Large language models (LLMs) have shown impressive capabilities, but still struggle with complex reasoning tasks requiring multiple steps. While prompt-based methods like Chain-of-Thought (CoT) can improve LLM reasoning at inference time, optimizing reasoning capabilities during training remains challenging. We introduce LaTent Reasoning Optimization (LaTRO), a principled framework that formulates reasoning as sampling from a latent distribution and optimizes it via variational approaches. LaTRO enables LLMs to concurrently improve both their reasoning process and ability to evaluate reasoning quality, without requiring external feedback or reward models. We validate LaTRO through experiments on GSM8K and ARC-Challenge datasets using multiple model architectures. On GSM8K, LaTRO improves zero-shot accuracy by an average of 12.5% over base models and 9.6% over supervised fine-tuning across Phi-3.5-mini, Mistral-7B, and Llama-3.1-8B. Our findings suggest that pre-trained LLMs possess latent reasoning capabilities that can be unlocked and enhanced through our proposed optimization approach in a self-improvement manner. The code of LaTRO is available at \url{https://github.com/SalesforceAIResearch/LaTRO}.

Preface: Three decades ago while working at Air Force Research Laboratory, I developed the first interactive Augmented Reality system, enabling users to reach out and touch a mixed world of real and virtual objects. I was so inspired by the reactions people had when they tried those early prototypes, I founded one of the first VR companies in 1993, Immersion Corp, and later founded the early AR company, Outland Research. Yes, I've been a believer for a long time. Looking forward, I expect augmented reality to become the platform of our lives, replacing smartphones as our primary means of accessing digital content. I still believe in the magical potential, but also fear the negative consequences. To paint a balanced picture of what our augmented lives will be like ten years from now, I've written the short narrative below. Like any fictional forecast it will not play out exactly like this, but I'm confident that the convergence of augmented reality and artificial intelligence will make much of this portrayal come true. It was a tiny room no larger than a walk-in closet. A small woman in a crisp white lab coat stood beside a large optometry machine, its smooth black surface covered in silver dials and knobs and levers. Flipping between settings she asked, "Better or worse?" "Better," rang a voice from behind the contraption. The woman pulled the machine forward, revealing Gordon Pines, squinting as the overhead lights suddenly came on. Balding with gray stubble, he looked older than his 68 years would suggest. That's because he was tired -- exhausted from the simple act of leaving his small apartment and venturing out into the busy city. Chicago had been his home for three decades but somehow it just didn't feel familiar anymore.

After a hugely successful first day at the Deep Learning Summit and Responsible AI Summit, we were back in Montreal for day two. Sportlogiq were up first on the Deep Learning stage with Bahar Pourbabaee, Machine Learning Team Lead, discussing some of the main challenges in developing and deploying deep learning algorithms at scale. The sheer size of this scale was reiterated with Bahar suggesting that they are processing more than 60,000 sports videos from different sources, all of which include many thousands of frames. Bahar's first example of Sportloqiq's latest work depicted that of a fast moving Premier League Soccer game, with examples showing the sheer depth of analysis suggesting that both decisions and non-decisions alongside their consequences can be looked at and scrutinised, whilst individual joints of players and their lateral movement was observed under the microscope. Bahar then continued to detail some of the problems with the visual perception of their learning representation model which included player/object detection, player/team identification, state estimation and data association.

Build really cool Lego robots, and then control them with your tablet? Lego Boost ($160 on Amazon) is aimed at kids ages 7 to 12, so I was a little worried that it would be too hard to enjoy with my son--he's 5 (and a half!) and just started kindergarten. But while I had to direct the complicated builds, controlling the finished robots and vehicles with the iPad app was well within his grasp--and so much fun for both of us. Lego Boost teaches kids in two ways: The builds themselves are a lesson in mechanics, showing how the motor, gears, and pieces fit together to make a moving robot. Plus, the tablet app is full of programming challenges to direct the robot to do what you want him to do.

Just like every other kid, I used to pretend that my toys were alive, especially my Lego creations. In my mind, those studded blocks became wild animals and knights in shinning armor. One of the five tutorial models is Frankie the Cat. Frankie can drink milk, move, and even fart. If only I had Lego Boost at my disposal, then my imagination would have actually come to life.

If you've ever wished your childhood Lego creations could come to life, your dreams are now closer to reality. Lego has just unveiled a new subbrand called Boost which promises to do just that. The base set contains a combination of sensors, motors and a unique companion app that teaches kids how to code so that they can program their new robot friends. Lego's Mindstorms could let you do this too, but that's a decidedly more advanced system aimed at young adults. Boost, on the other hand, is designed for kids ages seven and up.

Gem Spa, the narrow twenty-four-hour newsstand on St. Mark's Place, has served as a nerve center for generations of beats, hippies (undeterred by a sign reading, "No Combing of Hair--By Order of Health Dept"), rockers, and punks. The other day, Lily Tomlin, who is seventy-six, stopped by in the hope of getting an egg cream. Encountering a long line of customers waiting to buy magazines and lottery tickets, her personal assistant, Paul (burly, doting), shuffled her out. "I used to live up the street--this is back in the sixties--on Fifth between Second and Third," Tomlin said. She wore a navy overcoat, a silk scarf, and sunglasses.