ニュース

I’m not the sort of person who typically pays for a house cleaner. I’m too frugal, and I kind of enjoy putting on a podcast and going to town on my tub with a Magic Eraser. But when I saw an X post about a new humanoid-robot cleaning service in San Francisco, I immediately signed up for the waitlist. I wanted my apartment to feel like The Jetsons, even if just for an hour. I was also motivated by journalistic interest. In order to ensure an authentic experience, I didn’t tell Tau Robotics, the startup behind the X post, that I was a journalist—though I did use my real name, so it’s possible someone looked me up. Tau’s long-term goal is to make autonomous cleaning robots a reality, but they’re not quite there yet: For now, the robots are piloted remotely by humans, who watch what’s going on via cameras mounted in the robots’ heads and wrists. As the robots clean, they collect data that Tau will use to train up their more independent descendants. Every counter successfully wiped and T-shirt successfully folded will help the models that power the robots learn to do better on the next go-round. Because this data is so valuable to Tau, the company charges less for cleaning than one might expect for such a cutting-edge product. At the time I booked, Tau was offering $30 for an hourlong cleaning with a single robot; now the charge is $60 per weekly cleaning with two robots, or $100 for a one-time visit. But even at the cheaper rate of $30 per hour, Tau’s robots aren’t a great bargain. They are thorough but slow, so if you, like me, are tempted to stand around watching them for their entire visit, it would probably be a better use of your time to just take care of the cleaning yourself.  But the robots are physically capable of cleaning, and they’re undeniably cute. Tau has armed each four-and-a-half-foot-tall robot with a 3D-printed “head” with antennae that allow it to communicate with Tau’s headquarters, as well as two 3D-printed claw “hands” (the more humanlike alternatives apparently aren’t durable enough, and claws do surprisingly well for most cleaning tasks). Mine was wearing a Tau Robotics shirt, and grippy socks so that it didn’t scratch hardwood floors. With all the human support they need, the robots are closer to radio-controlled cars than Waymos. Many tasks, such as making the bed and folding laundry, require what’s called kinematic guidance: The human pilot performs the necessary motions, and sensors transmit those movements to the robot, which copies them. These kinematically guided tasks produce especially valuable training data. Tau’s AI models might not have solved laundry folding yet, but learning from the examples of robots that have completed the task in the real world is how they’ll get there. But according to Tau cofounder Cornelia Weinzierl, to whom I came clean as a journalist a few weeks after the cleaning, the robot can already complete some other tasks—such as vacuuming and mopping—without any human help. If Tau succeeds in using all the data it is collecting to train truly autonomous cleaning robots, they could even become San Francisco’s newest tech tourist attraction as Waymos become more commonplace countrywide. The cleaning robot that came to my building last month was already a bit of a spectacle. It arrived in a blue minivan, out of which it was helped by Weinzierl and another Tau employee. That’s when the robot encountered its first obstacle: the eight stairs leading up to the building’s front door. I’d come down to greet Weinzierl and the robot, so I had a front-row seat to its struggles with my steps. It climbed as if it were drunk. Each successful step was as likely as not to be followed by a stumbling backward stagger that erased all its progress—though, impressively, it never entirely lost its footing. When I later discussed my experience with Weinzierl, she explained that the robot sometimes walks strangely because it is operated remotely from Tau’s headquarters, and the short signal transmission delay can affect its coordination. “The amount of work and time that we spend just optimizing, for example, timings and connectivity is crazy,” she said. Inside, the carpeted staircase leading up to my apartment defeated the robot entirely, and Weinzierl had to lug it up the last few steps. This, it turned out, was not typical. As the robot cleaned, it repeatedly stalled out, something Weinzierl attributed to it overheating in the unseasonably hot weather—until she heard from one of her colleagues that the heat had shut down not the robot but the equipment that its human controller was using to watch its progress through my apartment. Without remote assistance, the robot couldn’t clean, and so it, and Weinzierl, departed before my booked hour was up. (Back in September, Weinzierl or another Tau representative remained with the robot throughout its visit, but she says that’s no longer necessary.) You can’t fault Tau on customer service, though. Weinzierl quickly emailed me to apologize, comp the visit, and offer me another on the house (minus whatever price I assign to my data privacy)—and that was all before I told her I was a journalist. I scheduled the next cleaning for about a month later, after a couple of trips out of town, and it went much more smoothly—the robot got up my stairs without much assistance, though still with the drunken stagger. In the hour it was there, it did a creditable job of cleaning my kitchen (though it didn’t brave the stove) and even saved me my most dreaded cleaning task, the toilet. That was its most impressive performance: Three times the head fell off its toilet cleaning wand, and three times it managed to reattach it by poking around with the handle until it made contact at just the right angle. The robot is still rough around the edges, which is to be expected for a product in beta.

