13 papers
Through a Compressed Lens: Investigating The Impact of Quantization on Factual Knowledge Recall
Qianli Wang, Mingyang Wang, Nils Feldhus +5
Quantization methods are widely used to accelerate inference and streamline the deployment of large language models (LLMs). Although quantization's effects on various LLM capabilit…
The Anatomy of an Edit: Mechanism-Guided Activation Steering for Knowledge Editing
Yuan Cao, Mingyang Wang, Hinrich Schütze
Large language models (LLMs) are increasingly used as knowledge bases, but keeping them up to date requires targeted knowledge editing (KE). However, it remains unclear how edits a…
Refusal Direction is Universal Across Safety-Aligned Languages
Xinpeng Wang, Mingyang Wang, Yihong Liu +2
Refusal mechanisms in large language models (LLMs) are essential for ensuring safety. Recent research has revealed that refusal behavior can be mediated by a single direction in ac…
On the Entity-Level Alignment in Crosslingual Consistency
Yihong Liu, Mingyang Wang, François Yvon +1
Multilingual large language models (LLMs) are expected to recall factual knowledge consistently across languages. However, the factors that give rise to such crosslingual consisten…
BlackboxNLP-2025 MIB Shared Task: Exploring Ensemble Strategies for Circuit Localization Methods
Philipp Mondorf, Mingyang Wang, Sebastian Gerstner +6
The Circuit Localization track of the Mechanistic Interpretability Benchmark (MIB) evaluates methods for localizing circuits within large language models (LLMs), i.e., subnetworks…
Tracing Multilingual Factual Knowledge Acquisition in Pretraining
Yihong Liu, Mingyang Wang, Amir Hossein Kargaran +5
Large Language Models (LLMs) are capable of recalling multilingual factual knowledge present in their pretraining data. However, most studies evaluate only the final model, leaving…