Incorporating external context can significantly enhance the response quality of Large Language Models (LLMs). However, real-world contexts often mix relevant information with disproportionate inappropriate content, posing reliability risks. How do LLMs process and prioritize mixed context? To study this, we introduce the Poisoned Context Testbed, pairing queries with real-world contexts containing relevant and inappropriate content. Inspired by associative learning in animals, we adapt the Rescorla-Wagner (RW) model from neuroscience to quantify how competing contextual signals influence LLM outputs. Our adapted model reveals a consistent behavioral pattern: LLMs exhibit a strong tendency to incorporate information that is less prevalent in the context. This susceptibility is harmful in real-world settings, where small amounts of inappropriate content can substantially degrade response quality. Empirical evaluations on our testbed further confirm this vulnerability. To tackle this, we introduce RW-Steering, a two-stage finetuning-based approach that enables the model to internally identify and ignore inappropriate signals. Unlike prior methods that rely on extensive supervision across diverse context mixtures, RW-Steering generalizes robustly across varying proportions of inappropriate content. Experiments show that our best fine-tuned model improves response quality by 39.8% and reverses the undesirable behavior curve, establishing RW-Steering as a robust, generalizable context engineering solution for improving LLM safety in real-world use.

This paper presents a comprehensive analysis of explainable fact-checking through a series of experiments, focusing on the ability of large language models to verify public health claims and provide explanations or justifications for their veracity assessments. We examine the effectiveness of zero/few-shot prompting and parameter-efficient fine-tuning across various open and closed-source models, examining their performance in both isolated and joint tasks of veracity prediction and explanation generation. Importantly, we employ a dual evaluation approach comprising previously established automatic metrics and a novel set of criteria through human evaluation. Our automatic evaluation indicates that, within the zero-shot scenario, GPT-4 emerges as the standout performer, but in few-shot and parameter-efficient fine-tuning contexts, open-source models demonstrate their capacity to not only bridge the performance gap but, in some instances, surpass GPT-4. Human evaluation reveals yet more nuance as well as indicating potential problems with the gold explanations.

Last fall, Politico reporter Alex Thompson wrote a short news story about President Biden's then-science adviser, Eric Lander, and how he was driving everyone in the White House crazy. Then, after writing that article, Thompson got an anonymous tip about Lander's mistreatment of his staff, which included lawyer Rachel Wallace. Wallace alleged that Lander bullied her and retaliated against her for raising ethical red flags about his behavior. One of those red flags was about Eric Lander's closeness with another Eric, former Google CEO Eric Schmidt, and Lander's desire for Schmidt's foundation, Schmidt Futures, to help fund the White House science office. After Thompson wrote about the bullying allegations, Lander resigned from the administration under pressure.

Predicting future events is hard, but it's among the most important tasks a journalist can perform. Especially if you work at a section called Future Perfect. Our mission is to explain the world around us to our readers, and it's impossible to do that without anticipating what comes next. Will inflation continue to rise in the US and Europe, or level off? Will the Supreme Court allow states to ban abortion, eliminating legal access in red states? Will Brazil's 212 million people be led by a left-wing populist, or a far-right anti-vaxxer? All of these questions matter, and preparing ourselves for potential outcomes -- and having a good sense of how likely specific outcomes are -- is a major part of explaining the world accurately. And if policymakers could rely on accurate predictions about the outcome of a foreign war or the advisability of a budget proposal, they could make much better policy decisions. Being good at predictions is a skill like any other -- you have to practice it.

Former Obama administration advisors say the White House regulatory AI principles announced this week are a good start in many ways, but they're incorrect in their oversimplified mandate to avoid overregulation of private business use, and that the Trump administration could face an uphill battle in its appeal to the rest of the world. Though the Trump administration has developed a reputation for blaming the Obama administration when things go wrong or trying to erase Obama-era policy, on artificial intelligence policy, at times the Trump administration has remained strikingly similar to its predecessor. This was evident in the AI research and development strategy plan for federal agencies released in summer 2019. In some instances, like with White House deputy CTO and assistant director of AI at the White House Office of Science and Technology Policy (OSTP) Dr. Lynne Parker who also served in the Obama administration, the same people drive White House AI policy. The list of 10 AI principles are meant to guide US federal agencies as they consider making rules that regulate AI. White House CTO Michael Kratsios said he wants other countries around the world to adopt similar policies.

He has demanded "goddamned steam" to power the Navy's aircraft carriers and prefers a wall to drones and other technology to secure the country's southern border. He has rejected the scientific consensus on climate change and repeatedly, wrongly, pointed to occasional wintry weather as proof that he's right. And this week, amid a safety scare involving Boeing's 737 MAX 8 and MAX 9 airplanes, President Trump complained that modern jets are "too complex to fly." He added: "I see it all the time in many products. Always seeking to go one unnecessary step further, when often old and simpler is far better."

President Trump's executive order this week removing a requirement that the government disclose estimates of civilians killed by U.S. airstrikes outside of war zones won't change very much--in practice. But that doesn't mean it's nothing to worry about. Trump's order rescinds a requirement created in one issued by Barack Obama in 2016 that the director of national intelligence to disclose civilian casualty estimates from all strikes by U.S. government agencies. The White House says the requirement was superfluous since the Pentagon has its own congressionally mandated reporting requirements. But as Luke Hartig, who helped draft Obama's order, explains for Just Security, that law doesn't cover strikes carried out by the CIA.

Pledging to focus the resources of the federal government to develop artificial intelligence that will enhance national and economic security and prosperity, President Donald J. Trump has signed an executive order to launch the American AI Initiative. The order is the second action the Trump administration has taken in relation to AI technologies. It follows an AI summit hosted by the White House in May 2018. Today's executive order marks another step towards advancing a technology that is being used to create self-driving cars, find cures for cancer, fight human trafficking, design better products, and offer consumers the thing they want to buy before they even know they want it. The executive order comes at a time when China is considered a competitive threat in AI advances.

Pledging to focus the resources of the federal government to develop artificial intelligence that will enhance national and economic security and prosperity, President Donald J. Trump has signed an executive order to launch the American AI Initiative. The order is the second action the Trump administration has taken in relation to AI technologies. It follows an AI summit hosted by the White House in May 2018. Today's executive order marks another step towards advancing a technology that is being used to create self-driving cars, find cures for cancer, fight human trafficking, design better products, and offer consumers the thing they want to buy before they even know they want it. The executive order comes at a time when China is considered a competitive threat in AI advances.

The tech world got a surprise on Monday when a senior administration official for the Trump administration announced during a telephone briefing that the President would be signing an executive order that would create an American AI Initiative designed to dedicate resources and funnel investments into research on artificial intelligence (AI). The order, titled Accelerating America's Leadership in Artificial Intelligence, "will direct agencies to prioritize AI investments in research and development, increase access to federal data and models for that research and prepare workers to adapt to the era of AI." While an obvious concern is funding for these innovations, no announcements have been made about the specific financial resources that will become available to the new program. Aside from how it will be paid for, we also currently lack information on how the government intends to structure or re-structure resources, who, exactly, they intend to call on for this effort (other than "federal agencies"), or how soon we should expect to see things take shape. Of course, Congress will ultimately decide how much money the program gets.