When serving LLM agents in production, the KV cache grows over turns, and mainstream serving engines (e.g., vLLM) rely on least recently used (LRU) eviction to manage memory pressure. While LRU is well-understood for OS pages and single-turn LLM serving, its behavior on agent workloads remains underexplored. We present an inverted age-importance phenomenon in agent workflows, where the oldest KV blocks (system prompts, tool outputs) are the most persistently attended, while the .newest blocks (chain-of-thought reasoning) decay fastest. As a result, recency-based eviction systematically discards the wrong blocks in agent workloads, producing two distinct overheads: (1) wasted prefill work on preempt-discard engines, and (2) wasted PCIe bandwidth on CPU-offload engines. We propose RoleKV, a training-free, lossless eviction policy that evicts KV blocks based on their semantic role in agent workflows. A lightweight Role Tagger assigns each block to a semantic role, and RoleKV then evicts by role-aware decay rather than recency. Unlike model-level KV compression methods that incur quality loss, RoleKV is a system-level policy that decides where blocks reside (GPU vs. CPU/disk), without affecting model outputs. RoleKV cuts preemptions by up to 64.3%, PCIe traffic by up to 63.8%, and improves end-to-end latency by up to 1.41× on 8B-70B models, with gains that grow on longer agent traces and persist on multi-GPU setups. This recasts KV eviction in production LLM serving from a recency problem to a role-aware retention problem.

Token-pruning has emerged as a primary focus in large language models (LLMs) to enhance model efficiency while preserving accuracy, especially for large sequence lengths. However, the eviction operation of token-pruning methods causes “holes” in KV tensors, posing two major challenges: (1) The shift operation, required to make the KV tensor contiguous, results in significant copy overheads; (2) The changes in position indices due to token eviction lead to increased computational requirements for Rotary Positional Encoding (RoPE). To address these issues, we introduce EQUIP, an EQUivariant preserving in-place token update mechanism that ensures the equivariance property of the operations performed in the attention computation. EQUIP offers two fundamental advantages: First, it combines eviction and a subsequent token insertion into an in-place replacement operation, which reduces the KV cache copy overheads significantly. Second, EQUIP reduces recomputation of rotation operations through a combination of in-place update, caching and a re-indexing strategy. Together, these optimizations enable EQUIP to achieve geomean speedups of 1.62× (or 1.47×) on CPU (GPU) over StreamingLLM, and 3.45× (or 1.86×) on CPU (GPU) over Heavy Hitters (H2O). EQUIP with Paged Attention achieves speedups of 4.18×(2.61×) on CPU (GPU) over auto-regressive baselines. EQUIP matches the model accuracy of baseline pruning methods while delivering superior performance.

Emotion lexicons provide information about associations between words and emotions. They have proven useful in analyses of reviews, literary texts, and posts on social media, among other things. We evaluate the feasibility of deriving emotion lexicons cross-lingually, especially for low-resource languages, from existing emotion lexicons in resource-rich languages. For this, we start out from very small corpora to induce cross-lingually aligned vector spaces. Our study empirically analyses the effectiveness of the induced emotion lexicons by measuring translation precision and correlations with existing emotion lexicons, along with measurements on a downstream task of sentence emotion prediction.