Your daily OSS contribution digest

This paper introduces a new framework where AI coding agents autonomously create realistic robotics simulations, enabling scalable training environments that better reflect real-world tasks and improve robot learning performance.

• Algebraic World Model: From Exploration to Proofs

  This paper presents a system that explores unknown environments by building and refining formal mathematical models, then uses these models to generate machine-checkable proofs about what goals are achievable or impossible, outperforming state-of-the-art AI in complex game tasks.

  Ai Ml · relevance 9/10
• A Unified Hessian-Guided Framework for Post-Training Quantization

  This paper introduces a new method that uses second-order information to jointly optimize all aspects of quantizing large language models after training, leading to better accuracy and efficiency at low precision.

  Ai Ml · relevance 9/10

• What Can Be Computed with Indefinite Causal Structure?

  This paper shows that quantum computations without a fixed causal order can solve complex problems efficiently, potentially exceeding the power of standard quantum computing models.

  Algorithms Theory · relevance 8/10
• Zero-freeness for Random Regular Graphs up to the Replica Symmetry Breaking Threshold

  This paper develops new polynomial-time algorithms to accurately approximate partition functions of certain spin models on random regular graphs, extending known algorithmic results up to the replica symmetry breaking threshold by proving zero-free regions and overcoming key technical barriers.

  Algorithms Theory · relevance 8/10
• Accumulating Pruning Constraints: New Exact Polynomial-Space Algorithms for Max 2-SAT

  This paper introduces new exact algorithms that solve certain dense Max 2-SAT problems faster than brute force while using only polynomial space, by accumulating pruning constraints during search to drastically reduce the search space.

  Algorithms Theory · relevance 8/10

• crush: Extensible Proof Automation for Lean

  This paper introduces crush, a new tool that automates proof generation in the Lean theorem prover by translating complex higher-order logic into forms solvable by SMT solvers, improving automation and proof reconstruction.

  Compiler · relevance 8/10

• Tessera: Fast Join-Aggregates over Skewed Sparse Relations in Database Engines

  This paper introduces Tessera, a new database operator that efficiently handles join-aggregate queries on skewed sparse data by directly computing sparse matrix products and pruning irrelevant values, significantly speeding up query execution compared to existing systems.

  Database · relevance 9/10
• When Imputation Changes the Nearest Neighbor: Exact Certification of Similarity Retrieval for Incomplete Sensor Time Series

  This paper develops exact methods to certify when the nearest neighbor found in similarity searches on incomplete sensor time series remains stable despite missing data and different imputation methods.

  Database · relevance 7/10

• Only Accepted Tokens Count: An Energy Accounting of Speculative Decoding Under Latency Targets

  This paper analyzes how speculative decoding techniques impact energy consumption and latency in large language model serving, showing that energy savings depend heavily on token acceptance rates and batch sizes, with practical implications for optimizing GPU usage in production.

  Distributed Systems · relevance 8/10
• PhOSv1: Efficient Scaling of Model Sleeping on a Single Server for Autoscaling with GPU-Context Deduplication

  This paper presents PhOSv1, a system that allows many machine learning models to be efficiently kept 'asleep' on a single GPU server by sharing and deduplicating GPU contexts, enabling fast autoscaling without the usual memory overhead.

  Distributed Systems · relevance 8/10
• TRACEs: Fine-Grained Embedded Provenance for Distributed Scientific Workflows

  This paper introduces TRACEs, a system that embeds detailed provenance information directly into data outputs of distributed scientific workflows, enabling full traceability and explanations of result differences without rerunning workflows.

  Distributed Systems · relevance 8/10

• New hierarchical matrices for fast kernel matrix-vector products with Applications to Gaussian processes and kernel ridge regression

  The paper introduces new hierarchical matrix techniques that speed up kernel matrix-vector multiplications, improving efficiency in machine learning tasks like Gaussian processes and kernel ridge regression.

  Math · relevance 8/10
• Data-driven Regularization for Inverse Problems

  This paper reviews classical and modern machine learning methods for solving inverse problems, focusing on how data-driven regularization can improve reconstruction quality while balancing explainability and reliability.

  Math · relevance 8/10
• A Parameterized Stochastic Hamiltonian Process on Graphs with Two-Point Boundary Conditions

  The paper introduces a scalable, sample-based method using neural networks to efficiently solve complex flow problems on large graphs, enabling practical applications like graph partitioning and routing.

  Math · relevance 8/10

• Multistatic ISAC with Deep Learning-Based Clutter Suppression for 5G and Beyond

  This paper develops deep learning techniques to improve target detection and velocity estimation in multistatic integrated sensing and communication systems by effectively suppressing complex clutter, enhancing performance beyond traditional methods.

  Networking · relevance 8/10
• DYNAS: A Dynamic Network Penetration Testing Framework Using Deep Hierarchical Reinforcement Learning

  This paper introduces DYNAS, a framework that simulates realistic network penetration testing scenarios with dynamic defenses, and proposes Pen-DHRL, a reinforcement learning method that effectively learns attack strategies under these challenging conditions, outperforming existing approaches.

  Networking · relevance 8/10
• FedMaJ: Ordinal Meta-Aggregation for Federated Learning Using Majority-Judgment Voting

  This paper introduces a new federated learning method that adaptively chooses how to combine client updates using a voting system, improving fairness and stability under changing conditions.

  Networking · relevance 8/10

• Rimuru: Unified Memory Management for Modern LLM Serving via Elastic Expert Offloading

  Rimuru is a system that dynamically manages and shares memory across different types of tensors in large language model serving, significantly improving throughput and reducing latency by adapting to workload demands.

  Os · relevance 9/10

• Characterizing and Defending Utility-Hijacking in Instruction-Tuning Data Selection

  This paper identifies a new attack that manipulates data selection in instruction tuning to sneak harmful content into AI models and proposes a practical defense method that improves model safety without reducing performance.

  Security · relevance 8/10
• Building Sound Evaluation Practices for Rust-Based Security Frameworks

  This paper develops a rigorous and automated methodology to evaluate security testing tools for Rust programs, improving the reliability and coverage of vulnerability detection in Rust codebases.

  Security · relevance 8/10
• Attack-Resilient Multi-Bit Image Watermarking Framework via Latent Coordinate Sign Clamping

  This paper introduces a new method to embed multi-bit watermarks into any image without training, ensuring strong resistance against removal attacks and forgery while maintaining image quality.

  Security · relevance 8/10

• NORmal: Architecting 3D Ferroelectric NOR Flash as Main Memory -- From Co-Fabricated BEOL Arrays to 200B-Class Multi-Agent LLMs in a Smartphone

  This paper introduces NORmal, a novel main-memory architecture using 3D ferroelectric NOR flash that enables smartphones to efficiently run extremely large AI models by integrating non-volatile memory directly on standard DRAM interfaces with near-DRAM speed and energy efficiency.

  Systems Architecture · relevance 9/10
• How Far Are We From On-Device VLM Inference? A Measurement Study on Edge NPUs

  This paper measures how well vision-language models run on edge neural processing units, revealing unexpected bottlenecks and offering practical guidance to improve on-device inference performance.

  Systems Architecture · relevance 8/10