In many real-world settings, regulations and economic incentives permit the sharing of models but not data across institutional boundaries. In such scenarios, practitioners might hope to adapt models to new domains, without losing performance on previous domains (so-called catastrophic forgetting). While any single model may struggle to achieve this goal, learning an ensemble of domain-specific experts offers the potential to adapt more closely to each individual institution. However, a core challenge in this context is determining which expert to deploy at test time. In this paper, we propose Generate to Discriminate (G2D), a domain-incremental continual learning method that leverages synthetic data to train a domain-discriminator that routes samples at inference time to the appropriate expert. Surprisingly, we find that leveraging synthetic data in this capacity is more effective than using the samples to \textit{directly} train the downstream classifier (the more common approach to leveraging synthetic data in the lifelong learning literature). We observe that G2D outperforms competitive domain-incremental learning methods on tasks in both vision and language modalities, providing a new perspective on the use of synthetic data in the lifelong learning literature.

Several recent works seek to develop foundation models specifically for medical applications, adapting general-purpose large language models (LLMs) and vision-language models (VLMs) via continued pretraining on publicly available biomedical corpora. These works typically claim that such domain-adaptive pretraining (DAPT) improves performance on downstream medical tasks, such as answering medical licensing exam questions. In this paper, we compare seven public "medical" LLMs and two VLMs against their corresponding base models, arriving at a different conclusion: all medical VLMs and nearly all medical LLMs fail to consistently improve over their base models in the zero-/few-shot prompting regime for medical question-answering (QA) tasks. For instance, across the tasks and model pairs we consider in the 3-shot setting, medical LLMs only outperform their base models in 12.1% of cases, reach a (statistical) tie in 49.8% of cases, and are significantly worse than their base models in the remaining 38.2% of cases. Our conclusions are based on (i) comparing each medical model head-to-head, directly against the corresponding base model; (ii) optimizing the prompts for each model separately; and (iii) accounting for statistical uncertainty in comparisons. While these basic practices are not consistently adopted in the literature, our ablations show that they substantially impact conclusions. Our findings suggest that state-of-the-art general-domain models may already exhibit strong medical knowledge and reasoning capabilities, and offer recommendations to strengthen the conclusions of future studies.

Several recent works seek to develop foundation models specifically for medical applications, adapting general-purpose large language models (LLMs) and vision-language models (VLMs) via continued pretraining on publicly available biomedical corpora. These works typically claim that such domain-adaptive pretraining (DAPT) improves performance on downstream medical tasks, such as answering medical licensing exam questions. In this paper, we compare ten public "medical" LLMs and two VLMs against their corresponding base models, arriving at a different conclusion: all medical VLMs and nearly all medical LLMs fail to consistently improve over their base models in the zero-/few-shot prompting and supervised fine-tuning regimes for medical question-answering (QA). For instance, across all tasks and model pairs we consider in the 3-shot setting, medical LLMs only outperform their base models in 22.7% of cases, reach a (statistical) tie in 36.8% of cases, and are significantly worse than their base models in the remaining 40.5% of cases. Our conclusions are based on (i) comparing each medical model head-to-head, directly against the corresponding base model; (ii) optimizing the prompts for each model separately in zero-/few-shot prompting; and (iii) accounting for statistical uncertainty in comparisons. While these basic practices are not consistently adopted in the literature, our ablations show that they substantially impact conclusions. Meanwhile, we find that after fine-tuning on specific QA tasks, medical LLMs can show performance improvements, but the benefits do not carry over to tasks based on clinical notes. Our findings suggest that state-of-the-art general-domain models may already exhibit strong medical knowledge and reasoning capabilities, and offer recommendations to strengthen the conclusions of future studies.

