Yu

Aurko Roy, Rohan Anil, Guangda Lai et al. · deepmind

Transformer models have recently emerged as one of the foundational models in natural language processing, and as a byproduct, there is significant recent interest and investment in scaling these models. However, the training and inference costs of these large Transformer language models are prohibitive, thus necessitating more research in identifying more efficient variants. In this work, we propose a simple yet effective modification to the Transformer architecture inspired by the literature in statistical language modeling, by augmenting the model with n-grams that are constructed from a discrete latent representation of the text sequence. We evaluate our model, the N-Grammer on language modeling on the C4 data-set as well as text classification on the SuperGLUE data-set, and find that it outperforms several strong baselines such as the Transformer and the Primer. We open-source our model for reproducibility purposes in Jax.

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri et al. · allen-ai, berkeley

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

Zhiyong Sun, Changbin, Yu

In this paper we study coupled dynamical systems and investigate dimension properties of the subspace spanned by solutions of each individual system. Relevant problems on \textit{collinear dynamical systems} and their variations are discussed recently by Montenbruck et. al. in \cite{collinear2017SCL}, while in this paper we aim to provide a unified analysis to derive the dimensional-invariance principles for networked coupled systems, and to generalize the invariance principles for networked systems with more general forms of coupling terms. To be specific, we consider two types of coupled systems, one with scalar couplings and the other with matrix couplings. Via the \textit{rank-preserving flow theory}, we show that any scalar-coupled dynamical system (with constant, time-varying or state-dependent couplings) possesses the dimensional-invariance principles, in that the dimension of the subspace spanned by the individual systems' solutions remains invariant. For coupled dynamical systems with matrix coefficients/couplings, necessary and sufficient conditions (for constant, time-varying and state-dependent couplings) are given to characterize dimensional-invariance principles. The proofs via a rank-preserving matrix flow theory in this paper simplify the analysis in \cite{collinear2017SCL}, and we also extend the invariance principles to the cases of time-varying couplings and state-dependent couplings. Furthermore, subspace-preserving property and signature-preserving flows are also developed for coupled networked systems with particular coupling terms. These invariance principles provide insightful characterizations to analyze transient behaviors and solution evolutions for a large family of coupled systems, such as multi-agent consensus dynamics, distributed coordination systems, formation control systems, among others.

Zhaolin Gao, Yu, Wang et al.

Prompt optimization improves language models without updating their weights by searching for a better system prompt, but its effectiveness varies widely across tasks. We study what makes a task amenable to prompt optimization. We show that the reward variance across different system prompts can be decomposed into two components: variance among responses, which captures generation stochasticity, and variance among system prompts, which captures differences in system prompt quality. Prompt optimization succeeds when variance among system prompts is sufficiently large, but fails when variance among responses dominates the variance of the system prompts. Surprisingly, we further show that scaling to more user prompts can hurt optimization by reducing variance among system prompts, especially on heterogeneous datasets where different user prompts favor different system prompts. Motivated by this insight, we propose $p1$, a simple user prompt filtering method that selects a small subset of user prompts with high variance across candidate system prompts. This subset of user prompts allows one to distinguish a good system prompt from a bad one, making system optimization easier. Experiments on reasoning benchmarks show that $p1$ substantially improves prompt optimization over training on the full dataset and outperforms strong baselines such as GEPA. Notably, training on only two prompts from AIME 24 yields a system prompt that generalizes well to other reasoning benchmarks.

Shuqing Luo, Pingzhi Li, Jie Peng et al.

Mixture-of-experts (MoE) architectures could achieve impressive computational efficiency with expert parallelism, which relies heavily on all-to-all communication across devices. Unfortunately, such communication overhead typically constitutes a significant portion of the total runtime, hampering the scalability of distributed training and inference for modern MoE models (consuming over $40\%$ runtime in large-scale training). In this paper, we first define collaborative communication to illustrate this intrinsic limitation, and then propose system- and algorithm-level innovations to reduce communication costs. Specifically, given a pair of experts co-activated by one token, we call them "collaborated", which comprises $2$ cases as intra- and inter-collaboration, depending on whether they are kept on the same device. Our pilot investigations reveal that augmenting the proportion of intra-collaboration can accelerate expert parallelism at scale. It motivates us to strategically optimize collaborative communication for accelerated MoE training and inference, dubbed Occult. Our designs are capable of either delivering exact results with reduced communication cost or controllably minimizing the cost with collaboration pruning, materialized by modified fine-tuning. Comprehensive experiments on various MoE-LLMs demonstrate that Occult can be faster than popular state-of-the-art inference or training frameworks (more than $1.5\times$ speed up across multiple tasks and models) with comparable or superior quality compared to the standard fine-tuning. Code is available at $\href{https://github.com/UNITES-Lab/Occult}{https://github.com/UNITES-Lab/Occult}$.

Ronilo Ragodos, Tong Wang, Lu Feng et al.

Machine learning models have been increasingly used in business research. However, most state-of-the-art machine learning models, such as deep neural networks and XGBoost, are black boxes in nature. Therefore, post hoc explainers that provide explanations for machine learning models by, for example, estimating numerical importance of the input features, have been gaining wide usage. Despite the intended use of post hoc explainers being explaining machine learning models, we found a growing trend in business research where post hoc explanations are used to draw inferences about the data. In this work, we investigate the validity of such use. Specifically, we investigate with extensive experiments whether the explanations obtained by the two most popular post hoc explainers, SHAP and LIME, provide correct information about the true marginal effects of X on Y in the data, which we call data-alignment. We then identify what factors influence the alignment of explanations. Finally, we propose a set of mitigation strategies to improve the data-alignment of explanations and demonstrate their effectiveness with real-world data in an econometric context. In spite of this effort, we nevertheless conclude that it is often not appropriate to infer data insights from post hoc explanations. We articulate appropriate alternative uses, the most important of which is to facilitate the proposition and subsequent empirical investigation of hypotheses. The ultimate goal of this paper is to caution business researchers against translating post hoc explanations of machine learning models into potentially false insights and understanding of data.

