Transfer learning typically leverages representations learned from a source domain to improve performance on a target task. A common approach is to extract features from a pre-trained model and directly apply them for target prediction. However, this strategy is prone to negative transfer where the source representation fails to align with the target distribution. In this article, we propose Residual Feature Integration (REFINE), a simple yet effective method designed to mitigate negative transfer. Our approach combines a fixed source-side representation with a trainable target-side encoder and fits a shallow neural network on the resulting joint representation, which adapts to the target domain while preserving transferable knowledge from the source domain. Theoretically, we prove that REFINE is sufficient to prevent negative transfer under mild conditions, and derive the generalization bound demonstrating its theoretical benefit. Empirically, we show that REFINE consistently enhances performance across diverse application and data modalities including vision, text, and tabular data, and outperforms numerous alternative solutions. Our method is lightweight, architecture-agnostic, and robust, making it a valuable addition to the existing transfer learning toolbox.
This work investigates the use of Wasserstein correlation -- a normalized measure of statistical dependence based on the Wasserstein distance between a joint distribution and the product of its marginals -- for unsupervised representation learning. Unlike, for example, contrastive methods, which naturally cluster classes in the latent space, we find that an (auto)encoder trained to maximize Wasserstein correlation between the input and encoded distributions instead acts as a compressor, reducing dimensionality while approximately preserving the topological and geometric properties of the input distribution. More strikingly, we show that Wasserstein correlation maximization can be used to arrive at an (auto)encoder -- either trained from scratch, or else one that extends a frozen, pretrained model -- that is approximately invariant to a chosen augmentation, or collection of augmentations, and that still approximately preserves the structural properties of the non-augmented input distribution. To do this, we first define the notion of an augmented encoder using the machinery of Markov-Wasserstein kernels. When the maximization objective is then applied to the augmented encoder, as opposed to the underlying, deterministic encoder, the resulting model exhibits the desired invariance properties. Finally, besides our experimental results, which show that even simple feedforward networks can be imbued with invariants or can, alternatively, be used to impart invariants to pretrained models under this training process, we additionally establish various theoretical results for optimal transport-based dependence measures. Code is available at https://github.com/keenan-eikenberry/wasserstein_correlation_maximization .
Editing presentation slides remains one of the most common and time-consuming tasks faced by millions of users daily, despite significant advances in automated slide generation. Existing approaches have successfully demonstrated slide editing via graphic user interface (GUI)-based agents, offering intuitive visual control. However, such methods often suffer from high computational cost and latency. In this paper, we propose Talk-to-Your-Slides, an LLM-powered agent designed to edit slides %in active PowerPoint sessions by leveraging structured information about slide objects rather than relying on image modality. The key insight of our work is designing the editing process with distinct high-level and low-level layers to facilitate interaction between user commands and slide objects. By providing direct access to application objects rather than screen pixels, our system enables 34.02% faster processing, 34.76% better instruction fidelity, and 87.42% cheaper operation than baselines. To evaluate slide editing capabilities, we introduce TSBench, a human-annotated dataset comprising 379 diverse editing instructions paired with corresponding slide variations in four categories. Our code, benchmark and demos are available at https://anonymous.4open.science/r/Talk-to-Your-Slides-0F4C.
The training of controllable text-to-video (T2V) models relies heavily on the alignment between videos and captions, yet little existing research connects video caption evaluation with T2V generation assessment. This paper introduces VidCapBench, a video caption evaluation scheme specifically designed for T2V generation, agnostic to any particular caption format. VidCapBench employs a data annotation pipeline, combining expert model labeling and human refinement, to associate each collected video with key information spanning video aesthetics, content, motion, and physical laws. VidCapBench then partitions these key information attributes into automatically assessable and manually assessable subsets, catering to both the rapid evaluation needs of agile development and the accuracy requirements of thorough validation. By evaluating numerous state-of-the-art captioning models, we demonstrate the superior stability and comprehensiveness of VidCapBench compared to existing video captioning evaluation approaches. Verification with off-the-shelf T2V models reveals a significant positive correlation between scores on VidCapBench and the T2V quality evaluation metrics, indicating that VidCapBench can provide valuable guidance for training T2V models. The project is available at https://github.com/VidCapBench/VidCapBench.
