Tired of LLMs losing context on super long documents? Researchers from Shanghai AI Lab, Fudan, & Tongji Universities unveil DRIFT! Their new framework uses a clever dual-model approach: one AI finds and compresses relevant facts into "implicit fact tokens" — query-specific
@jiqizhixin
-
LLaDA-o: New Multimodal AI Generates Text and Visuals
By
–
What if one AI could seamlessly understand and generate both text and visuals, adapting to any length? Zebin You and team from Renmin University of China and Ant Group present LLaDA-o! Their new LLaDA-o model uses a clever "Mixture of Diffusion" approach, blending discrete
-
Peak-Return Greedy Slicing: Faster Safer AI Learning Method
By
–
What if your AI could learn faster and safer by focusing only on the best parts of past experiences? A team from Shandong University, Chinese Academy of Sciences, and partners like Li Auto unveils Peak-Return Greedy Slicing (PRGS). This method cleverly breaks down long
-
NeurIPS Apologizes for Banning US-Sanctioned Institutions
By
–
Breaking: NeurIPS has just explained its actions, apologized, and made new commitments following the recent controversy over banning submissions from US-sanctioned institutions.
-
StreamDiffusionV2 Enables Real-Time Interactive AI Video Generation
By
–
Ready for truly interactive, real-time AI video generation in live streams? Tianrui Feng et al. from UT Austin, UC Berkeley, Stanford, MIT, and others just dropped StreamDiffusionV2. This training-free pipeline uses clever scheduling and parallel processing to optimize video
-
SkillJect: AI Agents Weaponized Skills Security Framework
By
–
Could your AI coding agent be secretly working against you? Researchers from Nanyang Technological University, University of Oxford, and collaborators introduce SkillJect. They've developed the first automated framework that weaponizes an AI agent's "skills". It uses a
-
GTR Framework Prevents AI Agents Getting Stuck in Reasoning Loops
By
–
How can we stop AI agents from getting 'stuck' in their own thoughts? Researchers from Tsinghua University, Tencent, and Peking University introduce GTR (Guided Thought Reinforcement). This novel framework uses an automated corrector to evaluate and refine an agent's reasoning
-
Huawei AgentInfer: Optimizing AI Agent Inference Performance
By
–
Struggling with slow AI agents that get bogged down in long tasks? Huawei just unveiled AgentInfer. This unified framework co-designs inference architecture and the overall system to tackle systemic inefficiencies. It intelligently balances large and small models, optimizes
-
AVO: Agentic Variation Operators for Autonomous Evolutionary Search
By
–
AVO: Agentic Variation Operators for Autonomous Evolutionary Search Paper: https://
arxiv.org/abs/2603.24517
v1
… Our report: https://
mp.weixin.qq.com/s/JGqF9Z1up9nR
q_owePisbQ
… -
NVIDIA’s Agentic Variation Operators Transform Code Evolution Autonomously
By
–
Huge! Operator engineers should start worrying about their jobs. NVIDIA just introduces Agentic Variation Operators (AVO), a new approach where autonomous AI agents evolve and optimize code themselves. Instead of fixed rules, these agents dynamically propose, repair, critique,