Concerns about the use of data and algorithms for decision-making have risen in concert with the ascendancy of data-driven methods. Prior research has investigated problems ranging from privacy violations to explainability and accountability [10] [12]. Here, we focus on fairness, which addresses discrimination and disparate treatment in data and algorithms [1]. Recently, fairness research has focused significantly on algorithmic fairness, which aims to investigate algorithms for unfairness and intervene on algorithms to prevent unfair outcomes. Prior work has developed metrics for fairness, methods for fair variable selection, and mechanisms to correct unfair models [14] [11]. Algorithmic fairness evolved out of concerns about the use of algorithms in high-stakes contexts, particularly those that have a history of disparity and discrimination. When training an algorithm on data from a discriminatory process, it is common for the algorithm to learn to replicate that discrimination, or'garbage in, garbage out'. Much work in algorithmic fairness focuses on detecting, evaluating, and correcting unfair algorithms under a garbage-in-garbage-out regime. However, little work focuses on assessing the extent to which the data-generating process itself is'garbage',

Replying to formal emails is time-consuming and cognitively demanding, as it requires crafting polite phrasing and providing an adequate response to the sender's demands. Although systems with Large Language Models (LLMs) were designed to simplify the email replying process, users still need to provide detailed prompts to obtain the expected output. Therefore, we proposed and evaluated an LLM-powered question-and-answer (QA)-based approach for users to reply to emails by answering a set of simple and short questions generated from the incoming email. We developed a prototype system, ResQ, and conducted controlled and field experiments with 12 and 8 participants. Our results demonstrated that the QA-based approach improves the efficiency of replying to emails and reduces workload while maintaining email quality, compared to a conventional prompt-based approach that requires users to craft appropriate prompts to obtain email drafts. We discuss how the QA-based approach influences the email reply process and interpersonal relationship dynamics, as well as the opportunities and challenges associated with using a QA-based approach in AI-mediated communication.

Large language models (LLMs) are known to struggle with consistently generating truthful responses. While various representation intervention techniques have been proposed, these methods typically apply a universal representation correction vector to all input queries, limiting their effectiveness against diverse queries in practice. In this study, we introduce TruthFlow, a novel method that leverages the Flow Matching technique for query-specific truthful representation correction. Specifically, TruthFlow first uses a flow model to learn query-specific correction vectors that transition representations from hallucinated to truthful states. Then, during inference, the trained flow model generates these correction vectors to enhance the truthfulness of LLM outputs. Experimental results demonstrate that TruthFlow significantly improves performance on open-ended generation tasks across various advanced LLMs evaluated on TruthfulQA. Moreover, the trained TruthFlow model exhibits strong transferability, performing effectively on other unseen hallucination benchmarks.

Diffusion policies trained via offline behavioral cloning have recently gained traction in robotic motion generation. While effective, these policies typically require a large number of trainable parameters. This model size affords powerful representations but also incurs high computational cost during training. Ideally, it would be beneficial to dynamically adjust the trainable portion as needed, balancing representational power with computational efficiency. For example, while overparameterization enables diffusion policies to capture complex robotic behaviors via offline behavioral cloning, the increased computational demand makes online interactive imitation learning impractical due to longer training time. To address this challenge, we present a framework, called DRIFT, that uses the Singular Value Decomposition to enable dynamic rank adjustment during diffusion policy training. We implement and demonstrate the benefits of this framework in DRIFT-DAgger, an imitation learning algorithm that can seamlessly slide between an offline bootstrapping phase and an online interactive phase. We perform extensive experiments to better understand the proposed framework, and demonstrate that DRIFT-DAgger achieves improved sample efficiency and faster training with minimal impact on model performance.

High-quality supervised fine-tuning (SFT) data are crucial for eliciting strong capabilities from pretrained large language models (LLMs). Typically, instructions are paired with multiple responses sampled from other LLMs, which are often out of the distribution of the target model to be fine-tuned. This, at scale, can lead to diminishing returns and even hurt the models' performance and robustness. We propose **GRAPE**, a novel SFT framework that accounts for the unique characteristics of the target model. For each instruction, it gathers responses from various LLMs and selects the one with the highest probability measured by the target model, indicating that it aligns most closely with the target model's pretrained distribution; it then proceeds with standard SFT training. We first evaluate GRAPE with a controlled experiment, where we sample various solutions for each question in UltraInteract from multiple models and fine-tune commonly used LMs like LLaMA3.1-8B, Mistral-7B, and Qwen2.5-7B on GRAPE-selected data. GRAPE significantly outperforms strong baselines, including distilling from the strongest model with an absolute gain of up to 13.8%, averaged across benchmarks, and training on 3x more data with a maximum performance improvement of 17.3%. GRAPE's strong performance generalizes to realistic settings. We experiment with the post-training data used for Tulu3 and Olmo-2. GRAPE outperforms strong baselines trained on 4.5 times more data by 6.1% and a state-of-the-art data selection approach by 3% on average performance. Remarkably, using 1/3 of the data and half the number of epochs, GRAPE enables LLaMA3.1-8B to surpass the performance of Tulu3-SFT by 3.5%.

