[ PAPERS // 2026-10-14 ]

This paper introduces a new large dataset and training method that provide dense progress rewards for long-horizon tasks, enabling more efficient training of autonomous agents that can solve complex, multi-step problems and generalize across domains.

This study analyzes real user conversations with ChatGPT about health in four Global South countries, revealing that about 20% of responses pose significant safety risks, especially when user-specific context is considered, highlighting potential dangers in relying on AI for medical advice.

This paper introduces Hybrid Reasoning Agents that combine internal latent reasoning with external text-based interaction, improving efficiency, transparency, and tool integration in AI agents.

This paper presents a new method to synthesize quantum circuits with fewer costly $T$ gates, improving efficiency for implementing arbitrary quantum operations and proving near-optimality of their approach.

This paper presents faster and more space-efficient algorithms for approximating solutions to path cover and traveling salesman problems with specific edge weights, improving previous approximation ratios and running times in both sublinear-time and semi-streaming models.

This paper presents a practical implementation of a recent theoretically fast algorithm for minimum-cost flow problems, showing competitive and sometimes superior performance compared to established solvers on benchmark datasets.

This paper presents ARCADE, a method to add advanced attribute-based access control to popular databases by externalizing authorization logic, improving flexibility and auditability with minimal performance impact.

This paper introduces HULK, a new algorithm that significantly speeds up bulk bitwise operations on encrypted data, enabling faster and more practical privacy-preserving computations in cloud applications.

This paper introduces FastEEP, a system that enables much faster and more efficient scaling of Mixture-of-Experts model serving by reusing runtime resources within instances instead of rebuilding them from scratch.

This paper introduces LunaTP, a new method that significantly speeds up key tensor operations in equivariant machine learning models for atomistic simulations, enabling much faster training and scalability on GPUs.

The paper presents a new unified mathematical method using semidefinite programming to estimate orientations of cryo-EM images for molecules with any symmetry, improving initial 3D model reconstruction from limited data.

This paper rigorously proves how particle density fluctuations in a two-dimensional asymmetric exclusion process spread faster than normal diffusion and converge to a specific stochastic heat equation, confirming long-standing physics predictions with exact scaling and noise behavior.

This paper proves that certain advanced MCMC methods can efficiently sample from complex heavy-tailed distributions in high dimensions, providing theoretical guarantees on their speed and introducing new geometric tools to understand these distributions.

This paper presents a method for predicting pedestrian movements by training models directly on vehicles using federated learning with Low-Rank Adaptation, reducing communication costs while maintaining high accuracy.

This paper presents HILO, a novel low-power radio frontend that uses tunnel diode oscillators to enable simultaneous dual-band wireless communication from a single circuit, improving reliability without increasing power consumption significantly.

This paper introduces Tanchun, a new system that detects large-scale HTTP(S) DDoS attacks using only flow-level data without decryption, by analyzing behavioral patterns over time to maintain accuracy even as traffic changes.

This paper extends quantum one-time programs to support stateful, controlled computations with limited queries, enabling reusable quantum tokens and more practical delegated computation schemes.

This paper presents a new method for a classical computer to efficiently verify quantum computations using only quasilinear resources, improving practical feasibility of quantum verification.

Tracekit is a system that securely logs and verifies the intentions, reasoning, and actions of autonomous coding agents to detect tampering and ensure trustworthy auditing.

This paper introduces ZTA-Q, an open-source RISC-V platform that accurately runs low-precision (INT8) CNN models from TensorFlow Lite on FPGA hardware, enabling detailed study of hardware approximations on model accuracy.

This paper presents a method to dynamically adjust the timing of individual operations in Tensor Processing Units running at low voltage, improving energy efficiency and performance without sacrificing accuracy.

This paper introduces CONFERM, a new scheduling method for coarse-grained reconfigurable arrays that better overlaps loop iterations and pipeline stages to significantly boost throughput and reduce configuration overhead.