Also Worth Noting — 2026-03-28
AI models can now improve their reasoning by teaching themselves without needing expensive human guidance.
Also Worth Noting
02 [Multimodal] Multimodal AI Models Judge Themselves for Better Reasoning Multimodal AI models can now improve their reasoning skills on their own using a new unsupervised self-evolution framework. This is impressive because traditional methods rely on expensive human-annotated data or a "teacher" model, which are difficult to scale. This breakthrough makes it cheaper and faster to develop powerful multimodal AI that can better understand and reason across text and images. link
03 [Video Gen] Measuring Physical Frame Rate in AI Videos A novel approach precisely measures the underlying physical time scale, or "motion pulse," within AI-generated videos. While current AI videos look realistic, they lack a consistent internal clock to match real-world speeds, making true physical simulation difficult. This breakthrough enables more accurate AI "world models" for applications like robotics, virtual reality, and scientific simulation, where consistent real-world timing is crucial. link
04 [Video Gen] CUA-Suite: Massive Video Demonstrations for Computer-Use Agents The CUA-Suite dataset offers massive human-annotated video demonstrations for training computer-use agents. It addresses a critical bottleneck by providing continuous video, not just sparse screenshots, which is crucial for building robust general-purpose agents. This will accelerate the development of agents capable of fully automating complex desktop workflows, making computers more efficient. link
05 [RAG] Self-Distillation Can Harm LLM Math Reasoning A study found that a training method called self-distillation can unexpectedly make large language models worse at mathematical reasoning tasks. This problem arises because self-distillation suppresses the model's crucial ability to express uncertainty or explore different reasoning paths, which are vital for solving complex problems. This discovery is important for developing more reliable AI that can handle intricate, multi-step challenges without losing its problem-solving capabilities. link
06 [Multimodal] UI-Voyager: Self-Evolving Agent Learns from Failed Mobile Tasks A new AI agent called UI-Voyager autonomously learns to interact with mobile app interfaces by self-evolving. This agent uniquely overcomes challenges by efficiently learning from its failed attempts and managing complex, multi-step tasks within apps. Such technology could lead to fully automated assistants that complete intricate workflows across various mobile applications. link