Jacob M. Springer

Suhas Kotha, Jacob Mitchell Springer, Aditi Raghunathan · cmu

We lack a systematic understanding of the effects of fine-tuning (via methods such as instruction-tuning or reinforcement learning from human feedback), particularly on tasks outside the narrow fine-tuning distribution. In a simplified scenario, we demonstrate that improving performance on tasks within the fine-tuning data distribution comes at the expense of capabilities on other tasks. We hypothesize that language models implicitly infer the task of the prompt and that fine-tuning skews this inference towards tasks in the fine-tuning distribution. To test this, we propose Conjugate Prompting, which artificially makes the task look farther from the fine-tuning distribution while requiring the same capability, and we find that this recovers some of the pretraining capabilities in our synthetic setup. Since real-world fine-tuning distributions are predominantly English, we apply conjugate prompting to recover pretrained capabilities in LLMs by simply translating the prompts to different languages. This allows us to recover in-context learning abilities lost via instruction tuning, natural reasoning capability lost during code fine-tuning, and, more concerningly, harmful content generation suppressed by safety fine-tuning in chatbots like ChatGPT.

Atharva Kulkarni, Jacob Mitchell Springer, Arjun Subramonian et al. · cmu, meta-ai

Geometric properties of Transformer weights, particularly the unembedding matrix, have been widely useful in language model interpretability research. Yet, their utility for estimating downstream performance remains unclear. In this work, we systematically investigate the relationship between model performance and the unembedding matrix geometry, particularly its effective rank. Our experiments, involving a suite of 108 OLMo-style language models trained under controlled variation, reveal several key findings. While the best-performing models often exhibit a high effective rank, this trend is not universal across tasks and training setups. Contrary to prior work, we find that low effective rank does not cause late-stage performance degradation in small models, but instead co-occurs with it; we find adversarial cases where low-rank models do not exhibit saturation. Moreover, we show that effective rank is strongly influenced by pre-training hyperparameters, such as batch size and weight decay, which in-turn affect the model's performance. Lastly, extending our analysis to other geometric metrics and final-layer representation, we find that these metrics are largely aligned, but none can reliably predict downstream performance. Overall, our findings suggest that the model's geometry, as captured by existing metrics, primarily reflects training choices rather than performance.

Ishaan Watts, Catherine Li, Sachin Goyal et al.

Pretraining optimizers are tuned to produce the strongest possible base model, on the assumption that a stronger starting point yields a stronger model after subsequent changes like post-training and quantization. This overlooks the geometry of the base model which controls how much of the base model's capabilities survive subsequent parameter updates. We study three pretraining optimization approaches that bias optimization toward flatter minima: Sharpness-Aware Minimization (SAM), large learning rates, and shortened learning rate annealing periods. Across model sizes ranging from 20M to 150M parameters, we find that these interventions consistently improve downstream performance after post-training on five common datasets with up to 80% less forgetting. These principles hold at scale: a short SAM mid-training phase applied to an existing OLMo-2-1B checkpoint reduces forgetting by 31% after MetaMath post-training and by 40% after 4-bit quantization.

Jacob Mitchell Springer, Madhu Advani, Lukas Aichberger et al.

Post-training (via supervised fine-tuning) improves instruction-following, but often induces semantic mode collapse by biasing models toward low-entropy fine-tuning data at the expense of the high-entropy pretraining distribution. Crucially, we find this trade-off worsens with scale. To close this semantic diversity gap, we propose annotation-anchored training, a principled method that enables models to adopt the preference-following behaviors of post-training without sacrificing the inherent diversity of pretraining. Our approach is simple: we pretrain on documents paired with semantic annotations, inducing a rich annotation distribution that reflects the full breadth of pretraining data, and we preserve this distribution during post-training. This lets us sample diverse annotations at inference time and use them as anchors to guide generation, effectively transferring pretraining's semantic richness into post-trained models. We find that models trained with annotation-anchored training can attain $6 \times$ less diversity collapse than models trained with SFT, and improve with scale.

Lawrence Feng, Gaurav R. Ghosal, Jacob Mitchell Springer et al.

