Tag: #llm-hallucination7 results
SRM-LoRA is a new research method presented at the ICML 2026 Workshop that uses sub-Riemannian metrics to reshape LoRA gradients. This approach effectively suppresses hallucination-prone update paths without increasing inference costs.
Slopsquatting is an emerging security threat where attackers register malicious packages hallucinated by AI coding assistants. This exploit bypasses traditional typosquatting defenses by leveraging the plausible-sounding names generated by LLMs.
Arxiv is implementing a one-year suspension policy for researchers who submit AI-generated content containing fabricated sources or errors. The move aims to maintain the integrity of the repository against the rise of 'AI slop'.
This research analyzes how AI answer engines handle conflict-related queries, revealing that models frequently hallucinate when source data is sparse. The study highlights the emerging threat of Generative Engine Optimization (GEO) in manipulating information environments.
TripAdvisor's AI-generated review summaries are inaccurately labeling unsafe hotels as 'spotless' and 'friendly'. This raises concerns about the reliability of LLM-based summarization in high-stakes consumer decision-making.
Researchers discovered that LLMs exhibit strong, version-specific biases when generating character names, often grouping them into recurring 'ensembles'. These hallucinated name clusters appear consistently across various AI-generated content, from fake papers to stock photo captions.
Fitbit整合的Gemini AI教練因提供不準確且令人困惑的健身建議而受到用戶批評。許多用戶對該功能的表現感到失望，甚至表示希望試用期儘快結束。