Qu; Siyuan Lu; Zhiyu Chen; Sheng Wang; Tao Lin,arXiv preprint; no venue stated,2026,https://arxiv.org/abs/2609.00891,2026-10-29,v1 (2026-09-01); latest v1,unverified,unverified,unverified,unverified,unverified,Per-head architecture-indexed mapper improving Heo et al.; same-family; cited by SwaT,pending,arXiv export API metadata only
yun2026,Prefill-Free Cross-Family KV Cache Transfer for Heterogeneous Multi-Agent LLMs (HeteroFold),Vincent-Daniel Yun; Woosang Lim; Haneul Yoo; Sungjoo Yoo; Murali Annavaram; Sai Praneeth Karimireddy,arXiv preprint; no venue stated,2026,https://arxiv.org/abs/2609.32259,2026-10-29,v1 (2026-09-26); latest v2 (2026-09-29),unverified,unverified,unverified,unverified,unverified,Direct competitor to SwaT: prefill-free cross-family transfer with frozen models (Llama-3.1-8B -> Ministral-3-14B),pending,arXiv export API metadata only
li2026,A Universal Context-Reuse Layer for Cross-Model KV Sharing,Yi Li; Dongming Jiang; Yi Zhao; Bingzhe Li,arXiv preprint; no venue stated,2026,https://arxiv.org/abs/2608.30963,2026-10-29,v1 (2026-08-31); latest v1,unverified,unverified,unverified,unverified,unverified,Serving-system layer translating KV state across models differing in scale/architecture/tokenizer/family,pending,arXiv export API metadata only
shi2025,KVComm: Enabling Efficient LLM Communication through Selective KV Sharing,Xiangyu Shi; Marco Chiesa; Gerald Q. Maguire; Dejan Kostic,ICLR 2026 (arXiv Comments),2025,https://arxiv.org/abs/2510.03346,2026-10-29,v1 (2025-10-02); latest v3 (2026-02-22),unverified,unverified,unverified,unverified,unverified,Selective layer-wise KV-pair sharing as inter-LLM channel; contemporaneous alternative to C2C,pending,arXiv export API metadata only
ye2025,KVCOMM: Online Cross-context KV-cache Communication for Efficient LLM-based Multi-agent Systems,Hancheng Ye; Zhengqi Gao; Mingyuan Ma; Qinsi Wang; Yuzhe Fu; Ming-Yu Chung; Yueqian Lin; Zhijian Liu; Jianyi Zhang; Danyang Zhuo; Yiran Chen,NeurIPS 2025 (arXiv Comments),2025,https://arxiv.org/abs/2510.12872,2026-10-29,v1 (2025-10-14); latest v2 (2025-11-01),https://github.com/FastMAS/KVCOMM,unverified,unverified,unverified,unverified,"Same-model multi-agent KV reuse with context-aware offsetting (~6.7x prefill speedup); efficiency baseline, not cross-model",pending,"arXiv export API metadata only; repo URL from arXiv Comments, not fetched"
peng2026,StateBridge: Training-free Hidden-state Alignment for Latent Communication in LLM Multi-Agent Systems,Yanwen Peng; Delvin Ce Zhang; Xi Wang; Nikolaos Aletras,COLM 2026 (arXiv Comments; 18 pages),2026,https://arxiv.org/abs/2608.13317,2026-10-29,v1 (2026-08-13); latest v1,unverified,unverified,unverified,unverified,unverified,Training-free orthogonal alignment of sender final-layer hidden states to receiver input space; cross-model latent channel without projectors,pending,arXiv export API metadata only
du2025,Enabling Agents to Communicate Entirely in Latent Space (Interlat),Zhuoyun Du; Runze Wang; Huiyu Bai; Zouying Cao; Xiaoyong Zhu; Yu Cheng; Bo Zheng; Wei Chen; Haochao Ying,Accepted to ACL 2026 (arXiv Comments),2025,https://arxiv.org/abs/2511.09149,2026-10-29,v1 (2025-11-12); latest v5 (2026-07-13),unverified,unverified,unverified,unverified,unverified,Last-hidden-state latent communication with learned compression; early latent-MAS baseline,pending,arXiv export API metadata only
