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:
- Anthropic ditches its core safety promise in the middle of an AI red line fight with the Pentagon
February 25, 2026
Anthropic, a company founded by OpenAI exiles worried about the dangers of AI, is loosening its core safety principle in response to competition. Instead of self-imposed guardrails constraining its development of AI models, Anthropic is adopting a nonbinding safety framework that it says can and will change. In a blog post Tuesday outlining its new policy, ...
- AWS says more than 600 FortiGate firewalls hit in AI-augmented campaign
February 23, 2026
Cybercriminals armed with off-the-shelf generative AI tools compromised more than 600 internet-exposed FortiGate firewalls across 55 countries in just over a month, according to a new incident report from AWS. The campaign, which ran from mid-January to mid-February, relied less on clever zero-days and more on the equivalent of trying every digital door handle – just ...
- OpenClaw AI agents targeted by infostealer malware for the first time
February 17, 2026
Thanks to its overnight success and widespread adoption, OpenClaw has painted a large target on its back and is now being attacked by infostealers, after security researchers Hudson Rock claimed to have seen a first-of-its-kind attack in the wild. OpenClaw (previously known as Clawdbot and Moltbot) is an open source AI assistant software designed to actually ...
- EU Parliament blocks AI tools over cyber, privacy fears
February 16, 2026
he European Parliament has disabled AI features on the work devices of lawmakers and their staff over cybersecurity and data protection concerns, according to an internal email seen by POLITICO. The chamber emailed its members on Monday to say it had disabled “built-in artificial intelligence features” on corporate tablets after its IT department assessed it couldn’t ...
- UNC1069 Targets Cryptocurrency Sector with New Tooling and AI-Enabled Social Engineering
February 9, 2026
North Korean threat actors continue to evolve their tradecraft to target the cryptocurrency and decentralized finance (DeFi) verticals. Mandiant recently investigated an intrusion targeting a FinTech entity within this sector, attributed to UNC1069, a financially motivated threat actor active since at least 2018. This investigation revealed a tailored intrusion resulting in the deployment of seven unique ...
- Viral AI, Invisible Risks: What OpenClaw Reveals About Agentic Assistants
February 6, 2026
The name OpenClaw might not immediately be recognizable, partly because it has undergone several name changes, from Clawdbot to Moltbot, then finally to OpenClaw. Yet one thing is certain: This new digital assistant feels genuinely groundbreaking. It remembers past interactions, keeps data on the user’s device, and adapts to individual preferences, making it feel like a ...