Maligned - December 18, 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. AI’s Logic Just Got a Major Upgrade 🧠

Forget elaborate designs; core improvements to Universal Transformers are making AI smarter at abstract reasoning. This new Universal Reasoning Model (URM) just blew past previous records on complex benchmarks like ARC-AGI, hitting 53.8% pass@1. It’s a fundamental step towards AI that can truly “think” through problems, not just pattern match.

Source: arXiv Link: https://arxiv.org/abs/2512.14693v1

2. Lifelike 3D Avatars from a Single Photo? Yes. 🗣️

VASA-3D just dropped, generating incredibly realistic, audio-driven 3D head avatars from a single image. It captures subtle expressions by translating VASA-1’s 2D realism to 3D, churning out free-viewpoint videos at up to 75 FPS. This isn’t just a parlor trick; it’s a massive leap for immersive digital interactions and AR/VR.

Source: arXiv Link: https://arxiv.org/abs/2512.14677v1

3. Generative AI Gets a Reality Check: Introducing MMGR 🧐

While video models look impressive, do they actually “understand” physics and logic? MMGR (Multi-Modal Generative Reasoning) is a new benchmark that exposes serious reasoning failures in models like Veo-3 and Sora-2, pushing evaluation beyond mere perceptual quality. It’s a crucial step to hold generative AI accountable for more than just looking good.

Source: arXiv Link: https://arxiv.org/abs/2512.14691v1

4. Robotics Simulation Gets a Physics Upgrade 🤖

Getting real human interactions into robot simulations without breaking physics is a nightmare – until CRISP. This method recovers physically plausible human motion and clean scene geometry from a single video, reducing motion tracking failures from 55% to under 7%. It’s a game-changer for training robots in realistic virtual environments and boosting real-to-sim transfer.

Source: arXiv Link: https://arxiv.org/abs/2512.14696v1

5. LLM Inference Just Got 4x Faster 🚀

Deploying large language models at scale is expensive and slow. Jacobi Forcing introduces a new parallel decoding method that dramatically speeds up LLM inference by up to 4x, with minimal impact on generation quality. This is a big deal for anyone running large models, slashing operational costs and latency for practical applications.

Source: arXiv Link: https://arxiv.org/abs/2512.14681v1

That’s it for today. Stay aligned. 🎯

Maligned - AI news without the BS