Chung-En, Yu, Hsuan-Chih et al.

To develop trustworthy Vision-Language Models (VLMs), it is essential to address adversarial robustness and hallucination mitigation, both of which impact factual accuracy in high-stakes applications such as defense and healthcare. Existing methods primarily focus on either adversarial defense or hallucination post-hoc correction, leaving a gap in unified robustness strategies. We introduce \textbf{Hydra}, an adaptive agentic framework that enhances plug-in VLMs through iterative reasoning, structured critiques, and cross-model verification, improving both resilience to adversarial perturbations and intrinsic model errors. Hydra employs an Action-Critique Loop, where it retrieves and critiques visual information, leveraging Chain-of-Thought (CoT) and In-Context Learning (ICL) techniques to refine outputs dynamically. Unlike static post-hoc correction methods, Hydra adapts to both adversarial manipulations and intrinsic model errors, making it robust to malicious perturbations and hallucination-related inaccuracies. We evaluate Hydra on four VLMs, three hallucination benchmarks, two adversarial attack strategies, and two adversarial defense methods, assessing performance on both clean and adversarial inputs. Results show that Hydra surpasses plug-in VLMs and state-of-the-art (SOTA) dehallucination methods, even without explicit adversarial defenses, demonstrating enhanced robustness and factual consistency. By bridging adversarial resistance and hallucination mitigation, Hydra provides a scalable, training-free solution for improving the reliability of VLMs in real-world applications.

Shuqing Luo, Ye Han, Pingzhi Li et al.

Mixture-of-Experts (MoE) architecture offers enhanced efficiency for Large Language Models (LLMs) with modularized computation, yet its inherent sparsity poses significant hardware deployment challenges, including memory locality issues, communication overhead, and inefficient computing resource utilization. Inspired by the modular organization of the human brain, we propose Mozart, a novel algorithm-hardware co-design framework tailored for efficient training of MoE-based LLMs on 3.5D wafer-scale chiplet architectures. On the algorithm side, Mozart exploits the inherent modularity of chiplets and introduces: (1) an expert allocation strategy that enables efficient on-package all-to-all communication, and (2) a fine-grained scheduling mechanism that improves communication-computation overlap through streaming tokens and experts. On the architecture side, Mozart adaptively co-locates heterogeneous modules on specialized chiplets with a 2.5D NoP-Tree topology and hierarchical memory structure. Evaluation across three popular MoE models demonstrates significant efficiency gains, enabling more effective parallelization and resource utilization for large-scale modularized MoE-LLMs.

Chung-En, Yu, Brian Jalaian et al.

Developing trustworthy intelligent vision systems for high-stakes domains, \emph{e.g.}, remote sensing and medical diagnosis, demands broad robustness without costly retraining. We propose \textbf{Visual Reasoning Agent (VRA)}, a training-free, agentic reasoning framework that wraps off-the-shelf vision-language models \emph{and} pure vision systems in a \emph{Think--Critique--Act} loop. While VRA incurs significant additional test-time computation, it achieves up to 40\% absolute accuracy gains on challenging visual reasoning benchmarks. Future work will optimize query routing and early stopping to reduce inference overhead while preserving reliability in vision tasks.

Yu, Qian, Wilson S. Geisler et al.

Previous studies have compared the brain and deep neural networks trained on image classification. Intriguingly, while some suggest that their representations are highly similar, others argued the opposite. Here, we propose a new approach to characterize the similarity of the decision strategies of two observers (models or brains) using decision variable correlation (DVC). DVC quantifies the correlation between decoded decisions on individual samples in a classification task and thus can capture task-relevant information rather than general representational alignment. We evaluate this method using monkey V4/IT recordings and models trained on image classification tasks. We find that model--model similarity is comparable to monkey--monkey similarity, whereas model--monkey similarity is consistently lower and, surprisingly, decreases with increasing ImageNet-1k performance. While adversarial training enhances robustness, it does not improve model--monkey similarity in task-relevant dimensions; however, it markedly increases model--model similarity. Similarly, pre-training on larger datasets does not improve model--monkey similarity. These results suggest a fundamental divergence between the task-relevant representations in monkey V4/IT and those learned by models trained on image classification tasks.

Qi, Yu, Xinran He et al.

Traditional anomaly detection on social media mostly focuses on individual point anomalies while anomalous phenomena usually occur in groups. Therefore it is valuable to study the collective behavior of individuals and detect group anomalies. Existing group anomaly detection approaches rely on the assumption that the groups are known, which can hardly be true in real world social media applications. In this paper, we take a generative approach by proposing a hierarchical Bayes model: Group Latent Anomaly Detection (GLAD) model. GLAD takes both pair-wise and point-wise data as input, automatically infers the groups and detects group anomalies simultaneously. To account for the dynamic properties of the social media data, we further generalize GLAD to its dynamic extension d-GLAD. We conduct extensive experiments to evaluate our models on both synthetic and real world datasets. The empirical results demonstrate that our approach is effective and robust in discovering latent groups and detecting group anomalies.