The rise of foundation models has transformed machine learning research, prompting efforts to uncover their inner workings and develop more efficient and reliable applications for better control. While significant progress has been made in interpreting Large Language Models (LLMs), multimodal foundation models (MMFMs) - such as contrastive vision-language models, generative vision-language models, and text-to-image models - pose unique interpretability challenges beyond unimodal frameworks. Despite initial studies, a substantial gap remains between the interpretability of LLMs and MMFMs. This survey explores two key aspects: (1) the adaptation of LLM interpretability methods to multimodal models and (2) understanding the mechanistic differences between unimodal language models and crossmodal systems. By systematically reviewing current MMFM analysis techniques, we propose a structured taxonomy of interpretability methods, compare insights across unimodal and multimodal architectures, and highlight critical research gaps.

With the integration of image modality, the semantic space of multimodal large language models (MLLMs) is more complex than text-only models, making their interpretability more challenging and their alignment less stable, particularly susceptible to low-quality data, which can lead to inconsistencies between modalities, hallucinations, and biased outputs. As a result, developing interpretability methods for MLLMs is crucial for improving alignment quality and efficiency. In text-only LLMs, Sparse Autoencoders (SAEs) have gained attention for their ability to interpret latent representations. However, extending SAEs to multimodal settings presents new challenges due to modality fusion and the difficulty of isolating cross-modal representations. To address these challenges, we introduce SAE-V, a mechanistic interpretability framework that extends the SAE paradigm to MLLMs. By identifying and analyzing interpretable features along with their corresponding data, SAE-V enables fine-grained interpretation of both model behavior and data quality, facilitating a deeper understanding of cross-modal interactions and alignment dynamics. Moreover, by utilizing cross-modal feature weighting, SAE-V provides an intrinsic data filtering mechanism to enhance model alignment without requiring additional models. Specifically, when applied to the alignment process of MLLMs, SAE-V-based data filtering methods could achieve more than 110% performance with less than 50% data. Our results highlight SAE-V's ability to enhance interpretability and alignment in MLLMs, providing insights into their internal mechanisms.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable advancements in multimodal understanding; however, their fundamental visual cognitive abilities remain largely underexplored. To bridge this gap, we introduce VisFactor, a novel benchmark derived from the Factor-Referenced Cognitive Test (FRCT), a well-established psychometric assessment of human cognition. VisFactor digitalizes vision-related FRCT subtests to systematically evaluate MLLMs across essential visual cognitive tasks including spatial reasoning, perceptual speed, and pattern recognition. We present a comprehensive evaluation of state-of-the-art MLLMs, such as GPT-4o, Gemini-Pro, and Qwen-VL, using VisFactor under diverse prompting strategies like Chain-of-Thought and Multi-Agent Debate. Our findings reveal a concerning deficiency in current MLLMs' fundamental visual cognition, with performance frequently approaching random guessing and showing only marginal improvements even with advanced prompting techniques. These results underscore the critical need for focused research to enhance the core visual reasoning capabilities of MLLMs. To foster further investigation in this area, we release our VisFactor benchmark at https://github.com/CUHK-ARISE/VisFactor.

The next token prediction loss is the dominant self-supervised training objective for large language models and has achieved promising results in a variety of downstream tasks. However, upon closer investigation of this objective, we find that it lacks an understanding of sequence-level signals, leading to a mismatch between training and inference processes. To bridge this gap, we introduce a contrastive preference optimization (CPO) procedure that can inject sequence-level information into the language model at any training stage without expensive human labeled data. Our experiments show that the proposed objective surpasses the next token prediction in terms of win rate in the instruction-following and text generation tasks.

Traditional evaluations of multimodal large language models (LLMs) have been limited by their focus on single-image reasoning, failing to assess crucial aspects like contextual understanding, reasoning stability, and uncertainty calibration. This study addresses these limitations by introducing a novel benchmark that integrates multi-image reasoning tasks with rejection-based evaluation and positional bias detection. To evaluate these dimensions, we further introduce entropy as a novel metric for quantifying reasoning consistency across reordered answer variants. We applied this benchmark to assess Grok 3, ChatGPT-4o, ChatGPT-o1, Gemini 2.0 Flash Experimental, DeepSeek Janus models, Qwen2.5-VL-72B-Instruct, QVQ-72B-Preview, and Pixtral 12B across eight visual reasoning tasks, including difference spotting and diagram interpretation. Our findings reveal ChatGPT-o1 leading in overall accuracy (82.5\%) and rejection accuracy (70.0\%), closely followed by Gemini 2.0 Flash Experimental (70.8\%). QVQ-72B-Preview demonstrated superior rejection accuracy (85.5\%). Notably, Pixtral 12B (51.7\%) showed promise in specific domains, while Janus models exhibited challenges in bias and uncertainty calibration, reflected in low rejection accuracies and high entropy scores. High entropy scores in Janus models (Janus 7B: 0.8392, Janus 1B: 0.787) underscore their susceptibility to positional bias and unstable reasoning, contrasting with the low entropy and robust reasoning of ChatGPT models. The study further demonstrates that model size is not the sole determinant of performance, as evidenced by Grok 3 underperformance despite its substantial parameter count. By employing multi-image contexts, rejection mechanisms, and entropy-based consistency metrics, this benchmark sets a new standard for evaluating multimodal LLMs, enabling a more robust and reliable assessment of next-generation AI systems.

