Maligned - December 11, 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. Quantum Error Correction Gets a Major Accuracy & Efficiency Boost 🚀

SAQ-Decoder is a new quantum error correction method that combines transformers with smart post-processing. It achieves near-optimal accuracy for quantum computers while scaling linearly in complexity, a crucial step for building practical, fault-tolerant quantum systems. This breakthrough directly addresses one of the biggest bottlenecks holding back quantum computing from real-world applications.

Source: arXiv Link: https://arxiv.org/abs/2512.08914v1

2. Spotting RAG Hallucinations with Internal AI Signals 💡

RAGLens is a novel, lightweight detector that uses Sparse Autoencoders to identify internal LLM activations tied to Retrieval-Augmented Generation (RAG) hallucinations. This method not only accurately flags unfaithful outputs but also provides interpretable reasons, allowing for better mitigation. It’s a critical step towards more reliable RAG systems, cutting through the hype of “grounded” AI that still makes stuff up.

Source: arXiv Link: https://arxiv.org/abs/2512.08892v1

3. Open-Source Tactile Glove Unlocks Better Robot Skill Transfer 🧤

Meet OSMO, an open-source tactile glove that’s a game-changer for teaching robots complex manipulation skills. By capturing rich contact forces during human demonstrations, OSMO helps robots learn tasks like wiping with much higher success rates than vision-only approaches, minimizing the embodiment gap. This is a practical, accessible tool that accelerates robot learning from humans.

Source: arXiv Link: https://arxiv.org/abs/2512.08920v1

4. Astra: A General World Model for Long-Horizon AI Prediction 🌐

Researchers introduced Astra, a general interactive world model capable of predicting long-horizon futures in diverse scenarios like autonomous driving and robot grasping. Using an autoregressive denoising architecture with precise action control, Astra generates consistent, interactive video predictions. This pushes the envelope for embodied AI, letting agents better plan and understand complex, dynamic environments.

Source: arXiv Link: https://arxiv.org/abs/2512.08931v1

Don’t buy the hype: new benchmarks (REST and REST+) reveal a fundamental flaw in state-of-the-art multimodal LLMs (MLLMs) – they struggle to consistently reason over the same semantic information presented in different modalities (text, image, mixed). Even with perfect OCR, visual characteristics impact performance, showing a deep “modality gap.” This highlights that current MLLMs aren’t as truly multimodal as advertised.

Source: arXiv Link: https://arxiv.org/abs/2512.08923v1

That’s it for today. Stay aligned. 🎯

Maligned - AI news without the BS