This study investigates the learnability of Recurrent Neural Networks (RNNs) in classifying structured formal languages, focusing on counter and Dyck languages. Traditionally, both first-order (LSTM) and second-order (O2RNN) RNNs have been considered effective for such tasks, primarily based on their theoretical expressiveness within the Chomsky hierarchy. However, our research challenges this notion by demonstrating that RNNs primarily operate as state machines, where their linguistic capabilities are heavily influenced by the precision of their embeddings and the strategies used for sampling negative examples. Our experiments revealed that performance declines significantly as the structural similarity between positive and negative examples increases. Remarkably, even a basic single-layer classifier using RNN embeddings performed better than chance. To evaluate generalization, we trained models on strings up to a length of 40 and tested them on strings from lengths 41 to 500, using 10 unique seeds to ensure statistical robustness. Stability comparisons between LSTM and O2RNN models showed that O2RNNs generally offer greater stability across various scenarios. We further explore the impact of different initialization strategies revealing that our hypothesis is consistent with various RNNs. Overall, this research questions established beliefs about RNNs' computational capabilities, highlighting the importance of data structure and sampling techniques in assessing neural networks' potential for language classification tasks. It emphasizes that stronger constraints on expressivity are crucial for understanding true learnability, as mere expressivity does not capture the essence of learning.

Large language models (LLMs) exhibit varying strengths and weaknesses across different tasks, prompting recent studies to explore the benefits of ensembling models to leverage their complementary advantages. However, existing LLM ensembling methods often overlook model compatibility and struggle with inefficient alignment of probabilities across the entire vocabulary. In this study, we empirically investigate the factors influencing ensemble performance, identifying model performance, vocabulary size, and response style as key determinants, revealing that compatibility among models is essential for effective ensembling. This analysis leads to the development of a simple yet effective model selection strategy that identifies compatible models. TE), a novel approach that efficiently combines models by focusing on the union of the top-k tokens from each model, thereby avoiding the need for full vocabulary alignment and reducing computational overhead. TE significantly enhances performance compared to existing methods, offering a more efficient framework for LLM ensembling. Large language models (LLMs) have demonstrated remarkable performance across a wide range of tasks and have shown promising results in real-world applications (OpenAI, 2023; Yang et al., 2024; Dubey et al., 2024). Given the diversity in data sources, model architectures, and training methods, LLMs exhibit varying strengths and weaknesses depending on the task at hand. Consequently, rather than relying solely on training an LLM from scratch, an alternative approach is to create an ensemble of LLMs. This method allows for leveraging the complementary advantages of different LLMs (Jiang et al., 2023b; Lu et al., 2024; Yu et al., 2024b). Existing model ensembling methods can be broadly categorized into three types: output-level, probability-level, and training-level approaches.

Drawing on the experiences of facilitators who have overseen 43 games over a four-year period, we illuminate recurring patterns, strategies, and decision-making processes observed during gameplay. Our analysis reveals key strategic considerations about AI development trajectories in this simulated environment, including: the destabilising effects of AI races, the crucial role of international cooperation in mitigating catastrophic risks, the challenges of aligning corporate and national interests, and the potential for rapid, transformative change in AI capabilities. We highlight places where we believe the game has been effective in exposing participants to the complexities and uncertainties inherent in AI governance. Key recurring gameplay themes include the emergence of international agreements, challenges to the robustness of such agreements, the critical role of cybersecurity in AI development, and the potential for unexpected crises to dramatically alter AI trajectories. By documenting these insights, we aim to provide valuable foresight for policymakers, industry leaders, and researchers navigating the complex landscape of AI development and governance.

