arXiv:2507.08762v3 Announce Type: replace
Abstract: We are interested in the relative
conditioning of the problem $y_0\mapsto \mathrm{e}^{tA}y_0$, i.e., the
relative conditioning of the action of the matrix exponential $\mathrm{e}%
^{tA}$ on a vector with respect to perturbations of this vector. The present paper is a qualitative study of the long-time behavior of this
conditioning. In other words, we are interested in studying the propagation to the solution $y(t)$ of perturbations of the initial value for a linear ordinary differential equation $y^\prime(t)=Ay(t)$, by measuring these perturbations with relative errors. We introduce three condition numbers: the first considers a specific initial value and a specific direction of perturbation; the second considers a specific initial value and the worst case by varying the direction of perturbation; and the third considers the worst case by varying both the initial value
and the direction of perturbation. The long-time behaviors of
these three condition numbers are studied.
Science Journals
arXiv:2605.16206v1 Announce Type: new
Abstract: We present fully kinetic two-dimensional, three-velocity-component (2D3V) PIConGPU simulations of a three-beam direct-drive interaction with a 15 $\mu$m solid-density cryogenic hydrogen cylinder, establishing a predictive numerical baseline for the operational DRACO ($\tau=30$ fs) and upcoming PENELOPE ($\tau=150$ fs) laser facilities at HZDR. The simulations resolve charge-separation fields on the order of 3 TV/m and reveal a robust kinematic bifurcation of the accelerated population into a fast (1-5 MeV) ion beam and a slower bulk (1-100 keV) flow. We demonstrate analytically and numerically that the charge-separation front ($v_{hb}$) is an intrinsically non-quasi-neutral electrostatic double layer that lies outside the closure assumptions of radiation-hydrodynamic models. A simple $2v_{hb}$ reflection scaling derived directly from the front trajectory tracks the centroid of the constant-energy fast-ion band under the impulsive 30 fs driver and the time-varying upper edge of the swept fast-ion band under the sustained 150 fs driver, across both intensities ($a_{0}=12.7$ and 22.0), establishing this non-thermal mechanism as the dominant acceleration pathway. We then scan an external axial magnetic field from 0 T to 10 kT. Laboratory-achievable 20 T fields leave all macroscopic observables unchanged; fields at the kT scale progressively magnetise the MeV hot-electron population, quench the laser-driven charge-separation mechanism, suppress the fast-ion band, and more than double the net-inward compression time of the short-pulse driver-while extending the outer target envelope. A geometric equivalence argument maps these kT-scale results onto larger-diameter cryogenic hydrogen jets.
arXiv:2605.16205v1 Announce Type: new
Abstract: Deploying compound LLM agents in adversarial, partially observable sequential environments requires navigating several design dimensions: (1) what the agent sees, (2) how it reasons, and (3) how tasks are decomposed across components. Yet practitioners lack guidance on which design choices improve performance versus merely increase inference costs. We present a controlled study of compound LLM agent design in CybORG CAGE-2, a cyber defense environment modeled as a Partially Observable Markov Decision Process (POMDP). Reward is non-positive, so all configurations operate in a failure-mitigation mode. Our evaluation spans five model families, six models, and twelve configurations (3,475 episodes) with token-level cost accounting. We vary context representation (raw observations vs. a deterministic state-tracking layer with compressed history), deliberation (self-questioning, self-critique, and self-improvement tools, with optional chain-of-thought prompting), and hierarchical decomposition (monolithic ReAct vs. delegation to specialized sub-agents). We find that: (1) Programmatic state abstraction delivers the largest returns per token spent (RPTS), improving mean return by up to 76% over raw observations. (2) Distributing deliberation tools across a hierarchy degrades performance relative to hierarchy alone for all five model families, reaching up to 3.4$\times$ worse mean return while using 1.8-2.7$\times$ more tokens. We call this destructive pattern a deliberation cascade. (3) Hierarchical decomposition without deliberation achieves the best absolute performance for most models, and context engineering is generally more cost-effective than deliberation. These findings suggest a design principle for structured adversarial POMDPs: invest in programmatic infrastructure and clean task decomposition rather than deeper per-agent reasoning, as these strategies can interfere when combined.
