Lean is an advanced proof assistant designed to facilitate formal theorem proving by providing a variety of interactive feedback. In this paper, we explore methodologies to leverage proof assistant feedback to augment the capabilities of large language models in constructing formal proofs. First, we deploy online reinforcement learning using Lean verification outcomes as the reward signal to improve the proof completion policy. This straightforward approach shows great promise in enhancing the model's alignment with the formal verification system. In addition, we propose RMaxTS, a variant of Monte-Carlo tree search that employs an intrinsic-reward-driven exploration strategy to generate diverse proof paths. The tree structure is organized to represent the transitions of intermediate tactic states, extracted from the compilation messages given by Lean's tactic mode. The intrinsic reward is constructed to incentivize the discovery of novel tactic states, which helps to to mitigate the sparse-reward problem inherent in proof search. These techniques lead to a more efficient planning scheme for formal proof generation, achieving new state-of-the-art results on both miniF2F and ProofNet benchmarks.

Reasoning is a fundamental capability of Large Language Models. While prior research predominantly focuses on enhancing narrow skills like math or code generation, improving performance on many other reasoning tasks remains challenging due to sparse and fragmented training data. To address this issue, we propose CodeI/O, a novel approach that systematically condenses diverse reasoning patterns inherently embedded in contextually-grounded codes, through transforming the original code into a code input-output prediction format. By training models to predict inputs/outputs given code and test cases entirely in natural language as Chain-of-Thought (CoT) rationales, we expose them to universal reasoning primitives—like logic flow planning, state-space searching, decision tree traversal, and modular decomposition—while decoupling structured reasoning from code-specific syntax and preserving procedural rigor. Experimental results demonstrate CodeI/O leads to consistent improvements across symbolic, scientific, logic, math & numerical, and commonsense reasoning tasks. By matching the existing ground-truth outputs or re-executing the code with predicted inputs, we can verify each prediction and further enhance the CoTs through multi-turn revision, resulting in CodeI/O++ and achieving higher performance. Our data and models will be publicly available.

Code generation is a longstanding challenge, aiming to generate a code snippet based on a natural language description. Usually, expensive text-code paired data is essential for training a code generation model. Recently, thanks to the success of pre-training techniques, large language models are trained on large unlabelled code corpora and perform well in generating code. In this paper, we investigate how to leverage an unlabelled code corpus to train a model for library-oriented code generation. Since it is a common practice for programmers to reuse third-party libraries, in which case the text-code paired data are harder to obtain due to the huge number of libraries. We observe that library-oriented code snippets are more likely to share similar code sketches. Hence, we present CERT with two steps: a sketcher generates the sketch, then a generator fills the details in the sketch. Both the sketcher and generator are continually pre-trained upon a base model using unlabelled data. Also, we carefully craft two benchmarks to evaluate library-oriented code generation named PandasEval and NumpyEval. Experimental results have shown the impressive performance of CERT. For example, it surpasses the base model by an absolute 15.67% improvement in terms of pass@1 on PandasEval. Our work is available at https://github.com/microsoft/PyCodeGPT.

Task-oriented virtual assistants are software systems that provide users with a natural language interface to complete domain-specific tasks. With the recent technological advances in natural language processing and machine learning, an increasing number of task-oriented virtual assistants have been developed. However, due to the well-known complexity and difficulties of the natural language understanding problem, it is challenging to manage a task-oriented virtual assistant software project. Meanwhile, the management and experience related to the development of virtual assistants are hardly studied or shared in the research community or industry, to the best of our knowledge. To bridge this knowledge gap, in this paper, we share our experience and the lessons that we have learned at managing a task-oriented virtual assistant software project at Microsoft. We believe that our practices and the lessons learned can provide a useful reference for other researchers and practitioners who aim to develop a virtual assistant system. Finally, we have developed a requirement management tool, named SpecSpace, which can facilitate the management of virtual assistant projects.

We study response generation for open domain conversation in chatbots. Existing methods assume that words in responses are generated from an identical vocabulary regardless of their inputs, which not only makes them vulnerable to generic patterns and irrelevant noise, but also causes a high cost in decoding. We propose a dynamic vocabulary sequence-to-sequence (DVS2S) model which allows each input to possess their own vocabulary in decoding. In training, vocabulary construction and response generation are jointly learned by maximizing a lower bound of the true objective with a Monte Carlo sampling method. In inference, the model dynamically allocates a small vocabulary for an input with the word prediction model, and conducts decoding only with the small vocabulary. Because of the dynamic vocabulary mechanism, DVS2S eludes many generic patterns and irrelevant words in generation, and enjoys efficient decoding at the same time. Experimental results on both automatic metrics and human annotations show that DVS2S can significantly outperform state-of-the-art methods in terms of response quality, but only requires 60% decoding time compared to the most efficient baseline.

As an effective way of learning node representations in networks, network embedding has attracted increasing research interests recently. Most existing approaches use shallow models and only work on static networks by extracting local or global topology information of each node as the algorithm input. It is challenging for such approaches to learn a desirable node representation on incomplete graphs with a large number of missing links or on dynamic graphs with new nodes joining in. It is even challenging for them to deeply fuse other types of data such as node properties into the learning process to help better represent the nodes with insufficient links. In this paper, we for the first time study the problem of network embedding on incomplete networks. We propose a Multi-View Correlation-learning based Deep Network Embedding method named MVC-DNE to incorporate both the network structure and the node properties for more effectively and efficiently perform network embedding on incomplete networks. Specifically, we consider the topology structure of the network and the node properties as two correlated views. The insight is that the learned representation vector of a node should reflect its characteristics in both views. Under a multi-view correlation learning based deep autoencoder framework, the structure view and property view embeddings are integrated and mutually reinforced through both self-view and cross-view learning. As MVC-DNE can learn a representation mapping function, it can directly generate the representation vectors for the new nodes without retraining the model. Thus it is especially more efficient than previous methods. Empirically, we evaluate MVC-DNE over three real network datasets on two data mining applications, and the results demonstrate that MVC-DNE significantly outperforms state-of-the-art methods.