paperlist.ai

AI conference papers

AI conference paper library

Search papers from 30+ top conferences, including NeurIPS, ICLR, ICML, CVPR, ACL and AAAI. Filter by keyword, then open any paper for its abstract, authors and links.

arXiv

Inferring Relative Consequences of Mechanical Ventilation from Observational Data Using Game-Based Comparisons

David J. Albers, Tell D. Bennett, Jana de Wiljes, George Hripcsak, Bradford J. Smith, Peter D. Sottile, J. N. Stroh

Identifying the effects of mechanical ventilation (MV) protocols in critical care requires analyzing data from heterogeneous patient-ventilator systems in the clinical decision-making environment. Multiscale interactions among these coupled components generate a high-dimensional state space that remains sparsely sampled despite extensive data collection. Analysis of existing data is essential for understanding current respiratory management practices and generating testable hypotheses about improvement. The scale and complexity of available data motivate the use of reinforcement learning (RL) to explore data-consistent counterfactual trajectories. However, formulating RL in practical applications requires a spatiotemporally dependent reward process that defines state-to-consequence relationships, their context dependence, and the delays over which consequences emerge. These poorly understood elements are not known \emph{a priori} and inferred from data via hypotheses. To that end, categorized observed states are contrasted according to their relative consequences by solving a game-based inverse problem that identifies a comparison model required for downstream probabilistic and stochastic methods such as reinforcement learning for seeking MV optimization and personalization. The inverted-game inference is validated on synthetic data to reveal potential caveats before proceeding to real-world ICU data applications that expose complexities of the data-generating process. Clinical data applications revealed that both breath-type consequences and their relative ordering are inherently context- and time-dependent, varying across patient subgroups, time, and comparison quantities, and effect timescale. The discussion includes potential developments toward a state transition model for simulating the effects of MV management actions using empirical data and game-inferred comparisons.

arXiv

When Memory Becomes Authority: Benchmarking Authority Collapse at the Memory Consolidation Boundary

Qiuyang Zhan, Rui Zhang, Sheng Guo, Lepeng Zhao, Zhuotao Liu

Persistent memory allows (self-evolving) LLM agents to adapt across tasks by consolidating heterogeneous interaction histories into reusable facts, preferences, observations, and rules. Yet consolidation also imposes an implicit authorization boundary: it determines whether stored information may later be consumed as a user fact, an attested observation, or a standing instruction. We identify authority collapse, in which consolidation preserves a claim while erasing the source constraints governing its authorized use, causing the stored memory to imply greater authority than its source permits. We introduce AuthMem-Bench, a controlled paired benchmark that holds the focal claim and downstream task fixed while varying only source authority. It evaluates write-time collapse, downstream authorization errors, and automatic authority preservation. Across seven consolidators based on widely used agent-memory systems and seven LLM backbones, we observe authority collapse in 48 of 49 evaluated configurations. In a controlled action-grounded evaluation, collapsed memories without authority metadata yield a mean unauthorized-action rate of 50.3%. In an end-to-end evaluation, automatically predicted and persisted authority labels reduce the observed unauthorized-action rate from 16.9% to 0.0%, while benign task success remains essentially unchanged. These findings show that memory-driven adaptation must preserve not only what was learned, but also the authority under which it may be reused.

arXiv

Where Reasoning Diverges: Localized Multi-Agent Debate for Multi-Hop Question Answering

Weijun Gao, Xiang Ding, Haoyang Liu, Tiancheng Xing

Multi-agent debate commonly exchanges complete rationales even when disagreements concern only a few intermediate claims. We introduce Localized Multi-Agent Debate (LMAD), an inference-time protocol that represents agent rationales as nodes, locates their earliest conflict, and restricts debate to the corresponding local segments. Guarded resolution extends a shared committed state so that later conflicts can be addressed without reopening accepted steps. We evaluate LMAD on four multi-hop question-answering benchmarks using ten backbones from four model families. Our method achieves the highest macro-averaged judge accuracy across all ten backbones, outperforming the strongest conventional baseline by up to 7.20 percentage points.

