The field of natural language processing is moving towards a deeper understanding of language models, with a focus on interpretability and robustness. Recent studies have explored the vulnerabilities of large language models to misinformation and the importance of monitoring their factual integrity. The use of causal masking on spatial data has also been investigated, with promising results. Furthermore, research has delved into the periodicity of information in natural language, the monitorability of chain-of-thought outputs, and the legibility of reasoning models. Noteworthy papers include 'Layer of Truth: Probing Belief Shifts under Continual Pre-Training Poisoning', which introduces a framework for probing belief dynamics in continually trained language models, and 'Causal Masking on Spatial Data: An Information-Theoretic Case for Learning Spatial Datasets with Unimodal Language Models', which demonstrates the viability of applying causal masking to spatial data.
Advances in Language Model Interpretability and Robustness
Sources
Layer of Truth: Probing Belief Shifts under Continual Pre-Training Poisoning
Causal Masking on Spatial Data: An Information-Theoretic Case for Learning Spatial Datasets with Unimodal Language Models
Identifying the Periodicity of Information in Natural Language
Measuring Chain-of-Thought Monitorability Through Faithfulness and Verbosity
Reasoning Models Sometimes Output Illegible Chains of Thought
Thought Branches: Interpreting LLM Reasoning Requires Resampling
ParaScopes: What do Language Models Activations Encode About Future Text?
Reversal Invariance in Autoregressive Language Models
Belief Dynamics Reveal the Dual Nature of In-Context Learning and Activation Steering
Priors in Time: Missing Inductive Biases for Language Model Interpretability
Accumulating Context Changes the Beliefs of Language Models
LLM Probing with Contrastive Eigenproblems: Improving Understanding and Applicability of CCS
Understanding New-Knowledge-Induced Factual Hallucinations in LLMs: Analysis, Solution, and Interpretation
Are language models aware of the road not taken? Token-level uncertainty and hidden state dynamics
Addressing divergent representations from causal interventions on neural networks