Long-context LLM inference is bottlenecked by ever-growing KV cache memory during decoding. HiKV introduces an algorithm-hardware co-design that compresses the cache hierarchically: first evicting unimportant tokens within a budget, then loading only significant elements of retained tokens, achieving compression beyond single-granularity methods. A dedicated accelerator with a reconfigurable importance sorter unifies both stages in one circuit. Evaluated on representative LLMs, HiKV delivers up to 7.95x speedup and 90% energy reduction with negligible accuracy loss. Applications include efficient long-context LLM serving infrastructure, edge and datacenter inference acceleration, and hardware-software co-design for scalable AI deployment.

Authors: Chao Fang, Jun Yin, Man Shi, Marian Verhelst

Paper: https://arxiv.org/abs/2607.22389v1

Vision-language model agents increasingly need to act on 3D scenes, not just describe them, yet existing benchmarks mostly score text responses or single-object tasks. SceneActBench introduces a unified agent-environment loop evaluating five 3D action tasks across 210 source instances and 520 task cases, using images, video frames, and 3D assets, scored against hidden ground truth with geometric metrics. Testing eleven proprietary VLM configurations revealed inconsistent performance (38.6-50.2 overall) with no model excelling across all tasks. Applications include benchmarking and improving embodied AI, robotics manipulation planning, and 3D-scene-grounded agent development for real-world spatial reasoning tasks.

Authors: Yifei Zhao, Xiangxin Zhou, Wenhao Yang, Jiaqi Tang, Pu Jian, Huanjin Yao, Jiarui Yao, Haowei Lin, Chunchao Guo, Zhuo Chen, Wenkai Lyu, Jianzhu Ma, Xueqian Wang, Wenxi Zhu

Paper: https://arxiv.org/abs/2607.22393v1

As LLM agents move into autonomous, tool-executing roles, reliability suffers from lack of ground truth and operational drift in open-ended environments. This framework introduces a self-calibration mechanism using an ARIMA forecaster to approximate ground truth without continuous human oversight, applied to profiling zero-knowledge workload resource usage in edge computing networks. It improved resource-prediction accuracy by 91.7% and prediction speed by 71.7% over baseline agents, while a novel ARIMA leaping algorithm ran 52% faster than standard ARIMA. Applications include decentralized infrastructure management, edge computing resource allocation, and building drift-resistant autonomous AI systems for infrastructure operations.

Authors: Fin Gentzen, Marla Grunewald, Iulisloi Zacarias, Mounir Bensalem, Admela Jukan

Paper: https://arxiv.org/abs/2607.22400v1

On-device fluid identification for microfluidic applications is challenging under varying flow, pressure, and temperature, and existing learning methods ignore underlying physics. PRIMS addresses this with a physics-aware multimodal Transformer combining physics-based sensor token vectorization, a viscosity-aware component synthesizer, and physics-guided attention fusion, embedding fluid mechanics directly into the architecture. On a five-fluid benchmark, PRIMS achieved 98.92% F1-score with just 0.46 million parameters, a 14x reduction versus prior Transformers, and stayed robust under unseen conditions. Applications include lab-on-chip diagnostics, industrial fluid monitoring, and lightweight, interpretable edge-deployed sensing systems for microfluidics.

Authors: Hai-Long Nguyen, Trung Thanh Nguyen, Lars Holm, Dennis Alveringh, Duc Viet Le

Paper: https://arxiv.org/abs/2607.22422v1

Large text-to-image diffusion models are usually opaque black boxes, limiting artists' ability to creatively manipulate them. This work reframes explainable AI for creative practice, proposing hands-on "model bending" through an interactive inspection interface built into ComfyUI, enabling layer selection and intervention. Applied to Stable Diffusion 1.5, the authors show that manipulating specific pipeline components produces consistent, learnable families of visual effects. Applications include generative art tools, creative AI education, artist-facing interpretability interfaces, and broader interaction design research into making large opaque models tangible, inspectable materials for hands-on artistic experimentation and intervention.