arXiv

A New Theory of Value for Post-AGI Economics

Keyun Ruan

Artificial general intelligence (AGI) may weaken scarcities in labour, expertise, information, and productive capability that underpin established theories of economic value. If cognitive work becomes widely automatable, market price, labour input, revealed preference, profit, and gross output may diverge sharply from human and societal benefit. This paper develops Flourishing Value Theory (FVT) as a foundation for post-AGI economics. FVT defines value as the counterfactual, distribution-sensitive contribution of a system, institution, asset, or intervention to the durable capabilities of persons and communities to flourish within social and planetary constraints. It treats societal value as multidimensional, agency-preserving, regenerative, and non-compensatory at critical thresholds. Drawing on the economics of AI, welfare and capability theory, automation, digital valuation, and ecological economics, the paper distinguishes value creation from value capture and retains price, profit, productivity, and GDP as partial signals rather than final measures of progress. It develops the shifts from scarcity to governed abundance, transaction to transformation, and zero-sum rivalry to positive-sum and infinite-game dynamics, with collective expansion of consciousness as an agency-preserving form of regenerative value. Building on Flourishing Metrics and Return on Flourishing (RoF), it proposes a layered architecture for firms, governments, work transitions, AI governance, and national accounting. The central post-AGI economic problem is not how to maximise output, but how to convert abundant intelligence into durable, fairly distributed human, societal, and planetary flourishing.

arXiv

Asking Questions the Right Way: A Multi-Agent Conversational System for Prompt Formulation in Complex Task Resolution

B. Sankar, Pawni Yadav, Srinidhi Ranjini Girish, Amogh A. S

Large language models (LLMs) are integral to complex intellectual tasks, yet output quality remains constrained by user-provided prompts. Iterative multi-turn prompting often leads to context degradation and diminishing cognitive returns. We present PAWNI (Prompt Architecture Wizard using Neural Intelligence), an agentic conversational interface of eight agents that transforms unstructured queries into structured prompts through guided question-and-answer dialogue informed by a self-evolving knowledge base. Rather than optimising the model's response, PAWNI optimises the question itself by front-loading intent clarification. We also propose a three-tier framework of 18 prompt elements across Essential, Enhancement, and Elevation categories. To evaluate system behaviour and validate a measurement protocol, we conducted an exploratory within-subjects study (N=4) across four complex tasks, integrating 32-channel EEG, NASA-TLX workload, and behavioural metrics. Participants produced more structurally complete prompts with PAWNI (42% to 91% of assessed elements), rated LLM outputs higher across all quality dimensions, and reported lower workload (39.6 vs. 21.7 NASA-TLX). Every participant reached satisfactory output in a single turn, compared to 1-12 turns unaided. While effect sizes are unstable due to sample size, direction consistency supports the hypothesis that optimising prompt formulation front-end is a critical lever for human-AI collaboration.

arXiv

When Prompts Control Robots: Prompt Injection Attacks in Multi-Agent Robotic Systems

Neha Nagaraja, Amisha Bagari, Hayretdin Bahsi

Large language models are increasingly integrated into autonomous robotic systems for task planning and control, but this integration exposes them to prompt injection attacks that can lead to unsafe decisions and physical harm. Multi-agent settings increase the risks through cross-agent contamination and broader attack surfaces. In this paper, we evaluate prompt injection attacks against an LLM-based multi-agent robotic system, considering both direct injections into task instructions and indirect injections through perception modules. In our experiments across varying attack-goal complexities and injection strategies in both single-agent and multi-agent settings, we show that prompt injection can induce adversarial actions while reducing task completion. We find that attacks can propagate from one agent to others through shared prompt structures, with impacts varying depending on prompt composition and the targeted agent. We further analyze how architectural changes affect LLM queries and, consequently, the attack success. To the best of our knowledge, this is the first study that systematically investigates prompt injection attacks in a multi-agent LLM-based robotic system.

