Internal Causal Mechanisms Robustly Predict Language Model Out-of-Distribution Behaviors

Interpretability research now offers a variety of techniques for identifying abstract internal mechanisms in neural networks. Can such techniques be used to predict how models will behave on out-of-distribution examples? In this work, we provide a positive answer to this question. Through a diverse set of language modeling tasks--including symbol manipulation, knowledge retrieval, and instruction following--we show that the most robust features for correctness prediction are those that play a distinctive causal role in the model's behavior. Specifically, we propose two methods that leverage causal mechanisms to predict the correctness of model outputs: counterfactual simulation (checking whether key causal variables are realized) and value probing (using the values of those variables to make predictions). Both achieve high AUC-ROC in distribution and outperform methods that rely on causal-agnostic features in out-of-distribution settings, where predicting model behaviors is more crucial. Our work thus highlights a novel and significant application for internal causal analysis of language models.

Inference and Verbalization Functions During In-Context Learning

Large language models (LMs) are capable of in-context learning from a few demonstrations (example-label pairs) to solve new tasks during inference. Despite the intuitive importance of high-quality demonstrations, previous work has observed that, in some settings, ICL performance is minimally affected by irrelevant labels (Min et al., 2022). We hypothesize that LMs perform ICL with irrelevant labels via two sequential processes: an inference function that solves the task, followed by a verbalization function that maps the inferred answer to the label space. Importantly, we hypothesize that the inference function is invariant to remappings of the label space (e.g., "true"/"false" to "cat"/"dog"), enabling LMs to share the same inference function across settings with different label words. We empirically validate this hypothesis with controlled layer-wise interchange intervention experiments. Our findings confirm the hypotheses on multiple datasets and tasks (natural language inference, sentiment analysis, and topic classification) and further suggest that the two functions can be localized in specific layers across various open-sourced models, including GEMMA-7B, MISTRAL-7B-V0.3, GEMMA-2-27B, and LLAMA-3.1-70B.

An Efficient Privacy-aware Split Learning Framework for Satellite Communications

In the rapidly evolving domain of satellite communications, integrating advanced machine learning techniques, particularly split learning, is crucial for enhancing data processing and model training efficiency across satellites, space stations, and ground stations. Traditional ML approaches often face significant challenges within satellite networks due to constraints such as limited bandwidth and computational resources. To address this gap, we propose a novel framework for more efficient SL in satellite communications. Our approach, Dynamic Topology Informed Pruning, namely DTIP, combines differential privacy with graph and model pruning to optimize graph neural networks for distributed learning. DTIP strategically applies differential privacy to raw graph data and prunes GNNs, thereby optimizing both model size and communication load across network tiers. Extensive experiments across diverse datasets demonstrate DTIP's efficacy in enhancing privacy, accuracy, and computational efficiency. Specifically, on Amazon2M dataset, DTIP maintains an accuracy of 0.82 while achieving a 50% reduction in floating-point operations per second. Similarly, on ArXiv dataset, DTIP achieves an accuracy of 0.85 under comparable conditions. Our framework not only significantly improves the operational efficiency of satellite communications but also establishes a new benchmark in privacy-aware distributed learning, potentially revolutionizing data handling in space-based networks.