World PulseNow
Newsletter

Quantifying Edits Decay in Fine-tuned LLMs

arXiv — cs.CL•Thursday, November 13, 2025 at 5:00:00 AM
NegativeArtificial Intelligence
The study titled 'Quantifying Edits Decay in Fine-tuned LLMs' explores how fine-tuning affects knowledge edits in large language models (LLMs). It addresses a critical question: do edits survive after fine-tuning? The findings indicate that edits do decay, with the extent of survival varying across different configurations. This decay poses challenges for knowledge editing methods, as it could necessitate re-editing fine-tuned models, thereby increasing costs. Additionally, the risk of propagating hidden malicious edits raises serious safety concerns. The research evaluated state-of-the-art editing methods like MEMIT and AlphaEdit, alongside fine-tuning approaches such as full-parameter, LoRA, and DoRA, across multiple LLMs. Notably, AlphaEdit edits were found to decay more than those from MEMIT. The study suggests a selective-layer fine-tuning method as a potential solution to mitigate these issues. As LLMs become increasingly integrated into various applications, understanding the im…
— via World Pulse Now AI Editorial System
Read Original

Was this article worth reading? Share it

Recommended Readings
Thinker: Training LLMs in Hierarchical Thinking for Deep Search via Multi-Turn Interaction
arXiv — cs.CL2 days ago
Thinker: Training LLMs in Hierarchical Thinking for Deep Search via Multi-Turn Interaction
PositiveArtificial Intelligence
The article presents Thinker, a hierarchical thinking model designed to enhance the reasoning capabilities of large language models (LLMs) through multi-turn interactions. Unlike previous methods that relied on end-to-end reinforcement learning without supervision, Thinker allows for a more structured reasoning process by breaking down complex problems into manageable sub-problems. Each sub-problem is represented in both natural language and logical functions, improving the coherence and rigor of the reasoning process.
Read full article
via arXiv — cs.CL
From Fact to Judgment: Investigating the Impact of Task Framing on LLM Conviction in Dialogue Systems
arXiv — cs.CL2 days ago
From Fact to Judgment: Investigating the Impact of Task Framing on LLM Conviction in Dialogue Systems
NeutralArtificial Intelligence
The article investigates the impact of task framing on the conviction of large language models (LLMs) in dialogue systems. It explores how LLMs assess tasks requiring social judgment, contrasting their performance on factual queries with conversational judgment tasks. The study reveals that reframing a task can significantly alter an LLM's judgment, particularly under conversational pressure, highlighting the complexities of LLM decision-making in social contexts.
Read full article
via arXiv — cs.CL
Expert-Guided Prompting and Retrieval-Augmented Generation for Emergency Medical Service Question Answering
arXiv — cs.CL2 days ago
Expert-Guided Prompting and Retrieval-Augmented Generation for Emergency Medical Service Question Answering
PositiveArtificial Intelligence
Large language models (LLMs) have shown potential in medical question answering but often lack the domain-specific expertise required in emergency medical services (EMS). The study introduces EMSQA, a dataset with 24.3K questions across 10 clinical areas and 4 certification levels, along with knowledge bases containing 40K documents and 2M tokens. It also presents Expert-CoT and ExpertRAG, strategies that enhance performance by integrating clinical context, resulting in improved accuracy and exam pass rates for EMS certification.
Read full article
via arXiv — cs.CL
Can LLMs Detect Their Own Hallucinations?
arXiv — cs.CL2 days ago
Can LLMs Detect Their Own Hallucinations?
PositiveArtificial Intelligence
Large language models (LLMs) are capable of generating fluent responses but can sometimes produce inaccurate information, referred to as hallucinations. A recent study investigates whether these models can recognize their own inaccuracies. The research formulates hallucination detection as a classification task and introduces a framework utilizing Chain-of-Thought (CoT) to extract knowledge from LLM parameters. Experimental results show that GPT-3.5 Turbo with CoT detected 58.2% of its own hallucinations, suggesting that LLMs can identify inaccuracies if they possess sufficient knowledge.
Read full article
via arXiv — cs.CL
PustakAI: Curriculum-Aligned and Interactive Textbooks Using Large Language Models
arXiv — cs.CL2 days ago
PustakAI: Curriculum-Aligned and Interactive Textbooks Using Large Language Models
PositiveArtificial Intelligence
PustakAI is a framework designed to create interactive textbooks aligned with the NCERT curriculum for grades 6 to 8 in India. Utilizing Large Language Models (LLMs), it aims to enhance personalized learning experiences, particularly in areas with limited educational resources. The initiative addresses challenges in adapting LLMs to specific curricular content, ensuring accuracy and pedagogical relevance.
Read full article
via arXiv — cs.CL
LaoBench: A Large-Scale Multidimensional Lao Benchmark for Large Language Models
arXiv — cs.CL2 days ago
LaoBench: A Large-Scale Multidimensional Lao Benchmark for Large Language Models
PositiveArtificial Intelligence
LaoBench is a newly introduced large-scale benchmark dataset aimed at evaluating large language models (LLMs) in the Lao language. It consists of over 17,000 curated samples that assess knowledge application, foundational education, and bilingual translation among Lao, Chinese, and English. The dataset is designed to enhance the understanding and reasoning capabilities of LLMs in low-resource languages, addressing the current challenges faced by models in mastering Lao.
Read full article
via arXiv — cs.CL
LoRaCompass: Robust Reinforcement Learning to Efficiently Search for a LoRa Tag
arXiv — cs.LG2 days ago
LoRaCompass: Robust Reinforcement Learning to Efficiently Search for a LoRa Tag
PositiveArtificial Intelligence
The Long-Range (LoRa) protocol is increasingly used in tags for mentally incapacitated persons (MIPs) to prevent them from going missing. A new study introduces LoRaCompass, a reinforcement learning model aimed at efficiently locating these LoRa tags in unknown environments. This model addresses challenges such as domain shift and signal fluctuation, which can lead to significant localization errors, by learning robust spatial representations from received signal strength indicators (RSSI).
Read full article
via arXiv — cs.LG
Scaling Latent Reasoning via Looped Language Models
arXiv — cs.CL2 days ago
Scaling Latent Reasoning via Looped Language Models
PositiveArtificial Intelligence
The article presents Ouro, a family of pre-trained Looped Language Models (LoopLM) designed to enhance reasoning capabilities during the pre-training phase. Unlike traditional models that rely on explicit text generation, Ouro incorporates iterative computation in latent space and an entropy-regularized objective for depth allocation. The models, Ouro 1.4B and 2.6B, demonstrate superior performance, matching results of larger state-of-the-art models while emphasizing improved knowledge manipulation rather than increased capacity.
Read full article
via arXiv — cs.CL