CryptoBench: A Dynamic Benchmark for Expert-Level Evaluation of LLM Agents in Cryptocurrency

Astra has been introduced as a pioneering multi-agent system designed for optimizing GPU kernel performance, addressing a long-standing challenge in high-performance computing and machine learning. This system leverages existing CUDA implementations from SGLang, a framework widely used for serving large language models (LLMs), marking a shift from traditional manual tuning methods.

A recent study has unveiled significant privacy risks associated with the Key-Value (KV) cache used in Large Language Model (LLM) inference. The research highlights that attackers can reconstruct sensitive user inputs from the KV-cache, demonstrating vulnerabilities through various attack vectors, including direct Inversion, Collision, and semantic-based Injection Attacks. To address these issues, the study proposes KV-Cloak, a novel defense mechanism designed to enhance privacy during LLM operations.

FairT2I has been introduced as an innovative framework aimed at addressing social biases in text-to-image generation, leveraging large language models (LLMs) for bias detection and attribute rebalancing. This framework operates without the need for extensive training, utilizing a mathematically grounded approach to enhance the generation process by adjusting attribute distributions based on user input.

ReSpace has been introduced as a generative framework for text-driven 3D indoor scene synthesis and editing, utilizing autoregressive language models to enhance scene representation and editing capabilities. This approach addresses limitations in current methods, such as oversimplified object semantics and restricted layouts, by providing a structured scene representation with explicit room boundaries.

The introduction of the Intelligent Multi-Agent Debate (iMAD) framework aims to enhance the efficiency and accuracy of Large Language Model (LLM) inference by selectively triggering structured debates among LLM agents. This approach addresses the computational costs and potential inaccuracies associated with traditional Multi-Agent Debate systems, which can degrade performance by overturning correct answers.

Recent research introduced SkyLadder, a novel pretraining strategy for large language models (LLMs) that optimizes context window scheduling. This approach transitions from short to long context windows, demonstrating improved performance and efficiency, particularly with models trained on 100 billion tokens.

LLM-NAS introduces a novel approach to Hardware-Aware Neural Architecture Search (HW-NAS), focusing on optimizing neural network designs for accuracy and latency while minimizing search costs. This method addresses the exploration bias observed in traditional LLM-driven approaches, which often limit the diversity of proposed architectures within a constrained search space.

ADORE, or Autonomous Domain-Oriented Relevance Engine, has been introduced as a novel framework aimed at improving relevance modeling in e-commerce search. It addresses challenges posed by traditional term-matching methods and the limitations of neural models, utilizing a combination of a Rule-aware Relevance Discrimination module, an Error-type-aware Data Synthesis module, and a Key-attribute-enhanced Knowledge Distillation module to enhance data generation and reasoning capabilities.