We introduce Pixtral-12B, a 12--billion-parameter multimodal language model. Pixtral-12B is trained to understand both natural images and documents, achieving leading performance on various multimodal benchmarks, surpassing a number of larger models. Unlike many open-source models, Pixtral is also a cutting-edge text model for its size, and does not compromise on natural language performance to excel in multimodal tasks. Pixtral uses a new vision encoder trained from scratch, which allows it to ingest images at their natural resolution and aspect ratio. This gives users flexibility on the number of tokens used to process an image. Pixtral is also able to process any number of images in its long context window of 128K tokens. Pixtral 12B substanially outperforms other open models of similar sizes (Llama-3.2 11B \& Qwen-2-VL 7B). It also outperforms much larger open models like Llama-3.2 90B while being 7x smaller. We further contribute an open-source benchmark, MM-MT-Bench, for evaluating vision-language models in practical scenarios, and provide detailed analysis and code for standardized evaluation protocols for multimodal LLMs. Pixtral-12B is released under Apache 2.0 license.

Training on model-generated synthetic data is a promising approach for finetuning LLMs, but it remains unclear when it helps or hurts. In this paper, we investigate this question for math reasoning via an empirical study, followed by building a conceptual understanding of our observations. First, we find that while the typical approach of finetuning a model on synthetic correct or positive problem-solution pairs generated by capable models offers modest performance gains, sampling more correct solutions from the finetuned learner itself followed by subsequent fine-tuning on this self-generated data $\textbf{doubles}$ the efficiency of the same synthetic problems. At the same time, training on model-generated positives can amplify various spurious correlations, resulting in flat or even inverse scaling trends as the amount of data increases. Surprisingly, we find that several of these issues can be addressed if we also utilize negative responses, i.e., model-generated responses that are deemed incorrect by a final answer verifier. Crucially, these negatives must be constructed such that the training can appropriately recover the utility or advantage of each intermediate step in the negative response. With this per-step scheme, we are able to attain consistent gains over only positive data, attaining performance similar to amplifying the amount of synthetic data by $\mathbf{8 \times}$. We show that training on per-step negatives can help to unlearn spurious correlations in the positive data, and is equivalent to advantage-weighted reinforcement learning (RL), implying that it inherits robustness benefits of RL over imitating positive data alone.

Trained models are often composed with post-hoc transforms such as temperature scaling (TS), ensembling and stochastic weight averaging (SWA) to improve performance, robustness, uncertainty estimation, etc. However, such transforms are typically applied only after the base models have already been finalized by standard means. In this paper, we challenge this practice with an extensive empirical study. In particular, we demonstrate a phenomenon that we call post-hoc reversal, where performance trends are reversed after applying these post-hoc transforms. This phenomenon is especially prominent in high-noise settings. For example, while base models overfit badly early in training, both conventional ensembling and SWA favor base models trained for more epochs. Post-hoc reversal can also suppress the appearance of double descent and mitigate mismatches between test loss and test error seen in base models. Based on our findings, we propose post-hoc selection, a simple technique whereby post-hoc metrics inform model development decisions such as early stopping, checkpointing, and broader hyperparameter choices. Our experimental analyses span real-world vision, language, tabular and graph datasets from domains like satellite imaging, language modeling, census prediction and social network analysis. On an LLM instruction tuning dataset, post-hoc selection results in > 1.5x MMLU improvement compared to naive selection. Code is available at https://github.com/rishabh-ranjan/post-hoc-reversal.

Self-training and contrastive learning have emerged as leading techniques for incorporating unlabeled data, both under distribution shift (unsupervised domain adaptation) and when it is absent (semi-supervised learning). However, despite the popularity and compatibility of these techniques, their efficacy in combination remains unexplored. In this paper, we undertake a systematic empirical investigation of this combination, finding that (i) in domain adaptation settings, self-training and contrastive learning offer significant complementary gains; and (ii) in semi-supervised learning settings, surprisingly, the benefits are not synergistic. Across eight distribution shift datasets (e.g., BREEDs, WILDS), we demonstrate that the combined method obtains 3--8% higher accuracy than either approach independently. We then theoretically analyze these techniques in a simplified model of distribution shift, demonstrating scenarios under which the features produced by contrastive learning can yield a good initialization for self-training to further amplify gains and achieve optimal performance, even when either method alone would fail.