Language models (LMs) hallucinate. We inquire: Can we detect and mitigate hallucinations before they happen? This work answers this research question in the positive, by showing that the internal representations of LMs provide rich signals that can be used for this purpose. We introduce FactCheckMate, which preemptively detects hallucinations by learning a classifier that predicts whether the LM will hallucinate, based on the model's hidden states produced over the inputs, before decoding begins. If a hallucination is detected, FactCheckMate then intervenes, by adjusting the LM's hidden states such that the model will produce more factual outputs. FactCheckMate provides fresh insights that the inner workings of LMs can be revealed by their hidden states. Practically, both the detection and mitigation models in FactCheckMate are lightweight, adding little inference overhead; FactCheckMate proves a more efficient approach for mitigating hallucinations compared to many post-hoc alternatives. We evaluate FactCheckMate over LMs of different scales and model families (including Llama, Mistral, and Gemma), across a variety of QA datasets from different domains. Our results demonstrate the effectiveness of leveraging internal representations for early hallucination detection and mitigation, achieving over 70% preemptive detection accuracy. On average, outputs generated by LMs with intervention are 34.4% more factual compared to those without intervention. The average overhead difference in the inference time introduced by FactCheckMate is around 3.16 seconds.

Introductory programming courses often emphasize mastering syntax and basic constructs before progressing to more complex and interesting programs. This bottom-up approach can be frustrating for novices, shifting the focus away from problem solving and potentially making computing less appealing to a broad range of students. The rise of generative AI for code production could partially address these issues by fostering new skills via interaction with AI models, including constructing high-level prompts and evaluating code that is automatically generated. In this experience report, we explore the inclusion of two prompt-focused activities in an introductory course, implemented across four labs in a six-week module. The first requires students to solve computational problems by writing natural language prompts, emphasizing problem-solving over syntax. The second involves students crafting prompts to generate code equivalent to provided fragments, to foster an understanding of the relationship between prompts and code. Most of the students in the course had reported finding programming difficult to learn, often citing frustrations with syntax and debugging. We found that self-reported difficulty with learning programming had a strong inverse relationship with performance on traditional programming assessments such as tests and projects, as expected. However, performance on the natural language tasks was less strongly related to self-reported difficulty, suggesting they may target different skills. Learning how to communicate with AI coding models is becoming an important skill, and natural language prompting tasks may appeal to a broad range of students.

Recent advancements in multimodal large language models (MLLMs) have aimed to integrate and interpret data across diverse modalities. However, the capacity of these models to concurrently process and reason about multiple modalities remains inadequately explored, partly due to the lack of comprehensive modality-wise benchmarks. We introduce OmniBench, a novel benchmark designed to rigorously evaluate models' ability to recognize, interpret, and reason across visual, acoustic, and textual inputs simultaneously. We define models capable of such tri-modal processing as omni-language models (OLMs). OmniBench is distinguished by high-quality human annotations, ensuring that accurate responses require integrated understanding and reasoning across all three modalities. Our main findings reveal that: i) most OLMs exhibit critical limitations in instruction-following and reasoning capabilities within tri-modal contexts; and ii) most baselines models perform poorly (below 50\% accuracy) even when provided with alternative textual representations of images or/and audio. These results suggest that the ability to construct a consistent context from text, image, and audio is often overlooked in existing MLLM training paradigms. To address this gap, we curate an instruction tuning dataset of 84.5K training samples, OmniInstruct, for training OLMs to adapt to multimodal contexts. We advocate for future research to focus on developing more robust tri-modal integration techniques and training strategies to enhance OLM performance across diverse modalities. The codes and live leaderboard could be found at https://m-a-p.ai/OmniBench.

Reinforcement learning from human feedback (RLHF) has demonstrated effectiveness in aligning large language models (LLMs) with human preferences. However, token-level RLHF suffers from the credit assignment problem over long sequences, where delayed rewards make it challenging for the model to discern which actions contributed to successful outcomes. This hinders learning efficiency and slows convergence. In this paper, we propose MA-RLHF, a simple yet effective RLHF framework that incorporates macro actions -- sequences of tokens or higher-level language constructs -- into the learning process. By operating at this higher level of abstraction, our approach reduces the temporal distance between actions and rewards, facilitating faster and more accurate credit assignment. This results in more stable policy gradient estimates and enhances learning efficiency within each episode, all without increasing computational complexity during training or inference. We validate our approach through extensive experiments across various model sizes and tasks, including text summarization, dialogue generation, question answering, and program synthesis. Our method achieves substantial performance improvements over standard RLHF, with performance gains of up to 30% in text summarization and code generation, 18% in dialogue, and 8% in question answering tasks. Notably, our approach reaches parity with vanilla RLHF 1.7x to 2x faster in terms of training time and continues to outperform it with further training. We will make our code and data publicly available at https://github.com/ernie-research/MA-RLHF .

