We assess the robustness of large language models (LLMs) on a simple task—fast-food order specification—in the face of user-initiated amendments. In amending an order, users must communicate, explicitly or implicitly, that a change is being made, the type of change, and what is being changed. We generate synthetic food orders with single amendments, controlling the type of amendment, which components are implicitly or incompletely specified, and the pragmatic inferences required to resolve them. On a dataset broadly covering the space of possible combinations, 7 flagship models average 97.7% accuracy. On a more targeted dataset with denser coverage of empirically difficult patterns, accuracy drops to 93.4%, with the hardest combinations dropping below 80% accuracy. We also test two further prompting methods: generating an order before interpreting amendments in a turn-taking interaction; and function-calling to parameterize explicit edit operations on the order itself. These somewhat improve performance (+0.2 pp and +3.2 pp respectively), but neither entirely eliminates the failure modes. We also find that asking clarification questions about implicit components improves accuracy on a subset of difficult examples (+5.4 pp), but models ask on only 11% of examples. Given the task’s deliberate simplicity—small fixed menu, a single standard order, with a single unambiguous amendment—these results illustrate the challenge of robustly performing the pragmatic inferences that real-world conversational interaction demands.

Conversational Recommendation Systems (CRSs) are a particularly interesting application for out-of-the-box LLMs due to their potential for eliciting user preferences and making recommendations in natural language across a wide set of domains. Somewhat surprisingly, we find however that in such a conversational application, the more questions a user answers about their preferences, the worse the model’s recommendations become. We demonstrate this phenomenon on a previously published dataset as well as two novel datasets which we contribute. We also explain why earlier benchmarks failed to detect this round-over-round performance loss, highlighting the importance of the evaluation strategy we use and expanding upon Li et al. (2023a). We also present preference elicitation and recommendation strategies that mitigate this degradation in performance, beating state-of-the-art results, and show how three underlying models, GPT-3.5, GPT-4, and Claude 3.5 Sonnet, differently impact these strategies. Our datasets and code are available at https://github.com/CtrlVGustavo/A-Flash- in-the-Pan-CRS.

How to usefully encode compositional task structure has long been a core challenge in AI. Recent work in chain of thought prompting has shown that for very large neural language models (LMs), explicitly demonstrating the inferential steps involved in a target task may improve performance over end-to-end learning that focuses on the target task alone. However, chain of thought prompting has significant limitations due to its dependency on huge pretrained LMs. In this work, we present compositional fine-tuning (CFT): an approach based on explicitly decomposing a target task into component tasks, and then fine-tuning smaller LMs on a curriculum of such component tasks. We apply CFT to recommendation tasks in two domains, world travel and local dining, as well as a previously studied inferential task (sports understanding). We show that CFT outperforms end-to-end learning even with equal amounts of data, and gets consistently better as more component tasks are modeled via fine-tuning. Compared with chain of thought prompting, CFT performs at least as well using LMs only 7.4% of the size, and is moreover applicable to task domains for which data are not available during pretraining.

Conversations aimed at determining good recommendations are iterative in nature. People often express their preferences in terms of a critique of the current recommendation (e.g., “It doesn’t look good for a date”), requiring some degree of common sense for a preference to be inferred. In this work, we present a method for transforming a user critique into a positive preference (e.g., “I prefer more romantic”) in order to retrieve reviews pertaining to potentially better recommendations (e.g., “Perfect for a romantic dinner”). We leverage a large neural language model (LM) in a few-shot setting to perform critique-to-preference transformation, and we test two methods for retrieving recommendations: one that matches embeddings, and another that fine-tunes an LM for the task. We instantiate this approach in the restaurant domain and evaluate it using a new dataset of restaurant critiques. In an ablation study, we show that utilizing critique-to-preference transformation improves recommendations, and that there are at least three general cases that explain this improved performance.

Intelligent systems require common sense, but automatically extracting this knowledge from text can be difficult. We propose and assess methods for extracting one type of commonsense knowledge, object-property comparisons, from pre-trained embeddings. In experiments, we show that our approach exceeds the accuracy of previous work but requires substantially less hand-annotated knowledge. Further, we show that an active learning approach that synthesizes common-sense queries can boost accuracy.