The stability of language model pre-training and its effects on downstream performance are still understudied. Prior work shows that the training process can yield significantly different results in response to slight variations in initial conditions, e.g., the random seed. Crucially, the research community still lacks sufficient resources and tools to systematically investigate pre-training stability, particularly for decoder-only language models. We introduce the PolyPythias, a set of 45 new training runs for the Pythia model suite: 9 new seeds across 5 model sizes, from 14M to 410M parameters, resulting in about 7k new checkpoints that we release. Using these new 45 training runs, in addition to the 5 already available, we study the effects of different initial conditions determined by the seed -- i.e., parameters' initialisation and data order -- on (i) downstream performance, (ii) learned linguistic representations, and (iii) emergence of training phases. In addition to common scaling behaviours, our analyses generally reveal highly consistent training dynamics across both model sizes and initial conditions. Further, the new seeds for each model allow us to identify outlier training runs and delineate their characteristics. Our findings show the potential of using these methods to predict training stability.

In this paper, we explore the nature of sudden breakthroughs in language model performance at scale, which stands in contrast to smooth improvements governed by scaling laws. While advocates of "emergence" view abrupt performance gains as capabilities unlocking at specific scales, others have suggested that they are produced by thresholding effects and alleviated by continuous metrics. We propose that breakthroughs are instead driven by continuous changes in the probability distribution of training outcomes, particularly when performance is bimodally distributed across random seeds. In synthetic length generalization tasks, we show that different random seeds can produce either highly linear or emergent scaling trends. We reveal that sharp breakthroughs in metrics are produced by underlying continuous changes in their distribution across seeds. Furthermore, we provide a case study of inverse scaling and show that even as the probability of a successful run declines, the average performance of a successful run continues to increase monotonically. We validate our distributional scaling framework on realistic settings by measuring MMLU performance in LLM populations. These insights emphasize the role of random variation in the effect of scale on LLM capabilities.

Language models (LMs), like other neural networks, often favor shortcut heuristics based on surface-level patterns. Although LMs behave like n-gram models early in training, they must eventually learn hierarchical syntactic representations to correctly apply grammatical rules out-of-distribution (OOD). In this work, we use case studies of English grammar to explore how complex, diverse training data drives models to generalize OOD. We construct a framework that unifies our understanding of random variation with training dynamics, rule selection with memorization, and data diversity with complexity. We show that these factors are nuanced, and that intermediate levels of diversity and complexity lead to inconsistent behavior across random seeds and to unstable training dynamics. Our findings emphasize the critical role of training data in shaping generalization patterns and illuminate how competing model strategies lead to inconsistent generalization outcomes across random seeds.

The rise of the term "mechanistic interpretability" has accompanied increasing interest in understanding neural models -- particularly language models. However, this jargon has also led to a fair amount of confusion. So, what does it mean to be "mechanistic"? We describe four uses of the term in interpretability research. The most narrow technical definition requires a claim of causality, while a broader technical definition allows for any exploration of a model's internals. However, the term also has a narrow cultural definition describing a cultural movement. To understand this semantic drift, we present a history of the NLP interpretability community and the formation of the separate, parallel "mechanistic" interpretability community. Finally, we discuss the broad cultural definition -- encompassing the entire field of interpretability -- and why the traditional NLP interpretability community has come to embrace it. We argue that the polysemy of "mechanistic" is the product of a critical divide within the interpretability community.

While the biases of language models in production are extensively documented, the biases of their guardrails have been neglected. This paper studies how contextual information about the user influences the likelihood of an LLM to refuse to execute a request. By generating user biographies that offer ideological and demographic information, we find a number of biases in guardrail sensitivity on GPT-3.5. Younger, female, and Asian-American personas are more likely to trigger a refusal guardrail when requesting censored or illegal information. Guardrails are also sycophantic, refusing to comply with requests for a political position the user is likely to disagree with. We find that certain identity groups and seemingly innocuous information, e.g., sports fandom, can elicit changes in guardrail sensitivity similar to direct statements of political ideology. For each demographic category and even for American football team fandom, we find that ChatGPT appears to infer a likely political ideology and modify guardrail behavior accordingly.

