Multi-agent debate - multiple instances of large language models discussing problems in turn-based interaction - has shown promise for solving knowledge and reasoning tasks. However, these methods show limitations, particularly when scaling them to longer reasoning chains. In this study, we unveil a new issue of multi-agent debate: discussions drift away from the initial problem over multiple turns. We define this phenomenon as problem drift and quantify its presence across ten tasks (i.e., three generative, three knowledge, three reasoning, and one instruction-following task). To identify the reasons for this issue, we perform a human study with eight experts on discussions suffering from problem drift, who find the most common issues are a lack of progress (35% of cases), low-quality feedback (26% of cases), and a lack of clarity (25% of cases). To systematically address the issue of problem drift, we propose DRIFTJudge, a method based on LLM-as-a-judge, to detect problem drift at test-time. We further propose DRIFTPolicy, a method to mitigate 31% of problem drift cases. Our study can be seen as a first step to understanding a key limitation of multi-agent debate, highlighting pathways for improving their effectiveness in the future.

A vast burn scar unfolds in drone footage of a landscape seared by massive wildfires north of Lake Tahoe. But amid the expanses of torched trees and gray soil, an unburnt island of lush green emerges. The patch of greenery was painstakingly engineered. A creek had been dammed, creating ponds that slowed the flow of water so the surrounding earth had more time to sop it up. A weblike system of canals helped spread that moisture through the floodplain.

This story was originally published by Wired and is reproduced here as part of the Climate Desk collaboration. For the first time in four centuries, it's good to be a beaver. Long persecuted for their pelts and reviled as pests, the dam-building rodents are today hailed by scientists as ecological saviors. Their ponds and wetlands store water in the face of drought, filter out pollutants, furnish habitat for endangered species, and fight wildfires. In California, Castor canadensis is so prized that the state recently committed millions to its restoration.

For the first time in four centuries, it's good to be a beaver. Long persecuted for their pelts and reviled as pests, the dam-building rodents are today hailed by scientists as ecological saviors. Their ponds and wetlands store water in the face of drought, filter out pollutants, furnish habitat for endangered species, and fight wildfires. In California, Castor canadensis is so prized that the state recently committed millions to its restoration. While beavers' benefits are indisputable, however, our knowledge remains riddled with gaps.

Starting today, Uber Eats customers in Fairfax, Virginia can get their next meal delivered by a robot. In an expansion of the company's existing partnership with Cartken, Uber has begun offering automated deliveries in the city's Mosaic District. With today's announcement, a select number of the more than 40 restaurants in the area have begun transporting their food aboard Cartken's six-wheeled robots. Among the restaurants participating in the pilot include Our Mom Eugenia, Pupatella and RASA. Uber has been testing autonomous delivery robots in a handful of markets throughout the US.

Fox News Flash top entertainment and celebrity headlines are here. Check out what clicked this week in entertainment. Amber Heard's former personal assistant accused the "Aquaman" actress of creating an abusive work environment in a videotaped deposition played Thursday during actor Johnny Depp's defamation trial against his ex-wife in Fairfax, Virginia. Depp, 58, is suing Heard, 35, for $50 million over an op-ed she wrote for the Washington Post alleging she was the victim of domestic abuse. Heard never identified Depp directly, but attorneys for the "Pirates of the Caribbean" actor say her allegations have negatively impacted Depp's career and relationship with his family.

UK-based security technology company Sophos (OTC:SPHHF) (LSE: SOPH) has announced an agreement to acquire Invincea's software business for up to $120 million in total consideration. Invincea has developed security software that utilizes machine learning and behavioral modeling to quickly detect and react to endpoint, perimeter threats and malware attacks. The technology is a necessary upgrade to the Sophos Central product line and also provides middle market focused Sophos with a foothold in the small business market through the addition of Invincea's customer base. Fairfax, Virginia-based Invincea was founded in 2009 by CEO Anup Ghosh to focus on endpoint threat detection, prevention, and analysis. The company provides security software to more than 25,000 small and medium enterprise customers through its antivirus'X' system that'combines machine learning and behavioral monitoring to eliminate endpoint security blind spots without sacrificing usability.'

Fairfax, VA, August 25 2016 – Invincea, the leader in machine learning for endpoint protection, announced today that its deep learning model for analyzing unknown malware is now fully integrated in the VirusTotal site. VirusTotal is a "service that analyzes suspicious files and URLs and facilitates the quick detection of viruses, worms, Trojans, and all kinds of malware"[1]. By integrating with VirusTotal, Invincea is pushing machine learning into mainstream cyber security solutions. By participating in the VirusTotal community, Invincea is continuing to fulfill their three key principles to security market transparency and accountability: participate in independent 3rd party testing, work towards commonly accepted standards, and avoid being a black box. As part of these principles, Invincea became one of the first next gen endpoint security companies to join the Anti-Malware Testing Standards Organization (AMTSO) in June.