WEBVTT Kind: captions Language: en-US 00:00:00.000 --> 00:00:00.142 [MUSIC] 00:00:00.142 --> 00:00:04.160 SOWBHAGYA SURESH: Hi, I’m Sowbhagya Suresh.  Welcome to Microsoft Research Focus. 00:00:04.160 --> 00:00:08.400 You've probably read our newsletter. And now  you can get highlights from across the research   00:00:08.400 --> 00:00:14.400 community at Microsoft here, on YouTube, and  major podcast platforms. Whether you read,   00:00:14.400 --> 00:00:19.760 watch, or listen to it, Research Focus is  your source for research news at Microsoft. 00:00:19.760 --> 00:00:25.280 Coming up in this edition: what everyday  AI use reveals about the way we work;   00:00:25.280 --> 00:00:30.480 how a new AI model could help scientists  better understand proteins; a new approach   00:00:30.480 --> 00:00:39.920 to improving weather forecasts weeks in  advance; and more. So here's what's new. 00:00:44.800 --> 00:00:48.466 More and more, AI agents are  handling tasks that are complex. 00:00:48.466 --> 00:00:48.480 [MUSIC FADES] 00:00:48.480 --> 00:00:53.920 Tasks that require multiple steps, decisions,  and interactions with external tools. But how   00:00:53.920 --> 00:00:58.080 can we count on AI agents to do the work  if their ability to adjust when things   00:00:58.080 --> 00:01:03.280 go wrong may require a lot of resources  with no guarantees of a better result? 00:01:03.280 --> 00:01:08.800 A new self-improving framework called ForkPilot  learns when it's worth revisiting earlier   00:01:08.800 --> 00:01:14.080 decisions to find a different approach and  adapts its strategy based on new outcomes. 00:01:14.080 --> 00:01:19.200 Tested across six benchmarks and leading  AI models, ForkPilot measures up to the   00:01:19.200 --> 00:01:24.640 best existing methods while reducing  AI processing demands by up to 59%.   00:01:24.640 --> 00:01:30.080 These results suggest that AI agents can solve  complex problems more reliably and efficiently,   00:01:30.080 --> 00:01:35.341 making them more practical and  cost-effective to deploy at scale. 00:01:35.341 --> 00:01:35.360 [TRANSITION MUSIC] 00:01:35.360 --> 00:01:41.440 Coding agents execute user queries by reasoning  about, editing, and running code. Understanding   00:01:41.440 --> 00:01:47.280 why these agents fail is critical if we want to  build more capable and reliable coding agents.   00:01:47.280 --> 00:01:52.000 Unfortunately, existing benchmarks  often make it difficult to identify   00:01:52.000 --> 00:01:54.800 which factors contribute to success or failure. 00:01:54.800 --> 00:01:58.880 CABRA is a framework for generating  coding challenges from scratch. It   00:01:58.880 --> 00:02:02.960 lets researchers increase complexity  in a controlled way and pinpoints   00:02:02.960 --> 00:02:05.520 the skills that are most difficult for agents. 00:02:05.520 --> 00:02:10.800 Across thousands of generated challenges, the  researchers found that agents often overcame   00:02:10.800 --> 00:02:16.480 coding problems by using tools. Yet the agents  still struggled when they needed to understand   00:02:16.480 --> 00:02:22.320 unfamiliar codebases. The findings suggest that  understanding existing code, rather than modifying   00:02:22.320 --> 00:02:29.341 it, remains one of the biggest obstacles  to developing more effective coding agents. 00:02:29.341 --> 00:02:29.360 [TRANSITION MUSIC] 00:02:29.360 --> 00:02:33.840 Computer-use agents are AI systems that  are designed to complete tasks that   00:02:33.840 --> 00:02:39.680 require clicking and typing. They can fill in  forms, process requests, and navigate websites.   00:02:39.680 --> 00:02:44.240 To do this effectively, they need to learn  the procedures that people use to get work   00:02:44.240 --> 00:02:48.960 done. User activity logs can offer a way  to capture this knowledge. But activity   00:02:48.960 --> 00:02:54.320 logs typically don't record the intent  behind each action and often combine tasks. 00:02:54.320 --> 00:02:58.720 The new framework TeleTune learns  reusable skills and workflows from   00:02:58.720 --> 00:03:04.640 these logs. It also infers the task goals  associated with the sequences of actions. 00:03:04.640 --> 00:03:09.120 In evaluations across software and  browser tasks, TeleTune outperformed   00:03:09.120 --> 00:03:14.880 existing approaches. By learning from existing  activity, TeleTune offers a scalable way to   00:03:14.880 --> 00:03:22.701 help AI agents adapt to unfamiliar software  and carry out complex, multi-step workflows 00:03:22.701 --> 00:03:22.720 [TRANSITION MUSIC] 00:03:22.720 --> 00:03:27.520 Workers across industries may appear to  be using AI in surprisingly similar ways,   00:03:27.520 --> 00:03:30.640 but a closer look reveals a different story. 