Graph Neural Networks (GNNs) have demonstrated their superiority in collaborative filtering, where the user-item (U-I) interaction bipartite graph serves as the fundamental data format. However, when graph-structured side information (e.g., multimodal similarity graphs or social networks) is integrated into the U-I bipartite graph, existing graph collaborative filtering methods fall short of achieving satisfactory performance. We quantitatively analyze this problem from a spectral perspective. Recall that a bipartite graph possesses a full spectrum within the range of [-1, 1], with the highest frequency exactly achievable at -1 and the lowest frequency at 1; however, we observe as more side information is incorporated, the highest frequency of the augmented adjacency matrix progressively shifts rightward. This spectrum shift phenomenon has caused previous approaches built for the full spectrum [-1, 1] to assign mismatched importance to different frequencies. To this end, we propose Spectrum Shift Correction (dubbed SSC), incorporating shifting and scaling factors to enable spectral GNNs to adapt to the shifted spectrum. Unlike previous paradigms of leveraging side information, which necessitate tailored designs for diverse data types, SSC directly connects traditional graph collaborative filtering with any graph-structured side information. Experiments on social and multimodal recommendation demonstrate the effectiveness of SSC, achieving relative improvements of up to 23% without incurring any additional computational overhead. Our code is available at https://github.com/yhhe2004/SSC-KDD.
When a retrieval system receives a query it has encountered before, previous relevance feedback, such as clicks or explicit judgments can help to improve retrieval results. However, the content of a previously relevant document may have changed, or the document might not be available anymore. Despite this evolved corpus, we counterfactually use these previously relevant documents as relevance signals. In this paper we proposed approaches to rewrite user queries and compare them against a system that directly uses the previous qrels for the ranking. We expand queries with terms extracted from the previously relevant documents or derive so-called keyqueries that rank the previously relevant documents to the top of the current corpus. Our evaluation in the CLEF LongEval scenario shows that rewriting queries with historical relevance feedback improves the retrieval effectiveness and even outperforms computationally expensive transformer-based approaches.
Extracting sentence embeddings from large language models (LLMs) is a practical direction, as it requires neither additional data nor fine-tuning. Previous studies usually focus on prompt engineering to guide LLMs to encode the core semantic information of the sentence into the embedding of the last token. However, the last token in these methods still encodes an excess of non-essential information, such as stop words, limiting its encoding capacity. To this end, we propose a Contrastive Prompting (CP) method that introduces an extra auxiliary prompt to elicit better sentence embedding. By contrasting with the auxiliary prompt, CP can steer existing prompts to encode the core semantics of the sentence, rather than non-essential information. CP is a plug-and-play inference-time intervention method that can be combined with various prompt-based methods. Extensive experiments on Semantic Textual Similarity (STS) tasks and downstream classification tasks demonstrate that our method can improve the performance of existing prompt-based methods across different LLMs. Our code will be released at https://github.com/zifengcheng/CP.
Mental-health stigma remains a pervasive social problem that hampers treatment-seeking and recovery. Existing resources for training neural models to finely classify such stigma are limited, relying primarily on social-media or synthetic data without theoretical underpinnings. To remedy this gap, we present an expert-annotated, theory-informed corpus of human-chatbot interviews, comprising 4,141 snippets from 684 participants with documented socio-cultural backgrounds. Our experiments benchmark state-of-the-art neural models and empirically unpack the challenges of stigma detection. This dataset can facilitate research on computationally detecting, neutralizing, and counteracting mental-health stigma. Our corpus is openly available at https://github.com/HanMeng2004/Mental-Health-Stigma-Interview-Corpus.