liu2024,DroidSpeak: KV Cache Sharing for Cross-LLM Communication and Multi-LLM Serving,Yuhan Liu; Yuyang Huang; Jiayi Yao; Shaoting Feng; Zhuohan Gu; Kuntai Du; Hanchen Li; Yihua Cheng; Junchen Jiang; Shan Lu; Madan Musuvathi; Esha Choukse,"arXiv; no venue in Comments (search snippet: NSDI 2026, unverified)",2024,https://arxiv.org/abs/2411.02820,2026-10-29,v1 (2024-11-05); latest v4 (2025-07-14),unverified,unverified,unverified,unverified,unverified,Same-architecture cross-LLM KV reuse (recompute few layers); precedent cited by SwaT,pending,arXiv export API metadata only
liu2026,Beyond tokens: a unified framework for latent communication in LLM-based multi-agent systems,Yingzhuo Liu,arXiv preprint (survey); no venue stated,2026,https://arxiv.org/abs/2606.05711,2026-10-29,v1 (2026-06-04); latest v3 (2026-07-15),unverified,unverified,unverified,unverified,unverified,Survey of 18 latent-communication methods 2024-2026; taxonomy for related work,pending,arXiv export API metadata only
villani2026,LatentPort: Beyond KV Cache - Cross-Model Transfer of Recurrent Memory in Hybrid Language Models: A 4B-to-9B Hybrid-State Handoff Without Target Prefix Replay,Simon P. Villani,"arXiv preprint (14 pages, Comments)",2026,https://arxiv.org/abs/2609.25053,2026-10-29,v1 (2026-09-07); latest v1,unverified,unverified,unverified,unverified,unverified,Transfers GDN recurrent state + KV across Qwen3.5 4B->9B hybrid siblings without prefix replay,pending,arXiv export API metadata only
kivi2024,KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache,"Zirui Liu, Jiayi Yuan, Hongye Jin, Shaochen Zhong, Zhaozhuo Xu, Vladimir Braverman, Beidi Chen, Xia Hu",ICML 2024 (stated in arXiv Comments and repo),2024,https://arxiv.org/abs/2402.02750,2026-10-29,v2 (2024-07-25),https://github.com/jy-yuan/KIVI,876b4d2d08e3b1d5f70d0969c299d8c7c42ddfb6 (2025-11-20),"none required (training-free; applies to HF Llama-2/3, Mistral, LongChat checkpoints)","PyTorch + custom CUDA extension (kivi_gemv) + Triton packing kernel + flash-attn; requirements.txt pins torch==2.1.2, transformers==4.36.2, flash-attn==2.5.6, triton==2.1.0 (README: compatible with transformers 4.43 since 2025-01); CUDA only; no Apple Silicon support",MIT,"asymmetric 2-bit KV quantization (keys per-channel grouped over G=32 tokens, values per-token, fp16 residual window R=128); question-agnostic lossy compression baseline; orthogonal to eviction",pending,"arXiv abstract + HTML v2, GitHub page, raw README/llama_kivi.py/requirements.txt/quant/setup.py/example.py, GitHub API fetched 2026-10-29"
snapkv2024,SnapKV: LLM Knows What You are Looking for Before Generation,"Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, Deming Chen",NeurIPS 2024 main track per proceedings.neurips.cc (secondary); NOT stated on arXiv page (no Comments) nor in repo (cites arXiv preprint),2024,https://arxiv.org/abs/2404.14469,2026-10-29,v2 (2024-06-17),https://github.com/FasterDecoding/SnapKV,e216ddc84c5bd210378cbdbbba12ba02102aa640 (2025-07-10),"none required (training-free monkey-patch; configs for Llama-2-chat, LongChat-7b-v1.5-32k, Mistral-7B-Instruct, Mixtral-8x7B-Instruct, LWM-Text-Chat-1M)","PyTorch + HF transformers monkey patches (Llama, Mistral, Mixtral); README tested transformers==4.37.0, requires transformers>=4.36 and flash-attn==2.4.0 (pyproject: transformers>=4.34, torch unpinned); CUDA via flash-attn; no Apple Silicon mention; core selection is plain PyTorch",Apache-2.0,question-aware eviction: last-window (16/32/64 tokens) attention votes + 1D pooling (kernel 5/7/13) + per-head top-k over prefix; the observation window is the end of the prompt (instruction/question); primary question-aware baseline,pending,"arXiv abstract + HTML v2, GitHub page, raw snapkv_utils.py/pyproject.toml/pred_snap.py/LongBench configs, GitHub API, NeurIPS proceedings page fetched 2026-10-29; repo not linked from arXiv page (found at expected URL, BibTeX matches)"
