In the emerging field of computational gastronomy, aligning culinary practices with scientifically supported nutritional goals is increasingly important. This study explores how large language models (LLMs) can be applied to optimize ingredient substitutions in recipes, specifically to enhance the phytochemical content of meals. Phytochemicals are bioactive compounds found in plants, which, based on preclinical studies, may offer potential health benefits. We fine-tuned models, including OpenAI's GPT-3.5, DaVinci, and Meta's TinyLlama, using an ingredient substitution dataset. These models were used to predict substitutions that enhance phytochemical content and create a corresponding enriched recipe dataset. Our approach improved Hit@1 accuracy on ingredient substitution tasks, from the baseline 34.53 plus-minus 0.10% to 38.03 plus-minus 0.28% on the original GISMo dataset, and from 40.24 plus-minus 0.36% to 54.46 plus-minus 0.29% on a refined version of the same dataset. These substitutions led to the creation of 1,951 phytochemically enriched ingredient pairings and 1,639 unique recipes. While this approach demonstrates potential in optimizing ingredient substitutions, caution must be taken when drawing conclusions about health benefits, as the claims are based on preclinical evidence. Future work should include clinical validation and broader datasets to further evaluate the nutritional impact of these substitutions. This research represents a step forward in using AI to promote healthier eating practices, providing potential pathways for integrating computational methods with nutritional science.

Grapheme-to-phoneme (G2P) conversion is critical in speech processing, particularly for applications like speech synthesis. G2P systems must possess linguistic understanding and contextual awareness of languages with polyphone words and context-dependent phonemes. Large language models (LLMs) have recently demonstrated significant potential in various language tasks, suggesting that their phonetic knowledge could be leveraged for G2P. In this paper, we evaluate the performance of LLMs in G2P conversion and introduce prompting and post-processing methods that enhance LLM outputs without additional training or labeled data. We also present a benchmarking dataset designed to assess G2P performance on sentence-level phonetic challenges of the Persian language. Our results show that by applying the proposed methods, LLMs can outperform traditional G2P tools, even in an underrepresented language like Persian, highlighting the potential of developing LLM-aided G2P systems.

Recently pre-trained Foundation Models (FMs) have been combined with Federated Learning (FL) to improve training of downstream tasks while preserving privacy. However, deploying FMs over edge networks with resource-constrained Internet of Things (IoT) devices is under-explored. This paper proposes a novel framework, namely, Federated Distilling knowledge to Prompt (FedD2P), for leveraging the robust representation abilities of a vision-language FM without deploying it locally on edge devices. This framework distills the aggregated knowledge of IoT devices to a prompt generator to efficiently adapt the frozen FM for downstream tasks. To eliminate the dependency on a public dataset, our framework leverages perclass local knowledge from IoT devices and linguistic descriptions of classes to train the prompt generator. Our experiments on diverse image classification datasets CIFAR, OxfordPets, SVHN, EuroSAT, and DTD show that FedD2P outperforms the baselines in terms of model performance.

Training and deploying reinforcement learning (RL) policies for robots, especially in accomplishing specific tasks, presents substantial challenges. Recent advancements have explored diverse reward function designs, training techniques, simulation-to-reality (sim-to-real) transfers, and performance analysis methodologies, yet these still require significant human intervention. This paper introduces an end-to-end framework for training and deploying RL policies, guided by Large Language Models (LLMs), and evaluates its effectiveness on bipedal robots. The framework consists of three interconnected modules: an LLM-guided reward function design module, an RL training module leveraging prior work, and a sim-to-real homomorphic evaluation module. This design significantly reduces the need for human input by utilizing only essential simulation and deployment platforms, with the option to incorporate human-engineered strategies and historical data. We detail the construction of these modules, their advantages over traditional approaches, and demonstrate the framework's capability to autonomously develop and refine controlling strategies for bipedal robot locomotion, showcasing its potential to operate independently of human intervention.

We examine the pre-training dynamics of language models, focusing on their ability to copy text from preceding context--a fundamental skill for various LLM applications, including in-context learning (ICL) and retrieval-augmented generation (RAG). We propose a novel perspective that Transformer-based language models develop copying abilities similarly to grokking, which refers to sudden generalization on test set long after the model fit to the training set. Our experiments yield three arguments: (1) The pre-training loss decreases rapidly, while the context copying ability of models initially lags and then abruptly saturates. (2) The speed of developing copying ability is independent of the number of tokens trained, similarly to how grokking speed is unaffected by dataset size as long as the data distribution is preserved. (3) Induction heads, the attention heads responsible for copying, form from shallow to deep layers during training, mirroring the development of circuits in deeper layers during grokking. We contend that the connection between grokking and context copying can provide valuable insights for more effective language model training, ultimately improving in-context performance. For example, we demonstrated that techniques that enhance grokking, such as regularization, either accelerate or enhance the development of context copying.