Authors: Ahmed M. Abuzuraiq, Philippe Pasquier

Paper: https://arxiv.org/abs/2607.22428v1

Robots typically communicate only through physical movement, unlike humans who also use shadows and silhouettes. This work builds a robotic system with a 21-degree-of-freedom soft-skinned hand and a learned shadow self-model that maps hand configurations to projected shadow appearance. Given a target shadow image or video, the robot optimizes hand poses through gradient-based search and collision-aware simulation to produce physically feasible, expressive shadows, demonstrated on sign language, shadow puppetry, and animal imitation. Applications include expressive human-robot interaction, assistive and educational robotics, entertainment and storytelling robotics, and broader research into non-morphological robot communication channels.

Authors: Danyang Yan, Boyuan Wang, Jiaxun Liu, Boyuan Chen

Paper: https://arxiv.org/abs/2607.22434v1

Hyperball-style optimizers, which normalize updates for scale-invariant networks, have shown strong large-scale training performance, but why remains unclear. This work derives an angular effective learning rate accounting for update angle, parameter norm, and update norm, then decomposes updates into radial and tangential components to explain optimizer behavior differences like why MuonH lags early but overtakes MuonWD later. Experiments reveal the difference stems from effective step-size evolution rather than superior update direction, and that careful learning-rate scheduling remains essential. Applications include informing optimizer design and training schedule choices for large-scale deep learning and foundation model pretraining.

Authors: Yihao Xiao, Jialong Sun, Zitian Gao, Zeming Wei, Chutian Wang, Ran Tao, Jiaye Teng, Bryan Dai

Paper: https://arxiv.org/abs/2607.22444v1

Enterprise AI agents often hold static, overly broad credentials for every possible task, expanding security risk. This work proposes dynamic least-privilege scoping via a three-source architecture combining role-based ceilings, a task-context classifier, and policy-derived prohibitions, ensuring agents only access credentials relevant to their current task. The authors release a validated synthetic dataset of 600 enterprise prompts labeled with minimum required permissions, achieving high human-agreement scores and reducing policy violations by 93% through iteration. Applications include enterprise AI security architecture, agent permission management systems, and providing behavioral signals for researching and mitigating LLM agent misuse and misalignment.

Authors: Halil Burak Noyan

Paper: https://arxiv.org/abs/2607.22445v1

This study asks whether general-purpose audio embeddings implicitly encode evolutionary relationships never explicitly trained into them. Testing four pretrained audio models on marine mammal and bird vocalizations, the authors found strong correlations between embedding-space distances and phylogenetic distance, particularly among cetaceans, robust even after controlling for frequency and dimensionality. Surprisingly, domain-specific bioacoustic models didn't outperform general-purpose ones on birds. Applications include bioacoustic monitoring, biodiversity and conservation research, taxonomic classification of species from sound, and broader insight into what unsupervised audio representations capture, informing model selection for ecological and evolutionary biology research.

Authors: Víctor Rincón Yepes

Paper: https://arxiv.org/abs/2607.22458v1

As generative AI reshapes programming education, instructors often infer student learning purely from classroom observation, missing crucial context. This study uses trio-ethnography, pairing two educators with differing teaching philosophies alongside an undergraduate student, across three dialogic sessions exploring AI-assisted learning. The student's firsthand narratives revealed learning processes invisible to instructors, prompting both educators to revise assumptions about assessment, transparency, and pedagogy. Applications include informing computing curriculum design, faculty development around AI-integrated teaching, and demonstrating collaborative reflective methodologies that other educational researchers could adopt to better understand student AI use beyond surface-level classroom behavior.

Authors: Jennie Ren, Jordan H. McDowell, Kyrie Zhixuan Zhou

Paper: https://arxiv.org/abs/2607.22463v1