Large Language Model's are instruction-finetuned to enhance their ability to follow user instructions and better comprehend input context. Still, they often struggle to follow the input context, especially when it contradicts model's parametric knowledge. This manifests as various failures, such as hallucinations where a model inserts outdated or unwarranted facts into its response. In this work, we observe an intriguing phenomenon: the context reliance of the model decreases as instruction finetuning progresses, $\textit{despite an initial expected increase}$. We call this phenomenon as the $\textbf{context-parametric inversion}$. This is surprising, as one would expect instruction tuning to improve the model's ability to follow input instructions. We observe this behavior on multiple general purpose instruction tuning datasets such as TULU, Alpaca and Ultrachat, across multiple model families like Llama, Mistral and Pythia. We perform various controlled studies to eliminate some simple hypothesis for this observed behavior and isolate what datapoints cause this counter-intuitive behavior. We then analyze the phenomenon theoretically, to explain why context reliance varies across the trajectory of finetuning. We tie the observed context-parametric inversion to the properties of the finetuning data, which provides us with some potential mitigation strategies that provide limited but insightful gains.

Vision Language Models (VLMs) have demonstrated strong capabilities across various visual understanding and reasoning tasks, driven by incorporating image representations into the token inputs of Large Language Models (LLMs). However, their real-world deployment is often constrained by high latency during inference due to the substantial compute required by the LLM to process the large number of input tokens, predominantly arising from the image. To reduce inference costs, one can either downsize the LLM or reduce the number of input tokens needed to represent the image, the latter of which has been the focus of many recent efforts around token compression. However, it is unclear what the optimal trade-off is given a fixed inference budget. We first characterize this optimal trade-off between the number of visual tokens and LLM parameters by establishing scaling laws that capture variations in performance with these two factors. Our results reveal a surprising trend: for visual reasoning tasks, the inference-optimal behavior in VLMs is achieved by using the largest LLM that fits within the inference budget while minimizing visual token count - often to a single token. While the token reduction literature has mainly focused on maintaining base model performance by modestly reducing the token count (e.g., $5-10\times$), our results indicate that the compute-optimal inference regime requires operating under even higher token compression ratios. Based on these insights, we take the first steps toward designing token compression algorithms tailored for high-compression settings, utilizing prompt-based compression of tokens. Our work underscores the performance and efficiency benefits of operating in low visual token regimes and the importance of developing tailored token reduction algorithms for such conditions.

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 \textbf{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.

As large language models (LLMs) are increasingly deployed in high-stakes settings, the risk of generating harmful or toxic content remains a central challenge. Post-hoc alignment methods are brittle: once unsafe patterns are learned during pretraining, they are hard to remove. In this work, we present a data-centric pretraining framework that builds safety into the model from the start. Our framework consists of four key steps: (i) Safety Filtering: building a safety classifier to classify webdata into safe and unsafe categories; (ii) Safety Rephrasing: we recontextualize unsafe webdata into safer narratives; (iii) Native Refusal: we synthetically generate pretraining datasets that actively teach models to refuse on unsafe content and the moral reasoning behind it, and (iv) Harmfulness-Tag annotated pretraining: we flag unsafe content during pretraining using a special token, and use it to steer models away from unsafe generations at inference-time. Our safety-pretrained models reduce attack success rates from 38.8% to 8.4% on standard LLM safety benchmarks with no performance degradation on general tasks.

Vision-language models (VLMs) are trained for thousands of GPU hours on carefully selected subsets of massive web scrapes. For instance, the LAION public dataset retained only about 10% of the total crawled data. In recent times, data curation has gained prominence with several works developing strategies to retain ‘high-quality’ subsets of ‘raw’ scraped data. However, these strategies are typically developed agnostic to the available compute for training. In this paper, we demonstrate that making filtering decisions independent of training compute is often suboptimal—well-curated data rapidly loses its utility when repeated, eventually decreasing below the utility of ‘unseen’ but ‘lower-quality’ data. While past research in neural scaling laws has considered web data to be homogenous, real data is not. Our work bridges this important gap in the literature by developing scaling laws that characterize the differing ‘utility’ of various data subsets, and accounting for how this diminishes for a data point at its ‘nth’ repetition. Our key message is that data curation can not be agnostic of the total compute a model will be trained for. Even without ever jointly training on multiple data buckets, our scaling laws enable us to estimate model performance under this dynamic trade-off between quality and repetition. This allows us to curate the best possible pool for achieving top performance on Datacomp at various compute budgets, carving out a pareto-frontier for data curation.

Language models generate responses by producing a series of tokens in immediate succession: the $(K+1)^{\rm th}$ token is an outcome of manipulating $K$ hidden vectors per layer, one vector per preceding token. What if instead we were to let the model manipulate say, $K+10$ hidden vectors, before it outputs the $(K+1)^{\rm th}$ token? We operationalize this idea by performing training and inference on language models with a (learnable) $\textit{pause}$ token, a sequence of which is appended to the input prefix. We then delay extracting the model's outputs until the last pause token is seen, thereby allowing the model to process extra computation before committing to an answer. We empirically evaluate $\textit{pause-training}$ on decoder-only models of 1B and 130M parameters with causal pretraining on C4, and on downstream tasks covering reasoning, question-answering, general understanding and fact recall. Our main finding is that inference-time delays show gains when the model is both pre-trained and finetuned with delays. For the 1B model, we witness gains on 8 of 9 tasks, most prominently, a gain of $18\\%$ EM score on the QA task of SQuAD, $8\\%$ on CommonSenseQA and $1\\%$ accuracy on the reasoning task of GSM8k. Our work raises a range of conceptual and practical future research questions on making delayed next-token prediction a widely applicable new paradigm.

