Large language models (LLMs) are trained on huge amounts of textual data, and concerns have been raised that the limits of such data may soon be reached. A potential solution is to train on synthetic data sampled from LLMs. In this work, we build on this idea and investigate the benefits of *contrastive decoding* for generating synthetic data. In a controlled setting, we experiment with sampling corpora using the relative difference between a GOOD and BAD model trained on the same original corpus of 100 million words. By amplifying the signal from a model that has better performance, we create a synthetic corpus and mix it with the original training data. Our findings show that training on a mixture of synthesized and real data improves performance on the language modeling objective and a range of downstream tasks.In particular, we see that training with a mix of synthetic data from contrastive decoding benefits tasks that require more *reasoning skills*, while synthetic data from traditional sampling helps more on tasks requiring surface-level *linguistic* capabilities.

Two central capabilities of language models (LMs) are: (i) drawing on prior knowledge about entities, which allows them to answer queries such as What’s the official language of Austria?, and (ii) adapting to new information provided in context, e.g., Pretend the official language of Austria is Tagalog., that is pre-pended to the question. In this article, we introduce targeted persuasion score (TPS), designed to quantify how persuasive a given context is to an LM where persuasion is operationalized as the ability of the context to alter the LM’s answer to the question. In contrast to evaluating persuasiveness only through a model’s most likely answer, TPS provides a more fine-grained view of model behavior. Based on the Wasserstein distance, TPS measures how much a context shifts a model’s original answer distribution towarda target distribution. Empirically, through aseries of experiments, we show that TPS captures a more nuanced notion of persuasiveness than previously proposed metrics.

To answer a question, language models often need to integrate prior knowledge learned during pretraining and new information presented in context. We hypothesize that models perform this integration in a predictable way across different questions and contexts: models will rely more on prior knowledge for questions about entities (e.g., persons, places, etc.) that they are more familiar with due to higher exposure in the training corpus, and be more easily persuaded by some contexts than others. To formalize this problem, we propose two mutual information-based metrics to measure a model’s dependency on a context and on its prior about an entity: first, the persuasion score of a given context represents how much a model depends on the context in its decision, and second, the susceptibility score of a given entity represents how much the model can be swayed away from its original answer distribution about an entity. We empirically test our metrics for their validity and reliability. Finally, we explore and find a relationship between the scores and the model’s expected familiarity with an entity, and provide two use cases to illustrate their benefits.

One strength of modern language models is their ability to incorporate information from a user-input context when answering queries. However, they are not equally sensitive to the subtle changes to that context.To quantify this, Du et al. (2024) gives an information-theoretic metric to measure such sensitivity. Their metric, susceptibility, is defined as the degree to which contexts can influence a model’s response to a query at a distributional level.However, exactly computing susceptibility is difficult and, thus, Du et al. (2024) falls back on a Monte Carlo approximation.Due to the large number of samples required, the Monte Carlo approximation is inefficient in practice. As a faster alternative, we propose Fisher susceptibility, an efficient method to estimate the susceptibility based on Fisher information.Empirically, we validate that Fisher susceptibility is comparable to Monte Carlo estimated susceptibility across a diverse set of query domains despite its being 70× faster.Exploiting the improved efficiency, we apply Fisher susceptibility to analyze factors affecting the susceptibility of language models.We observe that larger models are as susceptible as smaller ones.

Given the prompt “Rome is in”, can we steer a language model to flip its prediction of an incorrect token “France” to a correct token “Italy” by only multiplying a few relevant activation vectors with scalars? We argue that successfully intervening on a model is a prerequisite for interpreting its internal workings. Concretely, we establish a three-term objective: a successful intervention should flip the correct with the wrong token and vice versa (effectiveness), and leave other tokens unaffected (faithfulness), all while being sparse (minimality). Using gradient-based optimization, this objective lets us learn (and later evaluate) a specific kind of efficient and interpretable intervention: activation scaling only modifies the signed magnitude of activation vectors to strengthen, weaken, or reverse the steering directions already encoded in the model. On synthetic tasks, this intervention performs comparably with steering vectors in terms of effectiveness and faithfulness, but is much more minimal allowing us to pinpoint interpretable model components. We evaluate activation scaling from different angles, compare performance on different datasets, and make activation scalars a learnable function of the activation vectors themselves to generalize to varying-length prompts.

How much meaning influences gender assignment across languages is an active area of research in linguistics and cognitive science. We can view current approaches as aiming to determine where gender assignment falls on a spectrum, from being fully arbitrarily determined to being largely semantically determined. For the latter case, there is a formulation of the neo-Whorfian hypothesis, which claims that even inanimate noun gender influences how people conceive of and talk about objects (using the choice of adjective used to modify inanimate nouns as a proxy for meaning). We offer a novel, causal graphical model that jointly represents the interactions between a noun’s grammatical gender, its meaning, and adjective choice. In accordance with past results, we find a significant relationship between the gender of nouns and the adjectives that modify them. However, when we control for the meaning of the noun, the relationship between grammatical gender and adjective choice is near zero and insignificant.

Many popular feature-attribution methods for interpreting deep neural networks rely on computing the gradients of a model’s output with respect to its inputs. While these methods can indicate which input features may be important for the model’s prediction, they reveal little about the inner workings of the model itself. In this paper, we observe that the gradient computation of a model is a special case of a more general formulation using semirings. This observation allows us to generalize the backpropagation algorithm to efficiently compute other interpretable statistics about the gradient graph of a neural network, such as the highest-weighted path and entropy. We implement this generalized algorithm, evaluate it on synthetic datasets to better understand the statistics it computes, and apply it to study BERT’s behavior on the subject–verb number agreement task (SVA). With this method, we (a) validate that the amount of gradient flow through a component of a model reflects its importance to a prediction and (b) for SVA, identify which pathways of the self-attention mechanism are most important.