Recent advances in Generative Artificial Intelligence (GenAI) have transformed educational content creation, particularly in developing tutor training materials. However, biases embedded in AI-generated content--such as gender, racial, or national stereotypes--raise significant ethical and educational concerns. Despite the growing use of GenAI, systematic methods for detecting and evaluating such biases in educational materials remain limited. This study proposes an automated bias assessment approach that integrates the Contextualized Embedding Association Test with a prompt-engineered word extraction method within a Retrieval-Augmented Generation framework. We applied this method to AI-generated texts used in tutor training lessons. Results show a high alignment between the automated and manually curated word sets, with a Pearson correlation coefficient of r = 0.993, indicating reliable and consistent bias assessment. Our method reduces human subjectivity and enhances fairness, scalability, and reproducibility in auditing GenAI-produced educational content.
Knowledge editing, which aims to update the knowledge encoded in language models, can be deceptive. Despite the fact that many existing knowledge editing algorithms achieve near-perfect performance on conventional metrics, the models edited by them are still prone to generating original knowledge. This paper introduces the concept of "superficial editing" to describe this phenomenon. Our comprehensive evaluation reveals that this issue presents a significant challenge to existing algorithms. Through systematic investigation, we identify and validate two key factors contributing to this issue: (1) the residual stream at the last subject position in earlier layers and (2) specific attention modules in later layers. Notably, certain attention heads in later layers, along with specific left singular vectors in their output matrices, encapsulate the original knowledge and exhibit a causal relationship with superficial editing. Furthermore, we extend our analysis to the task of superficial unlearning, where we observe consistent patterns in the behavior of specific attention heads and their corresponding left singular vectors, thereby demonstrating the robustness and broader applicability of our methodology and conclusions. Our code is available here.
This paper presents the Duluth approach to the SemEval-2025 Task 7 on Multilingual and Crosslingual Fact-Checked Claim Retrieval. We implemented a TF-IDF-based retrieval system with experimentation on vector dimensions and tokenization strategies. Our best-performing configuration used word-level tokenization with a vocabulary size of 15,000 features, achieving an average success@10 score of 0.78 on the development set and 0.69 on the test set across ten languages. Our system showed stronger performance on higher-resource languages but still lagged significantly behind the top-ranked system, which achieved 0.96 average success@10. Our findings suggest that though advanced neural architectures are increasingly dominant in multilingual retrieval tasks, properly optimized traditional methods like TF-IDF remain competitive baselines, especially in limited compute resource scenarios.
Drafting patent claims is time-intensive, costly, and requires professional skill. Therefore, researchers have investigated large language models (LLMs) to assist inventors in writing claims. However, existing work has largely relied on datasets from the United States Patent and Trademark Office (USPTO). To enlarge research scope regarding various jurisdictions, drafting conventions, and legal standards, we introduce EPD, a European patent dataset. EPD presents rich textual data and structured metadata to support multiple patent-related tasks, including claim generation. This dataset enriches the field in three critical aspects: (1) Jurisdictional diversity: Patents from different offices vary in legal and drafting conventions. EPD fills a critical gap by providing a benchmark for European patents to enable more comprehensive evaluation. (2) Quality improvement: EPD offers high-quality granted patents with finalized and legally approved texts, whereas others consist of patent applications that are unexamined or provisional. Experiments show that LLMs fine-tuned on EPD significantly outperform those trained on previous datasets and even GPT-4o in claim quality and cross-domain generalization. (3) Real-world simulation: We propose a difficult subset of EPD to better reflect real-world challenges of claim generation. Results reveal that all tested LLMs perform substantially worse on these challenging samples, which highlights the need for future research.