arXiv:2605.16198v1 Announce Type: new
Abstract: We examine one particular dimension of AI governance: how to monitor and audit AI-enabled products and services throughout the AI development lifecycle, from pre-deployment testing to post-deployment auditing. Combining principles from formal methods with SoTA machine learning, we propose techniques that enable AI-enabled product and service developers, as well as third party AI developers and evaluators, to perform offline auditing and online (runtime) monitoring of product-specific (temporally extended) behavioral constraints such as safety constraints, norms, rules and regulations with respect to black-box advanced AI systems, notably LLMs. We further provide practical techniques for predictive monitoring, such as sampling-based methods, and we introduce intervening monitors that act at runtime to preempt and potentially mitigate predicted violations. Experimental results show that by exploiting the formal syntax and semantics of Linear Temporal Logic (LTL), our proposed auditing and monitoring techniques are superior to LLM baseline methods in detecting violations of temporally extended behavioral constraints; with our approach, even small-model labelers match or exceed frontier LLM judges. Our predictive and intervening monitors significantly reduce the violation rates of LLM-based agents while largely preserving task performance. We further show through controlled experiments that LLMs' temporal reasoning shows a pronounced degradation in accuracy with increasing event distance, number of constraints, and number of propositions.
arXiv:2503.00794v2 Announce Type: replace
Abstract: Accurate gait event detection is crucial for gait analysis, rehabilitation, and assistive technology, particularly in exoskeleton control, where precise identification of stance and swing phases is essential. This study evaluated the performance of seven kinematics-based methods and a Long Short-Term Memory (LSTM) model for detecting heel strike and toe-off events across 4363 gait cycles from 588 able-bodied subjects. The results indicated that while the Zeni et al. method achieved the highest accuracy among kinematics-based approaches, other methods exhibited systematic biases or required dataset-specific tuning. The LSTM model performed comparably to Zeni et al., providing a data-driven alternative without systematic bias. These findings highlight the potential of deep learning-based approaches for gait event detection while emphasizing the need for further validation in clinical populations and across diverse gait conditions. Future research will explore the generalizability of these methods in pathological populations, such as individuals with post-stroke conditions and knee osteoarthritis, as well as their robustness across varied gait conditions and data collection settings to enhance their applicability in rehabilitation and exoskeleton control.
arXiv:2605.15933v1 Announce Type: new
Abstract: We first show how the cohomology of some Bernstein-Gelfand-Gelfand (BGG) sequences that are important for the numerical analysis of partial differential equations, can be obtained through the construction of a long exact sequence connecting cohomology groups. Then we explain the extension of this result to the non-injective/surjective case through the systematic use of short exact sequences of complexes and their associated long exact sequences of cohomology groups. Finally an interpretation in terms of spectral sequences is given.
arXiv:2605.16197v1 Announce Type: new
Abstract: This position paper argues that safety and alignment cannot be achieved by constraining an external system: they must emerge from the co-regulatory design of the human--AI cognitive system as a whole ("AI as Part of Self"). Contemporary AI increasingly participates in attention allocation, reasoning, synthesis, and decision-making, shaping the very cognitive processes through which humans form beliefs, make decisions, and constitute their sense of self. Humans and AI occupy complementary epistemic roles under mutual constraint, forming a symbiotic cognitive unit whose co-regulation -- not the external control of either party alone -- is the proper locus of alignment. We identify the risks of unstructured delegation: deskilling, automation bias, transfer of epistemic authority, and oracle-style centralization of knowledge. Drawing on System~0 cognition theory, we further show that AI operates prior to conscious deliberation, shaping the pre-attentive infrastructures through which agency and trust are negotiated -- a level that conventional oversight cannot reach. We conclude with design principles for cognitive co-regulation addressed to ML engineers and governance bodies. The goal of this work is to guide human cognition toward resilience and epistemic agency at the foundation of human selfhood.
arXiv:2605.15925v1 Announce Type: new
Abstract: Let $\mathbb{F}_{p^m}$ be the field containing $p^m$ elements where $p$ is an odd prime and $m \in \mathbb{N}$.