pyramidkv2024,PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling,"Zefan Cai, Yichi Zhang, Bofei Gao, Yuliang Liu, Yucheng Li, Tianyu Liu, Keming Lu, Wayne Xiong, Yue Dong, Junjie Hu, Wen Xiao",COLM 2025 per Microsoft Research page (secondary); NOT stated on arXiv page (no Comments) nor in repos (BibTeX: CoRR 2024 / arXiv preprint),2024,https://arxiv.org/abs/2406.02069,2026-10-29,v4 (2025-05-15),"https://github.com/Zefan-Cai/KVCache-Factory (canonical, renamed from Zefan-Cai/PyramidKV on 2024-11-28); https://github.com/BartsLab/PyramidKV exists but is a 0-star fork snapshot",KVCache-Factory 68cd9551a63fedf362ddb0edb008badd04c24996 (2026-08-12); BartsLab fork 506c866a876bdcc7ca4f2f73aa7f856a8c666479 (2024-11-05),"none required (training-free; Llama-3-8B/70B-Instruct, Mistral-7B-Instruct-v0.2)",PyTorch + HF transformers; requirements.txt pins transformers==4.44.2 (README says >=4.41); flash-attn>=2.4.0.post1 optional (sdpa/eager supported); Llama + Mistral; batch size 1; no decoding-stage compression; no Apple Silicon mention; selection is plain PyTorch,MIT,"layer-wise budget pyramid (arithmetic from 2*base-base/beta at layer 0 down to base/beta at top layer, beta=20) on top of SnapKV-style question-aware selection (window alpha=8, kernel 7 maxpool in runner); question-aware; budget rule itself is content-independent",pending,"arXiv abstract + HTML v4, GitHub pages (BartsLab fork + KVCache-Factory), raw README/LICENSE/requirements.txt/run_longbench.py/pyramidkv_utils.py, GitHub API, Microsoft Research page fetched 2026-10-29; author list differs between arXiv v4 (Yucheng Li) and repo BibTeX/Microsoft page (Baobao Chang)"
h2o2023,H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models,"Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, Zhangyang Wang, Beidi Chen",NeurIPS 2023 (stated in repo description '[NeurIPS'23]'; arXiv page has no Comments field),2023,https://arxiv.org/abs/2306.14048,2026-10-29,v3 (2023-12-18),https://github.com/FMInference/H2O,ac75c2a8a9e76832b2a4139b9363373b56336bfb (2024-06-18),"none required (training-free eviction; OPT via FlexGen, LLaMA via HF; paper also GPT-NeoX)","PyTorch >=1.12; h2o_flexgen (FlexGen fork, OPT) and h2o_hf (HF transformers installed unpinned from git main; lm-eval; crfm-helm); simulation (attention masking) and real KV drop (utils_real_drop); batch size 1; CUDA assumed; no Apple Silicon mention; eviction logic is plain PyTorch",MIT badge in README only; NO LICENSE file in repository (GitHub API license=null) — unverified as formal license,"cumulative-attention eviction: per-head/per-layer accumulated softmax attention received + always-keep recent window; budget 20% of prompt split evenly heavy/recent (0.1/0.1); query-agnostic policy, question-informed only if the question is in the prefill; intermediate baseline between StreamingLLM and SnapKV",pending,"arXiv abstract + HTML v3 (2 of 3 chunks), GitHub page, raw README/h2o_hf README/h2o_flexgen README/utils_real_drop/modify_llama.py/utils_hh/modify_llama.py, GitHub API + contents listing fetched 2026-10-29"
