Large Reasoning Models (LRMs) excel at complex reasoning but are traditionally evaluated in static, "frozen world" settings: model responses are assumed to be instantaneous, and the context of a request is presumed to be immutable over the duration of the response. While generally true for short-term tasks, the "frozen world" assumption breaks down in modern reasoning tasks such as assistive programming, where models may take hours to think through problems and code may change dramatically from the time the model starts thinking to the model's final output. In this work, we challenge the frozen world assumption and evaluate LRM robustness under two realistic dynamic scenarios: interruptions, which test the quality of the model's partial outputs on a limited budget, and dynamic context, which tests model adaptation to in-flight changes. Across mathematics and programming benchmarks that require long-form reasoning, static evaluations consistently overestimate robustness: even state-of-the-art LRMs, which achieve high accuracy in static settings, can fail unpredictably when interrupted or exposed to changing context, with performance dropping by up to 60% when updates are introduced late in the reasoning process. Our analysis further reveals several novel failure modes, including reasoning leakage, where models fold the reasoning into their final answer when interrupted; panic, where under time pressure models abandon reasoning entirely and return incorrect answers; and self-doubt, where performance degrades while incorporating updated information. Project Page: http://dynamic-lm.github.io/

Differentially private text sanitization refers to the process of privatizing texts under the framework of Differential Privacy (DP), providing provable privacy guarantees while also empirically defending against adversaries seeking to harm privacy. Despite their simplicity, DP text sanitization methods operating at the word level exhibit a number of shortcomings, among them the tendency to leave contextual clues from the original texts due to randomization during sanitization $\unicode{x2013}$ this we refer to as $\textit{contextual vulnerability}$. Given the powerful contextual understanding and inference capabilities of Large Language Models (LLMs), we explore to what extent LLMs can be leveraged to exploit the contextual vulnerability of DP-sanitized texts. We expand on previous work not only in the use of advanced LLMs, but also in testing a broader range of sanitization mechanisms at various privacy levels. Our experiments uncover a double-edged sword effect of LLM-based data reconstruction attacks on privacy and utility: while LLMs can indeed infer original semantics and sometimes degrade empirical privacy protections, they can also be used for good, to improve the quality and privacy of DP-sanitized texts. Based on our findings, we propose recommendations for using LLM data reconstruction as a post-processing step, serving to increase privacy protection by thinking adversarially.

When the gravity acting on them is increased, locusts adapt. Locusts placed in a centrifuge to mimic the conditions of hypergravity grew tougher legs than those living normally – but not all of them survived the process. Many biological materials, such as bone and wood, can adapt and become stronger under physical strain, but it isn't clear whether animals with shell-like exoskeletons can adapt in the same way as those with internal skeletons. Karen Stamm and Jan-Henning Dirks at the City University of Applied Sciences in Bremen, Germany, studied this by placing locusts inside a specially designed centrifuge to stress-test their exoskeletons using simulated hypergravity. The locusts were assigned to one of four gravity conditions: 1g – which is typical gravity at sea level and didn't involve a centrifuge – and 3g, 5g or 8g conditions, all of which did involve centrifuging the insects.

Welcome to Pushing Buttons, the Guardian's gaming newsletter. If you'd like to receive it in your inbox every week, just pop your email in below – and check your inbox (and spam) for the confirmation email. These days, video games aren't often tied to a particular console or gadget – even if something like Halo or God of War originally appears on Xbox or PlayStation, it will often eventually make its way to PC, and the majority of games come out on everything at the same time. Unless you're playing a console-exclusive game like Returnal, specifically designed around the PlayStation 5 controller with its adaptive triggers and fancy haptic feedback, the experience of the game isn't shaped much by the hardware you play it on. You can play Minecraft with a controller or a mouse or on your phone, and it's still largely the same.

Identifying different foods on a tray can be a non-trivial task (Image source: Chris A. Tweten on Unsplash) Unfortunately, we don't always have sufficient time to enjoy a leisurely lunchtime meal. Suppose some of your workmates invite you to join them for lunch at a local fast-food restaurant. You all typically have a limited amount of time for your lunchtime break and you need to use this time wisely and efficiently. First, you have to get to the restaurant, either by walking or perhaps by taking a short drive. When you reach the restaurant, you have to select your food, pay for it, and eat it.

Using machines to augment human activity is nothing new. Egyptian hieroglyphs show the use of horse-drawn carriages even before 300 B.C. Ancient Indian literature such as "Silapadikaram" has described animals being used for farming. And one glance outside shows that today people use motorized vehicles to get around. Where in the past human beings have augmented ourselves in physical ways, now the nature of augmentation also is more intelligent. Again, all one needs to do is look to cars – engineers are seemingly on the cusp of self-driving cars guided by artificial intelligence.

Those are a few of the ideas pitched tonight at Mercedes-Benz's Innovation Challenge, a startup competition staged this evening at the Vance plant's training center. Mercedes-Benz opened its doors to firms from around the world to dream up solutions to challenges on the factory floor. And engineers and innovators came back this evening with their own ideas. Teams were split into startup companies and student & professional groups. Cash prizes will be awarded to winners of the challenge.

From video games such as Civilization to board games such as Risk, gaming offers more opportunities for world domination than you can rattle a sabre at. If you're a would-be Caesar, Napoleon or Vladimir Putin, there are countless releases that let you rule the globe with an iron fist. These games of conquest and diplomacy tend to be long and complicated, though, and while they make for some deep, absorbing contests, they can consume entire evenings in mammoth, multi-hour play sessions. Enter CIV: Carta Impera Victoria – an empire-building card game that aims to squeeze centuries of war, trade and politics into just 20 minutes. Unlike most releases in the genre, it comes without a giant board or masses of plastic soldiers.

Being able to go from idea to result with the least possible delay is key to doing good research. So yesterday someone told me you can build a (deep) neural network in 15 minutes in Keras. Of course, I didn't believe that at all. Last time I tried (maybe 2 years ago?) it was still quite some work, involving comprehensive knowledge of programming and math. That was some dead serious craftmanship.

Innovations with smartphones and computers may have flatlined in recent years, but 2016 has seen major leaps forward in other areas.