Large Language Models (LLMs) have exhibited remarkable reasoning capabilities, achieving impressive results across a wide range of tasks. Despite these advances, significant reasoning failures persist, occurring even in seemingly simple scenarios. To systematically understand and address these shortcomings, the authors of the paper present the first comprehensive survey dedicated to reasoning failures in LLMs.
The authors introduce a novel categorization framework that distinguishes reasoning into embodied and non-embodied types, with the latter further subdivided into informal (intuitive) and formal (logical) reasoning. In parallel, the authors classify reasoning failures along a complementary axis into three types: fundamental failures intrinsic to LLM architectures that broadly affect downstream tasks; application-specific limitations that manifest in particular domains; and robustness issues characterized by inconsistent performance across minor variations. For each reasoning failure, the authors provide a clear definition, analyze existing studies, explore root causes, and present mitigation strategies.
Read more…
Source: ARXIV, Cornell University
Sign up for the Cyber Security Review Newsletter
The latest cyber security news and insights delivered right to your inbox
Related:
- Enhancing Botnet Detection with AI using LLMs and Similarity Search
January 8, 2025
As botnets continue to evolve, so do the techniques required to detect them. While Transport Layer Security (TLS) encryption is widely adopted for secure communications, botnets leverage TLS to obscure command-and-control (C2) traffic. These malicious actors often have identifiable characteristics embedded within their TLS certificates, opening a potential pathway for advanced detection techniques. In first-of-its-kind research, ...
- AI-supported spear phishing fools more than 50% of targets
January 7, 2025
One of the first things everyone predicted when artificial intelligence (AI) became more commonplace was that it would assist cybercriminals in making their phishing campaigns more effective. Now, researchers have conducted a scientific study into the effectiveness of AI supported spear phishing, and the results line up with everyone’s expectations: AI is making it easier to ...
- Cybertruck driver used ChatGPT to plan Las Vegas attack
January 7, 2025
Police found a six-page manifesto on Matthew Livelsberger’s phone and said he used ChatGPT to plan his New Year’s Day bombing at the Trump International Hotel in Las Vegas, Sheriff Kevin McMahill said at a news conference Tuesday. A few of the entries posted in the application included “How much Tannerite is equivalent to 1 pound ...
- Top AI Trends from 2024 – A Look Back
January 3, 2025
2024 may go down as the year AI stopped being a technological novelty and became—more consequentially—a Fact of Life. Big names like Microsoft, Salesforce, and Intuit built AI into mainstream enterprise solutions; specialized AI apps and services sprung up for everything from copywriting to data analysis; and governments, think tanks, and regulators poured effort into ...
- Nigeria, South Africa, Algeria top targets for cyber attacks in 2024
January 1, 2025
In the first half of 2024, Nigeria saw 2,721 incidents, with the telecom sector, computer services sector, Data processing and hosting companies, and even local beauty salons having a fair dose of the attacks respectively. At the time, experts attributed the rise in cyberattacks to digital transformation initiatives the country was carrying out such as adoption ...
- Ukraine collects vast war data trove to train AI models
December 20, 2024
As the future of warfare pivots towards artificial intelligence, Ukraine is sitting on a valuable resource: millions of hours of footage from drones which can be used to train AI models to make decisions on the battlefield. AI has been deployed by both sides on the battlefield during Russia’s invasion of Ukraine to identify targets, scanning ...