Large web-crawled multimodal datasets have powered a slew of new methods for learning general-purpose visual representations, advancing the state of the art in computer vision and revolutionizing zero- and few-shot recognition. One crucial decision facing practitioners is how, if at all, to curate these ever-larger datasets. For example, the creators of the LAION-5B dataset chose to retain only image-caption pairs whose CLIP similarity score exceeded a designated threshold. In this paper, we propose a new state-of-the-art data filtering approach motivated by our observation that nearly $40\%$ of LAION's images contain text that overlaps significantly with the caption. Intuitively, such data could be wasteful as it incentivizes models to perform optical character recognition rather than learning visual features. However, naively removing all such data could also be wasteful, as it throws away images that contain visual features (in addition to overlapping text). Our simple and scalable approach, T-MARS (Text Masking and Re-Scoring), filters out only those pairs where the text dominates the remaining visual features---by first masking out the text and then filtering out those with a low CLIP similarity score of the masked image with original captions. Experimentally, T-MARS is the top ranked approach on Imagenet at ``medium scale'' of DataComp (a data filtering benchmark), and outperforms CLIP filtering by a margin of $6.5\%$ on ImageNet and $4.7\%$ on VTAB. Additionally, we show that the accuracy gains enjoyed by T-MARS linearly increase as data and compute are scaled exponentially.

Finetuning image-text models such as CLIP achieves state-of-the-art accuracies on a variety of benchmarks. However, recent works (Kumar et al., 2022; Wortsman et al., 2021) have shown that even subtle differences in the finetuning process can lead to surprisingly large differences in the final performance, both for in-distribution (ID) and out-of-distribution (OOD) data. In this work, we show that a natural and simple approach of mimicking contrastive pretraining consistently outperforms alternative finetuning approaches. Specifically, we cast downstream class labels as text prompts and continue optimizing the contrastive loss between image embeddings and class-descriptive prompt embeddings (contrastive finetuning). Our method consistently outperforms baselines across 7 distribution shift, 6 transfer learning, and 3 few-shot learning benchmarks. On WILDS-iWILDCam, our proposed approach FLYP outperforms the top of the leaderboard by 2.3% ID and 2.7% OOD, giving the highest reported accuracy. Averaged across 7 OOD datasets (2 WILDS and 5 ImageNet associated shifts), FLYP gives gains of 4.2% OOD over standard finetuning and outperforms current state-of-the-art (LP-FT) by more than 1 % both ID and OOD. Similarly, on 3 few-shot learning benchmarks, FLYP gives gains up to 4.6% over standard finetuning and 4.4% over the state-of-the-art. Thus we establish our proposed method of contrastive finetuning as a simple and intuitive state-of-the-art for supervised finetuning of image-text models like CLIP. Code is available at https://github.com/locuslab/FLYP.

Test-time adaptation (TTA) refers to adapting neural networks to distribution shifts, specifically with just access to unlabeled test samples from the new domain at test-time. Prior TTA methods optimize over unsupervised objectives such as the entropy of model predictions in TENT (Wang et al., 2021), but it is unclear what exactly makes a good TTA loss. In this paper, we start by presenting a surprising phenomenon: if we attempt to $\textit{meta-learn}$ the ``best'' possible TTA loss over a wide class of functions, then we recover a function that is $\textit{remarkably}$ similar to (a temperature-scaled version of) the softmax-entropy employed by TENT. This only holds, however, if the classifier we are adapting is trained via cross-entropy loss; if the classifier is trained via squared loss, a different ``best'' TTA loss emerges.To explain this phenomenon, we analyze test-time adaptation through the lens of the training losses's $\textit{convex conjugate}$. We show that under natural conditions, this (unsupervised) conjugate function can be viewed as a good local approximation to the original supervised loss and indeed, it recovers the ``best'' losses found by meta-learning. This leads to a generic recipe than be used to find a good TTA loss for $\textit{any}$ given supervised training loss function of a general class. Empirically, our approach dominates other TTA alternatives over a wide range of domain adaptation benchmarks. Our approach is particularly of interest when applied to classifiers trained with $\textit{novel}$ loss functions, e.g., the recently-proposed PolyLoss (Leng et al., 2022) function, where it differs substantially from (and outperforms) an entropy-based loss. Further, we show that our conjugate based approach can also be interpreted as a kind of self-training using a very specific soft label, which we refer to as the $\textit{conjugate pseudo-label}$. Overall, therefore, our method provides a broad framework for better understanding and improving test-time adaptation. Code is available at https://github.com/locuslab/tta_conjugate.

Classical approaches for one-class problems such as one-class SVM and isolation forest require careful feature engineering when applied to structured domains like images. State-of-the-art methods aim to leverage deep learning to learn appropriate features via two main approaches. The first approach based on predicting transformations (Golan & El-Yaniv, 2018; Hendrycks et al., 2019a) while successful in some domains, crucially depends on an appropriate domain-specific set of transformations that are hard to obtain in general. The second approach of minimizing a classical one-class loss on the learned final layer representations, e.g., DeepSVDD (Ruff et al., 2018) suffers from the fundamental drawback of representation collapse. In this work, we propose Deep Robust One Class Classification (DROCC) that is both applicable to most standard domains without requiring any side-information and robust to representation collapse. DROCC is based on the assumption that the points from the class of interest lie on a well-sampled, locally linear low dimensional manifold. Empirical evaluation demonstrates that DROCC is highly effective in two different one-class problem settings and on a range of real-world datasets across different domains: tabular data, images (CIFAR and ImageNet), audio, and time-series, offering up to 20% increase in accuracy over the state-of-the-art in anomaly detection. Code is available at https://github.com/microsoft/EdgeML