0 views collected around this technical thread.
The essay explores why AI systems produce hallucinations by viewing their reality as a vast fictional narrative built from human language data, arguing that their knowledge is bounded by the corpus they ingest, and reflecting on philosophical limits of language and truth.
DeepSeek’s R1 inference mode replaces linear chain‑of‑thought with a transparent, multi‑path tree reasoning system, offering layered analysis, intent understanding, memory management, emotion detection, and hallucination mitigation, illustrated through a practical example of buying authentic cigarettes and detailed technical breakdowns.
Two collaborative papers by Ant Insurance and Zhejiang University were accepted at AAAI 2025, introducing the MoLE decoding framework to reduce hallucination in large vision‑language models and the MHBench benchmark plus Motion Contrastive Decoding to address motion hallucination in video large language models, advancing reliable AI‑driven insurance claim processing.
This article explains why large language models produce plausible‑looking but false information, traces the problem to the supervised fine‑tuning stage, and outlines mitigation techniques such as knowledge interrogation, RLHF, and tool‑augmented search to reduce hallucinations.
Deploying large language models faces two critical challenges—machine deception, where AI generates plausible yet false content, and machine hallucination, where outputs are logically coherent but factually inaccurate—both undermining trust, and the article outlines their causes, impacts, and technical, ethical, and regulatory mitigation strategies.
The article explains how recent research tackles large language model hallucinations by combining retrieval‑augmented generation with reinforcement learning, achieving significant accuracy and reliability gains and paving the way for safe AI deployment in critical sectors such as finance and healthcare.
This article presents Task‑aware Decoding (TaD), a plug‑and‑play technique introduced by JD Tech and Tsinghua University and accepted at IJCAI 2024, which reduces intrinsic hallucinations in large language models by comparing pre‑ and post‑fine‑tuning outputs, and demonstrates its effectiveness combined with Retrieval‑Augmented Generation across various tasks.
TaD, a task‑aware decoding technique jointly developed by JD.com and Tsinghua University and presented at IJCAI 2024, leverages differences between pre‑ and post‑fine‑tuned LLM outputs to construct knowledge vectors, significantly reducing hallucinations across various models, tasks, and data‑scarce scenarios, especially when combined with RAG.
This article examines why large language models often produce hallucinations in industry question‑answering, defines the phenomenon, explores its data and training origins, proposes evaluation metrics, and presents practical strategies—including high‑quality fine‑tuning data, honest refusal mechanisms, advanced decoding methods, and external knowledge‑graph augmentation—to reduce hallucinations and improve reliability.
This article examines the challenges of using large language models for industry Q&A, defines hallucination phenomena, evaluates their causes and impact, and proposes a set of strategies—including high‑quality fine‑tuning data, honest alignment, advanced decoding, and external knowledge‑graph augmentation—to reduce hallucinations and improve answer reliability.
The article explains why large language models generate hallucinations—due to data errors, training conflicts, and inference uncertainty—and outlines data‑cleaning, model‑level feedback, knowledge augmentation, constraint techniques, and post‑processing methods such as the “Truth‑seeking” algorithm to mitigate the issue.
This article explains how to craft efficient prompts by combining clear instructions and questions, discusses prompt injection risks and mitigation with delimiters, addresses hallucinations, and introduces zero‑shot, few‑shot, and chain‑of‑thought prompting techniques for large language models.
The article critically examines GPT‑4 and similar transformer models, highlighting persistent hallucinations, outdated knowledge, insufficient domain coverage, lack of planning and memory, and proposes architectural extensions inspired by fast‑slow thinking and differentiable modules to overcome these fundamental constraints.