Decomposing the Contributions of In-Context Learning: Label Space, Format, and Discrimination
The performance improvement brought by in-context learning (ICL) can be decomposed into three factors: label space regulation, label format regulation, and discrimination power. ICL exhibits significant efficacy in regulating the label space and format, but has limited impact on improving the model's discriminative capability.
Exploring Concept Depth: How Large Language Models Acquire Knowledge at Different Layers
Different concepts are learned at different layers of large language models, with more difficult concepts being fully acquired at deeper layers.
Transparent Toolkit for Analyzing Transformer Language Models
The LM Transparency Tool provides a comprehensive framework for tracing back the behavior of Transformer-based language models to specific model components, enabling detailed analysis and interpretation of the decision-making process.
Detecting Non-Factual Content in Large Language Model Generations via Offline Consistency Checking and Probing
PINOSE, a method that trains a probing model on offline self-consistency checking results, can efficiently and effectively detect non-factual content generated by large language models without relying on human-annotated data.
On the Semantics of LM Latent Space: A Vocabulary-Defined Approach for Enhancing Language Model Performance and Interpretability
The authors propose a novel vocabulary-defined approach to analyze the semantics of language model latent space, which establishes a disentangled reference frame and enables effective model adaptation through semantic calibration.
Exploring Memory Characteristics in Large Language Models: Insights into the Interplay between Biological and Artificial Language Processing
Large language models exhibit key characteristics of human memory, such as primacy and recency effects, the influence of elaborations, and forgetting through interference rather than decay. These similarities suggest that the properties of human biological memory are reflected in the statistical structure of textual narratives, which is then captured by the language models.
Detecting and Forecasting Hallucinations in Large Language Models via State Transition Dynamics
Hallucinations in large language models can be effectively detected by analyzing the model's internal state transition dynamics during generation using tractable probabilistic models.
Tracing the Evolutionary Relationships and Performance Capabilities of Large Language Models through Phylogenetic Analysis
Applying phylogenetic algorithms to Large Language Models can reconstruct their evolutionary relationships and predict their performance on benchmarks, offering insights into model development and capabilities.
Measuring the Influence of Context and Prior Knowledge in Language Models
Language models integrate prior knowledge and new contextual information in predictable ways, relying more on prior knowledge for familiar entities and being more easily persuaded by some contexts than others.
Scope Ambiguities in Large Language Models: Insights into Semantic Structure and World Knowledge Representation
Large language models exhibit similar preferences to humans in interpreting scope ambiguous sentences and are sensitive to the presence of multiple readings in such sentences.
© 2024 by Linnk AI