arXiv

Optimising for Flourishing: Flourishing Metrics and Return on Flourishing as Success Criteria for Artificial Intelligence and Post-AGI Economic Systems

Keyun Ruan, Jonathan D. Teubner, John M. Bremen

Current evaluation frameworks for artificial intelligence focus mainly on capability, safety, and proxies such as adoption, engagement, efficiency, productivity, and financial return. These criteria are necessary but insufficient because they do not establish whether increasingly powerful systems improve or degrade human and planetary well-being. Through an integrative conceptual synthesis, we argue that human flourishing should serve as a primary success criterion for artificial intelligence, the global race to develop increasingly capable AI systems, and prospective post-AGI economic systems. We make three contributions. First, Flourishing Metrics provides an extensible framework spanning physical, emotional, financial, relational, spiritual, and planetary well-being, combining validated subjective measures with representative behavioural, organisational, community, and environmental indicators. Second, Return on Flourishing (RoF) extends return on investment by evaluating the counterfactual contribution of interventions, policies, and AI systems to flourishing relative to their resources, risks, and opportunity costs. Third, we develop distribution-sensitive safeguards and show how RoF could guide AI-enabled work redesign, institutional appraisal, assurance, and post-deployment monitoring through business pilots. We formalise flourishing as a dynamic system variable while emphasising the need for democratic specification, empirical calibration, independent validation, and protection against unacceptable losses within particular dimensions or stakeholder groups. RoF is proposed not as a universal reward function, but as a value-accounting and decision architecture for assessing whether intelligence, automation, and economic transformation generate durable human and planetary progress.

arXiv

Beyond Retrieval: Analytic Memory for Multimodal Agents

Zhoujin Tian, Hao Zhang, Yao Tian, Cheng Chen, Yakun Li, Lei Zhang, Xiaofang Zhou

Long-term multimodal memory must support not only retrieving relevant information but also computing over observations accumulated across interactions. Existing systems largely emphasize \emph{retrieval memory}, organizing interaction histories through summaries and indexes to return query-relevant information at multiple granularities, from high-level abstractions to underlying records. In this paper, we formulate \emph{analytic memory} as a complementary abstraction that organizes recurring multimodal observations into queryable structures supporting filtering, aggregation, ranking, and temporal comparison. We present AdaMM, a framework that jointly supports retrieval and analytic memory. Rather than relying on application-defined schemas, AdaMM extracts provenance-linked attribute-value observations from dialogue, images, and contextual metadata, discovers recurring field structures, and materializes them for analytical access. At inference time, a memory-aware planner decomposes queries into retrieval and analytic operations and routes each operation to the appropriate tools. Experiments on two long-term multimodal memory benchmarks, MemEye and MemGallery, show that AdaMM improves performance by up to 11.3\% and 6.9\%, respectively.

arXiv

MerchantBench: Benchmarking LLM Agents for Long-Term Coherence in E-Commerce Operations

Qiming Shi, Yulong Tao, Linbo Jin, Zhaolu Kang, Yibo Dou, Jiawen Zhu, Tianjun Pan, Shaokang Fu, Chengyu Wang, Siyue Li, Yaping Cheng, Di Weng, Chengfu Huo