In this article, we propose a unified approach to the study of skew constacyclic codes of length $np^s$ over the ring $R_k = \mathbb{F}_{p^m}[u]/\langle u^k \rangle,$ where $n, s, k \in \mathbb{N}$ and $\gcd(n, p)=1$.
Consider the skew polynomial ring $R_k[x;\Theta]$, where
$\Theta$ is an automorphism of $R_k$ such that $xa = \Theta(a)x$ for all $a \in R_k$. Let $f(x)$ be a central irreducible divisor of $x^{np^s} - \lambda$ of degree $l$ and multiplicity $j$ in $R_k[x;\Theta]$, where $\lambda $ is an invertible element in $R_k$. In this article, we study skew constacyclic codes of length \(np^s\) over \(R_k\), which reduces to the study of skew polycyclic codes of length $jl$ associated with a polynomial \(f(x)^j\).
Using the fact that skew polycyclic codes associated with a polynomial \(f(x)^j\) can be described by the left ideal structure of the quotient ring $R_k[x;\Theta]/\langle f(x)^{j}\rangle$, we investigate this class of codes for specific choices of $\Theta$. In particular, if $\lambda$ is an invertible element of $\mathbb{F}_{p^m}$, we classify all left ideals and establish an isomorphism between skew cyclic and skew constacyclic codes, under suitable conditions.
Furthermore, we provide a comprehensive analysis of skew constacyclic codes of length $3p^s$ over $R_k$. Finally, we examine skew cyclic and skew negacyclic codes of length $6p^s$ over $R_k$ using the factorization of $x^{6p^s} - 1$ and $x^{6p^s} + 1$, respectively; with a complete case-by-case analysis. Examples demonstrating codes with optimal parameters are also included.
arXiv:2605.16196v1 Announce Type: new
Abstract: In this paper, we investigate a bistatic multiple-input multiple-output (MIMO) integrated sensing and communication (ISAC) system over block-fading channels, focusing on the scenario where the sensing and communication receivers (Rxs) are co-located. Under the assumption of unknown channel state information (CSI) at the Rx, two schemes are considered: pilot sensing (PS) and data-aided sensing (DAS). The communication rate-sensing distortion functions for both schemes are characterized. For the DAS scheme, a closed-form asymptotic expression for the sensing distortion is derived by using random matrix theory (RMT). The asymptotic performance analysis explicitly quantifies the significant gains of the DAS scheme, revealing a strict $3$ dB effective SNR improvement in the low-SNR regime and a strictly faster performance scaling rate in the high-SNR limit compared to the PS scheme.
arXiv:2605.15924v1 Announce Type: new
Abstract: In virtual reality environments, the alignment of perceptual modalities is crucial for immersion and presence. In the AR domain, it is difficult to create such alignments because elements in the physical world are often beyond the user's control. However, recent advances in generative AI enable on-demand content creation, enabling highly reactive AR experiences. Combined with contextual information about the physical world, it has become possible to design experiences that seamlessly align with the user's environment. In this reflection paper, I emphasize the importance of "synchronized" realities for context-aware AR experiences, particularly in mobility scenarios. I present several examples of existing synchronized experiences and examine their commonalities and distinctions. Finally, I discuss opportunities and pitfalls of synchronizing AR experiences with the physical world.