Existing datasets available for crosslinguistic investigations have tended to focus on large amounts of data for a small group of languages or a small amount of data for a large number of languages. This means that claims based on these datasets are limited in what they reveal about universal properties of the human language faculty. While this has begun to change through the efforts of projects seeking to develop tagged corpora for a large number of languages, such efforts are still constrained by limits on resources. The current paper reports on a large automatically tagged parallel dataset which has been developed to partially address this issue. The taggedPBC contains more than 1,800 sentences of pos-tagged parallel text data from over 1,500 languages, representing 133 language families and 111 isolates, dwarfing previously available resources. The accuracy of tags in this dataset is shown to correlate well with both existing SOTA taggers for high-resource languages (SpaCy, Trankit) as well as hand-tagged corpora (Universal Dependencies Treebanks). Additionally, a novel measure derived from this dataset, the N1 ratio, correlates with expert determinations of word order in three typological databases (WALS, Grambank, Autotyp) such that a Gaussian Naive Bayes classifier trained on this feature can accurately identify basic word order for languages not in those databases. While much work is still needed to expand and develop this dataset, the taggedPBC is an important step to enable corpus-based crosslinguistic investigations, and is made available for research and collaboration via GitHub.
Predicting crash events is crucial for understanding crash distributions and their contributing factors, thereby enabling the design of proactive traffic safety policy interventions. However, existing methods struggle to interpret the complex interplay among various sources of traffic crash data, including numeric characteristics, textual reports, crash imagery, environmental conditions, and driver behavior records. As a result, they often fail to capture the rich semantic information and intricate interrelationships embedded in these diverse data sources, limiting their ability to identify critical crash risk factors. In this research, we propose TrafficSafe, a framework that adapts LLMs to reframe crash prediction and feature attribution as text-based reasoning. A multi-modal crash dataset including 58,903 real-world reports together with belonged infrastructure, environmental, driver, and vehicle information is collected and textualized into TrafficSafe Event Dataset. By customizing and fine-tuning LLMs on this dataset, the TrafficSafe LLM achieves a 42% average improvement in F1-score over baselines. To interpret these predictions and uncover contributing factors, we introduce TrafficSafe Attribution, a sentence-level feature attribution framework enabling conditional risk analysis. Findings show that alcohol-impaired driving is the leading factor in severe crashes, with aggressive and impairment-related behaviors having nearly twice the contribution for severe crashes compared to other driver behaviors. Furthermore, TrafficSafe Attribution highlights pivotal features during model training, guiding strategic crash data collection for iterative performance improvements. The proposed TrafficSafe offers a transformative leap in traffic safety research, providing a blueprint for translating advanced AI technologies into responsible, actionable, and life-saving outcomes.
Analysing the generalisation capabilities of relation extraction (RE) models is crucial for assessing whether they learn robust relational patterns or rely on spurious correlations. Our cross-dataset experiments find that RE models struggle with unseen data, even within similar domains. Notably, higher intra-dataset performance does not indicate better transferability, instead often signaling overfitting to dataset-specific artefacts. Our results also show that data quality, rather than lexical similarity, is key to robust transfer, and the choice of optimal adaptation strategy depends on the quality of data available: while fine-tuning yields the best cross-dataset performance with high-quality data, few-shot in-context learning (ICL) is more effective with noisier data. However, even in these cases, zero-shot baselines occasionally outperform all cross-dataset results. Structural issues in RE benchmarks, such as single-relation per sample constraints and non-standardised negative class definitions, further hinder model transferability.
The increasing context length of modern language models has created a need for evaluating their ability to retrieve and process information across extensive documents. While existing benchmarks test long-context capabilities, they often lack a structured way to systematically vary question complexity. We introduce KG-QAGen (Knowledge-Graph-based Question-Answer Generation), a framework that (1) extracts QA pairs at multiple complexity levels (2) by leveraging structured representations of financial agreements (3) along three key dimensions -- multi-hop retrieval, set operations, and answer plurality -- enabling fine-grained assessment of model performance across controlled difficulty levels. Using this framework, we construct a dataset of 20,139 QA pairs (the largest number among the long-context benchmarks) and open-source a part of it. We evaluate 13 proprietary and open-source LLMs and observe that even the best-performing models are struggling with set-based comparisons and multi-hop logical inference. Our analysis reveals systematic failure modes tied to semantic misinterpretation and inability to handle implicit relations.