Large language model agents are increasingly evaluated as autonomous tool users, yet most benchmarks focus on bounded tasks with immediate success criteria. Real-world deployments often require Long-Term Coherence, the capacity to preserve purposeful behavior across extended horizons while adapting decisions to accumulated evidence. Evaluating this capacity requires a persistent environment in which actions constrain future choices, feedback arrives at heterogeneous delays, and incoherent behavior produces measurable cumulative effects. Seller-side e-commerce provides a suitable setting for this evaluation through recurrent and interdependent decisions over Product Sourcing, Listing and Pricing Control, Cash-Flow Management, and Mixed-Latency Feedback Adaptation. We introduce MerchantBench, a 365-day order-level simulation grounded in 98,843 real e-commerce product records and equipped with 26 tools for agent interaction. MerchantBench couples promptly observable Upstream Supplier Events with delayed Downstream Order Outcomes, requiring agents to follow individual order lifecycles and revisit earlier decisions. We evaluate eight LLMs under two agent frameworks in 48 runs, each spanning 365 simulated days. Our results reveal a substantial gap between even the latest LLMs and human participants, with the best LLM configuration attaining only 27.3\% of the mean final net assets achieved by human participants.

arXiv

PAIChecker: Uncovering and Checking PR-Issue Misalignment in SWE-Bench-Like Benchmarks

Manyi Wang, Junjielong Xu, Pinjia He

SWE-bench-like benchmarks are widely used for evaluating LLM's issue resolution capability. They typically follow a common construction pipeline: each PR (Pull Request) is paired with its linked issue by extracting issue references from the PR description; the issue description is used as the problem statement, and the PR patch serves as the test oracle. However, due to the inherent complexity of developing and maintaining large repositories, such PR-Issue pairings are often misaligned in practice. In this work, we systematically study SWE-bench Verified instances, finding that 13.6% exhibit misalignment across five patterns in eleven fine-grained scenarios. To enable reliable and scalable construction of those benchmarks in the future, we propose PAIChecker, a multi-agent system for checking PR-Issue misalignment in SWE-bench-like benchmarks. Specifically, PAIChecker adopts a three-phase design that combines specific pattern identification, cross-agent label synthesis, and code-level validation, thereby enabling more accurate, generalizable, and progressively verified detection. Experiments on SWE-Gym and SWE-bench Multilingual show that PAIchecker achieves the best performance across all four LLM backbones, reaching up to 92.12% and 91.67% binary accuracy, respectively.

arXiv

SKILL-KD: Contrastive Skill Distillation for LLM Agents

Qiming Shi, Yibo Dou, Jiawen Zhu, Yulong Tao, Linbo Jin, Zhaolu Kang, Yunfan Zhou, Di Weng

Skill-based prompting has become a practical mechanism for improving large language model (LLM) agents, yet existing skill acquisition methods often treat skills as experience summaries, memory entries, or direct summaries of successful demonstrations. This creates a mismatch for weaker student agents: when a student fails because it lacks task knowledge or operational strategy, its failed trajectory may not contain enough evidence to infer the missing behavior, while the teacher trajectory may be too implicit to be internalized as reusable guidance. We propose SKILL-KD, a contrastive skill distillation framework that treats skills as an explicit distillation medium between agents of different capabilities. Given a student failure and the teacher trajectory on the same task, SKILL-KD distills their actionable discrepancy into a textual skill patch, evaluates the patch by re-running the student, and iteratively refines the patch when the student still fails. To prevent repeated local updates from causing skill drift, SKILL-KD further maintains trace-linked edit histories and performs Drift-Aware Skill Consolidation, deciding whether each patch should add a new rule, delete or modify an existing rule, or be skipped. Across five agent benchmarks and two student settings, SKILL-KD consistently improves frozen student agents over fixed-model adaptation baselines.

arXiv

SpatialCLI: Learning to Reason With Spatial Tools, Then Without Them

Yang Zhou, Zixuan Huang, Sunzhu Li, Zhuo Yang, Chen Zhang, Shunian Chen, Caijun Yan, Jianyao Xu, Shunyu Liu, Weijie Fu, Peiliang Li, Xiaozhi Chen, Yuxiang Cai