Large Language models (LLMs) have achieved encouraging results in tabular data generation. However, existing approaches require fine-tuning, which is computationally expensive. This paper explores an alternative: prompting a fixed LLM with in-context examples. We observe that using randomly selected in-context examples hampers the LLM's performance, resulting in sub-optimal generation quality. To address this, we propose a novel in-context learning framework: TabGen-ICL, to enhance the in-context learning ability of LLMs for tabular data generation. TabGen-ICL operates iteratively, retrieving a subset of real samples that represent the residual between currently generated samples and true data distributions. This approach serves two purposes: locally, it provides more effective in-context learning examples for the LLM in each iteration; globally, it progressively narrows the gap between generated and real data. Extensive experiments on five real-world tabular datasets demonstrate that TabGen-ICL significantly outperforms the random selection strategy. Specifically, it reduces the error rate by a margin of $3.5\%-42.2\%$ on fidelity metrics. We demonstrate for the first time that prompting a fixed LLM can yield high-quality synthetic tabular data. The code is provided in the \href{https://github.com/fangliancheng/TabGEN-ICL}{link}.

Existing navigation decision support systems often perform poorly when handling non-predefined navigation scenarios. Leveraging the generalization capabilities of large language model (LLM) in handling unknown scenarios, this research proposes a dual-core framework for LLM applications to address this issue. Firstly, through ReAct-based prompt engineering, a larger LLM core decomposes intricate navigation tasks into manageable sub-tasks, which autonomously invoke corresponding external tools to gather relevant information, using this feedback to mitigate the risk of LLM hallucinations. Subsequently, a fine-tuned and compact LLM core, acting like a first-mate is designed to process such information and unstructured external data, then to generates context-aware recommendations, ultimately delivering lookout insights and navigation hints that adhere to the International Regulations for Preventing Collisions at Sea (COLREGs) and other rules. Extensive experiments demonstrate the proposed framework not only excels in traditional ship collision avoidance tasks but also adapts effectively to unstructured, non-predefined, and unpredictable scenarios. A comparative analysis with DeepSeek-R1, GPT-4o and other SOTA models highlights the efficacy and rationality of the proposed framework. This research bridges the gap between conventional navigation systems and LLMs, offering a framework to enhance safety and operational efficiency across diverse navigation applications.

Providing students with detailed and timely grading feedback is essential for self-learning. While existing LLM-based grading systems are promising, most of them rely on one single model, which limits their performance. To address this, we propose Ensemble Tree-of-Thought (ToT), a framework that enhances LLM outputs by integrating multiple models. Using this framework, we develop a grading system. Ensemble ToT follows three steps: (1) analyzing LLM performance, (2) generating candidate answers, and (3) refining them into a final result. Based on this, our grading system first evaluates the grading tendencies of LLMs, then generates multiple results, and finally integrates them via a simulated debate. Experimental results demonstrate our approach's ability to provide accurate and explainable grading by effectively coordinating multiple LLMs.

Large Language Models (LLMs) have demonstrated potential for data science tasks via code generation. However, the exploratory nature of data science, alongside the stochastic and opaque outputs of LLMs, raise concerns about their reliability. While prior work focuses on benchmarking LLM accuracy, reproducibility remains underexplored, despite being critical to establishing trust in LLM-driven analysis. We propose a novel analyst-inspector framework to automatically evaluate and enforce the reproducibility of LLM-generated data science workflows - the first rigorous approach to the best of our knowledge. Defining reproducibility as the sufficiency and completeness of workflows for reproducing functionally equivalent code, this framework enforces computational reproducibility principles, ensuring transparent, well-documented LLM workflows while minimizing reliance on implicit model assumptions. Using this framework, we systematically evaluate five state-of-the-art LLMs on 1,032 data analysis tasks across three diverse benchmark datasets. We also introduce two novel reproducibility-enhancing prompting strategies. Our results show that higher reproducibility strongly correlates with improved accuracy and reproducibility-enhancing prompts are effective, demonstrating structured prompting's potential to enhance automated data science workflows and enable transparent, robust AI-driven analysis. Our code is publicly available.

Linear representation hypothesis posits that high-level concepts are encoded as linear directions in the representation spaces of LLMs. Park et al. (2024) formalize this notion by unifying multiple interpretations of linear representation, such as 1-dimensional subspace representation and interventions, using a causal inner product. However, their framework relies on single-token counterfactual pairs and cannot handle ambiguous contrasting pairs, limiting its applicability to complex or context-dependent concepts. We introduce a new notion of binary concepts as unit vectors in a canonical representation space, and utilize LLMs' (neural) activation differences along with maximum likelihood estimation (MLE) to compute concept directions (i.e., steering vectors). Our method, Sum of Activation-base Normalized Difference (SAND), formalizes the use of activation differences modeled as samples from a von Mises-Fisher (vMF) distribution, providing a principled approach to derive concept directions. We extend the applicability of Park et al. (2024) by eliminating the dependency on unembedding representations and single-token pairs. Through experiments with LLaMA models across diverse concepts and benchmarks, we demonstrate that our lightweight approach offers greater flexibility, superior performance in activation engineering tasks like monitoring and manipulation.