alphaXiv

Reproducing self-distillation papers with autoresearch agents.

The Shanghai Artificial Intelligence Laboratory developed Harness-of-Harness (HoH), a framework enabling multi-day autonomous software development by LLM-based agents through structured iterative...

Dense linear maps carry much of the parameter and computational burden of modern neural networks, yet their dense form leaves the organization of learned couplings implicit. Quantum many-body...

HAWKEYE is an open-source framework that enhances AI coding agents with hardware awareness to autonomously optimize GPU kernels, significantly reducing the effort to adapt software to new...

Researchers uncovered an architectural vulnerability in major proprietary large language model APIs where encrypted internal reasoning traces can be extracted in plaintext by weaker models within...

This is an expository talk written for the Bourbaki Seminar. After a brief introduction, Section 1 discusses in the categorical language the structure of the classical deterministic computations....

Researchers developed EMPATH, a multilingual auditor-judge benchmark to evaluate the safety of emotional-support chatbots in dynamic, multi-turn conversations. This framework, spanning US English...

A fixed-budget, cluster-aware evaluation standard is introduced for LLM-as-a-judge comparisons in multi-hop Retrieval-Augmented Generation (RAG) systems. This standard, when stress-tested, reveals...

RL fine-tuning small models to behave as recursive language models.