Large language models (LLMs) suffer from forgetting of upstream knowledge when fine-tuned. Despite efforts on mitigating forgetting, few have investigated how forgotten upstream examples are dependent on newly learned tasks. Insights on such dependencies enable efficient and targeted mitigation of forgetting. In this paper, we empirically analyze forgetting that occurs in N upstream examples of language modeling or instruction-tuning after fine-tuning LLMs on one of M new tasks, visualized in M N matrices. We show that the matrices are often well-approximated with low-rank matrices, indicating the dominance of simple associations between the learned tasks and forgotten upstream examples. Leveraging the analysis, we predict forgetting of upstream examples when fine-tuning LLMs on unseen tasks with matrix completion over the empirical associations. This enables fast identification of most forgotten examples without expensive inference on the entire upstream data. Despite simplicity, the approach outperforms prior approaches that learn semantic relationships of learned tasks and upstream examples with LMs. We demonstrate the practical utility of our analysis by showing statistically significantly reduced forgetting as we upweight predicted examples for replay during fine-tuning.

We know from prior work that LLMs encode social biases, and that this manifests in clinical tasks. In this work we adopt tools from mechanistic interpretability to unveil sociodemographic representations and biases within LLMs in the context of healthcare. Specifically, we ask: Can we identify activations within LLMs that encode sociodemographic information (e.g., gender, race)? We find that gender information is highly localized in middle MLP layers and can be reliably manipulated at inference time via patching. Such interventions can surgically alter generated clinical vignettes for specific conditions, and also influence downstream clinical predictions which correlate with gender, e.g., patient risk of depression. We find that representation of patient race is somewhat more distributed, but can also be intervened upon, to a degree. To our knowledge, this is the first application of mechanistic interpretability methods to LLMs for healthcare.

Recent advances in artificial intelligence have seen Large Language Models (LLMs) demonstrate notable proficiency in causal discovery tasks. This study explores the factors influencing the performance of LLMs in causal discovery tasks. Utilizing open-source LLMs, we examine how the frequency of causal relations within their pre-training corpora affects their ability to accurately respond to causal discovery queries. Our findings reveal that a higher frequency of causal mentions correlates with better model performance, suggesting that extensive exposure to causal information during training enhances the models' causal discovery capabilities. Additionally, we investigate the impact of context on the validity of causal relations. Our results indicate that LLMs might exhibit divergent predictions for identical causal relations when presented in different contexts. This paper provides the first comprehensive analysis of how different factors contribute to LLM performance in causal discovery tasks.