arXiv:2609.30906v2 Announce Type: replace Abstract: Large language models (LLMs) excel at natural language processing but struggle to interact with external environments. Tool learning provides a promising way to extend LLMs into actionable agents, where tool selection is a critical prerequisite for successful tool use. Existing work often assumes a small or predefined set of tools, leaving large-scale tool selection underexplored. Real-world repositories contain a vast and diverse array of tools, making it difficult for LLMs to effectively search, distinguish, and compose tools under context-length constraints. We identify large-scale tool selection as a new challenge for agentic reinforcement learning, highlighting that existing RL methods for knowledge-based question answering are inadequate for selecting tools while considering compatibility. To address this challenge, we propose ToolSearcher, a novel RL framework for effective multi-turn search and fine-grained optimization in large-scale tool selection. Specifically, we introduce category-constrained tool discrimination to improve the model’s ability to distinguish functionally similar tools, event-level search modeling to explicitly optimize the discovery of target tools during multi-turn search, and trajectory-aligned credit allocation to provide fine-grained reward signals for different stages of the search-selection process. Extensive experiments on large-scale tool selection benchmarks demonstrate that ToolSearcher consistently outperforms a set of strong baselines in challenging settings involving iterative search and complex tool composition.

arXiv:2611.07845v1 Announce Type: cross Abstract: Masked diffusion models generate text by reversing a stochastic masking process. Unlike autoregressive LLMs, which generate text one token at a time, they can reveal many tokens in parallel, at the price of an error, since tokens revealed together are drawn independently. We show that this trade-off maps onto the statistical physics of driven many-particle systems. (i) For any data, the error of parallel generation is the entropy production of the generation process relative to the reversed masking process. (ii) For Markov data, generation is a random sequential adsorption process with labelled gaps, which we solve exactly for any number of steps. (iii) For many steps, the error becomes the dissipation of a slowly driven system: the unmasking schedule plays the role of the driving protocol, the number of steps that of its duration, and the error is set by a thermodynamic metric that we obtain in closed form from the mutual information between tokens. Optimal schedules are geodesics of this metric. The exact solution shows how correlations in the text shape both the error and the quality of the output. The longer the correlations, the more the error concentrates at the start of generation. Schedules that start with small steps then gain more, both in path error and in perplexity, and need fewer steps to reach their asymptotic behaviour.

