Distributed optimization has become the default training paradigm in modern machine learning due to the growing scale of models and datasets. To mitigate communication overhead, local updates are often applied before global aggregation, resulting in a nested optimization approach with inner and outer steps. However, heavy-tailed stochastic gradient noise remains a significant challenge, particularly in attention-based models, hindering effective training. In this work, we propose TailOPT, an efficient framework designed to address heavy-tailed noise by leveraging adaptive optimization or clipping techniques. We establish convergence guarantees for the TailOPT framework under heavy-tailed noise with potentially unbounded gradient variance and local updates. Among its variants, we highlight a memory and communication efficient instantiation which we call $Bi^2Clip$, which performs coordinate-wise clipping at both the inner and outer optimizers, achieving adaptive-like performance (e.g., Adam) without the cost of maintaining or transmitting additional gradient statistics. Empirically, TailOPT, including $Bi^2Clip$, demonstrates superior performance on several language tasks and models, outperforming state-of-the-art methods.

High-quality, large-scale instructions are crucial for aligning large language models (LLMs), however, there is a severe shortage of instruction in the field of natural language understanding (NLU). Previous works on constructing NLU instructions mainly focus on information extraction (IE), neglecting tasks such as machine reading comprehension, question answering, and text classification. Furthermore, the lack of diversity in the data has led to a decreased generalization ability of trained LLMs in other NLU tasks and a noticeable decline in the fundamental model's general capabilities. To address this issue, we propose Hum, a large-scale, high-quality synthetic instruction corpus for NLU tasks, designed to enhance the NLU capabilities of LLMs. Specifically, Hum includes IE (either close IE or open IE), machine reading comprehension, text classification, and instruction generalist tasks, thereby enriching task diversity. Additionally, we introduce a human-LLMs collaborative mechanism to synthesize instructions, which enriches instruction diversity by incorporating guidelines, preference rules, and format variants. We conduct extensive experiments on 5 NLU tasks and 28 general capability evaluation datasets for LLMs. Experimental results show that Hum enhances the NLU capabilities of six LLMs by an average of 3.1\%, with no significant decline observed in other general capabilities.

Pre-trained multi-modal Vision-Language Models like CLIP are widely used offthe-shelf for a variety of applications. In this paper, we show that the common practice of individually exploiting the text or image encoders of these powerful multi-modal models is highly suboptimal for intra-modal tasks like imageto-image retrieval. We argue that this is inherently due to the CLIP-style intermodal contrastive loss that does not enforce any intra-modal constraints, leading to what we call intra-modal misalignment. To demonstrate this, we leverage two optimization-based modality inversion techniques that map representations from their input modality to the complementary one without any need for auxiliary data or additional trained adapters. We empirically show that, in the intra-modal tasks of image-to-image and text-to-text retrieval, approaching these tasks inter-modally significantly improves performance with respect to intramodal baselines on more than fifteen datasets. Additionally, we demonstrate that approaching a native inter-modal task (e.g. Finally, we show that incorporating an intra-modal term in the pre-training objective or narrowing the modality gap between the text and image feature embedding spaces helps reduce the intra-modal misalignment. In recent years the availability of massive, pre-trained Vision-Language Models (VLMs) has enabled a wide variety of applications ranging from zero-shot image segmentation (Zhou et al., 2022a; Lüddecke & Ecker, 2022) to visual question answering (Song et al., 2022; Parelli et al., 2023). These models are typically composed of independent image and text encoders which are simultaneously trained on massive corpora of image-text pairs to align the text and image embeddings of associated inputs. For example, the Contrastive Language-Image Pre-training (CLIP) model is trained on a corpus of 400M image-text pairs to map inputs from both modalities into a shared embedding space (Radford et al., 2021). CLIP is trained with an inter-modal contrastive loss that aims to maximize the similarity of corresponding image-text samples while minimizing the similarity with all the other examples within a batch. Despite CLIP's shared embedding space, visual and textual features lie in distinct regions. This phenomenon, known as the modality gap (Liang et al., 2022), originates from model initialization, and the inter-modal contrastive loss preserves and worsens it during training. Moreover, we note that CLIP's contrastive training strategy focuses on inter-modal (i.e. Consequently, the intra-image and intra-text similarities between CLIP representations might not faithfully correspond to those of the actual images or texts, as depicted in the left section of Figure 1. We refer to this issue as intra-modal misalignment. A simple experiment aimed at quantifying this problem is presented in Appendix B. These authors contributed equally to this work.

We propose a novel method to evaluate the theoretical limits of Transformers, allowing us to prove the first lower bounds against one-layer softmax Transformers with infinite precision. We establish those bounds for three tasks that require advanced reasoning. The first task, Match3 (Sanford et al., 2023), requires looking at all triples of positions. The second and third tasks address compositionality-based reasoning: one is composition of functions (Peng et al., 2024) and the other is composition of binary relations. We formally prove the inability of one-layer softmax Transformers to solve any of these tasks. In an attempt to overcome these limitations, we introduce Strassen attention and prove that with this mechanism a one-layer Transformer can in principle solve all these tasks. We also show that it enjoys sub-cubic running-time complexity, making it more scalable than similar previously proposed mechanisms, such as higher-order attention (Sanford et al., 2023). To complement our theoretical findings, we experimentally studied Strassen attention and compared it against standard (Vaswani et al, 2017), higher-order attention (Sanford et al., 2023) and triangular attention (Bergen et al. 2021). Our results help to disentangle all these attention mechanisms, highlighting their strengths and limitations. In particular, Strassen attention outperforms standard attention significantly on all the tasks. Altogether, understanding the theoretical limitations can guide research towards scalable attention mechanisms that improve the reasoning abilities of Transformers.