arXiv:2605.16194v1 Announce Type: new
Abstract: LLM agents routinely serve as first (and sometimes only) readers of academic papers, skimming for sub-claims, extracting reproducibility steps, and generalizing scope. Standard prose papers produce recurring failures in this role: sub-claims that cannot be cited at sub-paper granularity, scope overextension beyond what the paper tests, and figure commands buried in codebases rather than the paper itself. We propose `paper.json`, a companion JSON file that travels with the PDF and addresses each failure with a lightweight convention: stable claim IDs (C1), an explicit does-not-claim list (C2), exact per-figure shell commands (C3), and stable definition IDs (C5). A fifth convention (C4) holds that minimum viable compliance, hand-written JSON alongside the PDF, is achievable in under an hour for a finished paper without touching the human-readable output. C1, C2, C3, and C5 are open invitations: an agent that reads a compliant paper and acts on it produces evidence for or against them. This paper is itself compliant: `uv run validator.py paper.json --against paper.typ` passes. Repo: https://github.com/arquicanedo/paper-json
arXiv:2605.15671v1 Announce Type: cross
Abstract: Multimodal 3D MRI brain tumor segmentation is a pivotal step in radiotherapy target delineation, surgical planning and post-treatment assessment. Existing methods often assume artifact-free MRI images. However, inevitable patient motion during scanning introduces artifacts and blur that degrade boundary and texture features, leading to poor segmentation performance. To bridge this gap, we introduce Degradation-Aware Blur-Segmentation Net (DABSeg), a synchronous deblurring 3D multimodal MRI segmentation network that unifies blur removal and accurate segmentation. Specifically, we propose a feature-domain motion-deblurring stem to compensate for blur and rebalance intensity. Concurrently, the backbone network embeds a blur-aware cross-modal cross-attention module and multi-scale residual aggregation to yield effective modality complementarity. Notably, we optimize a joint loss that combines weighted Dice with a clear-reference reconstruction term, where imbalanced weights are applied to small targets to boost learning intensity and predictive stability for small lesions and border regions. Systematic comparisons and ablation experiments on the BraTS2020 dataset under both clear and degenerative conditions consistently demonstrate that DABSeg surpasses state-of-the-art methods in tumor Dice score and boundary precision. These results validate the effectiveness of degenerative-aware cross-task collaborative learning in improving the robustness and clinical utility of multi-modal 3D brain tumor segmentation under realistic degenerative conditions. The source code is available at https://github.com/YuchunWang24/DABSeg_ICPR
arXiv:2605.16101v1 Announce Type: cross
Abstract: Measurement is an important scientific activity. In most of science, including classical physics, is may be understood as a way of finding out about the physical world and representing the results numerically. No-go theorems show that measurement of quantum observables is not like that: the recorded outcome is typically created rather than revealed in a quantum measurement, in which case there is no objective fact about the observable's prior value. Other no-go theorems show that unitary quantum theory can generally neither explain nor even represent a unique recorded outcome, thereby threatening that outcome's objectivity. Methodological norms inherent in quantum physical practice nevertheless institute the objectivity, not only of unique recorded outcomes of quantum measurements, but also of non-quantum features of the world that physicists and other scientists take their models to represent.
arXiv:2605.16193v1 Announce Type: new
Abstract: Large language models (LLMs) are increasingly used to simulate human opinions and survey responses, but their ability to reproduce population responses across cultures remains limited. Existing persona-based prompting methods typically rely on sociodemographic or personality traits, which are only indirect proxies for the values that shape human responses. We propose a value-based persona construction method that derives textual descriptors from survey responses capturing core cultural dimensions. By sampling value profiles from target populations and aggregating LLM responses across personas, we obtain population-level predictions grounded in observed value distributions. We further introduce a calibration procedure that improves response diversity while preserving estimated opinions. We show that our approach reduces prediction error across countries, with the largest improvements observed in underrepresented populations. This substantially narrows the performance gap between countries aligned with dominant LLM priors and those that are less represented in training data, while also yielding response distributions that closely match human diversity.
