Get Started
Home
Topics
Search
Library
Research questionHow can long-reasoning KV caches retain useful context while reducing memory and eviction overhead?Extended reasoning creates long KV caches that consume substantial memory during inference. Eviction decisions may add overhead, while removing useful context can impair subsequent reasoning.
AI
AI Memory
Evaluation & Benchmarks
Inference Optimization
Reasoning
Research Paper
Technology
Latest papersRecent research connected to this question, newest first.BeaconKV: Key-Value Cache Compression Guided by Beacon Queries for Efficient Large Reasoning Model InferenceThe source studies training-free KV-cache compression during inference across four open-source large reasoning models and diverse reasoning benchmarks. Its evidence concerns memory reduction, reasoning accuracy, and throughput in those settings.research paper · Sep 4, 2026What Matters for Aggressive Decoding-Time KV Eviction? Temporal Aggregation and Ranking PreservationApplies to decoding-time KV-cache compression and eviction. The evidence covers multiple scoring and aggregation choices across six open-weight backbones on LongBench, LongBench-v2, and RULER, including periodic refresh and one-time frozen rankings; it reports ranking behavior, quality changes, and throughput. The findings do not establish that scoring quality is irrelevant in general.research paper · Sep 3, 2026Random Attention: Rethinking KV Cache Eviction for Efficient ReasoningThis paper evaluates score-free eviction that retains the prompt and samples cached reasoning tokens within each attention head. Evidence is limited to four language models, six reasoning tasks, and vLLM serving comparisons, so it does not establish behavior for other workloads, cache policies, or inference systems.research paper · Sep 3, 2026CONF-KV: Confidence-Aware KV Cache Eviction with Mixed-Precision Storage for Long-Horizon LLMThe source evaluates a confidence-aware cache policy combined with attention-based token ranking, a protected recent window, mixed-precision storage, blockwise attention, and per-layer budgets. Evidence covers four model families, generated lengths up to 4K, retrieval on Needle-in-a-Haystack up to 32K tokens, and 75 VisualWebArena tasks; the reported results do not establish behavior beyond these settings.research paper · May 24, 2026
Related questions
How can long-context LLM inference reduce KV-cache memory without losing attention-head-specific information?How can long-context language models reduce per-token reads from million-token KV caches without losing accuracy?How can LLM serving adapt KV-cache capacity as attention demand changes during long-output reasoning?How can low-bit KV caches save autoregressive decoding memory without losing long-context retrieval quality?