Mechanistic interpretability has made great strides in identifying neural network features (e.g., directions in hidden activation space) that mediate concepts(e.g., the birth year of a person) and enable predictable manipulation. Distributed alignment search (DAS) leverages supervision from counterfactual data to learn concept features within hidden states, but DAS assumes we can afford to conduct a brute force search over potential feature locations. To address this, we present HyperDAS, a transformer-based hypernetwork architecture that (1) automatically locates the token-positions of the residual stream that a concept is realized in and (2) constructs features of those residual stream vectors for the concept. In experiments with Llama3-8B, HyperDAS achieves state-of-the-art performance on the RAVEL benchmark for disentangling concepts in hidden states. In addition, we review the design decisions we made to mitigate the concern that HyperDAS (like all powerful interpretabilty methods) might inject new information into the target model rather than faithfully interpreting it.

Language Models (LMs) often struggle with linguistic understanding at the discourse level, even though discourse patterns such as coherence, cohesion, and narrative flow are prevalent in their pre-training data. Current methods address these challenges only after the pre-training phase, relying on expensive human annotated data to align the model. To improve the discourse capabilities of LMs already at the pre-training stage, we introduce DEPTH, an encoder-decoder model that learns to represent sentences using a discourse-oriented pre-training objective. DEPTH combines hierarchical sentence representations with two objectives: (1) Sentence Un-Shuffling, and (2) Span-Corruption. This approach trains the model to represent both sub-word-level and sentence-level dependencies over a massive amount of unstructured text. When trained either from scratch or continuing from a pre-trained T5 checkpoint, DEPTH learns semantic and discourse-level representations faster than T5, outperforming it in span-corruption loss despite the additional sentence-un-shuffling objective. Evaluations on the GLUE, DiscoEval, and NI benchmarks demonstrate DEPTH's ability to quickly learn diverse downstream tasks, which require syntactic, semantic, and discourse capabilities. Overall, our approach extends the discourse capabilities of T5, while minimally impacting other natural language understanding (NLU) capabilities in the resulting LM.

Estimating the log-likelihood of a given sentence under an autoregressive language model is straightforward: one can simply apply the chain rule and sum the log-likelihood values for each successive token. However, for masked language models (MLMs), there is no direct way to estimate the log-likelihood of a sentence. To address this issue, Salazar et al. (2020) propose to estimate sentence pseudo-log-likelihood (PLL) scores, computed by successively masking each sentence token, retrieving its score using the rest of the sentence as context, and summing the resulting values. Here, we demonstrate that the original PLL method yields inflated scores for out-of-vocabulary words and propose an adapted metric, in which we mask not only the target token, but also all within-word tokens to the right of the target. We show that our adapted metric (PLL-word-l2r) outperforms both the original PLL metric and a PLL metric in which all within-word tokens are masked. In particular, it better satisfies theoretical desiderata and better correlates with scores from autoregressive models. Finally, we show that the choice of metric affects even tightly controlled, minimal pair evaluation benchmarks (such as BLiMP), underscoring the importance of selecting an appropriate scoring metric for evaluating MLM properties.