Vision-language models (VLMs) are increasingly used in embodied agents to interpret visual inputs, reason about spatial relationships, and make task-level decisions based on that reasoning. However, a fundamental capability mismatch remains: general VLMs can reason about the overall task but often miss the visual details that determine success, while specialist vision models can capture those details but cannot translate them into task-level decisions. In this work, we propose SpatialCLI, a framework that teaches VLMs to reason with spatial tools and progressively internalize the specialist perceptual capabilities they provide. SpatialCLI proceeds in three stages: (1) Call exposes specialist vision models as spatial tools to augment the VLM's perception; (2) Learn uses Cold-Start SFT and agentic RL to improve tool use; and (3) Internalize verbalizes successful tool-use trajectories to internalize specialist perceptual capabilities. We further introduce SpatialCLI-Bench, a 516-example benchmark for compositional perception across localization, segmentation, depth, and pose. On MindCube, SpatialCLI raises Qwen3-VL-8B-Instruct from 29.3% to 84.6% with tools, surpassing GPT-5.6 Sol with tools (72.1%), while retaining 73.8% without tools after internalization.

arXiv

OPOD: On-Policy Omni Distillation

Tong Zhao, Yuyang Hu, Yutao Zhu, Reed Li, Haijin Liang, Haibo Shi, Yu Lu, Zhicheng Dou

Omni-modal models provide a unified interface for text, images, and audio. However, improving these abilities together remains difficult, as post-training on pooled multimodal data often fails to preserve the strengths of modality teachers. On-policy distillation (OPD) has recently become popular in model post-training. It samples responses from the current student and compares the teacher's and student's next-token distributions along those responses, yielding dense supervision while reducing the mismatch between training and inference. Despite these advantages, standard OPD does not readily extend to several modality teachers. Their guidance may favor conflicting changes to the shared model, while matching each teacher's next-token distribution can prevent the student from moving beyond that teacher. To address these challenges, we propose On-Policy Omni Distillation (OPOD), which consolidates text, image, and audio teachers into one omni model. OPOD routes each response to the corresponding teacher, controls the teachers independently, and applies guidance only when the teacher assigns a higher probability to the generated token. The selected teacher also evaluates answer confidence and whether the reasoning increases support for the answer. Extensive experiments on twelve benchmarks show that OPOD achieves the best average at three model scales, reaching 70.8, 51.7, and 46.2 and outperforming the strongest comparator by 2.1, 1.8, and 1.7 points. At 30B, it surpasses the base model and pooled RL training on all twelve benchmarks, and ranks first or second on eleven even when the teachers are included. Only the student is retained for deployment.

arXiv

Cura 1T: Specialized Model for Agentic Healthcare

actAVA AI, Haolin Chen, Leon Qi, Steve Brown, Deon Metelski, Tao Xia, Joonyul Lee, Qixuan Wang, Kevin Riley, Frank Wang, Weiran Yao

Healthcare AI agents handle patient consultation, clinical reasoning over text and images, interactive diagnosis, and electronic health record (EHR) tool use, yet specialized agentic models that cover these use cases together remain limited. These capabilities fail in different ways, and a narrow update for one task can degrade another. We present Cura 1T, a healthcare-specialized LLM built on the open-weight Kimi-K2.6 and trained through a human-gated recursive self-improvement (RSI) loop. Specifically, in each round, the RSI harness plans a target capability, trains the model, evaluates benchmark trajectories, and refines the data mixture from observed failures with targeted synthetic and curated examples rather than a single generic medical-data update. Across the healthcare evaluation suite, Cura 1T ranks at or near the top among frontier baselines while remaining competitive on out-of-domain reasoning and agentic benchmarks.

arXiv

InvFlowFD: Reference-Free and Background-Set-Free Perceptual Music Quality Metric with Flow Matching Inversion

