Your daily OSS contribution digest

This paper presents a method to train robotic multi-fingered hands to manipulate objects in-hand with high dexterity and safety, achieving robust real-world performance without human data and enabling complex tasks like spinning a lightbulb using only a wrist-mounted camera.

• EmbeddingGemma 2: Unified and Efficient Multimodal Representations

  EmbeddingGemma 2 is a versatile and efficient model that creates unified embeddings for multiple data types like text, images, audio, and video, outperforming larger specialized models while being adaptable to limited computing resources.

  Ai Ml · relevance 9/10
• Scaling Laws for Tool-Use Agents via Controllable Environment Synthesis

  This paper presents a method to generate synthetic training environments with controlled diversity that helps large language model agents better learn to use tools for complex tasks, and it derives scaling laws that guide how to optimize training data composition for improved generalization.

  Ai Ml · relevance 9/10

• Approximate Counting for Mixed $p$-Spin Glasses in Weak External Fields via Cluster Expansion

  This paper presents efficient algorithms to approximate the partition function of complex spin glass models under weak external fields, achieving polynomial-time accuracy guarantees using advanced combinatorial and probabilistic techniques.

  Algorithms Theory · relevance 7/10
• Constructive Arithmetic Regularity for Abelian Groups

  This paper presents the first efficient algorithm to construct the arithmetic regularity decomposition for subsets of finite Abelian groups beyond the binary vector space setting, enabling practical applications in additive combinatorics and related computational problems.

  Algorithms Theory · relevance 7/10
• Iterative Randomized Rounding for Better-Than-2 k-Median Approximation

  This paper improves the approximation ratio for the k-median clustering problem to just under 2 by refining iterative rounding techniques and allowing a few extra facilities, partially resolving a long-standing open question.

  Algorithms Theory · relevance 7/10

• Fast Approximate Line Nearest Neighbor Search

  This paper introduces a new method to efficiently find approximate nearest neighbors along a line query, improving search accuracy and speed for applications where variation along a certain direction is allowed, such as image or video synthesis.

  Database · relevance 8/10
• GraphRAGProv: GraphRAG Meets How-Provenance

  This paper introduces a system that not only answers questions using knowledge graphs but also tracks and explains exactly which pieces of source text contributed to each answer, improving transparency and trust.

  Database · relevance 7/10
• Constraining the Commit Boundary: A DIO/ODI-Informed Authorization Guardrail for Agentic Entity Resolution

  This paper proposes a framework to control when and how identity changes are authorized in automated entity resolution systems, aiming to reduce errors by enforcing strict evidence and policy checks before committing identity updates.

  Database · relevance 5/10

• PipeLink: Contract-Guided Linking of GPU Kernel Components

  This paper introduces PipeLink, a tool that enables separately compiled GPU kernel components to be linked together efficiently, achieving near-native performance and enabling flexible, reusable kernel compositions across different GPU architectures.

  Distributed Systems · relevance 9/10
• WorldPager: Zero-Stall Tiered Memory for Long-Horizon World Models on a Single Consumer GPU

  This paper presents WorldPager, a system that enables long-duration interactive world models on a single consumer GPU by efficiently managing memory across GPU, DRAM, and NVMe storage to avoid stalls and maintain consistent performance.

  Distributed Systems · relevance 8/10
• Request-Adaptive Execution for Efficient Hybrid LLM Inference

  This paper introduces Morph, a system that dynamically adapts the execution of hybrid large language models per request to improve inference speed and efficiency without sacrificing output quality.

  Distributed Systems · relevance 8/10

• AMOR: Adaptive Modal Operator Refinement for Operator Learning on Arbitrary Meshes and Irregular 3D Geometries

  This paper introduces AMOR, a new operator learning method that works efficiently on irregular 3D meshes and complex geometries, outperforming existing Fourier-based methods in many cases and scaling better to higher dimensions.

  Math · relevance 8/10
• Quantum circuits of elastic wave equations via Schrödingerisation

  This paper develops explicit quantum circuits to simulate elastic wave equations efficiently by transforming them into Schrödinger-type equations, potentially enabling faster and more accurate modeling of wave propagation in materials.

  Math · relevance 8/10
• Multifidelity variational inference for derivative-free PDE-constrained inverse problems

  This paper introduces a new method to efficiently solve complex inverse problems constrained by PDEs without needing derivative information, by combining high- and low-fidelity models to reduce computational cost and variance in the estimation process.

  Math · relevance 8/10

• Keeping Inference in the Radio: Design and Measurement of a Spiking Accelerator for mmWave ISAC Activity Recognition

  This paper demonstrates a low-power spiking neural network accelerator for recognizing human activity from mmWave radio signals, achieving high accuracy and energy efficiency suitable for always-on sensing in wireless devices.

  Networking · relevance 8/10

• SecDMP: Provably Secure Data Memory-Dependent Prefetching (Extended Version)

  This paper presents SecDMP, a new hardware defense that prevents secret data leaks caused by data memory-dependent prefetchers in CPUs, achieving strong security guarantees with minimal performance impact.

  Security · relevance 9/10
• Penny for Your Prompts: A Measurement Study of Contextual Leakage in AI Chat Platforms

  This paper reveals that many AI chatbot platforms unintentionally share users' private conversation data and personal information with third-party trackers, raising significant privacy concerns.

  Security · relevance 8/10
• Revisiting Privacy Risks in Quantized Deep Neural Networks

  This paper studies how quantizing deep neural networks affects their vulnerability to privacy attacks that try to reveal if specific data was used in training, finding that typical quantization offers little privacy benefit and existing defenses are limited.

  Security · relevance 8/10

• SLICE-CiM: Selective Layer-wise Inference Control Engine with Confidence-Based Early Exit for SRAM based Digital Compute-in-Memory

  This paper introduces a set of hardware control modules that improve energy efficiency and speed for SRAM-based compute-in-memory CNN inference by selectively controlling precision and enabling early exit based on confidence, achieving significant power savings without retraining.

  Systems Architecture · relevance 8/10
• Kernel Performance Engineering under Physical Non-Idealities

  This paper reveals how physical hardware limitations affect GPU kernel performance and proposes a new clustering method that improves throughput by better managing these constraints, with benefits also shown on other accelerators.

  Systems Architecture · relevance 8/10