Maligned - November 27, 2025

AI news without the BS

Here’s what actually matters in AI today. No fluff, no hype - just 5 developments worth your time.

Today’s Top 5 AI Developments

1. Infinite Video: No More Temporal Limits 🎬

Infinity-RoPE just dropped, allowing autoregressive video models to generate infinite, controllable footage. It smashes limits on temporal horizon, maintains fine-grained action control throughout long generations, and even enables cinematic scene transitions within a single stream. This is a major leap for continuous video creation and virtual experiences, bypassing core bottlenecks without retraining.

Source: arXiv Link: https://arxiv.org/abs/2511.20649v1

2. Agent LLMs: Talking in Latent Space, Not Just Text 🤝

Forget text-based chat for AI agents. LatentMAS enables LLMs to collaborate directly within their continuous latent space, making multi-agent systems faster and more expressive. This training-free framework boosts reasoning quality by up to 14.6% while cutting token usage by 70-83% and speeding up inference 4x, showing the real power of silent, efficient AI teamwork.

Source: arXiv Link: https://arxiv.org/abs/2511.20639v1

3. VLMs See in 3D: Finally, Real-World Perception 🌎

LocateAnything3D brings true multi-object 3D detection to Vision-Language Models, turning 3D perception into a “next-token prediction” task. By mimicking human reasoning (2D object -> 3D properties), it achieves state-of-the-art results on Omni3D, enabling VLMs to understand spatial relationships and act in the physical world without specialized heads. This is a crucial step for embodied AI and robotics.

Source: arXiv Link: https://arxiv.org/abs/2511.20648v1

4. Medical AI: Catching Everything, in Real-Time ⚕️

MedROV introduces the first real-time, open-vocabulary object detection model for diverse medical imaging. It can spot novel structures across nine modalities, outperforming existing foundation models by a huge margin (40 mAP50 improvement) while running at 70 FPS. This is a game-changer for faster, more comprehensive medical diagnostics, moving beyond limited, pre-defined object sets.

Source: arXiv Link: https://arxiv.org/abs/2511.20650v1

5. Aligning Generative AI: No More Trade-Offs? ⚖️

MapReduce LoRA and Reward-aware Token Embedding (RaTE) are tackling the “alignment tax” in generative models, letting you optimize for multiple human preferences without sacrificing one for another. This framework significantly boosts quality across modalities (e.g., 36% on GenEval for T2I, 43% helpfulness for LLMs) by training and merging preference-specific LoRA experts. It’s a major step towards truly versatile and user-controlled generative AI.

Source: arXiv Link: https://arxiv.org/abs/2511.20629v1

That’s it for today. Stay aligned. 🎯

Maligned - AI news without the BS