streamingllm2023,Efficient Streaming Language Models with Attention Sinks,"Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, Mike Lewis",ICLR 2024 (stated in arXiv Comments and repo description),2023,https://arxiv.org/abs/2309.17453,2026-10-29,v4 (2024-04-07),https://github.com/mit-han-lab/streaming-llm,2e5042606d69933d88fbf909bd77907456b9b4dd (2024-07-11),"none required (training-free; Llama-2, MPT, Falcon, Pythia/GPT-NeoX; demo lmsys/vicuna-13b-v1.3); sink-token pretrained 160M models not released on page","PyTorch (unpinned), Python 3.8, transformers==4.33.0 pinned; pos-shift patches for llama/gpt_neox/falcon, MPT via ALiBi; CUDA demo; no Apple Silicon mention (third-party iPhone integration cited in News); cache logic is pure tensor slicing",MIT,position-only eviction: keep start_size=4 attention sinks + rolling recent window (recent_size 512 default / 2000 demo / 2044 paper); keys cached un-rotated and re-rotated with positions 0..kv_len-1 within the cache; canonical question-blind baseline,pending,"arXiv abstract + HTML v4, GitHub page, raw README/kv_cache.py/pos_shift/modify_llama.py/enable_streaming_llm.py/run_streaming_llama.py/setup.py, GitHub API fetched 2026-10-29"
llmlingua2_2024,LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression,"Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang, Menglin Xia, Xufang Luo, Jue Zhang, Qingwei Lin, Victor Rühle, Yuqing Yang, Chin-Yew Lin, H. Vicky Zhao, Lili Qiu, Dongmei Zhang",Findings of ACL 2024 (arXiv Comments: Accepted at Findings of ACL 2024),2024,https://arxiv.org/abs/2403.12968,2026-10-29,arXiv v2 (2024-08-12); pip llmlingua 0.2.2 (2024-04-09),https://github.com/microsoft/LLMLingua,5a4c78ae18ab17a98cf997e8259354e546081d64,"HF: microsoft/llmlingua-2-xlm-roberta-large-meetingbank (0.56B F32, 2235.8 MB, rev ebaba9b, mit); microsoft/llmlingua-2-bert-base-multilingual-cased-meetingbank (0.18B F32, 709.4 MB, rev 5f0c827, apache-2.0)",torch + transformers (AutoModelForTokenClassification); device_map cuda (default) / cpu / mps verified in prompt_compressor.py,"MIT (repo, pip); checkpoints mit / apache-2.0",Token-classification extractive compressor used as the prompt-compression baseline (rate / target_token),pending,"page fetched (arXiv, GitHub, HF API, PyPI) + source code read"
longbench2024,"LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding","Yushi Bai, Xin Lv, Jiajie Zhang, Hongchang Lyu, Jiankai Tang, Zhidian Huang, Zhengxiao Du, Xiao Liu, Aohan Zeng, Lei Hou, Yuxiao Dong, Jie Tang, Juanzi Li","ACL 2024 Long Papers pp. 3119-3137, DOI 10.18653/v1/2024.acl-long.172",2024,https://aclanthology.org/2024.acl-long.172/,2026-10-29,arXiv v2 (2024-06-19); HF dataset zai-org/LongBench sha 5e628be450b7e67fb7ae6e201bd6d8f7056f7672 (2024-12-18),https://github.com/THUDM/LongBench,2e00731f8d0bff23dc4325161044d0ed8af94c1e,"HF dataset THUDM/LongBench -> zai-org/LongBench; subsets qasper 200, multifieldqa_en 150, hotpotqa 200, 2wikimqa 200, musique 200; metric F1; context field included",datasets library (JSON); evaluation code in repo,"Code MIT; dataset card has no license field (unverified); upstream: Qasper CC BY 4.0, HotpotQA CC BY-SA 4.0, MuSiQue CC BY 4.0, 2WikiMultihopQA repo Apache-2.0 (data unverified), MultiFieldQA-en unverified","QA evaluation set (F1) with explicit long context for compression / transfer experiments; v2 (arXiv 2412.15204, 503 MCQ, 8k-2M words, apache-2.0) is not a drop-in replacement",pending,"page fetched (arXiv, ACL Anthology, GitHub raw, HF API, 4 upstream license pages)"