Keeping large foundation models up to date on latest data is inherently expensive. To avoid the prohibitive costs of constantly retraining, it is imperative to continually train these models. This problem is exacerbated by the lack of any large scale continual learning benchmarks or baselines. We introduce the first set of web-scale Time-Continual (TiC) benchmarks for training vision-language models: TiC-DataCompt, TiC-YFCC, and TiC-RedCaps with over 12.7B timestamped image-text pairs spanning 9 years (2014--2022). We first use our benchmarks to curate various dynamic evaluations to measure temporal robustness of existing models. We show OpenAI's CLIP (trained on data up to 2020) loses $\approx 8\%$ zero-shot accuracy on our curated retrieval task from 2021--2022 compared with more recently trained models in OpenCLIP repository. We then study how to efficiently train models on time-continuous data. We demonstrate that a simple rehearsal-based approach that continues training from the last checkpoint and replays old data reduces compute by $2.5\times$ when compared to the standard practice of retraining from scratch.

Despite the emergence of principled methods for domain adaptation under label shift, their sensitivity to shifts in class conditional distributions is precariously under explored. Meanwhile, popular deep domain adaptation heuristics tend to falter when faced with label proportions shifts. While several papers modify these heuristics in attempts to handle label proportions shifts, inconsistencies in evaluation standards, datasets, and baselines make it difficult to gauge the current best practices. In this paper, we introduce RLSbench, a large-scale benchmark for relaxed label shift, consisting of $>$500 distribution shift pairs spanning vision, tabular, and language modalities, with varying label proportions. Unlike existing benchmarks, which primarily focus on shifts in class-conditional $p(x|y)$, our benchmark also focuses on label marginal shifts. First, we assess 13 popular domain adaptation methods, demonstrating more widespread failures under label proportion shifts than were previously known. Next, we develop an effective two-step meta-algorithm that is compatible with most domain adaptation heuristics: (i) pseudo-balance the data at each epoch; and (ii) adjust the final classifier with target label distribution estimate. The meta-algorithm improves existing domain adaptation heuristics under large label proportion shifts, often by 2--10\% accuracy points, while conferring minimal effect ($<$0.5\%) when label proportions do not shift. We hope that these findings and the availability of RLSbench will encourage researchers to rigorously evaluate proposed methods in relaxed label shift settings. Code is publicly available at https://github.com/acmi-lab/RLSbench.

When deploying a model, it is important to be confident in how it will perform under inevitable distribution shift. Standard methods for achieving this include data dependent uniform convergence bounds (Ben-David et al., 2006, Mansour et al., 2009) (typically vacuous in practice) or assuming a precise model of how the distribution can shift (Chen et al., 2022, Rahimian and Mehrotra, 2019, Rosenfeld et al., 2021). Unfortunately, it is difficult or impossible to determine how severely these assumptions are violated by real data ("all models are wrong"), so practitioners usually cannot trust such bounds with confidence. To better estimate test performance in the wild, some recent work instead tries to directly predict accuracy of neural networks using unlabeled data from the test distribution of interest, (Baek et al., 2022, Garg et al., 2022, Lu et al., 2023). While these methods predict the test performance surprisingly well, they lack pointwise trustworthiness and verifiability: their estimates are good on average over all distribution shifts, but they provide no guarantee or signal of the quality of any individual prediction (here, each point is a single test distribution, for which a method predicts a classifier's average accuracy). Because of the opaque conditions under which these methods work, it is also difficult to anticipate their failure cases--indeed, it is reasonably common for them to substantially overestimate test accuracy for a particular shift, which is problematic when optimistic deployment can be costly or catastrophic.