Scaling data and model size has been proven effective for boosting the performance of large language models. In addition to training-time scaling, recent studies have revealed that increasing test-time computational resources can further improve performance. In this work, we introduce Aggregation Fine-Tuning (AFT), a supervised finetuning paradigm where the model learns to synthesize multiple draft responses, referred to as proposals, into a single, refined answer, termed aggregation. At inference time, a propose-and-aggregate strategy further boosts performance by iteratively generating proposals and aggregating them. Empirical evaluations on benchmark datasets show that AFT-trained models substantially outperform standard SFT. Notably, an AFT model, fine-tuned from Llama3.1-8B-Base with only 64k data, achieves a 41.3% LC win rate on AlpacaEval 2, surpassing significantly larger LLMs such as Llama3.1-405B-Instruct and GPT4. By combining sequential refinement and parallel sampling, the propose-and-aggregate framework scales inference-time computation in a flexible manner. Overall, These findings position AFT as a promising approach to unlocking additional capabilities of LLMs without resorting to increasing data volume or model size.

Social media platforms give rise to an abundance of posts and comments on every topic imaginable. Many of these posts express opinions on various aspects of society, but their unfalsifiable nature makes them ill-suited to fact-checking pipelines. In this work, we aim to distill such posts into a small set of narratives that capture the essential claims related to a given topic. Understanding and visualizing these narratives can facilitate more informed debates on social media. As a first step towards systematically identifying the underlying narratives on social media, we introduce PAPYER, a fine-grained dataset of online comments related to hygiene in public restrooms, which contains a multitude of unfalsifiable claims. We present a human-in-the-loop pipeline that uses a combination of machine and human kernels to discover the prevailing narratives and show that this pipeline outperforms recent large transformer models and state-of-the-art unsupervised topic models.

USA Today's Jefferson Graham suggests tips on the best way to shop through Amazon's home assistant, Alexa. LOS ANGELES -- Alexa is great for listening to music and answering trivia questions, but when it comes to shopping...well, that's another story. If you have one of the classic Echo speakers, you know that there are no visuals to see the product in question and once Alexa makes suggestions, it's hard to comparison shop or get the best prices. But Amazon wants you to buy this way, and throws some special deals that can only be accessed by asking Alexa for them. On Prime Day, many specials will be Alexa-only.

It's hard to overstate Amazon's online retail dominance. With 76 percent market share of online retail, it's as if the 95-96 Chicago Bulls entered your local rec league. No one can challenge Amazon today, but a newly announced partnership between Google and Walmart--allowing you to order groceries with from the latter with Google Assistant, or online via Google Express, starting late September--may ultimately present a threat. In Walmart, Google adds a retail behemoth to its Google Express service, an online shopping bazaar in need of an anchor. In Google, Walmart gains a foothold in the voice-enabled future of commerce.

Gierad Laput, a Ph.D. student at Carnegie Mellon University, demonstrates how his team's universal sensor picks up the sound from a handheld vacuum. Gierad Laput, a Ph.D. student at Carnegie Mellon University, demonstrates how his team's universal sensor picks up the sound from a handheld vacuum. John Essey and family live in a modest, two-story home on a tree-lined street in the suburbs north of Pittsburgh. From the outside, it looks like any other house in the neighborhood, but this house has a brain. Doors unlock, [it] kinda sets the mood for the rest of the house too, turns on lights, sets the thermostat accordingly," Essey says. Essey is an engineer at Uber and an early adopter of the Internet of things. He can control his lights with his Amazon Echo or an array of touchpad sensors he's installed throughout the home. Sensors tell him when there's water in the basement or a leak under the sink. While Essey's setup might sound a little like science fiction it's a prototype of the future. Some critics are worried these devices won't be secure and that companies will use them to spy on us to make money. Gierad Laput, a Ph.D. student at Carnegie Mellon University, says as the Internet of things becomes more engrained in our daily lives, there are a couple of ways people are turning ordinary homes into smart homes. "One way is basically to buy all the appliances, smart oven, smart dishwasher, smart microwave, smart toaster, all these things," Laput says. But that stuff is really expensive. Smart refrigerators can cost $3,000 or more. And Laput said those devices don't always talk to each other, especially if they're made by different manufacturers. The other way is to get sensors, and put them on everything you want to monitor. "But then those get really unwieldy and you've got all these things sticking around and they look ugly and socially obtrusive," Laput says. So Laput and his team wanted to see if they could build just one sensor that could monitor a whole range of activity in a room. The board senses about a dozen different facets of its environment: vibrations, sounds, light color and so on. The sensor communicates wirelessly with a computer, which interprets everything it picks up. Laput demonstrated how the sensor works by turning on a blender. Laput turned on a light, and the screen said, "light on." Laput says he imagines both domestic and commercial applications for such a sensor. It could tell you that you left your stove on or that you're almost out of paper towels in the bathroom at the restaurant you own. But critics say there's a catch. "Surveillance is now the business model of the Internet.

Amazon's range of smart speakers and their artificial intelligence assistant Alexa have proved to be a huge sales hit. But the product is still a shadow of what the man in charge - Dave Limp - and indeed their owners, hope it will become. "We have thousands of engineers inside Amazon adding to [its] capability every day and then another tens of thousands of developers adding to the skills," he tells the BBC. "The thing I am sure of is that this time next year she will be significantly more intelligent than she is now, and that sometime in the future we will hit our goal of reinventing the Star Trek computer." It's a lofty goal, especially since any attempt to go beyond commanding a weather update or asking for the lights to be switched on is currently asking for trouble.

Many people worry about the impact of artificial intelligence on the world of work. They fear robots could replace workers. What if it's actually managers who are at greatest risk of replacement, leaving frontline staff members to supervise themselves? Artificial intelligence -- the kind of machines that genuinely learn -- are not yet a reality in the workplace. But augmented intelligence, the use of computing power to surface patterns in data without humans having to look at it, is already here.

Google is making a play for a bigger piece of the voice-based search market and it is looking to use its ability to drive online and offline sales to gain market share. Google today at Shoptalk announced that it using artificial intelligence to make it easier for users of its voice-activated Google Home smart speaker devices to buy online. Among the features it is rolling out are the ability to buy multiple items from multiple merchants in a single transaction through the Google Assistant via Google Express, the search giant's delivery service that's available in an app and online. A consumer also has the ability to add items to her shopping list that's accessible across multiple surfaces such as her Home device and smartphone. Today's announcement comes a little more than four months after Google launched Google Home on Nov. 4 and a little more than a month after it initially enabled consumers to shop with the Google Assistant on Google Home.

People always express a distaste for lawyers but nobody really elaborates as to why. The main reason is that lawyers know you have to use their services so they do stupid isht like charge ridiculous rates and bill you just to have a conversation with them. When we think about lawyers, we often think about that dude in the bathroom at the nightclub who is trying to hand you paper towels after you wash your hands. He's trying to interject himself into your life in hopes that you'll tip him a dollar for handing you a paper towel and offering you some cologne. He's nothing special because any other dude like him could offer the exact same service.

When I go to Google Photos and search my photos for'skyline', it finds me this picture of the New York skyline I took in August, without me having labelled it! When I search for'cathedral', Google's neural networks find me pictures of cathedrals & churches I've seen. But of course, neural networks aren't magic–nothing is! I recently read a paper, "Explaining and Harnessing Adversarial Examples", that helped demystify neural networks a little for me. The paper explains how to force a neural network to make really egregious mistakes. It does this by exploiting the fact that the network is simpler (more linear!) than you might expect. It's important to understand that this doesn't explain all (or even most) kinds of mistakes neural networks make. There are a lot of possible mistakes!