The hallucinations of large language models (LLMs) are increasingly mitigated by allowing LLMs to search for information and to ground their answers in real sources. Unfortunately, LLMs often struggle with posing the right search queries, especially when dealing with complex or otherwise indirect topics. Observing that LLMs can learn to search for relevant facts by $\textit{trying}$ different queries and learning to up-weight queries that successfully produce relevant results, we introduce $\underline{Le}$arning to $\underline{Re}$trieve by $\underline{T}$rying (LeReT), a reinforcement learning framework that explores search queries and uses preference-based optimization to improve their quality. LeReT can improve the absolute retrieval accuracy by up to 29\% and the downstream generator evaluations by 17\%. The simplicity and flexibility of LeReT allows it to be applied to arbitrary off-the-shelf retrievers and makes it a promising technique for improving general LLM pipelines.

This paper is the first attempt at providing a holistic view of the Chrome Web Store (CWS). We leverage historical data provided by ChromeStats to study global trends in the CWS and security implications. We first highlight the extremely short life cycles of extensions: roughly 60% of extensions stay in the CWS for one year. Second, we define and show that Security-Noteworthy Extensions (SNE) are a significant issue: they pervade the CWS for years and affect almost 350 million users. Third, we identify clusters of extensions with a similar code base. We discuss how code similarity techniques could be used to flag suspicious extensions. By developing an approach to extract URLs from extensions' comments, we show that extensions reuse code snippets from public repositories or forums, leading to the propagation of dated code and vulnerabilities. Finally, we underline a critical lack of maintenance in the CWS: 60% of the extensions in the CWS have never been updated; half of the extensions known to be vulnerable are still in the CWS and still vulnerable 2 years after disclosure; a third of extensions use vulnerable library versions. We believe that these issues should be widely known in order to pave the way for a more secure CWS.

The diversity of contexts in which large language models (LLMs) are deployed requires the ability to modify or customize default model behaviors to incorporate nuanced requirements and preferences. A convenient interface to specify such model adjustments is high-level verbal feedback, such as “Don’t use emojis when drafting emails to my boss.” However, while writing high-level feedback is far simpler than collecting annotations for reinforcement learning from human feedback (RLHF), we find that simply prompting a model with such feedback leads to $\textbf{overgeneralization}$–applying feedback in contexts where it is not relevant. We propose a new method Contextualized Critiques with Constrained Preference Optimization (C3PO) to learn from high-level verbal feedback while reducing overgeneralization compared to current work. C3PO uses a piece of high-level feedback to generate a small synthetic preference dataset to specify when and how the feedback should (and should not) be applied. It then fine-tunes the model in accordance with the synthetic preference data while minimizing the divergence from the original model for prompts where the feedback does not apply. Our experimental results indicate that our approach effectively applies verbal feedback to relevant scenarios while preserving existing behaviors for other contexts more than current methods. For both human- and GPT-4-generated high-level feedback, C3PO effectively adheres to the given feedback comparably to in-context baselines while reducing overgeneralization by 30%.