Alon Ziv, Harel Pogoda, Yossi Adi

Existing reference-free methods for evaluating music perceptual quality alleviate the need for paired noisy-clean data, but they still rely on a background set, which is used to compute aggregated statistics of clean audio samples. In this work, we propose a novel approach that eliminates this requirement, achieving background-set-free and reference-free quality estimation using only a pre-trained Flow Matching backbone. We demonstrate that unconditional Flow Matching inversion via simple Euler integration is sufficient to detect various artificial distortions and accurately rank music generation models against human perceptual judgments. We introduce InvFlowFD, which performs flow inversion and compares a group of inverted samples to the prior distribution. We evaluate our method against prior work, quantitatively and with a thorough human study. Results suggest that InvFlowFD is highly correlated with human perception of sound distortions, as well as generative models' quality, while being more flexible and less restrictive than existing metrics.

arXiv

Out-Of-The-Loop Multi-Fidelity Bayesian Optimization

Gustavo Sutter, Hao Wang, Luis Ricardez-Sandoval, Pascal Poupart, Agustinus Kristiadi

Black-box optimization is a ubiquitous problem in science and engineering, often dealing with expensive objective functions with cheaper lower-fidelity proxies available. Multi-fidelity Bayesian optimization (MF-BO) is a principled approach to this problem, leveraging correlations across different fidelities when querying the objective. However, for many important MF-BO tasks, the true highest-fidelity function is prohibitively expensive to be part of the optimization loop. Nevertheless, practitioners often have gold standard data (observations of the highest-fidelity function) obtained from previous experiments that might provide information for the current task. For instance, in molecular optimization, chemists often pick the top-$k$ candidate molecules using various computer simulations, and later reveal their true objective function values. In this work, we demonstrate the suboptimality of standard MF-BO algorithms in the real-world scenarios above, even under ideal assumptions. Next, we mitigate this problem by incorporating historical high-fidelity data accompanied by task descriptors---which can be explicitly given or extracted from unstructured metadata. We demonstrate the effectiveness of our methods on synthetic functions, as well as real-world problems in chemistry and hyperparameter optimization.

arXiv

Think Fast: Estimating No-CoT Task-Completion Time Horizons of Frontier AI Models

Dewi Gould, Francis Rhys Ward, Anders Cairns Woodruff, Rauno Arike, Josh Hills, Alex Serrano, Ida Caspary, Jason Ross Brown, Jo J. Jiao, Patrick Leask, Twm Stone, Ram Potham, Ionut Gabriel Stan, Harry Mayne, Simeon Hellsten, Shubhorup Biswas, Ariana Azarbal, William L. Anderson, Elle Najt, Ryan Greenblatt, Julian Stastny

Many efforts to ensure frontier AI models are safe rely on monitoring their chain-of-thought (CoT) reasoning. If models become able to perform sufficiently complex reasoning internally, without explicit thinking tokens, this would undermine such oversight. We measure how well frontier models reason without CoT across a suite of over 30,000 questions spanning 43 benchmarks in domains including math, coding, puzzles, causality, theory-of-mind, and strategic reasoning. To compare models against humans, we estimate the $50\%$-task-completion time horizon (TH): the human time required for tasks a model completes with $50\%$ success rate. We complement this with a $50\%$ reasoning token horizon: the minimum number of o3-mini reasoning tokens needed for tasks a model solves with $50\%$ success rate. We find that the no-CoT $50\%$ TH of frontier models has been doubling roughly every year over the past six years, with GPT-5.5's TH reaching over 3 minutes and reasoning token horizon exceeding 1,500 tokens. Our median estimates predict that frontier no-CoT THs could exceed 7 minutes by 2028, and 25 minutes by 2030, though these projections carry substantial uncertainty. We recommend frontier developers track this explicitly.

arXiv

Language model agents show in-group trust bias invisible to standard behavioural audits