A sweeping bit of legislation designed to cut red tape could speed up the buildout of energy infrastructure in the US. With the dust barely settled on this week’s momentous midterms, the Senate is set to vote on the Bipartisan American Affordability and Jobs Act (BAAJA) on November 9. It’s a plan that could finally start an overhaul of the slow, complicated process for getting energy projects approved, as part of something called permitting reform. People on both sides of the political aisle have been pushing for this type of change for some years.  The key question is whether this will cut greenhouse-gas emissions from the US energy sector or push them higher still.   The federal permitting system is designed to protect the environment against industrial pollution from large projects like power plants, transmission lines, and natural-gas pipelines. Today, permitting is governed by a range of laws, including the National Environmental Policy Act and the Clean Water Act. And if there’s one thing you might know about the current state of permitting in the US it’s that it’s slow and incredibly complicated. So in its 417 pages, BAAJA attempts to tighten things up. The bill has provisions that narrow the sorts of projects that need to apply for federal permits and others that enact deadlines in hopes of speeding up a process that today can take years. (There are also specific provisions about transmission planning, mining, and the rest of the kitchen sink; this thing is dense.) One thing to know about this bill is that it’s largely technology neutral: It opens the doors for the buildout of projects involving both fossil fuels and renewables. So there’s a weird mishmash of legislators and outside groups supporting it. The bill’s four sponsors in the Senate are two Republicans and two Democrats. The White House has backed this permitting reform bill, with its Council of Economic Advisors finding that it would reduce review times and have economic benefits like boosting GDP. Clean energy groups like the American Clean Power Association have largely come out in favor of the bill too, with the group’s analysis finding that it could unlock 350 gigawatts of new clean energy for the grid. Some environmental groups, however, oppose the bill, or parts of it. An assessment from the Natural Resources Defense Council says the parts about transmission could have substantial benefits for the grid, but the narrowing of environmental reviews could be damaging for public and private lands. Many in the climate world have long been calling for permitting reform, arguing that renewables stand to benefit more than fossil fuels. The thinking is that solar and wind power are some of the cheapest sources available today, so if we unlock the ability to build more energy, they are likely to win out economically. A few early analyses have tried to quantify just what BAAJA would mean for emissions in the US. One, from the think tank Rhodium Group, found that in the best-case scenario, the bill would cut economy-wide emissions by 4%, reduce annual energy bills by $78 per household, and add 200 gigawatts of new clean energy by 2040. In the worst case, emissions would stay about the same. Another report, from the nonprofit research group Resources for the Future, found that the bill could cut power-sector emissions 6% by 2041 and save more than $350 billion on system costs by midcentury. But there are some concerns that this permitting bill’s real effect would be to open the floodgates for new natural-gas plants. As data centers have put more pressure on the grid, a wave of such plants have been announced. “You get down to a trade-off between essentially growing the clean electricity industry on the one hand and then, as a compromise, growing the fossil-fuel industry on the other,” says Kevin Kircher, an energy modeler and assistant professor of mechanical engineering at Purdue. To quantify that trade-off, Kircher also modeled the potential emissions effects of the bill and found that if BAAJA had been in place in 2025, energy sector emissions would have been higher than the actual baseline. Health and mortality risks associated with air pollution would also have increased. This bill is scheduled for a Senate vote on November 9, well before the next Congress takes control. It’s not clear what the dramatic victories for Democrats might mean for the legislation, which was expected to pass the Senate with largely Republican support if a few Democrats got on board. Democrats could decide to wait for next year, when they’ll have control of both chambers. But any deal would ultimately need to win Trump’s signature, limiting what a future bill might be able to change.  In a world where permitting new projects is easier, we’ll see more clean energy, but we’ll see more natural gas, too. And while prices could get lower, that could push demand for electricity higher, leading to more emissions overall. Simplifying the permitting system and cutting red tape is a worthy goal. But when it comes to emissions, it still really matters what, exactly, you want to build.

arXiv:2608.10459v2 Announce Type: replace Abstract: As LLM-generated content becomes more sophisticated, detection systems for distinguishing those texts from human-written text must operate at scale while handling diverse writing styles, domains, languages, and generator models. Input-only encoder detectors are suitable for practical deployment setting, but standard binary classification supplies only the class label and does not explicitly organize the substantial variation within either class. We propose MD-ProTector, which represents each class with multiple trainable reference vectors in the encoder embedding space, referred to as prototypes. These prototypes provide separate decision boundaries for different groups of texts within the same class. However, adding multiple prototypes alone does not determine which variation each prototype should represent. MD-ProTector addresses this problem with Prototype Positioning loss, which separates class-level structure from the within-class variation that differentiates individual prototypes. Evaluated across five settings from three large-scale benchmarks covering domain, generator, language, and adversarial variation, MD-ProTector achieves the highest AvgRec on MAGE CDCM and RAID and the highest AUROC and lowest FPR95 on RAID among the compared encoder-based methods.