Large-scale numerical simulations often produce high-dimensional gridded data that is challenging to process for downstream applications. A prime example is numerical weather prediction, where atmospheric processes are modeled using discrete gridded representations of the physical variables and dynamics. Uncertainties are assessed by running the simulations multiple times, yielding ensembles of simulated fields as a high-dimensional stochastic representation of the forecast distribution. The high-dimensionality and large volume of ensemble datasets poses major computing challenges for subsequent forecasting stages. Data-driven dimensionality reduction techniques could help to reduce the data volume before further processing by learning meaningful and compact representations. However, existing dimensionality reduction methods are typically designed for deterministic and single-valued inputs, and thus cannot handle ensemble data from multiple randomized simulations. In this study, we propose novel dimensionality reduction approaches specifically tailored to the format of ensemble forecast fields. We present two alternative frameworks, which yield low-dimensional representations of ensemble forecasts while respecting their probabilistic character. The first approach derives a distribution-based representation of an input ensemble by applying standard dimensionality reduction techniques in a member-by-member fashion and merging the member representations into a joint parametric distribution model. The second approach achieves a similar representation by encoding all members jointly using a tailored variational autoencoder. We evaluate and compare both approaches in a case study using 10 years of temperature and wind speed forecasts over Europe. The approaches preserve key spatial and statistical characteristics of the ensemble and enable probabilistic reconstructions of the forecast fields.

With the integration of Multimodal large language models (MLLMs) into robotic systems and various AI applications, embedding emotional intelligence (EI) capabilities into these models is essential for enabling robots to effectively address human emotional needs and interact seamlessly in real-world scenarios. Existing static, text-based, or text-image benchmarks overlook the multimodal complexities of real-world interactions and fail to capture the dynamic, multimodal nature of emotional expressions, making them inadequate for evaluating MLLMs' EI. Based on established psychological theories of EI, we build EmoBench-M, a novel benchmark designed to evaluate the EI capability of MLLMs across 13 valuation scenarios from three key dimensions: foundational emotion recognition, conversational emotion understanding, and socially complex emotion analysis. Evaluations of both open-source and closed-source MLLMs on EmoBench-M reveal a significant performance gap between them and humans, highlighting the need to further advance their EI capabilities. All benchmark resources, including code and datasets, are publicly available at https://emo-gml.github.io/.

The generalization of language models (LMs) is undergoing active debates, contrasting their potential for general intelligence with their struggles with basic knowledge composition (e.g., reverse/transition curse). This paper uncovers the phenomenon of linear correlations in LMs during knowledge composition. For explanation, there exists a linear transformation between certain related knowledge that maps the next token prediction logits from one prompt to another, e.g., "X lives in the city of" $\rightarrow$ "X lives in the country of" for every given X. This mirrors the linearity in human knowledge composition, such as Paris $\rightarrow$ France. Our findings indicate that the linear transformation is resilient to large-scale fine-tuning, generalizing updated knowledge when aligned with real-world relationships, but causing hallucinations when it deviates. Empirical results suggest that linear correlation can serve as a potential identifier of LM's generalization. Finally, we show such linear correlations can be learned with a single feedforward network and pre-trained vocabulary representations, indicating LM generalization heavily relies on the latter.

Multimodal Federated Learning (MFL) enables multiple clients to collaboratively train models on multimodal data while ensuring clients' privacy. However, modality and task heterogeneity hinder clients from learning a unified representation, weakening local model generalization, especially in MFL with mixed modalities where only some clients have multimodal data. In this work, we propose an Adaptive prototype-based Multimodal Federated Learning (AproMFL) framework for mixed modalities and heterogeneous tasks to address the aforementioned issues. Our AproMFL transfers knowledge through adaptively-constructed prototypes without a prior public dataset. Clients adaptively select prototype construction methods in line with tasks; server converts client prototypes into unified multimodal prototypes and aggregates them to form global prototypes, avoid clients keeping unified labels. We divide the model into various modules and only aggregate mapping modules to reduce communication and computation overhead. To address aggregation issues in heterogeneity, we develop a client relationship graph-based scheme to dynamically adjust aggregation weights. Extensive experiments on representative datasets evidence effectiveness of AproMFL.

We investigate algorithmic decision problems where agents can respond strategically to the decision maker's (DM) models. The demand for clear and actionable explanations from DMs to (potentially strategic) agents continues to rise. While prior work often treats explanations as full model disclosures, explanations in practice might convey only partial information, which can lead to misinterpretations and harmful responses. When full disclosure of the predictive model is neither feasible nor desirable, a key open question is how DMs can use explanations to maximise their utility without compromising agent welfare. In this work, we explore well-known local and global explanation methods, and establish a necessary condition to prevent explanations from misleading agents into self-harming actions. Moreover, with conditional homogeneity, we establish that action recommendation (AR)-based explanations are sufficient for non-harmful responses, akin to the revelation principle in information design. To operationalise AR-based explanations, we propose a simple algorithm to jointly optimise the predictive model and AR policy to balance DM outcomes with agent welfare. Our empirical results demonstrate the benefits of this approach as a more refined strategy for safe and effective partial model disclosure in algorithmic decision-making.

The goal of this paper is threefold: i) to present some practical aspects of using full-text version of Corpus of Historical American English (COHA), the largest diachronic multi-genre corpus of the English language, in the investigation of a linguistic trend of change; ii) to test a widely held assumption that sentence length in written English has been steadily decreasing over the past few centuries; iii) to point to a possible link between the changes in sentence length and changes in the English syntactic usage. The empirical proof of concept for iii) is provided by the decline in the frequency of the non-finite purpose subordinator in order to. Sentence length, genre and the likelihood of occurrence of in order to are shown to be interrelated.