We introduce Phi-4-Mini and Phi-4-Multimodal, compact yet highly capable language and multimodal models. Phi-4-Mini is a 3.8-billion-parameter language model trained on high-quality web and synthetic data, significantly outperforming recent open-source models of similar size and matching the performance of models twice its size on math and coding tasks requiring complex reasoning. This achievement is driven by a carefully curated synthetic data recipe emphasizing high-quality math and coding datasets. Compared to its predecessor, Phi-3.5-Mini, Phi-4-Mini features an expanded vocabulary size of 200K tokens to better support multilingual applications, as well as group query attention for more efficient long-sequence generation. Phi-4-Multimodal is a multimodal model that integrates text, vision, and speech/audio input modalities into a single model. Its novel modality extension approach leverages LoRA adapters and modality-specific routers to allow multiple inference modes combining various modalities without interference. For example, it now ranks first in the OpenASR leaderboard to date, although the LoRA component of the speech/audio modality has just 460 million parameters. Phi-4-Multimodal supports scenarios involving (vision + language), (vision + speech), and (speech/audio) inputs, outperforming larger vision-language and speech-language models on a wide range of tasks. Additionally, we experiment to further train Phi-4-Mini to enhance its reasoning capabilities. Despite its compact 3.8-billion-parameter size, this experimental version achieves reasoning performance on par with or surpassing significantly larger models, including DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-Llama-8B.

We introduce phi-3-mini, a 3.8 billion parameter language model trained on 3.3 trillion tokens, whose overall performance, as measured by both academic benchmarks and internal testing, rivals that of models such as Mixtral 8x7B and GPT-3.5 (e.g., phi-3-mini achieves 69% on MMLU and 8.38 on MT-bench), despite being small enough to be deployed on a phone. The innovation lies entirely in our dataset for training, a scaled-up version of the one used for phi-2, composed of heavily filtered publicly available web data and synthetic data. The model is also further aligned for robustness, safety, and chat format. We also provide some initial parameter-scaling results with a 7B and 14B models trained for 4.8T tokens, called phi-3-small and phi-3-medium, both significantly more capable than phi-3-mini (e.g., respectively 75% and 78% on MMLU, and 8.7 and 8.9 on MT-bench). Moreover, we also introduce phi-3-vision, a 4.2 billion parameter model based on phi-3-mini with strong reasoning capabilities for image and text prompts.

Recent advances in Named Entity Recognition (NER) show that document-level contexts can significantly improve model performance. In many application scenarios, however, such contexts are not available. In this paper, we propose to find external contexts of a sentence by retrieving and selecting a set of semantically relevant texts through a search engine, with the original sentence as the query. We find empirically that the contextual representations computed on the retrieval-based input view, constructed through the concatenation of a sentence and its external contexts, can achieve significantly improved performance compared to the original input view based only on the sentence. Furthermore, we can improve the model performance of both input views by Cooperative Learning, a training method that encourages the two input views to produce similar contextual representations or output label distributions. Experiments show that our approach can achieve new state-of-the-art performance on 8 NER data sets across 5 domains.

Knowledge distillation is a critical technique to transfer knowledge between models, typically from a large model (the teacher) to a smaller one (the student). The objective function of knowledge distillation is typically the cross-entropy between the teacher and the student's output distributions. However, for structured prediction problems, the output space is exponential in size; therefore, the cross-entropy objective becomes intractable to compute and optimize directly. In this paper, we derive a factorized form of the knowledge distillation objective for structured prediction, which is tractable for many typical choices of the teacher and student models. In particular, we show the tractability and empirical effectiveness of structural knowledge distillation between sequence labeling and dependency parsing models under four different scenarios: 1) the teacher and student share the same factorization form of the output structure scoring function; 2) the student factorization produces smaller substructures than the teacher factorization; 3) the teacher factorization produces smaller substructures than the student factorization; 4) the factorization forms from the teacher and the student are incompatible. Deeper and larger neural networks have led to significant improvement in accuracy in various tasks, but they are also more computationally expensive and unfit for resource-constrained scenarios such as online serving. An interesting and viable solution to this problem is knowledge distillation (KD) (Buciluǎ et al., 2006; Ba & Caruana, 2014; Hinton et al., 2015), which can be used to transfer the knowledge of a large model (the teacher) to a smaller model (the student). In the field of natural language processing, for example, KD has been successfully applied to compress massive pretrained language models such as BERT (Devlin et al., 2019) and XLM-R (Conneau et al., 2020) into much smaller and faster models without significant loss in accuracy (Tang et al., 2019; Sanh et al., 2019; Tsai et al., 2019; Mukherjee & Hassan Awadallah, 2020).

Pretrained contextualized embeddings are powerful word representations for structured prediction tasks. Recent work found that better word representations can be obtained by concatenating different types of embeddings. However, the selection of embeddings to form the best concatenated representation usually varies depending on the task and the collection of candidate embeddings, and the ever-increasing number of embedding types makes it a more difficult problem. In this paper, we propose Automated Concatenation of Embeddings (ACE) to automate the process of finding better concatenations of embeddings for structured prediction tasks, based on a formulation inspired by recent progress on neural architecture search. Specifically, a controller alternately samples a concatenation of embeddings, according to its current belief of the effectiveness of individual embedding types in consideration for a task, and updates the belief based on a reward. We follow strategies in reinforcement learning to optimize the parameters of the controller and compute the reward based on the accuracy of a task model, which is fed with the sampled concatenation as input and trained on a task dataset. Empirical results on 6 tasks and 23 datasets show that our approach outperforms strong baselines and achieves state-of-the-art performance with fine-tuned embeddings in the vast majority of evaluations.

The linear-chain Conditional Random Field (CRF) model is one of the most widely-used neural sequence labeling approaches. Exact probabilistic inference algorithms such as the forward-backward and Viterbi algorithms are typically applied in training and prediction stages of the CRF model. However, these algorithms require sequential computation that makes parallelization impossible. In this paper, we propose to employ a parallelizable approximate variational inference algorithm for the CRF model. Based on this algorithm, we design an approximate inference network that can be connected with the encoder of the neural CRF model to form an end-to-end network, which is amenable to parallelization for faster training and prediction. The empirical results show that our proposed approaches achieve a 12.7-fold improvement in decoding speed with long sentences and a competitive accuracy compared with the traditional CRF approach.