Your daily OSS contribution digest

This paper introduces a model that simultaneously generates realistic video and corresponding 3D motion for any rigged character, enabling animation replay in 3D scenes with camera alignment and control.

• GPT-6-Astra Understands Touch and Outperform Vision--Tactile--Language--Action Models

  This paper introduces GPT-6-Astra, a general-purpose multimodal model that uses tactile feedback to significantly improve robot manipulation tasks without needing specialized retraining, achieving near state-of-the-art success rates.

  Ai Ml · relevance 9/10
• Racer: Hunting GPU Kernel Bugs with Symbolic Equivalence Checking

  This paper introduces Racer, a tool that uses symbolic equivalence checking to precisely verify GPU kernel correctness across all inputs, uncovering subtle bugs missed by traditional testing.

  Ai Ml · relevance 9/10

• Exact Envy-Free Cake Cutting with Polynomial Expected Queries

  This paper presents a randomized algorithm that can fairly divide a resource among multiple agents without envy, using only a polynomial expected number of queries about their preferences.

  Algorithms Theory · relevance 8/10
• Dynamic Distances Without Matrix Inverse

  This paper presents a new, faster algorithm for dynamically maintaining exact shortest distances in changing directed graphs by using a novel algebraic approach based on the Frobenius normal form, avoiding previous limitations tied to matrix inversion.

  Algorithms Theory · relevance 8/10
• Birthday Estimators Are Uniformly Optimal for Renyi Entropy

  This paper proves that a simple method based on counting collisions is the best possible way to estimate Rényi entropy across all settings, providing tight guarantees on how many samples are needed.

  Algorithms Theory · relevance 7/10

• What is a Good Projection?: Correctness and Separation Results of End-Point Projections

  This paper defines general criteria for projecting global communication protocols into local types that guarantee correctness and deadlock-freedom, analyzes existing algorithms against these criteria, and provides new modular implementations to help developers choose reliable projection methods.

  Compiler · relevance 7/10
• ExeBench-FHD: Addressing Fidelity Issues in Decompiler Evaluation

  This paper identifies key issues in a popular benchmark for evaluating AI-based decompilers and introduces an improved version that more accurately measures decompiler performance, leading to better evaluation results.

  Compiler · relevance 7/10
• A General Construction of Strict Models in HoTT

  The paper develops a new method within homotopy type theory to convert approximate models of type theory into fully strict models, enabling more reliable semantic interpretations.

  Compiler · relevance 4/10

• NousDB - A Fast Query Engine for Semantic Analytics

  NousDB is a new fast query engine that integrates AI-powered semantic operators into data queries, achieving large speedups and better optimization compared to existing systems.

  Database · relevance 9/10
• Local Inference of LLMs on Long Contexts: A Data Systems Perspective

  This paper analyzes why running large language models with long context locally on single GPUs is inefficient, revealing that self-attention underutilizes GPU resources and identifying data management bottlenecks that limit performance.

  Database · relevance 8/10
• SWatcher+: A Holistic Approach to Deploying Learned Cardinality Estimators in a Dynamic Setting

  This paper presents SWatcher+, a system that improves database query optimization by dynamically adapting learned cardinality estimators to changing data and workloads without costly retraining.

  Database · relevance 8/10

• BitMTP: When Multi-Token Prediction Meets Low-Bit Large Language Models

  This paper introduces BitMTP, a new method that combines ultra-low-bit large language models with multi-token prediction to significantly speed up text generation without losing quality, enabling faster and more efficient deployment on various GPUs.

  Distributed Systems · relevance 9/10
• The Token-Energy Law of LLM Decode: Active Bytes and Realized Time Account for Joules per Token

  This paper presents a simple, accurate model that predicts the GPU energy cost per token generated by large language models, helping optimize energy efficiency by relating it to memory usage and decoding time.

  Distributed Systems · relevance 8/10
• Discount Yourself: Self-inclusive Freshness-weighted Gossip for Decentralised Learning under Client Outages

  This paper introduces a new method to improve decentralized federated learning when devices go offline and come back, by better handling outdated model information including the device's own stale model, leading to more accurate and stable learning.

  Distributed Systems · relevance 8/10

• Fast and scalable multi-GPU sliced FFT solvers for asymptotically compatible discretizations of 3D nonlocal diffusion problems

  This paper presents a new multi-GPU algorithm that significantly speeds up large-scale 3D nonlocal diffusion simulations by combining sliced FFT techniques with efficient GPU communication, enabling simulations with billions of variables.

  Math · relevance 8/10
• Near Optimal Finite Model POMDP Approximation of Controlled Diffusion with Partial Information

  The paper develops a method to approximate complex continuous-time control problems with partial information by simpler discrete models, proving that solutions to these approximations are nearly optimal for the original problem.

  Math · relevance 7/10
• First-Order Mean Field Games with Time-Measurable and $C ^{1,1}$ Hamiltonians: Analysis and Approximation

  This paper develops and analyzes a numerical method to solve certain complex Mean Field Game models with irregular time data and less smooth Hamiltonians, proving convergence and uniqueness under some conditions.

  Math · relevance 7/10

• ORCA: Onboard Reflective Collaborative Agents for Reliable Task Offloading in LEO Satellite Networks

  This paper introduces ORCA, a decentralized agent-based system that enables LEO satellites to reliably offload computing tasks onboard despite changing network conditions and malicious nodes, improving efficiency and robustness without central coordination.

  Networking · relevance 8/10
• Entanglement Distribution in Quantum Switch Networks: Memory Limits and Achievability

  This paper identifies the minimal quantum memory needed at a central switch to distribute entangled states among multiple users and presents an efficient protocol that significantly reduces memory usage compared to previous methods.

  Networking · relevance 7/10
• Agentic Communication Networks: When Agents Become Network Entities

  This paper proposes a new networking framework that treats autonomous agents as fundamental network entities, enabling more effective coordination, identity management, and task execution across distributed systems.

  Networking · relevance 7/10

• Context-Aware Prompt Injection Detection via Small Language Models in RAG Pipelines: nokast-secureRAG

  This paper introduces a small language model that acts as a semantic filter to detect malicious prompt injections in retrieval-augmented generation systems, significantly improving detection accuracy while maintaining practical latency.

  Security · relevance 8/10
• FedGraph: Defending Federated Large Language Model Fine-Tuning Against Backdoor Attacks via Graph-Based Aggregation

  This paper introduces FedGraph, a new method that uses graph-based analysis to detect and block malicious updates during federated fine-tuning of large language models, improving security against backdoor attacks while preserving model performance.

  Security · relevance 8/10
• I Prove, Therefore I Am: Spatiotemporal Multi-Party Computation

  This paper introduces a new type of secure multiparty computation that incorporates physical location and time data, ensuring that computations respect real-world spatiotemporal constraints while maintaining privacy and security.

  Security · relevance 8/10

• RPolyPlan: A Unified Global Floorplanning and Legalization Framework for Rectilinear Soft Modules

  This paper introduces RPolyPlan, a new method that directly optimizes flexible rectilinear module shapes during chip floorplanning, improving area use and wirelength compared to previous approaches.

  Systems Architecture · relevance 8/10
• DiffRMST: A GPU-Accelerated Fully Differentiable Rectilinear Minimum Spanning Tree Wirelength Model for Global Placement

  This paper introduces DiffRMST, a new differentiable model for estimating wirelength in chip placement that improves accuracy and speed by enabling gradient-based optimization and GPU acceleration.

  Systems Architecture · relevance 8/10