Large language models (LLMs) show remarkable potential to act as computer agents, enhancing human productivity and software accessibility in multi-modal tasks that require planning and reasoning. However, measuring agent performance in realistic environments remains a challenge since: (i) most benchmarks are limited to specific modalities or domains (e.g. text-only, web navigation, Q&A, coding) and (ii) full benchmark evaluations are slow (on order of magnitude of days) given the multi-step sequential nature of tasks. To address these challenges, we introduce the Windows Agent Arena: a reproducible, general environment focusing exclusively on the Windows operating system (OS) where agents can operate freely within a real Windows OS and use the same wide range of applications, tools, and web browsers available to human users when solving tasks. We adapt the OSWorld framework (Xie et al., 2024) to create 150+ diverse Windows tasks across representative domains that require agent abilities in planning, screen understanding, and tool usage. Our benchmark is scalable and can be seamlessly parallelized in Azure for a full benchmark evaluation in as little as 20 minutes. To demonstrate Windows Agent Arena's capabilities, we also introduce a new multi-modal agent, Navi. Our agent achieves a success rate of 19.5% in the Windows domain, compared to 74.5% performance of an unassisted human. Navi also demonstrates strong performance on another popular web-based benchmark, Mind2Web. We offer extensive quantitative and qualitative analysis of Navi's performance, and provide insights into the opportunities for future research in agent development and data generation using Windows Agent Arena. Webpage: https://microsoft.github.io/WindowsAgentArena Code: https://github.com/microsoft/WindowsAgentArena

Post-training large language models (LLMs) to develop reasoning capabilities has proven effective across diverse domains, such as mathematical reasoning and code generation. However, existing methods primarily focus on improving task-specific reasoning but have not adequately addressed the model's generalization capabilities across a broader range of reasoning tasks. To tackle this challenge, we introduce Critical Planning Step Learning (CPL), which leverages Monte Carlo Tree Search (MCTS) to explore diverse planning steps in multi-step reasoning tasks. Based on long-term outcomes, CPL learns step-level planning preferences to improve the model's planning capabilities and, consequently, its general reasoning capabilities. Furthermore, while effective in many scenarios for aligning LLMs, existing preference learning approaches like Direct Preference Optimization (DPO) struggle with complex multi-step reasoning tasks due to their inability to capture fine-grained supervision at each step. We propose Step-level Advantage Preference Optimization (Step-APO), which integrates an advantage estimate for step-level preference pairs obtained via MCTS into the DPO. This enables the model to more effectively learn critical intermediate planning steps, thereby further improving its generalization in reasoning tasks. Experimental results demonstrate that our method, trained exclusively on GSM8K and MATH, not only significantly improves performance on GSM8K (+10.5%) and MATH (+6.5%), but also enhances out-of-domain reasoning benchmarks, such as ARC-C (+4.0%), BBH (+1.8%), MMLU-STEM (+2.2%), and MMLU (+0.9%).

Machine learning models make mistakes, yet sometimes it is difficult to identify the systematic problems behind the mistakes. Practitioners engage in various activities, including error analysis, testing, auditing, and red-teaming, to form hypotheses of what can go (or has gone) wrong with their models. To validate these hypotheses, practitioners employ data slicing to identify relevant examples. However, traditional data slicing is limited by available features and programmatic slicing functions. In this work, we propose SemSlicer, a framework that supports semantic data slicing, which identifies a semantically coherent slice, without the need for existing features. SemSlicer uses Large Language Models to annotate datasets and generate slices from any user-defined slicing criteria. We show that SemSlicer generates accurate slices with low cost, allows flexible trade-offs between different design dimensions, reliably identifies under-performing data slices, and helps practitioners identify useful data slices that reflect systematic problems.

Ensuring content compliance with community guidelines is crucial for maintaining healthy online social environments. However, traditional human-based compliance checking struggles with scaling due to the increasing volume of user-generated content and a limited number of moderators. Recent advancements in Natural Language Understanding demonstrated by Large Language Models unlock new opportunities for automated content compliance verification. This work evaluates six AI-agents built on Open-LLMs for automated rule compliance checking in Decentralized Social Networks, a challenging environment due to heterogeneous community scopes and rules. Analyzing over 50,000 posts from hundreds of Mastodon servers, we find that AI-agents effectively detect non-compliant content, grasp linguistic subtleties, and adapt to diverse community contexts. Most agents also show high inter-rater reliability and consistency in score justification and suggestions for compliance. Human-based evaluation with domain experts confirmed the agents' reliability and usefulness, rendering them promising tools for semi-automated or human-in-the-loop content moderation systems.

With the increasing demand for practical applications of Large Language Models (LLMs), many attention-efficient models have been developed to balance performance and computational cost. However, the adversarial robustness of these models remains under-explored. In this work, we design a framework to investigate the trade-off between efficiency, performance, and adversarial robustness of LLMs and conduct extensive experiments on three prominent models with varying levels of complexity and efficiency -- Transformer++, Gated Linear Attention (GLA) Transformer, and MatMul-Free LM -- utilizing the GLUE and AdvGLUE datasets. The AdvGLUE dataset extends the GLUE dataset with adversarial samples designed to challenge model robustness. Our results show that while the GLA Transformer and MatMul-Free LM achieve slightly lower accuracy on GLUE tasks, they demonstrate higher efficiency and either superior or comparative robustness on AdvGLUE tasks compared to Transformer++ across different attack levels. These findings highlight the potential of simplified architectures to achieve a compelling balance between efficiency, performance, and adversarial robustness, offering valuable insights for applications where resource constraints and resilience to adversarial attacks are critical.