Messi H. J. Lee

Language-model agents are moving from single-user assistants into persistent networks that build trust and reputation with one another, and the same models increasingly control physically embodied robots as well as software. Here we show that five widely used open-weight reasoning models develop an in-group trust bias the moment group membership becomes visible to them, even when the groups are arbitrary labels with no real-world meaning: in a 20-agent simulation, agents direct 53.6-54.6% of their trust-building actions toward in-group targets against a 47.4% base rate expected by chance, a shift present in every model tested and confirmed by three independent statistical checks and an instruction-rewording robustness test. This bias is easy for current evaluation practice to miss, because it operates through which agent receives an action rather than which action is chosen - a channel invisible to the aggregate behaviour-log audits that are the standard way multi-agent AI systems are evaluated today. A resource-scarcity manipulation, intended to test whether competition intensifies the bias, instead reduced it in three of five models; we trace this to an artifact of how scarcity was enforced, not to a failure of the underlying mechanism. Group-contingent social dynamics are therefore already present in the models multi-agent AI systems are built from, and auditing practice built around single-model, single-decision evaluation cannot detect them.

arXiv

Spatiotemporal Graph Transformer for Traffic Intelligence in Edge Computing

Laha Ale, Letian Lin, Na Cao, Zheng Ma, Peng Yu

Accurate traffic forecasting is essential for proactive resource management in edge computing, where service demand evolves dynamically across both space and time. In practical cellular edge systems, traffic exhibits strong spatial correlations among neighboring service regions and long-range temporal dependencies driven by user mobility and application behavior. Existing recurrent forecasting approaches can capture short-term dynamics but often struggle to model long-horizon traffic evolution under non-stationary conditions. To address this challenge, we propose a spatiotemporal graph Transformer framework that jointly models spatial interactions and temporal dependencies for traffic forecasting in edge computing. The framework employs graph neural networks to capture spatial correlations among service regions and leverages Transformer-based self-attention to learn long-range temporal patterns from historical traffic observations. By decoupling spatial representation learning from temporal reasoning, the proposed approach provides an effective mechanism for large-scale spatiotemporal traffic modeling. Extensive experiments on a real-world cellular network dataset demonstrate that the proposed graph Transformer consistently outperforms recurrent graph-based baselines, including GCN-RNN, GCN-LSTM, and GCN-GRU models, across multiple forecasting horizons. The resulting forecasts enable more effective proactive resource provisioning and reduce overload risk compared with reactive management strategies. These results highlight the potential of graph-enhanced attention mechanisms for building intelligent and adaptive edge computing systems.

arXiv

Spend Bits Where Queries Look: KV Cache Vector Quantization with Attention-Preserving Transforms

Samuel Fernández-Menduiña, Amir Ziashahabi, Eduardo Pavez, Antonio Ortega, Salman Avestimehr

Long-context LLM decoding reads the key-value (KV) cache at every step. Loading it takes longer than computing attention over it, so throughput is bandwidth-bound. Hence, reducing the cache size can raise both decoding speed and serving capacity. The challenge is to reduce cache size while preserving the attention products, keeping reconstruction cheap, and using a fixed per-token bit count. At two bits per element, the most competitive methods rely on orthogonal transforms. However, existing techniques are either data-oblivious or use the query statistics without deriving the transform from a distortion criterion. Moreover, they rely on transforms built on top of random or Hadamard rotations, which equalize variances across entries rather than compacting energy, and fixed-width scalar quantizers, which are suboptimal at low rates. In this paper, we formulate KV cache quantization as a transform coding problem in which distortion is the error in the attention products. We derive closed-form optimal transforms for keys and values from calibration statistics, under a high-resolution model. We show that the optimal key transform is not orthogonal and satisfies a generalized Parseval relation: the attention-aware distortion becomes mean-squared error (MSE) in the transform domain. Thus, we can use MSE-optimal vector quantizers applied directly to the transformed key coefficients. To meet the fixed-width layout requirement, we show that grouping coefficients into equal-volume partitions makes equal-size codebooks attain the variable-rate optimum under the same high-resolution model. At two bits per element, our method, termed NOVA-KV, recovers most of the long-context retrieval accuracy lost by scalar quantization methods at comparable throughput.

Keep browsing

More papers from top AI conferences

30+ top conferences, by conference, year, and topic.