Knowledge Distillation for Enhancing Walmart E-commerce Search Relevance Using Large Language Models

Ensuring the products displayed in e-commerce search results are relevant to users queries is crucial for improving the user experience. With their advanced semantic understanding, deep learning models have been widely used for relevance matching in search tasks. While large language models (LLMs) offer superior ranking capabilities, it is challenging to deploy LLMs in real-time systems due to the high-latency requirements. To leverage the ranking power of LLMs while meeting the low-latency demands of production systems, we propose a novel framework that distills a high performing LLM into a more efficient, low-latency student model. To help the student model learn more effectively from the teacher model, we first train the teacher LLM as a classification model with soft targets. Then, we train the student model to capture the relevance margin between pairs of products for a given query using mean squared error loss. Instead of using the same training data as the teacher model, we significantly expand the student model dataset by generating unlabeled data and labeling it with the teacher model predictions. Experimental results show that the student model performance continues to improve as the size of the augmented training data increases. In fact, with enough augmented data, the student model can outperform the teacher model. The student model has been successfully deployed in production at Walmart.com with significantly positive metrics.

Phase transition in the knapsack problem

We examine the phase transition phenomenon for the Knapsack problem from both a computational and a human perspective. We first provide, via an empirical and a theoretical analysis, a characterization of the phenomenon in terms of two instance properties; normalised capacity and normalised profit. Then, we show evidence that average time spent by human decision makers in solving an instance peaks near the phase transition. Given the ubiquity of the Knapsack problem in every-day life, a better understanding of its structure can improve our understanding not only of computational techniques but also of human behavior, including the use and development of heuristics and occurrence of biases.

Supervisory Control for Behavior Composition

We relate behavior composition, a synthesis task studied in AI, to supervisory control theory from the discrete event systems field. In particular, we show that realizing (i.e., implementing) a target behavior module (e.g., a house surveillance system) by suitably coordinating a collection of available behaviors (e.g., automatic blinds, doors, lights, cameras, etc.) amounts to imposing a supervisor onto a special discrete event system. Such a link allows us to leverage on the solid foundations and extensive work on discrete event systems, including borrowing tools and ideas from that field. As evidence of that we show how simple it is to introduce preferences in the mapped framework.

Reasoning about Agent Programs using ATL-like Logics

We propose a variant of Alternating-time Temporal Logic (ATL) grounded in the agents' operational know-how, as defined by their libraries of abstract plans. Inspired by ATLES, a variant itself of ATL, it is possible in our logic to explicitly refer to "rational" strategies for agents developed under the Belief-Desire-Intention agent programming paradigm. This allows us to express and verify properties of BDI systems using ATL-type logical frameworks.

Qualitative Approximate Behavior Composition

The behavior composition problem involves automatically building a controller that is able to realize a desired, but unavailable, target system (e.g., a house surveillance) by suitably coordinating a set of available components (e.g., video cameras, blinds, lamps, a vacuum cleaner, phones, etc.) Previous work has almost exclusively aimed at bringing about the desired component in its totality, which is highly unsatisfactory for unsolvable problems. In this work, we develop an approach for approximate behavior composition without departing from the classical setting, thus making the problem applicable to a much wider range of cases. Based on the notion of simulation, we characterize what a maximal controller and the "closest" implementable target module (optimal approximation) are, and show how these can be computed using ATL model checking technology for a special case. We show the uniqueness of optimal approximations, and prove their soundness and completeness with respect to their imported controllers.