Maligned - January 15, 2026

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. The Real Test: AI for Software Engineering Work 💻

Forget coding challenges; can AI actually build and debug real-world software? The new APEX-SWE benchmark dives into complex integration and observability tasks, finding that while frontier models like Gemini 3 Pro are making progress, they still struggle with the nuanced epistemic reasoning needed for practical software engineering. This benchmark is crucial for assessing AI’s true utility beyond academic metrics.

Source: arXiv Link: https://arxiv.org/abs/2601.08806v1

2. LLMs Just Got Smarter with “Multiplex Thinking” 🧠

Chain-of-Thought is old news. Researchers just unveiled “Multiplex Thinking,” a new reasoning mechanism that allows LLMs to “think softly” by sampling and aggregating multiple plausible next steps simultaneously. This leads to significantly more effective reasoning on tough tasks, producing shorter sequences and outperforming traditional methods, marking a real leap in LLM problem-solving capabilities.

Source: arXiv Link: https://arxiv.org/abs/2601.08808v1

3. Invisible AI Watermarks? Don’t Trust Them Yet 🚫

Just as platforms roll out invisible watermarking for AI-generated images, new research (RAVEN) exposes a critical vulnerability. By reformulating watermark removal as a novel view synthesis problem, attackers can effectively erase these invisible marks without access to the detector or watermark knowledge. This is a serious blow to content authenticity efforts and the integrity of AI-generated media.

Source: arXiv Link: https://arxiv.org/abs/2601.08832v1

4. Unmasking Political Bias in Large Language Models 🏛️

AI isn’t politically neutral, and now we have a robust way to measure it. New research introduces a methodology using real parliamentary voting records to benchmark LLM political biases across multiple countries. The findings consistently show state-of-the-art models displaying left-leaning or centrist tendencies, often with negative biases toward right-conservative parties – a critical insight for ethical AI deployment.

Source: arXiv Link: https://arxiv.org/abs/2601.08785v1

5. Fighting AI Deepfakes: Smarter Detection is Here 🖼️

While watermarks are failing, the fight against AI-generated misinformation gets a boost with the Multi-Cue Aggregation Network (MCAN). This new framework significantly improves AI-generated image detection by integrating diverse cues (spatial, frequency, chromaticity). MCAN proves more robust and generalizes better across various image generators, offering a much-needed upgrade in identifying synthetic content.

Source: arXiv Link: https://arxiv.org/abs/2601.08790v1

That’s it for today. Stay aligned. 🎯

Maligned - AI news without the BS