00:03:30.640 --> 00:03:35.520 Microsoft researchers analyzed aggregated  Copilot Chat activity across a broad range   00:03:35.520 --> 00:03:40.000 of customer categories and found that  editing documents led all classified   00:03:40.000 --> 00:03:42.400 activities, accounting for more than 20 percent. 00:03:42.400 --> 00:03:47.120 But when researchers set aside the 20 most  common activities, differences between   00:03:47.120 --> 00:03:52.000 customer categories became substantially  more pronounced. The most common uses of   00:03:52.000 --> 00:03:57.360 AI may be hiding important differences in how  organizations use it for more specialized work. 00:03:57.360 --> 00:04:02.240 For business leaders, the takeaway is clear:  don’t assume you have the same AI needs as   00:04:02.240 --> 00:04:07.360 others in your industry. Finding the right  opportunities means looking beyond usage   00:04:07.360 --> 00:04:15.261 statistics to understand what employees actually  do, how they work, and what outcomes matter most. 00:04:15.261 --> 00:04:15.280 [TRANSITION MUSIC] 00:04:15.280 --> 00:04:20.320 A new study shows how Shiksha Copilot—a  human-in-the-loop AI tool—could transform   00:04:20.320 --> 00:04:23.040 teacher workflows in Global South classrooms. 00:04:23.040 --> 00:04:26.480 The study in low-resource schools  in India found the tool reduced   00:04:26.480 --> 00:04:31.280 the time teachers spent on bilingual  lesson planning and bureaucracy. That   00:04:31.280 --> 00:04:35.040 gave them more time to craft more  engaging classroom activities. 00:04:35.040 --> 00:04:40.480 More than a thousand teachers participated, and  many reported reduced stress levels and a positive   00:04:40.480 --> 00:04:45.840 impression of how AI can help them. These teachers  actively engaged with Shiksha Copilot, using their   00:04:45.840 --> 00:04:51.360 professional expertise to adapt and repurpose  materials to suit their own classrooms. 00:04:51.360 --> 00:04:52.720 [TRANSITION MUSIC] 00:04:52.720 --> 00:04:59.200 Ten-day weather forecasts have become remarkably  dependable. But what about the weeks that follow? 00:04:59.200 --> 00:05:03.360 The two-to-six-week window remains a  major challenge for forecasting models,   00:05:03.360 --> 00:05:07.440 which is a problem. That’s when a wide  range of critical decisions get made   00:05:07.440 --> 00:05:11.760 related to farming, managing wildfires,  and preparing for extreme weather. 00:05:11.760 --> 00:05:16.720 A new machine-learning framework called  Probabilistic Bias Correction, or PBC, is   00:05:16.720 --> 00:05:22.000 helping close that gap. PBC learns from a weather  model's past mistakes and helps correct them. 00:05:22.000 --> 00:05:24.000 And the early results are promising. 00:05:24.000 --> 00:05:27.840 PBC took top honors in a global  competition focused on improving   00:05:27.840 --> 00:05:31.440 extended-range forecasting. It's  also being deployed in Kenya,   00:05:31.440 --> 00:05:35.040 where it's helping improve critical  monsoon forecasts for farmers. 00:05:35.040 --> 00:05:36.880 This development could help make weather   00:05:36.880 --> 00:05:41.440 predictions more useful across  a broad range of applications. 00:05:42.800 --> 00:05:43.600 [TRANSITION MUSIC] 00:05:43.600 --> 00:05:47.200 BioEmu2, a new foundation model for  understanding biomolecular structure,   00:05:47.200 --> 00:05:50.880 dynamics, and function, is now  available in public preview. 00:05:50.880 --> 00:05:56.320 Developed by the Microsoft Research AI for  Science team, BioEmu2 builds on Biomolecular   00:05:56.320 --> 00:06:01.920 Emulator, a deep-learning model designed to help  scientists better understand how proteins work. 00:06:01.920 --> 00:06:06.240 That understanding could have important  implications for medicine. By learning   00:06:06.240 --> 00:06:11.600 more about how proteins function, scientists  could move closer to developing more effective   00:06:11.600 --> 00:06:17.040 medications that enhance beneficial protein  activity or prevent proteins from causing harm. 00:06:17.040 --> 00:06:22.840 Go to the BioEmu2 Microsoft Research GitHub  repository to explore the model and learn more. 00:06:22.840 --> 00:06:23.760 [MUSIC] 00:06:23.760 --> 00:06:25.520 And that’s your Research Focus. 00:06:25.520 --> 00:06:28.800 From more capable AI agents  to better weather forecasting,   00:06:28.800 --> 00:06:32.640 these are just a few of the developments  researchers are exploring at Microsoft. 00:06:32.640 --> 00:06:36.720 Want to learn more and stay connected?  Explore the latest research, discover   00:06:36.720 --> 00:06:40.720 the people behind it, and keep tabs  on the ideas shaping what's next by   00:06:40.720 --> 00:06:43.600 following Microsoft Research  on major social platforms. 00:06:43.600 --> 00:06:45.760 I'm  Sowbhagya Suresh. Thanks so much for tuning   00:06:45.760 --> 00:06:51.520 in. We’ll see you in the next  Microsoft Research Focus. 00:06:51.520 --> 00:06:52.362 [MUSIC FADES]