Modern conversational agents like ChatGPT and Alexa+ rely on predefined policies specifying metadata, response styles, and tool-usage rules. As these LLM-based systems expand to support diverse business and user queries, such policies, often implemented as in-context prompts, are becoming increasingly complex and lengthy, making faithful adherence difficult and imposing large fixed computational costs. With the rise of multimodal agents, policies that govern visual and multimodal behaviors are critical but remain understudied. Prior prompt-compression work mainly shortens task templates and demonstrations, while existing policy-alignment studies focus only on text-based safety rules. We introduce Multimodal Policy Internalization (MPI), a new task that internalizes reasoning-intensive multimodal policies into model parameters, enabling stronger policy-following without including the policy during inference. MPI poses unique data and algorithmic challenges. We build two datasets spanning synthetic and real-world decision-making and tool-using tasks and propose TriMPI, a three-stage training framework. TriMPI first injects policy knowledge via continual pretraining, then performs supervised finetuning, and finally applies PolicyRollout, a GRPO-style reinforcement learning extension that augments rollouts with policy-aware responses for grounded exploration. TriMPI achieves notable gains in end-to-end accuracy, generalization, and robustness to forgetting. As the first work on multimodal policy internalization, we provide datasets, training recipes, and comprehensive evaluations to foster future research. Project page: https://mikewangwzhl.github.io/TriMPI.

In many scenarios, humans prefer a text-based representation of quantitative data over numerical, tabular, or graphical representations. The attractiveness of textual summaries for complex data has inspired research on data-to-text systems. While there are several data-to-text tools for time series, few of them try to mimic how humans summarize for time series. In this paper, we propose a model to create human-like text descriptions for time series. Our system finds patterns in time series data and ranks these patterns based on empirical observations of human behavior using utility estimation. Our proposed utility estimation model is a Bayesian network capturing interdependencies between different patterns. We describe the learning steps for this network and introduce baselines along with their performance for each step. The output of our system is a natural language description of time series that attempts to match a human's summary of the same data.