Fusing speech into pre-trained language model (SpeechLM) usually suffers from inefficient encoding of long-form speech and catastrophic forgetting of pre-trained text modality. Leveraging speech-text alignments, our approach segments and compresses speech features to match the granularity of text embeddings. Additionally, we introduce a two-stage training pipeline that includes the distillation and fine-tuning phases to mitigate catastrophic forgetting. In this work, we focus on integrating speech into pre-trained language models (SpeechLMs). A straightforward approach is to transcribe speech into text and use these transcriptions as prompts for large language models (Huang et al., 2023); however, such cascaded systems suffer from error propagation, higher latency, and cannot leverage raw speech information like emotion, speaker identity, and other paralinguistic cues (Faruqui & Hakkani-Tür, 2021; Lin et al., 2022; Kim et al., 2024). Speech representations can be integrated into pre-trained language models mainly through two approaches. The first method involves using connector modules that align speech representations with the language model's input space without modifying the model's existing vocabulary. These connector-based techniques typically incorporate a compression module to shorten the speech features, enhancing efficiency.

Controlling AUVs can be challenging because of the effect of complex non-linear hydrodynamic forces acting on the robot, which, unlike ground robots, are significant in water and cannot be ignored. The problem is especially challenging for small AUVs for which the dynamics can change significantly with payload changes and deployments under different water conditions. The common approach to AUV control is a combination of passive stabilization with added buoyancy on top and weights on the bottom, and a PID controller tuned for simple and smooth motion primitives. However, the approach comes at the cost of sluggish controls and often the need to re-tune controllers with configuration changes. We propose a fast (trainable in minutes), reinforcement learning based approach for full 6 degree of freedom (DOF) control of an AUV, enabled by a new, highly parallelized simulator for underwater vehicle dynamics. We demonstrate that the proposed simulator models approximate hydrodynamic forces with enough accuracy that a zero-shot transfer of the learned policy to a real robot produces performance comparable to a hand-tuned PID controller. Furthermore, we show that domain randomization on the simulator produces policies that are robust to small variations in vehicle's physical parameters.

A novel Learning-by-Education Node Community framework (LENC) for Collaborative Knowledge Distillation (CKD) is presented, which facilitates continual collective learning through effective knowledge exchanges among diverse deployed Deep Neural Network (DNN) peer nodes. These DNNs dynamically and autonomously adopt either the role of a student, seeking knowledge, or that of a teacher, imparting knowledge, fostering a collaborative learning environment. The proposed framework enables efficient knowledge transfer among participating DNN nodes as needed, while enhancing their learning capabilities and promoting their collaboration. LENC addresses the challenges of handling diverse training data distributions and the limitations of individual DNN node learning abilities. It ensures the exploitation of the best available teacher knowledge upon learning a new task and protects the DNN nodes from catastrophic forgetting. Additionally, it innovates by enabling collaborative multitask knowledge distillation, while addressing the problem of task-agnostic continual learning, as DNN nodes have no information on task boundaries. Experimental evaluation on a proof-of-concept implementation demonstrates the LENC framework's functionalities and benefits across multiple DNN learning and inference scenarios. The conducted experiments showcase its ability to gradually maximize the average test accuracy of the community of interacting DNN nodes in image classification problems, by appropriately leveraging the collective knowledge of all node peers. The LENC framework achieves state-of-the-art performance in on-line unlabelled CKD.

We present MM1.5, a new family of multimodal large language models (MLLMs) designed to enhance capabilities in text-rich image understanding, visual referring and grounding, and multi-image reasoning. Building upon the MM1 architecture, MM1.5 adopts a data-centric approach to model training, systematically exploring the impact of diverse data mixtures across the entire model training lifecycle. This includes high-quality OCR data and synthetic captions for continual pre-training, as well as an optimized visual instruction-tuning data mixture for supervised fine-tuning. Our models range from 1B to 30B parameters, encompassing both dense and mixture-of-experts (MoE) variants, and demonstrate that careful data curation and training strategies can yield strong performance even at small scales (1B and 3B). Additionally, we introduce two specialized variants: MM1.5-Video, designed for video understanding, and MM1.5-UI, tailored for mobile UI understanding. Through extensive empirical studies and ablations, we provide detailed insights into the training processes and decisions that inform our final designs, offering valuable guidance for future research in MLLM development.

AI-based systems such as language models can replicate and amplify social biases reflected in their training data. Among other questionable behavior, this can lead to LM-generated text--and text suggestions--that contain normatively inappropriate stereotypical associations. In this paper, we consider the question of how "debiasing" a language model impacts stories that people write using that language model in a predictive text scenario. We find that (n=414), in certain scenarios, language model suggestions that align with common social stereotypes are more likely to be accepted by human authors. Conversely, although anti-stereotypical language model suggestions sometimes lead to an increased rate of anti-stereotypical stories, this influence is far from sufficient to lead to "fully debiased" stories.