Large language models (LLMs), such as o1 from OpenAI, have demonstrated remarkable reasoning capabilities. o1 generates a long chain-of-thought (LongCoT) before answering a question. LongCoT allows LLMs to analyze problems, devise plans, reflect, and backtrack effectively. These actions empower LLM to solve complex problems. After the release of o1, many teams have attempted to replicate its LongCoT and reasoning capabilities. In terms of methods, they primarily rely on knowledge distillation with data from existing models with LongCoT capacities (e.g., OpenAI-o1, Qwen-QwQ, DeepSeek-R1-Preview), leaving significant uncertainties on systematically developing such reasoning abilities. In terms of data domains, these works focus narrowly on math while a few others include coding, limiting their generalizability. This paper introduces a novel approach to enable LLM's LongCoT capacity without distillation from o1-like models or expensive human annotations, where we bootstrap LongCoT (BOLT) from a standard instruct model. BOLT involves three stages: 1) LongCoT data bootstrapping with in-context learning on a standard instruct model; 2) LongCoT supervised finetuning; 3) online training to further refine LongCoT capacities. In BOLT, only a few in-context examples need to be constructed during the bootstrapping stage; in our experiments, we created 10 examples, demonstrating the feasibility of this approach. We use Llama-3.1-70B-Instruct to bootstrap LongCoT and apply our method to various model scales (7B, 8B, 70B). We achieve impressive performance on a variety of benchmarks, Arena-Hard, MT-Bench, WildBench, ZebraLogic, MATH500, which evaluate diverse task-solving and reasoning capabilities.

Despite the remarkable capabilities of large language models across various tasks, their continued scaling faces a critical challenge: the scarcity of high-quality pretraining data. While model architectures continue to evolve, the natural language data struggles to scale up. To tackle this bottleneck, we propose \textbf{MA}ssive \textbf{G}enre-\textbf{A}udience~(MAGA) reformulation method, which systematic synthesizes diverse, contextually-rich pretraining data from existing corpus. This work makes three main contributions: (1) We propose MAGA reformulation method, a lightweight and scalable approach for pretraining corpus expansion, and build a 770B tokens MAGACorpus. (2) We evaluate MAGACorpus with different data budget scaling strategies, demonstrating consistent improvements across various model sizes (134M-13B), establishing the necessity for next-generation large-scale synthetic pretraining language models. (3) Through comprehensive analysis, we investigate prompt engineering's impact on synthetic training collapse and reveal limitations in conventional collapse detection metrics using validation losses. Our work shows that MAGA can substantially expand training datasets while maintaining quality, offering a reliably pathway for scaling models beyond data limitations.

The proliferation of NLP-powered language technologies, AI-based natural language generation models, and English as a mainstream means of communication among both native and non-native speakers make the output of AI-powered tools especially intriguing to linguists. This paper investigates how Grammarly and ChatGPT affect the English language regarding wordiness vs. conciseness. A case study focusing on the purpose subordinator in order to is presented to illustrate the way in which Grammarly and ChatGPT recommend shorter grammatical structures instead of longer and more elaborate ones. Although the analysed sentences were produced by native speakers, are perfectly correct, and were extracted from a language corpus of contemporary English, both Grammarly and ChatGPT suggest more conciseness and less verbosity, even for relatively short sentences. The present article argues that technologies such as Grammarly not only mirror language change but also have the potential to facilitate or accelerate it.

Load forecasting is essential for the efficient, reliable, and cost-effective management of power systems. Load forecasting performance can be improved by learning the similarities among multiple entities (e.g., regions, buildings). Techniques based on multi-task learning obtain predictions by leveraging consumption patterns from the historical load demand of multiple entities and their relationships. However, existing techniques cannot effectively assess inherent uncertainties in load demand or account for dynamic changes in consumption patterns. This paper proposes a multi-task learning technique for online and probabilistic load forecasting. This technique provides accurate probabilistic predictions for the loads of multiple entities by leveraging their dynamic similarities. The method's performance is evaluated using datasets that register the load demand of multiple entities and contain diverse and dynamic consumption patterns. The experimental results show that the proposed method can significantly enhance the effectiveness of current multi-task learning approaches across a wide variety of load consumption scenarios.

During the 33rd International Joint Conference on Artificial Intelligence (IJCAI), held in Jeju, I had the opportunity to meet with one of the keynote speakers, and winner of the 2024 IJCAI Computers and Thought Award, Professor Nisarg Shah. I asked him about his research, the role of theory in machine learning research, fairness and safety guarantees, regulation, conference reviews, and advice for those just starting out on their research journey. Could you start by telling us about yourself, your career, and your education? Nisarg Shah (NS): I grew up in India and went to IIT Bombay for my undergraduate. Ever since then, I knew that I wanted to go into higher education and academia. I actually did do an industrial placement after my undergrad, and I got a job offer that was very lucrative and would have been more lucrative than doing a PhD. However, that [money] is not why I wanted to do my PhD. I wanted to do my PhD because I was genuinely curious about different questions in this field, and I wanted to study more about them and have fun while doing it.