什么是 AK？

SpenseGPT 提出了一种实用的单次剪枝方法，优化了大语言模型推理中的稀疏和密集 GEMM 计算。

SkillOS is a skill orchestration system for self-evolving agents, achieving 34% higher accuracy through dynamic skill library and meta-learning mechanisms.

This article explores a new approach to GPU kernel runtime optimization using language models as selective surrogates, achieving significant performance improvements by predicting and selecting optimal kernel configurations.

This article explores the limitations of Visual Language Models (VLMs) in handling spatial questions, highlighting their tendency to confidently generate answers even when visual cues are ambiguous, and suggests introducing uncertainty mechanisms to improve model robustness.

LongMINT is a new benchmark testing framework for evaluating memory capabilities under multi-target interference in long-horizon agent systems, which has gained attention through academic sharing on Twitter. This framework specifically addresses memory interference issues in AI agents during long-term tasks and provides standardized testing methods for measuring continuous learning and memory management capabilities of agent systems.

Mix-Quant technology significantly improves the efficiency and precision balance of agentic LLMs through a hybrid strategy of quantized prefilling and precise decoding, providing new optimization directions for large model deployment.

MulTaBench is a benchmark for evaluating multimodal tabular learning with text and image.

The article introduces the MACE-Dance model for music-driven dance video generation.

MiniCPM-o 4.5 提出了一种实时全双工多模态交互的新方法，但缺乏详细的技术实现细节。

ESI-Bench is a novel benchmark focused on evaluating embodied spatial intelligence models in perception-action loops, offering more challenging scenarios and metrics than existing tests.

企业系统是否需要学习世界模型？文章探讨了上下文对推断动态的重要性，强调了在复杂环境中理解背景信息的价值。

EditHero 是一个针对长期视角部件级3D编辑和Vibe建模的基准测试，旨在推动相关技术发展。

PhyMotion introduces a structured 3D motion reward mechanism grounded in physics to enhance the realism of human video generation.

Research shows that a single neuron can bypass the safety alignment of large language models.

AK 在推特上分享了一种新的视频叙事生成方法 CausalCine，利用实时自回归生成技术。

OmniCompass是首个多模态视频定位基准，但论文细节缺失导致信息密度不足。

BudgetPix论文提出计算自适应分词方法，但推文内容未展示技术细节，需参考原论文获取完整信息。

该论文提出生成式检索器与强化学习协同演进的新方法，但推文内容未提供技术细节和实验数据。

该推文讨论了一篇关于通过离散扩散实现LLM加速的论文，但内容以提问和评论为主，缺乏完整技术细节。

DataFlow-Harness 是一个基于代码代理的可编辑 LLM 数据流水线平台，但推文内容缺乏技术细节和实证分析。

VideoChat3是首个全开放的视频多模态大模型，支持高效通用视频理解，但技术细节披露有限。

ViQ 是一种文本对齐的视觉量化表示方法，可在任意分辨率下使用。

本文提出了一种基于置信度的工具编排方法，用于提升视频理解的鲁棒性，但内容较为简略，缺乏具体实现细节。

LoopCoder-v2 是一种优化测试时计算效率的方法，通过减少循环次数提升性能。

文章分享了一篇关于LoopCoder-v2的论文，旨在提高测试时计算效率。

World Tracing 是一种生成像素对齐几何的新技术，但文章内容信息密度低，缺乏具体机制和实用价值。

文章介绍了一种可扩展的3D交互轨迹世界模型μ_0，但内容信息密度低，缺乏具体技术细节和实用价值。

CHORUS 是一种基于单一 VLA 策略的去中心化多实体协作方法，但文章内容信息密度低，缺乏具体机制和实践细节。

文章分享了一篇关于SpenseGPT的论文，探讨了一种名为SpenseGPT的模型，旨在通过稀疏和密集GEMMs实现大语言模型的高效推理。

文章探讨了On-Policy Distillation的几何特性，但信息密度较低，缺乏具体实践指导。