mlsanityData to 7 Oct 2026

Efficient reasoning and agent inference

172 papers5 this weekHigh confidence in this grouping: the papers form one clear topicHigh confidence in this grouping: the papers form one clear topic · About

Keywords: reasoning, cache, compression, agents, agent

Key facts. Efficient reasoning and agent inference: 172 papers on arXiv since 2024 (data through 7 Oct 2026). 22 of its 172 mapped papers (13%) cite “R-KV: Redundancy-aware KV Cache Compression for Reasoning Models” (2025, arXiv 2505.24133). 20 of its 172 mapped papers (12%) cite “Dynamic Early Exit in Reasoning Models” (2025, arXiv 2504.15895). 20 new papers 24 Sep–7 Oct.

What’s new

24 Sep – 7 Oct 2026

20 new papers, 24 Sep – 7 Oct; 5 this week.

  1. This week · 5
  2. Persistent Memory in Multi-Agent LLM Inference: What It Costs, What It Buys, and When You Can Tell
  3. Can Agent Harnesses and Inference Engines Hear Each Other? The HEAR Protocol for Agentic LLM Serving
  4. The Optimization Landscape of Learning Compacted Context Models
  5. Request Order Matters: Cache-History Sensitivity in Selective KV-Cache Reuse for Rolling Agents
  6. Spend Bytes on Breadth: Precision-Count Trade-offs for Decode-Time KV Compression in Long Chain-of-Thought Reasoning
  7. Earlier · 15
  8. AvoKV-E: Payload-Aware KV Cache Eviction for Long Reasoning also fits: KV cache optimization
  9. Persistent Context Graphs for Efficient Memory Compaction in LLM Agents
  10. Taming Speculative Search for Test-Time Scaling in LLM Serving also fits: Speculative decoding
  11. SparseEngine: Sparse-First Inference Engine
  12. FOCUS: Training-Free Decision-Preserving Context Compression for LLM Agents
All 20
  1. Adapting Context Compression for Long-Horizon Agents with Counterfactual Continuations
  2. KV-streams for Efficient Compaction in Agentic Reinforcement Learning
  3. AgentPerfBench: A Benchmarking and Evaluation Suite for Inference Performance of Agentic LLMs also fits: Serving systems and request scheduling
  4. KVCMAS: Efficient KV cache Correction for Shared Context in Multi-Agent Systems
  5. PReCache: Efficient KV Cache Sharing for Multi-LoRA Agents via Low-Rank Precomputation and Neutral Reconstruction
  6. Beyond Token Savings: A Systematic Study of Context Compression in LLM Agents
  7. Prefill-Free Cross-Family KV Cache Transfer for Heterogeneous Multi-Agent LLMs
  8. Receiver-Conditioned Latent Communication gives 94% CacheBack
  9. ActKV: Efficient LLM Agents through Action-Guided KV Cache Management
  10. When Can Agents Forget Their Reasoning? ICLR for Long-Horizon Agent Context Compression

Foundations

Most cited by papers here

  1. R-KV: Redundancy-aware KV Cache Compression for Reasoning Models2025cited by 22 here
  2. Dynamic Early Exit in Reasoning Models2025cited by 20 here
  3. KVFlow: Efficient Prefix Caching for Accelerating LLM-Based Multi-Agent Workflows2025cited by 20 here
  4. Autellix: An Efficient Serving Engine for LLM Agents as General Programs2025cited by 18 here
  5. Continuum: Efficient and Robust Multi-Turn LLM Agent Scheduling with KV Cache Time-to-Live2025cited by 15 here

Lineage

The papers this direction builds on, oldest first.

Each paper on the line cites the one before.

