In this section, we provide further details for our data modeling. We note the difficulties of appropriately modeling the terminal variable which is a binary variable compared to the rest of the dimensions which are continuous for the environments we investigate. This is particularly challenging for "expert" datasets where early termination is rare. An immediate advantage of sampling data from a generative model is compression. As we discuss in Appendix B.3, sampling is fast ER provides high levels of dataset compression without sacrificing downstream performance in offline reinforcement learning.

Real-world evaluation of perception-based planning models for robotic systems, such as autonomous vehicles, can be safely and inexpensively conducted offline, i.e. by computing model prediction error over a pre-collected validation dataset with ground-truth annotations. However, extrapolating from offline model performance to online settings remains a challenge. In these settings, seemingly minor errors can compound and result in test-time infractions or collisions. This relationship is understudied, particularly across diverse closed-loop metrics and complex urban maneuvers. In this work, we revisit this undervalued question in policy evaluation through an extensive set of experiments across diverse conditions and metrics. Based on analysis in simulation, we find an even worse correlation between offline and online settings than reported by prior studies, casting doubts on the validity of current evaluation practices and metrics for driving policies. Next, we bridge the gap between offline and online evaluation. We investigate an offline metric based on epistemic uncertainty, which aims to capture events that are likely to cause errors in closed-loop settings. The resulting metric achieves over 13% improvement in correlation compared to previous offline metrics. We further validate the generalization of our findings beyond the simulation environment in real-world settings, where even greater gains are observed.

Standard offline evaluations for language models -- a series of independent, state-less inferences made by models -- fail to capture how language models actually behave in practice, where personalization fundamentally alters model behavior. For instance, identical benchmark questions to the same language model can produce markedly different responses when prompted to a state-less system, in one user's chat session, or in a different user's chat session. In this work, we provide empirical evidence showcasing this phenomenon by comparing offline evaluations to field evaluations conducted by having 800 real users of ChatGPT and Gemini pose benchmark and other provided questions to their chat interfaces.

Recommender systems often must maximize a primary objective while ensuring secondary ones satisfy minimum thresholds, or "guardrails." This is critical for maintaining a consistent user experience and platform ecosystem, but enforcing these guardrails despite orthogonal system changes is challenging and often requires manual hyperparameter tuning. We introduce the Automated Constraint Targeting (ACT) framework, which automatically finds the minimal set of hyperparameter changes needed to satisfy these guardrails. ACT uses an offline pairwise evaluation on unbiased data to find solutions and continuously retrains to adapt to system and user behavior changes. We empirically demonstrate its efficacy and describe its deployment in a large-scale production environment.

Error correction is an important capability when applying large language models (LLMs) to facilitate user typing on mobile devices. In this paper, we use LLMs to synthesize a high-quality dataset of error correction pairs to evaluate and improve LLMs for mobile applications. We first prompt LLMs with error correction domain knowledge to build a scalable and reliable addition to the existing data synthesis pipeline. We then adapt the synthetic data distribution to match the mobile application domain by reweighting the samples. The reweighting model is learnt by predicting (a handful of) live A/B test metrics when deploying LLMs in production, given the LLM performance on offline evaluation data and scores from a small privacy-preserving on-device language model. Finally, we present best practices for mixing our synthetic data with other data sources to improve model performance on error correction in both offline evaluation and production live A/B testing.

When a Commit Message Generation (CMG) system is integrated into the IDEs and other products at JetBrains, we perform online evaluation based on user acceptance of the generated messages. However, performing online experiments with every change to a CMG system is troublesome, as each iteration affects users and requires time to collect enough statistics. On the other hand, offline evaluation, a prevalent approach in the research literature, facilitates fast experiments but employs automatic metrics that are not guaranteed to represent the preferences of real users. In this work, we describe a novel way we employed to deal with this problem at JetBrains, by leveraging an online metric - the number of edits users introduce before committing the generated messages to the VCS - to select metrics for offline experiments. To support this new type of evaluation, we develop a novel markup collection tool mimicking the real workflow with a CMG system, collect a dataset with 57 pairs consisting of commit messages generated by GPT-4 and their counterparts edited by human experts, and design and verify a way to synthetically extend such a dataset. Then, we use the final dataset of 656 pairs to study how the widely used similarity metrics correlate with the online metric reflecting the real users' experience. Our results indicate that edit distance exhibits the highest correlation with the online metric, whereas commonly used similarity metrics such as BLEU and METEOR demonstrate low correlation. This contradicts the previous studies on similarity metrics for CMG, suggesting that user interactions with a CMG system in real-world settings differ significantly from the responses by human labelers within controlled environments. We release all the code and the dataset to support future research in the field: https://jb.gg/cmg-evaluation.

Multi-objective alignment from human feedback (MOAHF) in large language models (LLMs) is a challenging problem as human preferences are complex, multifaceted, and often conflicting. Recent works on MOAHF considered a-priori multi-objective optimization (MOO), where human preferences are known at training or inference time. In contrast, when human preferences are unknown or difficult to quantify, a natural approach is to cover the Pareto front by multiple diverse solutions. We propose an algorithm HaM for learning diverse LLM policies that maximizes their hypervolume. This is the first application of a-posteriori MOO to MOAHF. HaM is computationally and space efficient, and empirically superior across objectives such as harmlessness, helpfulness, humor, faithfulness, and hallucination, on various datasets.