ruler2024,RULER: What's the Real Context Size of Your Long-Context Language Models?,"Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, Yang Zhang, Boris Ginsburg",COLM 2024 (arXiv Comments); GitHub BibTeX still cites arXiv preprint,2024,https://arxiv.org/abs/2404.06654,2026-10-29,arXiv v3 (2024-08-06),https://github.com/NVIDIA/RULER,c3f5e3b4f87f97e048793bb510a3a6b19a46bf3a,No checkpoints; synthetic data generator (haystack: repeat / essay (Paul Graham) / needle; needles words / numbers / uuids); QA from SQuAD and HotpotQA,Python generator + model-agnostic inference scripts (vLLM/HF/TRT-LLM per repo),Apache-2.0 (repo); generated-data license depends on upstream sources (unverified),"Synthetic long-context tasks (NIAH single/multi-key/multi-value/multi-query, variable_tracking, common_words_extraction, freq_words_extraction, qa) for controlled compression stress tests",pending,"page fetched (arXiv, GitHub API, raw README)"
mlxlm2026,mlx-lm: LLM inference and fine-tuning with MLX (Apple silicon),Apple ML Explore team (ml-explore),software release v0.32.2 (2026-10-16),2026,https://github.com/ml-explore/mlx-lm,2026-10-29,v0.32.2 (GitHub release 2026-10-16T02:49:02Z; PyPI mlx-lm 0.32.2),https://github.com/ml-explore/mlx-lm,6b572ab1b531dd8fd2d13f6f36a89c2eccda18ea,N/A (framework); prompt caches saved as .safetensors via mlx_lm.cache_prompt / save_prompt_cache,MLX on macOS (mlx>=0.32.2 Darwin); extras cuda12 / cuda13 / cpu; Python >=3.11; transformers>=5.7.0,MIT,"Local KV-cache backend: mlx_lm.cache_prompt, --prompt-cache-file, --max-kv-size, --kv-bits, --kv-group-size, --quantized-kv-start, make_prompt_cache, save_prompt_cache, load_prompt_cache, trim_prompt_cache, KVCache, QuantizedKVCache, RotatingKVCache",pending,"page fetched (README, releases, GitHub/PyPI API) + source read at HEAD"
hf_kvcache_docs,Transformers documentation: Cache strategies (KV cache),Hugging Face,"online documentation, Transformers v5.21.1",2026,https://huggingface.co/docs/transformers/en/kv_cache,2026-10-29,docs v5.21.1 (links on page),https://github.com/huggingface/transformers,unverified,N/A (framework docs),PyTorch; quantized cache backends hqq / quanto / fp8; offloading to CPU; torch.compile with StaticCache,Apache-2.0 (Transformers library; license not restated on the docs page),"Documents DynamicCache, StaticCache, QuantizedCache (hqq/quanto/fp8), EncoderDecoderCache, cache_implementation static / offloaded / offloaded_static / quantized; in-process prefix reuse via Prefill a cache (prefix caching) with copy.deepcopy, Iterative generation, crop(); cache_position and to-disk serialisation not mentioned on the page",pending,"page fetched (kv_cache + internal/generation_utils, full text)"
anthropic_prompt_caching,Claude Developer Platform documentation: Prompt caching (+ Models overview),Anthropic,online documentation (living page),2026,https://platform.claude.com/docs/en/build-with-claude/prompt-caching,2026-10-29,page state on 2026-10-29 (no version number shown),unverified,unverified,N/A (hosted API),"Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud (Google managed), Microsoft Foundry",proprietary documentation (no license stated),"cache_control ephemeral, up to 4 breakpoints, prefix order tools->system->messages, 20-block lookback; minimum 512 / 1024 / 2048 / 4096 tokens by model; TTL 5 min default or 1h; usage fields cache_creation_input_tokens, cache_read_input_tokens, cache_creation.ephemeral_5m_input_tokens / ephemeral_1h_input_tokens; writes 1.25x (5m) / 2x (1h), reads 0.1x (0.025x Fable 5.1 / Mythos 5.1; 0.05x Opus 5.5 / Sonnet 5.5); current IDs claude-fable-5-1 ($10/$50), claude-opus-5-5 ($4/$20), claude-sonnet-5-5 ($2/$10), claude-haiku-5-5 (from $0.10/$0.50)",pending,"page fetched (both pages, full text)"
li2024,Prompt Compression for Large Language Models: A Survey,"Zongqian Li, Yinhong Liu, Yixuan Su, Nigel Collier",arXiv (Comments empty); NAACL 2025 Long per search snippet (unverified),2024,https://arxiv.org/abs/2410.12388,2026-10-29,arXiv v2 (2024-10-17),unverified,unverified,N/A (survey),N/A,arXiv (license not checked),Taxonomy of hard vs soft prompt compression; situates LLMLingua-2,pending,arXiv page fetched; venue from search snippet only