arXiv:2605.15923v1 Announce Type: new
Abstract: Modern image encoders achieve high generalization by decoupling semantic meaning from resolution, an ability yet to be fully realized in the 3D domain. We investigate the failure of 3D point cloud encoders to achieve similar generalization and find that existing models are highly sensitive to sampling resolution and scale changes, leading to significant performance degradation. This sensitivity is a major bottleneck for real-world deployment in robotics, as it suggests models overfit to specific quantization densities and object scales rather than learning invariant semantic features. To mitigate this dependency, we propose Invaria, a point cloud encoder that achieves scale and density invariance through next-resolution prediction and receptive field calibration. While our objective is not the explicit generation of high-resolution point clouds, we find that this training objective encourages the model to learn robust, structural invariants. The resulting encoder achieves significant performance gains during resolution shifts while maintaining high efficiency through a compact model size and reduced token requirements. Specifically, on ScanNet, Invaria achieves a 56.0\% higher mIoU at 3$\times$ lower resolution and a 20\% improvement when the objects scale is reduced by a factor of 3. These gains are achieved with a 45\% smaller model size and an average reduction of 40\% in input tokens.
arXiv:2605.16067v1 Announce Type: new
Abstract: We propose a variational quantum classifier operating on high dimensional deep representations via amplitude encoding, stabilized by a learnable classical pre encoding layer.By combining normalized amplitude embeddings with bounded quantum observables, the resulting model induces a structured and smooth hypothesis class with controlled sensitivity to input variations. Model reliability is assessed using SAFE-AI metrics derived from the Cramer von Mises divergence, enabling consistent evaluation across accuracy, robustness, and explainability dimensions. Empirical results show that the proposed quantum model provides competitive predictive performance compared with strong classical baselines while exhibiting a more balanced SAFE reliability profile, with improved robustness to noise and stability under structured feature removal. These findings suggest that variational quantum circuits offer a principled mechanism for stability oriented SAFE learning in safety critical settings.
arXiv:2605.15921v1 Announce Type: new
Abstract: Object removal aims to eliminate specified objects from images while plausibly inpainting the affected regions with background content. Current training-free methods typically block attention to object regions within self-attention layers during the image generation process, leveraging surrounding background information to restore the image. However, indiscriminate suppression of self-attention in the vacated areas can degrade generation quality, as the model must simultaneously reconstruct background content in these regions. To solve this conflict, we propose AdaEraser, an adaptive framework that dynamically modulates attention based on the estimated presence of target object concepts. Through analysis of self-attention map evolution across denoising timesteps before and during removal, we develop a token-wise adaptive attention suppression strategy. This approach enables progressive perception of object removal throughout the denoising process, with the suppression strength in self-attention layers adjusted adaptively. Extensive experiments demonstrate that AdaEraser achieves superior performance in object removal, outperforming even training-based methods.
arXiv:2605.16065v1 Announce Type: new
Abstract: 3D Gaussian Splatting (3D-GS) enables real-time 3D scene reconstruction but lacks robust segmentation for editing tasks such as object removal, extraction, and recoloring. Existing approaches that lift 2D segmentations to the 3D domain suffer from view inconsistencies and coarse masks. In this paper, we propose a novel framework that leverages the Segment Anything Model High Quality (SAM-HQ) to generate accurate 2D masks, addressing the limitations of the standard SAM in boundary fidelity and fine-structure preservation. To achieve robust 3D segmentation of any target object in a given scene, we introduce a prior-guided label reassignment method that assigns labels to 3D Gaussians by enforcing multiview consistency with learned priors. Our approach achieves state-of-the-art segmentation accuracy and enables interactive, real-time object editing while maintaining high visual fidelity. Qualitative results demonstrate superior boundary preservation and practical utility in Virtual Reality (VR) and robotics, advancing 3D scene editing.