Common narratives about automation often pit new technologies against workers. The introduction of advanced machine tools, industrial robots, and AI have all been met with concern that technological progress will mean fewer jobs. However, workers themselves offer a more optimistic, nuanced perspective. Drawing on a far-reaching 2024 survey of more than 9,000 workers across nine countries, this paper finds that more workers report potential benefits from new technologies like robots and AI for their safety and comfort at work, their pay, and their autonomy on the job than report potential costs. Workers with jobs that ask them to solve complex problems, workers who feel valued by their employers, and workers who are motivated to move up in their careers are all more likely to see new technologies as beneficial. In contrast to assumptions in previous research, more formal education is in some cases associated with more negative attitudes toward automation and its impact on work. In an experimental setting, the prospect of financial incentives for workers improve their perceptions of automation technologies, whereas the prospect of increased input about how new technologies are used does not have a significant effect on workers' attitudes toward automation.

Reinforcement learning from human feedback (RLHF) has become the leading approach for fine-tuning large language models (LLM). However, RLHF has limitations in multi-task learning (MTL) due to challenges of reward hacking and extreme multi-objective optimization (i.e., trade-off of multiple and/or sometimes conflicting objectives). Applying RLHF for MTL currently requires careful tuning of the weights for reward model and data combinations. This is often done via human intuition and does not generalize. In this work, we introduce a novel post-training paradigm which we called Constrained Generative Policy Optimization (CGPO). The core of CGPO is Mixture of Judges (MoJ) with cost-efficient constrained policy optimization with stratification, which can identify the perfect blend in RLHF in a principled manner. It shows strong empirical results with theoretical guarantees, does not require extensive hyper-parameter tuning, and is plug-and-play in common post-training pipelines. Together, this can detect and mitigate reward hacking behaviors while reaching a pareto-optimal point across an extremely large number of objectives. Our empirical evaluations demonstrate that CGPO significantly outperforms standard RLHF algorithms like PPO and DPO across various tasks including general chat, STEM questions, instruction following, and coding. Specifically, CGPO shows improvements of 7.4% in AlpacaEval-2 (general chat), 12.5% in Arena-Hard (STEM & reasoning), and consistent gains in other domains like math and coding. Notably, PPO, while commonly used, is prone to severe reward hacking in popular coding benchmarks, which CGPO successfully addresses. This breakthrough in RLHF not only tackles reward hacking and extreme multi-objective optimization challenges but also advances the state-of-the-art in aligning general-purpose LLMs for diverse applications.

A practical application of quantum machine learning in real-world scenarios in the short term remains elusive, despite significant theoretical efforts. Image classification, a common task for classical models, has been used to benchmark quantum algorithms with simple datasets, but only few studies have tackled complex real-data classification challenges. In this work, we address such a gap by focusing on the classification of satellite images, a task of particular interest to the earth observation (EO) industry. We first preprocess the selected intrincate dataset by reducing its dimensionality. Subsequently, we employ neural quantum kernels (NQKs)- embedding quantum kernels (EQKs) constructed from trained quantum neural networks (QNNs)- to classify images which include solar panels. We explore both $1$-to-$n$ and $n$-to-$n$ NQKs. In the former, parameters from a single-qubit QNN's training construct an $n$-qubit EQK achieving a mean test accuracy over 86% with three features. In the latter, we iteratively train an $n$-qubit QNN to ensure scalability, using the resultant architecture to directly form an $n$-qubit EQK. In this case, a test accuracy over 88% is obtained for three features and 8 qubits. Additionally, we show that the results are robust against a suboptimal training of the QNN.

Federated learning (FL) is an appealing paradigm that allows a group of machines (a.k.a. clients) to learn collectively while keeping their data local. However, due to the heterogeneity between the clients' data distributions, the model obtained through the use of FL algorithms may perform poorly on some client's data. Personalization addresses this issue by enabling each client to have a different model tailored to their own data while simultaneously benefiting from the other clients' data. We consider an FL setting where some clients can be adversarial, and we derive conditions under which full collaboration fails. Specifically, we analyze the generalization performance of an interpolated personalized FL framework in the presence of adversarial clients, and we precisely characterize situations when full collaboration performs strictly worse than fine-tuned personalization. Our analysis determines how much we should scale down the level of collaboration, according to data heterogeneity and the tolerable fraction of adversarial clients. We support our findings with empirical results on mean estimation and binary classification problems, considering synthetic and benchmark image classification datasets.

As LLMs become capable of complex tasks, there is growing potential for personalized interactions tailored to the subtle and idiosyncratic preferences of the user. We present a public benchmark, PersonalLLM, focusing on adapting LLMs to provide maximal benefits for a particular user. Departing from existing alignment benchmarks that implicitly assume uniform preferences, we curate open-ended prompts paired with many high-quality answers over which users would be expected to display heterogeneous latent preferences. Instead of persona-prompting LLMs based on high-level attributes (e.g., user's race or response length), which yields homogeneous preferences relative to humans, we develop a method that can simulate a large user base with diverse preferences from a set of pre-trained reward models. Our dataset and generated personalities offer an innovative testbed for developing personalization algorithms that grapple with continual data sparsity--few relevant feedback from the particular user--by leveraging historical data from other (similar) users. We explore basic in-context learning and meta-learning baselines to illustrate the utility of PersonalLLM and highlight the need for future methodological development. Our dataset is available at https://huggingface.co/datasets/namkoong-lab/PersonalLLM