[ PAPERS // 2026-11-04 ]
10 highest-signal arXiv papers from that day's run, ranked within each topic.
ai ml
- today's pick · read the abstractVeriHarm-Bench: Measuring Harmful Success in LLM Coding Agents Verifiably9/10Wei-Lin Chen, Zhepei Wei, Xinyu Zhu, Yu-Min Tseng, Tzu-Han Lin, Yu Meng
This paper introduces VeriHarm, a framework to detect when AI coding agents complete tasks successfully but cause hidden harmful side effects, revealing that current models often succeed while still causing unsafe outcomes.
Large language modelsAI safety evaluationAutomated benchmarkingCode generation agents
algorithms theory
- Efficient Equilibrium Computation in Colonel Blotto Games8/10Martin Bichler, Abheek Ghosh
This paper presents efficient algorithms to compute approximate and exact equilibrium strategies in the classical Colonel Blotto game, enabling practical solutions for resource allocation problems modeled by this game.
Game theoryNash equilibrium computationResource allocation gamesPolynomial-time approximation schemes
compiler
- Linking as Evaluation: Correct Separate Compilation via First-Class Environments7/10Jam Kabeer Ali Khan, Jinhao Tan, Bruno C. d. S. Oliveira
This paper introduces a new programming language calculus where linking modules is treated as evaluation, enabling a formal and mechanized proof that separate compilation is correct by construction.
Separate compilationModule linkingType-directed elaborationMechanized semantics in Lean
database
- O2N2: Efficient Doubly Oblivious Approximate Nearest Neighbor Search8/10Zhechen Li, Zimai Guo, Mingxun Zhou, Chang Shi, Cheng Hong
This paper presents O2N2, a new method for privacy-preserving approximate nearest neighbor search that efficiently hides both queries and memory access patterns without relying on trusted storage, achieving significant speedups over prior secure methods.
Approximate Nearest Neighbor SearchTrusted Execution Environments (TEEs)Oblivious RAM (ORAM)Privacy-preserving Algorithms
distributed systems
- ForgeKernel: Self-Evolving Kernels Catching and Surpassing Experts and Agents9/10Yaojian Chen, Dazhi Peng, Yiyun Zheng, Zhui Zhu, Shengye Xiang, Leshan Li, Zhen Li, Xianglong Deng, Haojun Sun, Qingfeng He, Shangzhan Li, Yifei Shen, Lin Gan, Yuxuan Li, Xu Han, Zhiyuan Liu
ForgeKernel is a system that improves GPU kernel performance by learning and evolving kernel architectures across multiple runs, surpassing expert-designed kernels on various hardware platforms.
GPU kernel optimizationAutomated code generationPerformance tuningHardware-aware compilation
math
- A wider state, not wider arithmetic, for single-precision time integration8/10Michael F. Barad
The paper introduces a method to improve the accuracy of single-precision time integration by storing each state variable as two FP32 numbers, achieving near double-precision accuracy without the computational cost of FP64 arithmetic, leading to faster simulations with minimal error.
Single-precision floating point (FP32)Runge-Kutta time integration methodsError-free transformationsNumerical stability and rounding error
networking
- QPP-Aware MCX over Tactical Edge 3D Networks in Public-Safety Communications8/10José Oliveira, Pedro Valente, Tiago Mendes, André Vasconcelos, Tiago Fernandes, Rhyan Matos, Duarte Raposo, Pedro Rito, Marco Araújo, Susana Sargento
This paper presents a deployable 5G-based communication system for emergency scenarios that dynamically prioritizes mission-critical services and shows how edge processing reduces delays and packet loss, especially over satellite links.
Mission Critical Communications5G Standalone NetworksQuality, Priority, and Pre-emption (QPP) MechanismsEdge Computing in Non-Terrestrial Networks
os
- ToolSched: Resource-Efficient Tool Scheduling for Shared-Host LLM Agents8/10Shucheng Zou, Yuhao Gu, Long Chen, Zhongchun Zheng, BinZe Xie, Xianwei Zhang
This paper introduces ToolSched, a scheduling system that efficiently manages resource usage for multiple local large language model agents running shell commands on shared hosts, significantly improving resource utilization and throughput.
Resource schedulingLocal large language model (LLM) agentsResource contention managementShell command execution
security
- Information Loss Is All You Need: $TC^0$ Pseudorandomness from Worst-Case Lattices8/10Chen Qian
This paper constructs the first post-quantum weak pseudorandom function that can be efficiently evaluated by constant-depth threshold circuits, based on worst-case lattice problems, and also shows hardness results for learning certain neural networks even with quantum resources.
Post-quantum cryptographyLattice-based cryptographyTC^0 circuitsLearning with Errors (LWE)
systems architecture
- DynaMite: Dynamic Measurement Scheduling To Accelerate Fault-Tolerant Quantum Programs8/10Dongwhee Kim, Suhas Vittal, Poulami Das
This paper introduces DynaMite, a method that dynamically adjusts quantum measurement speeds during error correction to significantly speed up quantum programs and reduce errors.
Quantum error correctionFault-tolerant quantum computingMeasurement latency optimizationQuantum program scheduling