Sampling-based search, a simple paradigm for utilizing test-time compute, involves generating multiple candidate responses and selecting the best one--typically by having models self-verify each response for correctness. In this paper, we study the scaling trends governing sampling-based search. Among our findings is that simply scaling up a minimalist implementation of sampling-based search, using only random sampling and direct self-verification, provides a practical inference method that, for example, elevates the reasoning capabilities of Gemini v1.5 Pro above that of o1-Preview on popular benchmarks. We partially attribute the scalability of sampling-based search to a phenomenon of implicit scaling, where sampling a larger pool of responses in turn improves self-verification accuracy. We further identify two useful principles for improving self-verification capabilities with test-time compute: (1) comparing across responses provides helpful signals about the locations of errors and hallucinations, and (2) different model output styles are useful for different contexts--chains of thought are useful for reasoning but harder to verify. We also find that, though accurate verification can be elicited, frontier models demonstrate remarkably weak out-of-box verification capabilities and introduce a benchmark to measure progress on these deficiencies.

Time is running out inside the Gamebox, and you've got to jump or die. You're partnered up with your friends, who must agree on questions about a series of images projected to the left and right of you. Each of you wears a visor with sensors on top of it; the box can tell where you are and how you move. In this challenge, called "Glass Bridge," your team must decide, as seconds tick away, the answer to questions like, "Which side had the most birds?" To vote, teammates jump on circles representing the left or right side.

Here is a brief, not-nearly-complete list of Warner Bros. characters that appear in the movie Space Jam: A New Legacy: Harry Potter, Harley Quinn, Rick & Morty, Yogi Bear, Fred Flintstone, Space Ghost, the Matrix, Superman, Batman, King Kong, the Pink Panther, Pennywise the killer clown, the droogs from A Clockwork Orange, the Night King from Game of Thrones, and Rosey, the robot maid from The Jetsons. The complete roster runs to well over 100 entries, but this sampling should be enough to give you the flavor of what a random grab-bag of intellectual properties the movie presents. If the first Space Jam, released 25 years ago, was a brand summit between the Looney Tunes and the NBA, with Michael Jordan acting as the chief negotiator, its supercharged successor both literally and figuratively opens the vaults, zapping LeBron James into the "Warner 3000 Serververse," where all of the media conglomerate's holdings exist on the same plane. A New Legacy's villain and chief instigator is Don Cheadle's Al G. Rhythm, a Warner Bros. algorithm determined to get public recognition for his overlooked accomplishments. But what's noteworthy about the movie's garbage-dump of WB properties is just how arbitrary and non-algorithmic it feels. There's no apparent logic to what's included and what's left out, who makes the cut and who gets left to molder in some forgotten corner of the digital domain.

Watching movies is one of the most popular entertainments among people. Every year, a huge amount of money goes to the movie industry to release movies to the market. In this paper, we propose a multimodal model to predict the likability of movies using textual, visual and product features. With the help of these features, we capture different aspects of movies and feed them as inputs to binary and multi-class classification and regression models to predict IMDB rating of movies at early steps of production. We also propose our own dataset consisting of about 15000 movie subtitles along with their metadata and poster images. We achieve 76% and 63% weighted F-score for binary and multiclass classification respectively, and 0.7 mean square error for the regression model.

The poor quality of drive-thru ordering may be an old joke (and a staple of comedy movies), but it's also a problem that could benefit from a high-tech overhaul. Machine learning and voice recognition can ease the many pain points of this encounter, contends Denver technology entrepreneur Rob Carpenter, the CEO of Valyant AI. Carpenter's company has developed an artificial intelligence platform that automates fast-food customer service, order-ahead, drive-thru and in-store sales, with technology in development to integrate more directly with point of sale systems. Valyant AI was a recent finalist at a developer program at Visa, and is reportedly in discussions with McDonald's, Walmart and advisors from Yum Brands. Carpenter did not identify his clients, saying the first deployment would come in about four weeks.

New research can predict how plots, images, and music affect your emotions while watching a movie. Every time I think AI can't surprise me anymore, new research arrives to prove me wrong. Yesterday, scientists at the MIT Media Lab announced that they've taught a machine how to manipulate our emotions–a technology that they believe can help filmmakers create more engrossing movies and TV. In a blog post published in collaboration with strategic consulting firm McKinsey & Company, the researchers said that they used a deep neural network to watch thousands of small slices of video--movies, TV, and short online features. For each slice, the neural network guessed which were the different elements that made a moment emotionally special, constructing an emotional arc. To test their accuracy, the team got human volunteers to watch the same clips, tagging their reaction and labeling which elements--from the music to the dialogue to the type of imagery shown on screen–had a stronger weight in their emotional response.

Physical photographs aren't the only way Firewatch will invade the real world. Developer Campo Santo recently revealed a partnership with production house Good Universe (Neighbors and Last Vegas) to make a movie based on the indie game about a fire lookout in a Wyoming forest, according to The Hollywood Reporter. No other details are available at the time, but fingers crossed that some enterprising Ford dealership doesn't repurpose the movie's eventual trailer for a summer sales event. Miss the game when it came out earlier this year on PlayStation 4 and PC? If the reason was because you only have an Xbox One, well, now you can fix that as the game graced Microsoft's console last week -- replete with a temporarily exclusive free-roam mode.

Apple says their voice recognition app Siri is improved. USA TODAY's Jefferson Graham puts Siri through a few tests to find out. Siri's closest Indian restaurant turned out to be in San Jose, California, just a 6 hour drive from Los Angeles (Photo: Robert Hanashiro) LOS ANGELES -- Hey, Siri: if you can't find me an Indian restaurant just three miles from my office, what can you get right? With the IOS 10 mobile operating system upgrade, Apple has promised us that its five-year-old personal digital assistant Siri would be better than ever, and would expand beyond the core Apple apps to also work with third-party apps like LinkedIn, Uber and Pinterest. The results of my tests is that Siri is slightly improved, but only if you take the time to enunciate correctly and slowly, and teach the voice-activated assistant what you want.

Nintendo may be making a return to the big screen, its president revealed in an interview with Japan's Asahi Shimbun Monday. The company is exploring the potential of films based on popular video game franchises such as "Super Mario Bros." and "The Legend of Zelda." Another possibility is drawing on the material to create other digital content, a firm representative told Reuters. Nintendo is in talks with several production companies, with the goal of releasing its first new movie within the next three years. The production and its successors could introduce the firm's characters to a broader audience, which could bolster its video game business.