Measuring nonlinear feature interaction is an established This paper investigates Shapley interactions in a number of approach to understanding complex patterns tasks and architectures. We use Shapley interactions as a of attribution in many models. In this paper, case study to illustrate the importance of grounding model we use Shapley Taylor interaction indices interpretations in the underlying structure of the data and the (STII) to analyze the impact of underlying data target models. To this end, we draw connections between structure on model representations in a variety of interaction metrics and various structural properties of the modalities, tasks, and architectures. Considering data in each setting: syntax, tokenization, and idiomatic linguistic structure in masked and auto-regressive expressions in masked and autoregressive language models language models (MLMs and ALMs), we find (MLMs and ALMs, respectively); phoneme articulation differences that STII increases within idiomatic expressions in speech models; and distinctions between edges, and that MLMs scale STII with syntactic distance, foreground, and background pixels in image classifiers. After relying more on syntax in their nonlinear structure introducing our approach to Shapley interactions, we than ALMs do. Our speech model findings apply them in a variety of settings and find the following.

Most interpretability research in NLP focuses on understanding the behavior and features of a fully trained model. However, certain insights into model behavior may only be accessible by observing the trajectory of the training process. We present a case study of syntax acquisition in masked language models (MLMs) that demonstrates how analyzing the evolution of interpretable artifacts throughout training deepens our understanding of emergent behavior. In particular, we study Syntactic Attention Structure (SAS), a naturally emerging property of MLMs wherein specific Transformer heads tend to focus on specific syntactic relations. We identify a brief window in pretraining when models abruptly acquire SAS, concurrent with a steep drop in loss. This breakthrough precipitates the subsequent acquisition of linguistic capabilities. We then examine the causal role of SAS by manipulating SAS during training, and demonstrate that SAS is necessary for the development of grammatical capabilities. We further find that SAS competes with other beneficial traits during training, and that briefly suppressing SAS improves model quality. These findings offer an interpretation of a real-world example of both simplicity bias and breakthrough training dynamics.

The impact of randomness on model training is poorly understood. How do differences in data order and initialization actually manifest in the model, such that some training runs outperform others or converge faster? Furthermore, how can we interpret the resulting training dynamics and the phase transitions that characterize different trajectories? To understand the effect of randomness on the dynamics and outcomes of neural network training, we train models multiple times with different random seeds and compute a variety of metrics throughout training, such as the $L_2$ norm, mean, and variance of the neural network's weights. We then fit a hidden Markov model (HMM) over the resulting sequences of metrics. The HMM represents training as a stochastic process of transitions between latent states, providing an intuitive overview of significant changes during training. Using our method, we produce a low-dimensional, discrete representation of training dynamics on grokking tasks, image classification, and masked language modeling. We use the HMM representation to study phase transitions and identify latent "detour" states that slow down convergence.

The question of how to probe contextual word representations for linguistic structure in a way that is both principled and useful has seen significant attention recently in the NLP literature. In our contribution to this discussion, we argue for a probe metric that reflects the fundamental trade-off between probe complexity and performance: the Pareto hypervolume. To measure complexity, we present a number of parametric and non-parametric metrics. Our experiments using Pareto hypervolume as an evaluation metric show that probes often do not conform to our expectations -- e.g., why should the non-contextual fastText representations encode more morpho-syntactic information than the contextual BERT representations? These results suggest that common, simplistic probing tasks, such as part-of-speech labeling and dependency arc labeling, are inadequate to evaluate the linguistic structure encoded in contextual word representations. This leads us to propose full dependency parsing as a probing task. In support of our suggestion that harder probing tasks are necessary, our experiments with dependency parsing reveal a wide gap in syntactic knowledge between contextual and non-contextual representations.

The generalization of machine learning (ML) models to out-of-distribution (OOD) examples remains a key challenge in extracting information from upcoming astronomical surveys. Interpretability approaches are a natural way to gain insights into the OOD generalization problem. We use Centered Kernel Alignment (CKA), a similarity measure metric of neural network representations, to examine the relationship between representation similarity and performance of pre-trained Convolutional Neural Networks (CNNs) on the CAMELS Multifield Dataset. We find that when models are robust to a distribution shift, they produce substantially different representations across their layers on OOD data. However, when they fail to generalize, these representations change less from layer to layer on OOD data. We discuss the potential application of similarity representation in guiding model design, training strategy, and mitigating the OOD problem by incorporating CKA as an inductive bias during training.