arXiv:2605.15719v1 Announce Type: new
Abstract: Wearable devices are widely used for continuous health monitoring, yet reliable sleep tracking on emerging platforms remains underexplored due to reliance on proprietary algorithms and device-specific activity representations. We present a lightweight and reproducible sleep tracking pipeline that operates directly on raw accelerometer signals. The method converts data into epoch-level activity features, applies temporal smoothing and normalized scoring, and performs sleep/wake classification using a globally calibrated threshold. We calibrate the model on the Multilevel Monitoring of Activity and Sleep in Healthy People (MMASH) dataset and evaluate it in a cross-device study using the WeBe wearable platform and a commercial ActiGraph device. On MMASH, the method achieves a mean absolute error of 41.6 minutes in Total Sleep Time (TST), with onset and offset errors of 6.3 and 7.4 minutes. On real-world WeBe data from three participants across five sessions, it achieves a mean TST error of 27.4 minutes and onset and offset errors of 13.9 and 8.0 minutes. In contrast, a commercial ActiGraph pipeline shows larger discrepancies relative to ground truth. These results demonstrate accurate and generalizable sleep tracking using a simple and reproducible pipeline.
arXiv:2605.15916v1 Announce Type: new
Abstract: Parameter-efficient fine-tuning (PEFT) has emerged as an critical technique for adapting large-scale foundation models across natural language processing and computer vision. While existing methods such as low-rank adaptations achieve parameter efficiency via low-rank weight updates, they are limited in their ability to preserve the geometric structure of pretrained representations. We introduce Low-rank Compositional Orthogonal fine-tuning (LoCO), a novel PEFT method that constructs orthogonal transformations through low-rank skew-symmetric matrices and compositional rotation chains. We propose an approximation scheme that enables fully parallel computation of compositional rotations, making the approach practical for high-dimensional feature spaces. Our method maintains low computational complexity while maintaining orthogonality with controlled approximation error. We validate LoCO across diverse domains, including diffusion transformer fine-tuning, vision transformer adaptation, and language model adaptation. Our method demonstrates superior or competitive performance compared to both existing orthogonal and non-orthogonal methods.
arXiv:2503.02597v3 Announce Type: replace
Abstract: Recent Multimodal Large Language Models (MLLMs) have demonstrated significant progress in perceiving and reasoning over multimodal inquiries, ushering in a new research era for foundation models. However, vision-language misalignment in MLLMs has emerged as a critical challenge, where the textual responses generated by these models are not factually aligned with the given text-image inputs. Existing efforts to address vision-language misalignment have focused on developing specialized vision-language connectors or leveraging visual instruction tuning from diverse domains. In this paper, we tackle this issue from a fundamental yet unexplored perspective by revisiting the core architecture of MLLMs. Most MLLMs are typically built on decoder-only LLMs consisting of a causal attention mechanism, which limits the ability of the earlier modalities (e.g., images) to incorporate information from the latter modalities (e.g., text). To address this problem a MLLM that unlocks causal attention into our proposed modality-mutual attention (MMA) to enable image tokens to attend to text tokens. This simple yet effective design allows MMA to achieve state-of-the-art performance in 12 multimodal understanding benchmarks (+6.2% on average across 3 LLMs backbones) without introducing additional parameters. Our MMA design is intended to be generic, allowing for applications across various modalities, and scalable to accommodate diverse multimodal scenarios.
arXiv:2312.05975v3 Announce Type: replace
Abstract: Explainability is a vital aspect of modern AI for real-world impact and usability. The main objective of this paper is to emphasise the need to understand the predictions of Computer Vision models, specifically Convolutional Neural Network (CNN) models. Existing methods for explaining CNN predictions are largely based on Gradient-weighted Class Activation Maps (Grad-CAM) and focus solely on a single target class; this assumption about the target class selection neglects a large portion of the predictor CNN's prediction process. In this paper, we present an exhaustive methodology, called Fused Multi-class Gradient-weighted Class Activation Map (FM-G-CAM), that considers multiple top-predicted classes and provides a holistic explanation of the predictor CNN's rationale. We also provide a detailed mathematical and algorithmic description of our method. Furthermore, alongside a concise comparison of existing methods, we compare FM-G-CAM with Grad-CAM, quantitatively and qualitatively highlighting its benefits through real-world practical use cases. Finally, we present an open-source Python library with an FM-G-CAM implementation to conveniently generate saliency maps for CNN-based model predictions.
