Model fusing has always been an important topic, especially in an era where large language models (LLM) and multi-modal language models (MLM) with different architectures, parameter sizes and training pipelines, are being created all the time. In this work, we propose a post-hoc framework, aiming at fusing heterogeneous models off-the-shell, which we call \textit{likelihood composition}, and the basic idea is to compose multiple models' likelihood distribution when doing a multi-choice visual-question-answering task. Here the core concept, \textit{likelihood}, is actually the log-probability of the candidate answer. In \textit{likelihood composition}, we introduce some basic operations: \textit{debias}, \textit{highlight}, \textit{majority-vote} and \textit{ensemble}. By combining (composing) these basic elements, we get the mixed composition methods: \textit{mix-composition}. Through conducting comprehensive experiments on 9 VQA datasets and 10 MLMs, we prove the effectiveness of \textit{mix-composition} compared with simple \textit{ensemble} or \textit{majority-vote} methods. In this framework, people can propose new basic composition methods and combine them to get the new mixed composition methods. We hope our proposed \textit{likelihood composition} can provide a new perspective of fusing heterogeneous models and inspire the exploration under this framework.

This work provides an explanatory view of how LLMs can apply moral reasoning to both criticize and defend sexist language. We assessed eight large language models, all of which demonstrated the capability to provide explanations grounded in varying moral perspectives for both critiquing and endorsing views that reflect sexist assumptions. With both human and automatic evaluation, we show that all eight models produce comprehensible and contextually relevant text, which is helpful in understanding diverse views on how sexism is perceived. Also, through analysis of moral foundations cited by LLMs in their arguments, we uncover the diverse ideological perspectives in models' outputs, with some models aligning more with progressive or conservative views on gender roles and sexism. Based on our observations, we caution against the potential misuse of LLMs to justify sexist language. We also highlight that LLMs can serve as tools for understanding the roots of sexist beliefs and designing well-informed interventions. Given this dual capacity, it is crucial to monitor LLMs and design safety mechanisms for their use in applications that involve sensitive societal topics, such as sexism.

Learning from multimodal datasets can leverage complementary information and improve performance in prediction tasks. A commonly used strategy to account for feature correlations in high-dimensional datasets is the latent variable approach. Several latent variable methods have been proposed for multimodal datasets. However, these methods either focus on extracting the shared component across all modalities or on extracting both a shared component and individual components specific to each modality. To address this gap, we propose a Multi-Modal Fission Learning (MMFL) model that simultaneously identifies globally joint, partially joint, and individual components underlying the features of multimodal datasets. Unlike existing latent variable methods, MMFL uses supervision from the response variable to identify predictive latent components and has a natural extension for incorporating incomplete multimodal data. Through simulation studies, we demonstrate that MMFL outperforms various existing multimodal algorithms in both complete and incomplete modality settings. We applied MMFL to a real-world case study for early prediction of Alzheimers Disease using multimodal neuroimaging and genomics data from the Alzheimers Disease Neuroimaging Initiative (ADNI) dataset. MMFL provided more accurate predictions and better insights into within- and across-modality correlations compared to existing methods.

Code generation aims to automatically generate code from input requirements, significantly enhancing development efficiency. Recent large language models (LLMs) based approaches have shown promising results and revolutionized code generation task. Despite the promising performance, LLMs often generate contents with hallucinations, especially for the code generation scenario requiring the handling of complex contextual dependencies in practical development process. Although previous study has analyzed hallucinations in LLM-powered code generation, the study is limited to standalone function generation. In this paper, we conduct an empirical study to study the phenomena, mechanism, and mitigation of LLM hallucinations within more practical and complex development contexts in repository-level generation scenario. First, we manually examine the code generation results from six mainstream LLMs to establish a hallucination taxonomy of LLM-generated code. Next, we elaborate on the phenomenon of hallucinations, analyze their distribution across different models. We then analyze causes of hallucinations and identify four potential factors contributing to hallucinations. Finally, we propose an RAG-based mitigation method, which demonstrates consistent effectiveness in all studied LLMs. The replication package including code, data, and experimental results is available at https://github.com/DeepSoftwareAnalytics/LLMCodingHallucination

Human language, while aimed at conveying meaning, inherently carries ambiguity. It poses challenges for speech and language processing, but also serves crucial communicative functions. Efficiently solve ambiguity is both a desired and a necessary characteristic. The lexical meaning of a word in context can be determined automatically by Word Sense Disambiguation (WSD) algorithms that rely on external knowledge often limited and biased toward English. When adapting content to other languages, automated translations are frequently inaccurate and a high degree of expert human validation is necessary to ensure both accuracy and understanding. The current study addresses previous limitations by introducing a new resource for Spanish WSD. It includes a sense inventory and a lexical dataset sourced from the Diccionario de la Lengua Espa\~nola which is maintained by the Real Academia Espa\~nola. We also review current resources for Spanish and report metrics on them by a state-of-the-art system.