In this work, we investigate the performance of LLMs on a new task that requires combining discussion with background knowledge for summarization. This aims to address the limitation of outside observer confusion in existing dialogue summarization systems due to their reliance solely on discussion information. To achieve this, we model the task output as background and opinion summaries and define two standardized summarization patterns. To support assessment, we introduce the first benchmark comprising high-quality samples consistently annotated by human experts and propose a novel hierarchical evaluation framework with fine-grained, interpretable metrics. We evaluate 12 LLMs under structured-prompt and self-reflection paradigms. Our findings reveal: (1) LLMs struggle with background summary retrieval, generation, and opinion summary integration. (2) Even top LLMs achieve less than 69% average performance across both patterns. (3) Current LLMs lack adequate self-evaluation and self-correction capabilities for this task.
We propose SLOT (Sample-specific Language Model Optimization at Test-time), a novel and parameter-efficient test-time inference approach that enhances a language model's ability to more accurately respond to individual prompts. Existing Large Language Models (LLMs) often struggle with complex instructions, leading to poor performances on those not well represented among general samples. To address this, SLOT conducts few optimization steps at test-time to update a light-weight sample-specific parameter vector. It is added to the final hidden layer before the output head, and enables efficient adaptation by caching the last layer features during per-sample optimization. By minimizing the cross-entropy loss on the input prompt only, SLOT helps the model better aligned with and follow each given instruction. In experiments, we demonstrate that our method outperforms the compared models across multiple benchmarks and LLMs. For example, Qwen2.5-7B with SLOT achieves an accuracy gain of 8.6% on GSM8K from 57.54% to 66.19%, while DeepSeek-R1-Distill-Llama-70B with SLOT achieves a SOTA accuracy of 68.69% on GPQA among 70B-level models. Our code is available at https://github.com/maple-research-lab/SLOT.
We present Team asdfo123's submission to the LLMSR@XLLM25 shared task, which evaluates large language models on producing fine-grained, controllable, and interpretable reasoning processes. Systems must extract all problem conditions, decompose a chain of thought into statement-evidence pairs, and verify the logical validity of each pair. Leveraging only the off-the-shelf Meta-Llama-3-8B-Instruct, we craft a concise few-shot, multi-turn prompt that first enumerates all conditions and then guides the model to label, cite, and adjudicate every reasoning step. A lightweight post-processor based on regular expressions normalises spans and enforces the official JSON schema. Without fine-tuning, external retrieval, or ensembling, our method ranks 5th overall, achieving macro F1 scores on par with substantially more complex and resource-consuming pipelines. We conclude by analysing the strengths and limitations of our approach and outlining directions for future research in structural reasoning with LLMs. Our code is available at https://github.com/asdfo123/LLMSR-asdfo123.
Large Language Models (LLMs) exhibit remarkable capabilities across various tasks, yet guiding them to follow desired behaviours during inference remains a significant challenge. Activation steering offers a promising method to control the generation process of LLMs by modifying their internal activations. However, existing methods commonly intervene in the model's behaviour using steering vectors generated by the model itself, which constrains their effectiveness to that specific model and excludes the possibility of leveraging powerful external expert models for steering. To address these limitations, we propose ExpertSteer, a novel approach that leverages arbitrary specialized expert models to generate steering vectors, enabling intervention in any LLMs. ExpertSteer transfers the knowledge from an expert model to a target LLM through a cohesive four-step process: first aligning representation dimensions with auto-encoders to enable cross-model transfer, then identifying intervention layer pairs based on mutual information analysis, next generating steering vectors from the expert model using Recursive Feature Machines, and finally applying these vectors on the identified layers during inference to selectively guide the target LLM without updating model parameters. We conduct comprehensive experiments using three LLMs on 15 popular benchmarks across four distinct domains. Experiments demonstrate that ExpertSteer significantly outperforms established baselines across diverse tasks at minimal cost.