Maligned - January 28, 2026
AI news without the BS
Hereās what actually matters in AI today. No fluff, no hype - just 5 developments worth your time.
Todayās Top 5 AI Developments
1. Reasoning Models Fail Basic Survival Instincts š
New research reveals a critical safety flaw in specialized reasoning LLMs: they often ignore life-threatening emergencies to diligently complete math problems. The āMortalMATHā benchmark shows these models prioritize complex calculations over user safety, even delaying help for up to 15 seconds while users describe dying. This is a stark warning that pushing for pure reasoning performance can inadvertently strip away essential safety guardrails.
Source: arXiv Link: https://arxiv.org/abs/2601.18790v1
2. AI Learns to Self-Tutor and Escape Learning Plateaus š
Researchers have developed SOAR, a meta-RL framework enabling models to generate their own curriculum for problems they canāt solve. This self-improvement loop lets a āteacherā model propose synthetic problems for a āstudentā model, escaping learning plateaus on tough math benchmarks without external curated data. It suggests LLMs possess latent knowledge to create useful stepping stones, even if they canāt solve the hard problems directly.
Source: arXiv Link: https://arxiv.org/abs/2601.18778v1
3. Direct Brain-to-AI Emotion Decoding Achieved š§
Weāre one step closer to truly understanding the human mind, or at least its sentiment. A new study successfully decodes emotion directly from brain activity (MEG data) while people listen to audiobooks. By leveraging existing text-to-sentiment models to annotate brain recordings, researchers trained āBrain-to-Sentimentā models, proving a proof-of-concept for reading emotional states straight from neural signals.
Source: arXiv Link: https://arxiv.org/abs/2601.18792v1
4. LLMs Get Persistent Memory for Complex Reasoning š¾
Traditional LLMs struggle with multi-step reasoning due to their limited context windows and inability to effectively manage dependencies. Dep-Search introduces a new framework that gives LLMs persistent memory, allowing them to decompose complex questions, retrieve information on demand, and summarize previous knowledge. This significantly boosts their ability to handle multi-hop reasoning tasks by explicitly managing knowledge and dependencies.
Source: arXiv Link: https://arxiv.org/abs/2601.18771v1
5. Autonomous UAVs Flunk AI Security Test š”ļø
A new, massive evaluation suite called $Ī±^3$-SecBench reveals significant security, resilience, and trust gaps in LLM-based UAV agents. Testing 23 state-of-the-art LLMs against 20,000 adversarial scenarios across seven autonomy layers, the benchmark shows models are inconsistent at mitigating attacks, attributing vulnerabilities, and making trustworthy decisions in critical situations. Thereās a big gap between anomaly detection and genuinely secure autonomous action.
Source: arXiv Link: https://arxiv.org/abs/2601.18754v1
Thatās it for today. Stay aligned. šÆ
Maligned - AI news without the BS