  1. 2023H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Modelscited by 44 here
  2. 2023Model Tells You What to Discard: Adaptive KV Cache Compression for LLMscited by 20 here
  3. 2024KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantizationcited by 28 here
  4. 2024KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cachecited by 28 here
  5. 2024Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inferencecited by 18 here
  6. 2025R-KV: Redundancy-aware KV Cache Compression for Reasoning Modelscited by 22 here
18 more it builds on, not in this chain
  1. 2022Fast Inference from Transformers via Speculative Decodingcited by 13 here
  2. 2023Scissorhands: Exploiting the Persistence of Importance Hypothesis for LLM KV Cache Compression at Test Timecited by 23 here
  3. 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 80 here
  4. 2023LongLLMLingua: Accelerating and Enhancing LLMs in Long Context Scenarios via Prompt Compressioncited by 17 here
  5. 2024DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Servingcited by 15 here
  6. 2024Taming Throughput-Latency Tradeoff in LLM Inference with Sarathi-Servecited by 14 here
  7. 2024LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compressioncited by 19 here
  8. 2024Cost-Efficient Large Language Model Serving for Multi-turn Conversations with CachedAttentioncited by 12 here
  9. 2024SnapKV: LLM Knows What You are Looking for Before Generationcited by 52 here
  10. 2024CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusioncited by 23 here
  11. 2024PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funnelingcited by 27 here
  12. 2024Quest: Query-Aware Sparsity for Efficient Long-Context LLM Inferencecited by 25 here
  13. 2024DroidSpeak: KV Cache Sharing for Cross-LLM Communication and Multi-LLM Servingcited by 11 here
  14. 2025Autellix: An Efficient Serving Engine for LLM Agents as General Programscited by 18 here
  15. 2025Dynamic Early Exit in Reasoning Modelscited by 20 here
  16. 2025KVFlow: Efficient Prefix Caching for Accelerating LLM-Based Multi-Agent Workflowscited by 20 here
  17. 2025LMCache: An Efficient KV Cache Layer for Enterprise-Scale LLM Inferencecited by 15 here
  18. 2025Continuum: Efficient and Robust Multi-Turn LLM Agent Scheduling with KV Cache Time-to-Livecited by 15 here

The 24 most cited of 389.

Browse

30 of 172, ranked 6–35

  1. 6DroidSpeak: KV Cache Sharing for Cross-LLM Communication and Multi-LLM Serving202411
  2. 7Reasoning Path Compression: Compressing Generation Trajectories for Efficient LLM Reasoning20259
  3. 8LazyEviction: Lagged KV Eviction with Attention Pattern Observation for Efficient Long Reasoning20259
  4. 9ThinKV: Thought-Adaptive KV Cache Compression for Efficient Reasoning Models20259
  5. 10KVCOMM: Online Cross-context KV-cache Communication for Efficient LLM-based Multi-agent Systems20259
  6. 11ACON: Optimizing Context Compression for Long-horizon LLM Agents20258
  7. 12TokenCake: A KV-Cache-centric Serving Framework for LLM-based Multi-Agent Applications20257
  8. 13TriAttention: Efficient Long Reasoning with Trigonometric KV Compression20266
  9. 14SideQuest: Model-Driven KV Cache Management for Long-Horizon Agentic Reasoning20265
  10. 15LRAgent: Efficient KV Cache Sharing for Multi-LoRA LLM Agents20264
All 30
  1. 16ThunderAgent: A Simple, Fast and Program-Aware Agentic Inference System20264
  2. 17Entropy After </Think> for reasoning model early exiting20254
  3. 18CONCUR: High-Throughput Agentic Batch Inference of LLM via Congestion-Based Concurrency Control20263
  4. 19ForesightKV: Optimizing KV Cache Eviction for Reasoning Models by Learning Long-Term Contribution20263
  5. 20Agent Memory Below the Prompt: Persistent Q4 KV Cache for Multi-Agent LLM Inference on Edge Devices20263
  6. 21TokenDance: Scaling Multi-Agent LLM Serving via Collective KV Cache Sharing20263
  7. 22MEMENTO: Teaching LLMs to Manage Their Own Context20263
  8. 23Leyline: KV Cache Directives for Agentic Inference20263
  9. 24Value-Aware Stochastic KV Cache Eviction for Reasoning Models20263
  10. 25IntentKV: Cross-Turn Intent-Aware KV Cache Pruning for Agent Inference20263
  11. 26RelayCaching: Accelerating LLM Collaboration via Decoding KV Cache Reuse20262
  12. 27LongFlow: Efficient KV Cache Compression for Reasoning Models20262
  13. 28Efficient LLM Serving for Agentic Workflows: A Data Systems Perspective20262
  14. 29MemDecay: Region-Aware KV Cache Eviction for Efficient LLM Agent Inference20262
  15. 30FlashThink: An Early Exit Method For Efficient Reasoning20252
  16. 31Think Clearly: Improving Reasoning via Redundant Token Pruning20252
  17. 32ReasonCache: Accelerating Large Reasoning Model Serving through KV Cache Sharing20252
  18. 33Batch Query Processing and Optimization for Agentic Workflows20252
  19. 34Breadcrumbs Reasoning: Memory-Efficient Reasoning with Compression Beacons20252
  20. 35LYNX: Learning Dynamic Exits for Confidence-Controlled Reasoning20252