liu2025,KV Cache Compression for Inference Efficiency in LLMs: A Review,"Yanyu Liu, Jingying Fu, Sixiang Liu, Yitian Zou, You Fu, Jiehan Zhou, Shouhua Zhang",arXiv preprint (Comments: 12 pages),2025,https://arxiv.org/abs/2508.06297,2026-10-29,arXiv v1 (2025-08-08),unverified,unverified,N/A (review),N/A,arXiv (license not checked),Review of selective / quantisation / attention-based KV-cache compression,pending,arXiv page fetched
javidnia2025,"Key, Value, Compress: A Systematic Exploration of KV Cache Compression Techniques","Neusha Javidnia, Bita Darvish Rouhani, Farinaz Koushanfar",IEEE CICC 2025 (invited; arXiv Comments),2025,https://arxiv.org/abs/2503.11816,2026-10-29,arXiv v3 (2025-04-22),unverified,unverified,N/A,N/A,arXiv (license not checked),Taxonomy and accuracy / latency evaluation of KV-cache compression methods,pending,arXiv page fetched
gao2025,Rethinking Key-Value Cache Compression Techniques for Large Language Model Serving,"Wei Gao, Xinyu Zhou, Peng Sun, Tianwei Zhang, Yonggang Wen",MLSys 2025 (arXiv Comments),2025,https://arxiv.org/abs/2503.24000,2026-10-29,arXiv v1 (2025-03-31),unverified,unverified,open-source tools announced (URL not verified),N/A,arXiv (license not checked),Shows KV compression can lengthen outputs and raise end-to-end latency in serving; benchmark caveats for our cost model,pending,arXiv page fetched
ye2025,KVCOMM: Online Cross-context KV-cache Communication for Efficient LLM-based Multi-agent Systems,"Hancheng Ye, Zhengqi Gao, Mingyuan Ma, Qinsi Wang, Yuzhe Fu, Ming-Yu Chung, Yueqian Lin, Zhijian Liu, Jianyi Zhang, Danyang Zhuo, Yiran Chen",NeurIPS 2025 (arXiv Comments: Accepted for publication in NeurIPS2025),2025,https://arxiv.org/abs/2510.12872,2026-10-29,arXiv v2 (2025-11-01),unverified,unverified,unverified,GPU (H100 reported),arXiv (license not checked),"Closest prior work: training-free cross-agent KV-cache reuse with context-aware offsets, ~6.7x prefill speedup in 3-agent setting",pending,arXiv page fetched
bian2026,TokenDance: Scaling Multi-Agent LLM Serving via Collective KV Cache Sharing,"Zhuohang Bian, Feiyang Wu, Chengrui Zhang, Hangcheng Dong, Yun Liang, Youwei Zhuo",arXiv preprint under review,2026,https://arxiv.org/abs/2604.03143,2026-10-29,arXiv v1 (2026-04-03),unverified,unverified,unverified,GPU serving (vLLM comparison),arXiv (CC BY 4.0 per HTML header in search snippet; unverified),All-gather KV sharing across agents with sparse diffs against a master copy; up to 17.5x lower per-agent KV storage,pending,arXiv page fetched
jeon2026,LRAgent: Efficient KV Cache Sharing for Multi-LoRA LLM Agents,"Hyesung Jeon, Hyeongju Ha, Jae-Joon Kim",ICML 2026 Poster (arXiv journal-ref),2026,https://arxiv.org/abs/2602.01053,2026-10-29,arXiv v2 (2026-05-31),unverified,unverified,unverified,GPU,arXiv (license not checked),Shares the base-model part of the KV cache across LoRA-adapted agents; low-rank adapter-specific part,pending,arXiv page fetched
chang2026,MemCollab: Cross-Model Memory Collaboration via Contrastive Trajectory Distillation,"Yurui Chang, Yiran Wu, Qingyun Wu, Lu Lin",arXiv preprint (no venue stated),2026,https://arxiv.org/abs/2603.23234,2026-10-29,arXiv v2 (2026-05-28),unverified,unverified,unverified,N/A,arXiv (license not checked),Cross-agent / cross-backbone shared memory via contrastive trajectory distillation; knowledge transfer at memory level rather than KV level,pending,arXiv page fetched