Key Value Indicators (KVIs) provide a decision oriented view of a service by summarizing how operational performance translates into stakeholder value, risk, and outcomes. However, in many domains KVIs are difficult to compute in practice because they require selecting relevant KVI categories, defining measurable Key Performance Indicators (KPIs), collecting KPI values, and applying consistent calculation logic, all of which is typically performed manually and inconsistently from unstructured service documentation. This paper presents KPI2KVI, a tool that transforms a natural language service description into computed KVI estimates by orchestrating a deterministic multi agent workflow powered by Large Language Models (LLMs) that (i) elicits missing service context, (ii) extracts and finalizes relevant KVI categories from a taxonomy, (iii) generates service specific KPIs with units and descriptions, (iv) collects KPI values through an interactive dialogue and also supports intelligent estimation for KPI values that are unavailable, and (v) computes interval valued KVI outputs (minimum, exact, maximum) with traceable explanations for each KVI code. Simulations with representative service descriptions demonstrate that KPI2KVI consistently produces a complete end to end mapping from description to KVI intervals and provides transparent calculation narratives that support post hoc auditing and interactive advisory queries.
Figures & tables
Fig. 1: KPI2KVI pipeline: from service interview to Key Value Indicator category selection, Key Performance Indicator generation and value collection, per-Key Value Indicator calculation with explicit bounds, and post-hoc advising, with shared memory storing reusable artifacts across stages.
Fig. 2: TheKPI2KVI architecture: a streaming frontend/backend setup with a workflow controller that discovers agents, executes them via an LLM provider, and manages cross-agent state and reusable structured artifacts.
Fig. 3: Simulation results for the four method variants. (A) KVI instability across repeated runs versus calculation complexity (formula depth / #KPIs per KVI). (B) end-to-end success rate (runs without verifier-flagged errors) versus the number of requested KVI categories. (C) category-selection accuracy (F1) versus taxonomy quality q (coverage/correctness). Curves show mean over 10 runs; shaded regions show variance.
Multi-turn LLM agents fan short queries into long trajectories of tool calls, search results, and intermediate reasoning. Both KV memory and KV read bandwidth grow by orders of magnitude across a single trajectory, making the key-value (KV) cache, not parameter compute, the dominant serving bottleneck for long-horizon agents. We introduce IntentKV, learned KV pruning that keeps the base LLM frozen. IntentKV maintains a session-level QueryMemory of cross-turn intent, scores live history tokens with a memory-attention rule, and adds a zero-initialized residual head with cross-attention over current-query K-vectors. To stay composable with prefix caches, eviction is a slot-map redirection: dropped positions route to a sentinel dead slot while surviving K/V rows, RoPE phases, and slot identities stay in place. IntentKV matches the no-pruning full-cache baseline with almost no accuracy drop under tight KV budgets: at an 8k KV budget, mean peak request tokens drop 23.9% on Qwen3-8B and 30.7% on Qwen2.5-14B. On the 100 longest BCP queries that all methods complete on Qwen2.5-14B, IntentKV-8k further cuts worst-case peak request tokens from 92.3k to 20.5k, a 77.8% reduction, and worst-case raw KV reads from 411M to 31M, a 92.6% reduction.
Key Point Analysis (KPA) aims to identify a concise set of key points that summarize a collection of arguments together with their prevalence. We argue that KPA is fundamentally a structured prediction problem that requires recovering semantic groupings, generating representative key points, ensuring coverage, and estimating prevalence. Under this formulation, we show that existing KPA benchmarks suffer from limitations in grouping quality, redundancy, coverage, and argument-key point mappings, causing ceiling violation and selection failure in reference-based evaluation. To support future research on true KPA, we introduce a structure-aware, distribution-sensitive benchmark built via a human-in-the-loop re-annotation. Human and LLM evaluations consistently show that the resulting structures yield more coherent groupings, higher-quality key points, better coverage, and more reliable prevalence estimates than existing annotations. We further release several annotation resources to support research on KPA evaluation, argument-key point matching, explainable KPA, and LLM-as-a-judge methodologies, and outline a research agenda for true KPA.
Agentic serving can consume orders of magnitude more tokens than chatbot workloads, stressing both KV-cache capacity and decode-time bandwidth. Most KV-eviction methods score cached keys against representative queries drawn from the most recent tokens, assuming future attention resembles recent attention. We show that agentic generation violates this assumption: future queries form a mixture over think, act, tool, and others phases, and principal-angle analysis shows these components occupy measurably different query subspaces, so recency representatives systematically undervalue keys that upcoming phases will need. We propose AGENTKV, which maintains a small query buffer per phase and scores cached keys against their union. We further implement AGENTKV in a persistent multi-turn serving path that carries compressed KV state across turns and compacts retained KV pages online. Across two models, six task domains, and three KV budgets each, AGENTKV improves task score by 5.5 points on average over R-KV and 5.3 over Tri-attention. Relative to upstream full-KV SGLang, AGENTKV improves output-token throughput by up to 1.80x. Code: https://github.com/LiuTaowen-Tony/agentkv.
Taowen Tony Liu, Jeffrey T. H. Wong, Can Xiao +3
Imperial College London, London, United Kingdom · Columbia University, New York, NY, USA