Large Vision-Language Models (LVLMs) have shown remarkable performance on many visual-language tasks. However, these models still suffer from multimodal hallucination, which means the generation of objects or content that violates the images. Many existing work detects hallucination by directly judging whether an object exists in an image, overlooking the association between the object and semantics. To address this issue, we propose Hierarchical Feedback Learning with Vision-enhanced Penalty Decoding (HELPD). This framework incorporates hallucination feedback at both object and sentence semantic levels. Remarkably, even with a marginal degree of training, this approach can alleviate over 15% of hallucination. Simultaneously, HELPD penalizes the output logits according to the image attention window to avoid being overly affected by generated text. HELPD can be seamlessly integrated with any LVLMs. Our experiments demonstrate that the proposed framework yields favorable results across multiple hallucination benchmarks. It effectively mitigates hallucination for different LVLMs and concurrently improves their text generation quality.

In an era where large language models (LLMs) are increasingly integrated into a wide range of everyday applications, research into these models' behavior has surged. However, due to the novelty of the field, clear methodological guidelines are lacking. This raises concerns about the replicability and generalizability of insights gained from research on LLM behavior. In this study, we discuss the potential risk of a replication crisis and support our concerns with a series of replication experiments focused on prompt engineering techniques purported to influence reasoning abilities in LLMs. We tested GPT-3.5, GPT-4o, Gemini 1.5 Pro, Claude 3 Opus, Llama 3-8B, and Llama 3-70B, on the chain-of-thought, EmotionPrompting, ExpertPrompting, Sandbagging, as well as Re-Reading prompt engineering techniques, using manually double-checked subsets of reasoning benchmarks including CommonsenseQA, CRT, NumGLUE, ScienceQA, and StrategyQA. Our findings reveal a general lack of statistically significant differences across nearly all techniques tested, highlighting, among others, several methodological weaknesses in previous research. We propose a forward-looking approach that includes developing robust methodologies for evaluating LLMs, establishing sound benchmarks, and designing rigorous experimental frameworks to ensure accurate and reliable assessments of model outputs.

Random Feature Model (RFM) with a nonlinear activation function is instrumental in understanding training and generalization performance in high-dimensional learning. While existing research has established an asymptotic equivalence in performance between the RFM and noisy linear models under isotropic data assumptions, empirical observations indicate that the RFM frequently surpasses linear models in practical applications. To address this gap, we ask, "When and how does the RFM outperform linear models?" In practice, inputs often have additional structures that significantly influence learning. Therefore, we explore the RFM under anisotropic input data characterized by spiked covariance in the proportional asymptotic limit, where dimensions diverge jointly while maintaining finite ratios. Our analysis reveals that a high correlation between inputs and labels is a critical factor enabling the RFM to outperform linear models. Moreover, we show that the RFM performs equivalent to noisy polynomial models, where the polynomial degree depends on the strength of the correlation between inputs and labels. Our numerical simulations validate these theoretical insights, confirming the performance-wise superiority of RFM in scenarios characterized by strong input-label correlation.

Theoretical and empirical evidence suggests that joint graph embedding algorithms induce correlation across the networks in the embedding space. In the Omnibus joint graph embedding framework, previous results explicitly delineated the dual effects of the algorithm-induced and model-inherent correlations on the correlation across the embedded networks. Accounting for and mitigating the algorithm-induced correlation is key to subsequent inference, as sub-optimal Omnibus matrix constructions have been demonstrated to lead to loss in inference fidelity. This work presents the first efforts to automate the Omnibus construction in order to address two key questions in this joint embedding framework: the correlation-to-OMNI problem and the flat correlation problem. In the flat correlation problem, we seek to understand the minimum algorithm-induced flat correlation (i.e., the same across all graph pairs) produced by a generalized Omnibus embedding. Working in a subspace of the fully general Omnibus matrices, we prove both a lower bound for this flat correlation and that the classical Omnibus construction induces the maximal flat correlation. In the correlation-to-OMNI problem, we present an algorithm -- named corr2Omni -- that, from a given matrix of estimated pairwise graph correlations, estimates the matrix of generalized Omnibus weights that induces optimal correlation in the embedding space. Moreover, in both simulated and real data settings, we demonstrate the increased effectiveness of our corr2Omni algorithm versus the classical Omnibus construction.

Brain-related research topics in artificial intelligence have recently gained popularity, particularly due to the expansion of what multimodal architectures can do from computer vision to natural language processing. Our main goal in this work is to explore the possibilities and limitations of these architectures in spoken text decoding from non-invasive fMRI recordings. Contrary to vision and textual data, fMRI data represent a complex modality due to the variety of brain scanners, which implies (i) the variety of the recorded signal formats, (ii) the low resolution and noise of the raw signals, and (iii) the scarcity of pretrained models that can be leveraged as foundation models for generative learning. These points make the problem of the non-invasive decoding of text from fMRI recordings very challenging. In this paper, we propose and end-to-end multimodal LLM for decoding spoken text from fMRI signals. The proposed architecture is founded on (i) an encoder derived from a specific transformer incorporating an augmented embedding layer for the encoder and a better-adjusted attention mechanism than that present in the state of the art, and (ii) a frozen large language model adapted to align the embedding of the input text and the encoded embedding of brain activity to decode the output text. A benchmark in performed on a corpus consisting of a set of interactions human-human and human-robot interactions where fMRI and conversational signals are recorded synchronously. The obtained results are very promising, as our proposal outperforms the evaluated models, and is able to generate text capturing more accurate semantics present in the ground truth. The implementation code is provided in https://github.com/Hmamouche/brain_decode.