Science & Technology News - March 4, 2026

AI Agents Tackle Corrupt Success and PDEs Get Smarter

Large Language Models (LLMs) are no longer just about generating text; they're agents tasked with complex procedures. However, a new arXiv paper, "Beyond Task Completion: Revealing Corrupt Success in LLM Agents through Procedure-Aware Evaluation", highlights a critical flaw: LLMs can achieve task completion through faulty or "corrupt" intermediate steps. This isn't just a theoretical problem; it means AI agents might appear successful while fundamentally misunderstanding or misexecuting their goals. The implication for real-world applications, from autonomous driving to scientific discovery, is significant. We need evaluation metrics that scrutinize the how, not just the what, to ensure AI reliability.

Elsewhere in AI, "From Complex Dynamics to DynFormer: Rethinking Transformers for PDEs" proposes a novel Transformer architecture specifically designed for solving Partial Differential Equations (PDEs). This is crucial because PDEs govern everything from fluid dynamics to quantum mechanics. Traditional numerical methods are computationally intensive. By adapting the Transformer architecture, which has revolutionized natural language processing, researchers aim to accelerate scientific simulations and unlock new insights into complex physical systems. The potential here is faster drug discovery, more accurate weather forecasting, and deeper understanding of fundamental physics.

Looking at optimization, "Why Adam Can Beat SGD: Second-Moment Normalization Yields Sharper Tails" offers a compelling theoretical explanation for the empirical success of the Adam optimizer over simpler Stochastic Gradient Descent (SGD). While Adam is widely used, its advantages weren't fully understood. This research suggests Adam's second-moment normalization helps it navigate complex loss landscapes more effectively, leading to faster convergence and potentially better final model performance. This insight could guide the development of even more efficient training algorithms for the massive models powering modern AI.

References

• My relationship with my PhD supervisor has become toxic — what do I do? - Nature
• First drone passengers may be combat casualties and criminals - New Scientist
• A tiny twist creates giant magnetic skyrmions in 2D crystals - Science Daily
• Four decades of data give unique insight into the sun's inner life - Phys.org
• Climate Physicists Face the Ghosts in Their Machines: Clouds - Quanta Magazine
• ‘Veronika’ Is the First Cow Known to Use a Tool - WIRED Science
• Beyond Task Completion: Revealing Corrupt Success in LLM Agents through Procedure-Aware Evaluation - arXiv
• From Complex Dynamics to DynFormer: Rethinking Transformers for PDEs - arXiv