Weixin Cai

Felix Faltings, Michel Galley, Baolin Peng et al. · microsoft-research

Users interact with text, image, code, or other editors on a daily basis. However, machine learning models are rarely trained in the settings that reflect the interactivity between users and their editor. This is understandable as training AI models with real users is not only slow and costly, but what these models learn may be specific to user interface design choices. Unfortunately, this means most of the research on text, code, and image generation has focused on non-interactive settings, whereby the model is expected to get everything right without accounting for any input from a user who may be willing to help. We introduce a new Interactive Text Generation task that allows training generation models interactively without the costs of involving real users, by using user simulators that provide edits that guide the model towards a given target text. We train our interactive models using Imitation Learning, and our experiments against competitive non-interactive generation models show that models trained interactively are superior to their non-interactive counterparts, even when all models are given the same budget of user inputs or edits.

Minghui Zhang, Alex Sokolov, Weixin Cai et al. · microsoft-research

Natural language generation (NLG) is one of the most impactful fields in NLP, and recent years have witnessed its evolution brought about by large language models (LLMs). As the key instrument for writing assistance applications, they are generally prone to replicating or extending offensive content provided in the input. In low-resource data regime, they can also lead to repetitive outputs. Usually, offensive content and repetitions are mitigated with post-hoc methods, including n-gram level blocklists, top-k and nucleus sampling. In this paper, we apply non-exact repetition suppression using token and sequence level unlikelihood loss, and further explore the framework of unlikelihood training objective in order to jointly endow the model with abilities to avoid generating offensive words and phrases from the beginning. Finally, with comprehensive experiments, we demonstrate that our proposed methods work exceptionally in controlling the repetition and content quality of LLM outputs.

Shirley Wu, Michel Galley, Baolin Peng et al.

Large Language Models are typically trained with next-turn rewards, limiting their ability to optimize for long-term interaction. As a result, they often respond passively to ambiguous or open-ended user requests, failing to help users reach their ultimate intents and leading to inefficient conversations. To address these limitations, we introduce CollabLLM, a novel and general training framework that enhances multiturn human-LLM collaboration. Its key innovation is a collaborative simulation that estimates the long-term contribution of responses using Multiturn-aware Rewards. By reinforcement fine-tuning these rewards, CollabLLM goes beyond responding to user requests, and actively uncovers user intent and offers insightful suggestions-a key step towards more human-centered AI. We also devise a multiturn interaction benchmark with three challenging tasks such as document creation. CollabLLM significantly outperforms our baselines with averages of 18.5% higher task performance and 46.3% improved interactivity by LLM judges. Finally, we conduct a large user study with 201 judges, where CollabLLM increases user satisfaction by 17.6% and reduces user spent time by 10.4%.

MohammadReza Davari, Utkarsh Garg, Weixin Cai et al.

An increasing number of NLP applications interact with large language models (LLMs) through black-box APIs, making prompt engineering critical for controlling model outputs. While recent Automatic Prompt Optimization (APO) methods iteratively refine prompts using model-generated feedback, textual gradients, they primarily focus on error correction and neglect valuable insights from correct predictions. This limits both their effectiveness and efficiency. In this paper, we propose a novel APO framework centered on enhancing the feedback mechanism. We reinterpret the textual gradient as a form of negative reinforcement and introduce the complementary positive reinforcement to explicitly preserve beneficial prompt components identified through successful predictions. To mitigate the noise inherent in LLM-generated feedback, we introduce a technique called feedback diversification, which aggregates multiple feedback signals, emphasizing consistent, actionable advice while filtering out outliers. Motivated by the rapid evolution and diversity of available LLMs, we also formalize Continual Prompt Optimization (CPO), addressing the practical challenge of efficiently migrating optimized prompts between different model versions or API providers. Our experiments reveal that naive prompt migration often degrades performance due to loss of critical instructions. In contrast, our approach consistently outperforms strong baselines, achieving significant accuracy improvements, faster convergence, and lower computational costs in both standard and migration scenarios.

Yao Dou, Michel Galley, Baolin Peng et al.

Large language models (LLMs) are increasingly used in interactive applications, and human evaluation remains the gold standard for assessing their performance in multi-turn conversations. Since human studies are costly, time-consuming, and hard to reproduce, recent work explores using LLMs to simulate users for automatic assistant evaluation. However, there is no benchmark or systematic study to evaluate whether these simulated users are reliable stand-ins for real users. To address this, we introduce SimulatorArena, a benchmark of 909 annotated human-LLM conversations on two interactive tasks -- math tutoring and document creation. SimulatorArena evaluates simulators based on how closely their messages match human behavior and how well their assistant ratings align with human judgments. Experiments on various simulator methods show that simulators conditioned on user profiles, capturing traits like background and message styles, align closely with human judgments. They reach Spearman's $ρ$ of 0.7 on both tasks, providing a practical, scalable alternative to human evaluation. Using the best simulator for each task, we benchmark 18 assistants, including the latest LLMs such as GPT-5, Claude 4.1 Opus, and Gemini 2.5 Pro.

Jeremy R. Coyle, Nima S. Hejazi, Ivana Malenica et al.

Targeted Learning is a subfield of statistics that unifies advances in causal inference, machine learning and statistical theory to help answer scientifically impactful questions with statistical confidence. Targeted Learning is driven by complex problems in data science and has been implemented in a diversity of real-world scenarios: observational studies with missing treatments and outcomes, personalized interventions, longitudinal settings with time-varying treatment regimes, survival analysis, adaptive randomized trials, mediation analysis, and networks of connected subjects. In contrast to the (mis)application of restrictive modeling strategies that dominate the current practice of statistics, Targeted Learning establishes a principled standard for statistical estimation and inference (i.e., confidence intervals and p-values). This multiply robust approach is accompanied by a guiding roadmap and a burgeoning software ecosystem, both of which provide guidance on the construction of estimators optimized to best answer the motivating question. The roadmap of Targeted Learning emphasizes tailoring statistical procedures so as to minimize their assumptions, carefully grounding them only in the scientific knowledge available. The end result is a framework that honestly reflects the uncertainty in both the background knowledge and the available data in order to draw reliable conclusions from statistical analyses - ultimately enhancing the reproducibility and rigor of scientific findings.

Weixin Cai, Nima S. Hejazi, Alan E. Hubbard

Current statistical inference problems in areas like astronomy, genomics, and marketing routinely involve the simultaneous testing of thousands -- even millions -- of null hypotheses. For high-dimensional multivariate distributions, these hypotheses may concern a wide range of parameters, with complex and unknown dependence structures among variables. In analyzing such hypothesis testing procedures, gains in efficiency and power can be achieved by performing variable reduction on the set of hypotheses prior to testing. We present in this paper an approach using data-adaptive multiple testing that serves exactly this purpose. This approach applies data mining techniques to screen the full set of covariates on equally sized partitions of the whole sample via cross-validation. This generalized screening procedure is used to create average ranks for covariates, which are then used to generate a reduced (sub)set of hypotheses, from which we compute test statistics that are subsequently subjected to standard multiple testing corrections. The principal advantage of this methodology lies in its providing valid statistical inference without the \textit{a priori} specifying which hypotheses will be tested. Here, we present the theoretical details of this approach, confirm its validity via a simulation study, and exemplify its use by applying it to the analysis of data on microRNA differential expression.