arXiv:2605.16179v1 Announce Type: new
Abstract: Agricultural landscape segmentation in the Global South is challenging as it is characterized by fragmented plots, high intra-class variance, and a scarcity of labeled training data. Recent advances in segmentation have been made by Multimodal Large Language Models (MLLMs). However, current approaches encounter critical context length bottlenecks and a domain alignment gap in understanding satellite features. We address these limitations through MAgSeg, a novel, decoder-free MLLM segmentation approach. MAgSeg is an architecturally efficient approach that enables standard MLLMs to perform segmentation of complex smallholder agricultural landscapes from high-resolution satellite imagery, without requiring auxiliary vision decoders. We introduce a novel instruction tuning data format designed to enable scalable fine-tuning and post-training on high resolution satellite imagery, which enables MAgSeg to learn from the global context of the image while generating text tokens for only a patch within the image. Extensive evaluations on datasets spanning three countries in the Global South demonstrate that MAgSeg significantly outperforms state-of-the-art MLLM baselines, offering a scalable solution to map smallholder agricultural environments.
arXiv:2605.16175v1 Announce Type: new
Abstract: Pediatric critical care is a dynamic, high-stakes process involving constant monitoring and adjustments in life-saving treatments. Modeling these interventions is crucial for effective decision support. To address the challenges of high complexity and data scarcity in pediatric Extracorporeal Membrane Oxygenation (ECMO), we frame clinical decision-making as learning to act from trajectories, i.e., imitation learning that learns action models from observational data, with a key feature that actions are not directly observed. We consider TabPFN, a recent transformer-based approach for tabular data, and traditional baselines including XGBoost and Multi-Layer Perceptrons(MLPs) on real-world pediatric ECMO data to learn the action models. We find that the TabPFN-based approach consistently outperforms these classical baselines, supporting its use as a strong clinician-behavior baseline for pediatric ECMO decision support.
arXiv:2503.07518v2 Announce Type: replace
Abstract: Large Language Models (LLMs) rely on the Key-Value (KV) Cache to store token history, enabling efficient decoding of tokens. As the KV-Cache grows, it becomes a major memory and computation bottleneck. However, there is an opportunity to alleviate this bottleneck, prior research has shown that only a small subset of tokens contribute meaningfully to each decoding step. A key challenge in finding these critical tokens is that they are dynamic, and heavily input query-dependent. Existing methods either risk quality by evicting tokens permanently, or retain the full KV-Cache but rely on retrieving chunks of tokens and many existing KV-Cache sparsity methods rely on inaccurate proxies for token importance. To address these limitations, we introduce TokenButler, a high-granularity, query-aware predictor that learns to identify these critical tokens. TokenButler predicts low-dimensional importance queries at a fixed depth stride, and combines them with a learned projection of the real KV-cache keys to score tokens cheaply, enabling dynamic per-token selection under a fixed budget while preserving the full KV cache. We train TokenButler by distilling the model's masked causal attention distributions, optimizing a lightweight predictor with minimal parameter overhead. We evaluate TokenButler on a novel synthetic small-context co-referential retrieval task, demonstrating near-oracle accuracy where existing methods fail. Furthermore, TokenButler achieves competitive or superior performance on long-context benchmarks (RULER, LongBench), up to $\approx1.6\times$ on-GPU speedup using our proposed *prediction interval with neighbor fetching* that amortizes predictor cost while maintaining accuracy within $\approx$1.1\%, and up to 7.6$\times$ reduction in latency compared to Dense Attention with CPU offloading. Code is available: https://github.com/abdelfattah-lab/TokenButler