How can we train models whose post-trained capabilities survive subsequent fine-tuning? Rather than focusing on downstream interventions to mitigate forgetting of upstream capabilities, we study how upstream training choices - that is, the manner in which a capability is acquired - shape how robustly that capability is retained. We investigate this question in a controlled three-stage language-model pipeline: pretraining, post-training to acquire a target capability, and downstream fine-tuning on a new objective. Across 135M and 1B models, two post-training domains, and two downstream fine-tuning tasks, we find that immediate post-training performance does not reliably predict retention after subsequent fine-tuning: training recipes that look equivalent immediately after post-training can retain the target capability very differently after subsequent fine-tuning. In particular, early exposure - mixing post-training data into pretraining - consistently improves the frontier between retained upstream performance and downstream performance. In compute-matched experiments, where the target data must be allocated between pretraining and post-training, we find that the optimum lies at neither extreme. Together with our other empirical and theoretical findings, this supports the view that post-training drives immediate specialization while early exposure improves robustness to later forgetting. Replay and dropout, typically used to mitigate forgetting as it occurs during fine-tuning, provide complementary gains to early exposure when applied during post-training. Our findings suggest that robustness to subsequent fine-tuning should be treated as a first-class objective of upstream training, addressed preventatively through choices like early exposure rather than reactively during fine-tuning itself.

Jacob Mitchell Springer, Suhas Kotha, Daniel Fried et al. · cmu

Bidirectional models are considered essential for strong text embeddings. Recent approaches to adapt autoregressive language models (LMs) into strong text embedding models have largely had the requirement to modify the LM architecture to be bidirectional. We challenge this premise by introducing "echo embeddings" which converts autoregressive LMs into high quality text embedding models without changing the architecture or requiring fine-tuning. By repeating the input and extracting embeddings from the repeated tokens -- which have access to all original tokens -- echo embeddings improve over classical LM embeddings by over 5% in zero-shot settings. Our zero-shot embeddings nearly match those obtained by bidirectionally-converted LMs that undergo additional masked-language modeling training. Echo embeddings are also compatible with supervised fine-tuning, matching or outperforming bidirectionally-converted LMs in an apples-to-apples comparison, even with an identical compute budget during training and inference. Overall, repetition is a simple and effective strategy to circumvent the need for bidirectional attention in embedding models, paving the way towards a unified architecture for all NLP tasks.

Jacob Mitchell Springer, Sachin Goyal, Kaiyue Wen et al. · cmu

Large language models are pre-trained on ever-growing token budgets under the assumption that better pre-training performance translates to improved downstream models. In this work, we challenge this assumption and show that extended pre-training can make models harder to fine-tune, leading to degraded final performance. We term this phenomenon catastrophic overtraining. For example, the instruction-tuned OLMo-1B model pre-trained on 3T tokens leads to over 2% worse performance on multiple standard LLM benchmarks than its 2.3T token counterpart. Through controlled experiments and theoretical analysis, we show that catastrophic overtraining arises from a systematic increase in the broad sensitivity of pre-trained parameters to modifications, including but not limited to fine-tuning. Our findings call for a critical reassessment of pre-training design that considers the downstream adaptability of the model.

Taeyoun Kim, Jacob Springer, Aditi Raghunathan et al. · allen-ai, cmu

In retrieval augmented generation (RAG) systems, each individual component -- the LLM, embedder, and corpus -- could introduce biases in the form of skews towards outputting certain perspectives or identities. In this work, we study the conflict between biases of each component and their relationship to the overall bias of the RAG system, which we call bias conflict. Examining both gender and political biases as case studies, we show that bias conflict can be characterized through a linear relationship among components despite its complexity in 6 different LLMs. Through comprehensive fine-tuning experiments creating 120 differently biased embedders, we demonstrate how to control bias while maintaining utility and reveal the importance of reverse-biasing the embedder to mitigate bias in the overall system. Additionally, we find that LLMs and tasks exhibit varying sensitivities to the embedder bias, a crucial factor to consider for debiasing. Our results underscore that a fair RAG system can be better achieved by carefully controlling the bias of the embedder rather than increasing its fairness.

Jacob Mitchell Springer, Vaibhav Adlakha, Siva Reddy et al. · cmu, mila

Recent progress in developing general purpose text embedders has been driven by training on ever-growing corpora of synthetic LLM-generated data. Nonetheless, no publicly available synthetic dataset exists, posing a barrier to studying its role for generalization. To address this issue, we first reproduce and publicly release the synthetic data proposed by Wang et al. (Mistral-E5). Our synthetic data is high quality and leads to consistent improvements in performance. Next, we critically examine where exactly synthetic data improves model generalization. Our analysis reveals that benefits from synthetic data are sparse and highly localized to individual datasets. Moreover, we observe trade-offs between the performance on different categories and data that benefits one task, degrades performance on another. Our findings highlight the limitations of current synthetic data approaches for building general-purpose embedders and challenge the notion that training on synthetic data leads to more robust embedding models across tasks.