Alzheimer's disease (AD) is a progressive neurodegenerative disorder with increasing prevalence among the aging population, necessitating early and accurate diagnosis for effective disease management. In this study, we present a novel hybrid deep learning framework that integrates both 2D Convolutional Neural Networks (2D-CNN) and 3D Convolutional Neural Networks (3D-CNN), along with a custom loss function and volumetric data augmentation, to enhance feature extraction and improve classification performance in AD diagnosis. According to extensive experiments, AlzhiNet outperforms standalone 2D and 3D models, highlighting the importance of combining these complementary representations of data. The depth and quality of 3D volumes derived from the augmented 2D slices also significantly influence the model's performance. The results indicate that carefully selecting weighting factors in hybrid predictions is imperative for achieving optimal results. Our framework has been validated on the Magnetic Resonance Imaging (MRI) from Kaggle and MIRIAD datasets, obtaining accuracies of 98.9% and 99.99%, respectively, with an AUC of 100%. Furthermore, AlzhiNet was studied under a variety of perturbation scenarios on the Alzheimer's Kaggle dataset, including Gaussian noise, brightness, contrast, salt and pepper noise, color jitter, and occlusion. The results obtained show that AlzhiNet is more robust to perturbations than ResNet-18, making it an excellent choice for real-world applications. This approach represents a promising advancement in the early diagnosis and treatment planning for Alzheimer's disease.

Voice assistants, such as Siri and Google Assistant, typically model audio and text separately, resulting in lost speech information and increased complexity. Recent efforts to address this with end-to-end Speech Large Language Models (LLMs) trained with supervised finetuning (SFT) have led to models "forgetting" capabilities from text-only LLMs. Our work proposes an alternative paradigm for training Speech LLMs without instruction data, using the response of a text-only LLM to transcripts as self-supervision. Importantly, this process can be performed without annotated responses. We show that our Distilled Voice Assistant (DiVA) generalizes to Spoken Question Answering, Classification, and Translation. Furthermore, we show that DiVA better meets user preferences, achieving a 72% win rate compared with state-of-the-art models like Qwen 2 Audio, despite using >100x less training compute. Figure 1: Training Pipeline for Distilled Voice Assistant (DiVA), Red indicates trainable components while Blue indicates frozen pretrained modules. DiVA modifies a text-only LLM into a general purpose Speech LLM by using the model's own responses to transcribed speech as self-supervision. As Large Language Models (LLMs) capabilities increase, so does the value of bringing these capabilities to new modalities, including audio and speech (Shu et al., 2023; Wang et al., 2023; Gong et al., 2023). Speech is a natural interaction surface for language technology (Murad et al., 2019), offering measurable efficiency gains for users (Ruan et al., 2018). One straightforward method of integrating speech with LLMs is to feed audio to an Automatic Speech Recognition (ASR) model and produce a text transcription for the LLM to use. All authors besides first and last sorted alphabetically.

To make large language models (LLMs) more helpful across diverse cultures, it is essential to have effective cultural knowledge benchmarks to measure and track our progress. Effective benchmarks need to be robust, diverse, and challenging. We introduce CulturalBench: a set of 1,227 human-written and human-verified questions for effectively assessing LLMs' cultural knowledge, covering 45 global regions including the underrepresented ones like Bangladesh, Zimbabwe, and Peru. Questions - each verified by five independent annotators - span 17 diverse topics ranging from food preferences to greeting etiquettes. We evaluate models on two setups: CulturalBench-Easy and CulturalBench-Hard which share the same questions but asked differently. We find that LLMs are sensitive to such difference in setups (e.g., GPT-4o with 27.3% difference). Compared to human performance (92.6% accuracy), CulturalBench-Hard is more challenging for frontier LLMs with the best performing model (GPT-4o) at only 61.5% and the worst (Llama3-8b) at 21.4%. Moreover, we find that LLMs often struggle with tricky questions that have multiple correct answers (e.g., What utensils do the Chinese usually use?), revealing a tendency to converge to a single answer. Our results also indicate that OpenAI GPT-4o substantially outperform other proprietary and open source models in questions related to all but one region (Oceania). Nonetheless, all models consistently underperform on questions related to South America and the Middle East.