Labyrinth 1.1 introduces a new sub-protocol to ensure end-to-end encrypted messages survive device loss, switching, and long sign-in gaps, …
Researches
TeamTR introduces trust-region fine-tuning to prevent shared-context drift, a critical failure mode in multi-agent LLM systems, …
This explainer clarifies recent LLM benchmark results, addressing claims of 0% scores and detailing actual performance on complex software …
This paper reveals that instruction-tuned LLMs can produce fair outputs while still retaining causally potent and asymmetrically biased …
This paper proposes 'face density' as a novel metric to quantify data complexity, particularly for instance counting tasks in computer …
Mistral AI's Vox-Trainer is a new multimodal model capable of understanding and generating both spoken audio and text, with accessible …
This paper introduces an actor-verifier AI architecture that enhances reliability and interpretability in safety-critical systems by having …
This paper introduces a novel method to train LLMs to internally recognize their own hallucinations by distilling weak, external …
RAGEN-2 identifies and measures 'reasoning collapse' in multi-turn LLM agents, where internal thought processes degrade despite initial task …
SymptomWise proposes a framework that enhances AI reliability and interpretability by separating natural language understanding (handled by …
Google's TurboQuant algorithm slashes LLM KV cache memory by 6x and delivers up to 8x attention speedup with zero accuracy loss, …
MTA-Agent introduces a modular, multi-turn agent framework that enhances Multimodal Large Language Models (MLLMs) by integrating specialized …