Jacob M. Springer, Melanie Mitchell, Garrett T. Kenyon

Adversarial examples for neural network image classifiers are known to be transferable: examples optimized to be misclassified by a source classifier are often misclassified as well by classifiers with different architectures. However, targeted adversarial examples -- optimized to be classified as a chosen target class -- tend to be less transferable between architectures. While prior research on constructing transferable targeted attacks has focused on improving the optimization procedure, in this work we examine the role of the source classifier. Here, we show that training the source classifier to be "slightly robust" -- that is, robust to small-magnitude adversarial examples -- substantially improves the transferability of class-targeted and representation-targeted adversarial attacks, even between architectures as different as convolutional neural networks and transformers. The results we present provide insight into the nature of adversarial examples as well as the mechanisms underlying so-called "robust" classifiers.

Jacob M. Springer, Melanie Mitchell, Garrett T. Kenyon

Neural networks trained on visual data are well-known to be vulnerable to often imperceptible adversarial perturbations. The reasons for this vulnerability are still being debated in the literature. Recently Ilyas et al. (2019) showed that this vulnerability arises, in part, because neural network classifiers rely on highly predictive but brittle "non-robust" features. In this paper we extend the work of Ilyas et al. by investigating the nature of the input patterns that give rise to these features. In particular, we hypothesize that in a neural network trained in a standard way, non-robust features respond to small, "non-semantic" patterns that are typically entangled with larger, robust patterns, known to be more human-interpretable, as opposed to solely responding to statistical artifacts in a dataset. Thus, adversarial examples can be formed via minimal perturbations to these small, entangled patterns. In addition, we demonstrate a corollary of our hypothesis: robust classifiers are more effective than standard (non-robust) ones as a source for generating transferable adversarial examples in both the untargeted and targeted settings. The results we present in this paper provide new insight into the nature of the non-robust features responsible for adversarial vulnerability of neural network classifiers.

Jacob M. Springer, Bryn Marie Reinstadler, Una-May O'Reilly

Neural networks are well-known to be vulnerable to imperceptible perturbations in the input, called adversarial examples, that result in misclassification. Generating adversarial examples for source code poses an additional challenge compared to the domains of images and natural language, because source code perturbations must retain the functional meaning of the code. We identify a striking relationship between token frequency statistics and learned token embeddings: the L2 norm of learned token embeddings increases with the frequency of the token except for the highest-frequnecy tokens. We leverage this relationship to construct a simple and efficient gradient-free method for generating state-of-the-art adversarial examples on models of code. Our method empirically outperforms competing gradient-based methods with less information and less computational effort.

Jacob M. Springer, Garrett T. Kenyon

Efforts to improve the learning abilities of neural networks have focused mostly on the role of optimization methods rather than on weight initializations. Recent findings, however, suggest that neural networks rely on lucky random initial weights of subnetworks called "lottery tickets" that converge quickly to a solution. To investigate how weight initializations affect performance, we examine small convolutional networks that are trained to predict n steps of the two-dimensional cellular automaton Conway's Game of Life, the update rules of which can be implemented efficiently in a 2n+1 layer convolutional network. We find that networks of this architecture trained on this task rarely converge. Rather, networks require substantially more parameters to consistently converge. In addition, near-minimal architectures are sensitive to tiny changes in parameters: changing the sign of a single weight can cause the network to fail to learn. Finally, we observe a critical value d_0 such that training minimal networks with examples in which cells are alive with probability d_0 dramatically increases the chance of convergence to a solution. We conclude that training convolutional neural networks to learn the input/output function represented by n steps of Game of Life exhibits many characteristics predicted by the lottery ticket hypothesis, namely, that the size of the networks required to learn this function are often significantly larger than the minimal network required to implement the function.

Jacob M. Springer, Charles S. Strauss, Austin M. Thresher et al.

Although deep learning has shown great success in recent years, researchers have discovered a critical flaw where small, imperceptible changes in the input to the system can drastically change the output classification. These attacks are exploitable in nearly all of the existing deep learning classification frameworks. However, the susceptibility of deep sparse coding models to adversarial examples has not been examined. Here, we show that classifiers based on a deep sparse coding model whose classification accuracy is competitive with a variety of deep neural network models are robust to adversarial examples that effectively fool those same deep learning models. We demonstrate both quantitatively and qualitatively that the robustness of deep sparse coding models to adversarial examples arises from two key properties. First, because deep sparse coding models learn general features corresponding to generators of the dataset as a whole, rather than highly discriminative features for distinguishing specific classes, the resulting classifiers are less dependent on idiosyncratic features that might be more easily exploited. Second, because deep sparse coding models utilize fixed point attractor dynamics with top-down feedback, it is more difficult to find small changes to the input that drive the resulting representations out of the correct attractor basin.