Most recommendation benchmarks evaluate how well a model imitates user behavior. In financial advisory, however, observed actions can be noisy or short-sighted under market volatility and may conflict with a user's long-term goals. Treating what users chose as the sole ground truth, therefore, conflates behavioral imitation with decision quality. We introduce Conv-FinRe, a conversational and longitudinal benchmark for stock recommendation that evaluates LLMs beyond behavior matching. Given an onboarding interview, step-wise market context, and advisory dialogues, models must generate rankings over a fixed investment horizon. Crucially, Conv-FinRe provides multi-view references that distinguish descriptive behavior from normative utility grounded in investor-specific risk preferences, enabling diagnosis of whether an LLM follows rational analysis, mimics user noise, or is driven by market momentum. We build the benchmark from real market data and human decision trajectories, instantiate controlled advisory conversations, and evaluate a suite of state-of-the-art LLMs. Results reveal a persistent tension between rational decision quality and behavioral alignment: models that perform well on utility-based ranking often fail to match user choices, whereas behaviorally aligned models can overfit short-term noise. The dataset is publicly released on Hugging Face. https://huggingface.co/collections/TheFinAI/conv-finre, and the codebase is available on GitHub. https://github.com/The-FinAI/Conv-FinRe.

Going beyond simple text processing, financial auditing requires detecting semantic, structural, and numerical inconsistencies across large-scale disclosures. As financial reports are filed in XBRL, a structured XML format governed by accounting standards, auditing becomes a structured information extraction and reasoning problem involving concept alignment, taxonomy-defined relations, and cross-document consistency. Although large language models (LLMs) show promise on isolated financial tasks, their capability in professional-grade auditing remains unclear. We introduce FinAuditing, a taxonomy-aligned, structure-aware benchmark built from real XBRL filings. It contains 1,102 annotated instances averaging over 33k tokens and defines three tasks: Financial Semantic Matching (FinSM), Financial Relationship Extraction (FinRE), and Financial Mathematical Reasoning (FinMR). Evaluations of 13 state-of-the-art LLMs reveal substantial gaps in concept retrieval, taxonomy-aware relation modeling, and consistent cross-document reasoning. These findings highlight the need for realistic, structure-aware benchmarks. We release the evaluation code1 and dataset2 publicly, and the task currently serves as the official benchmark of an ongoing public evaluation contest3.

Despite Spanish's pivotal role in the global finance industry, a pronounced gap exists in Spanish financial natural language processing (NLP) and application studies compared to English, especially in the era of large language models (LLMs).To bridge this gap, we unveil Toisón de Oro, the first bilingual framework that establishes instruction datasets, finetuned LLMs, and evaluation benchmark for financial LLMs in Spanish joint with English.We construct a rigorously curated bilingual instruction dataset including over 144K Spanish and English samples from 15 datasets covering 7 tasks.Harnessing this, we introduce FinMA-ES, an LLM designed for bilingual financial applications.We evaluate our model and existing LLMs using FLARE-ES, the first comprehensive bilingual evaluation benchmark with 21 datasets covering 9 tasks.The FLARE-ES benchmark results

As an integral part of people's daily lives, social media is becoming a rich source for automatic mental health analysis. As traditional discriminative methods bear poor generalization ability and low interpretability, the recent large language models (LLMs) have been explored for interpretable mental health analysis on social media, which aims to provide detailed explanations along with predictions in zero-shot or few-shot settings. The results show that LLMs still achieve unsatisfactory classification performance in a zero-shot/few-shot manner, which further significantly affects the quality of the generated explanations. Domain-specific finetuning is an effective solution, but faces two critical challenges: 1) lack of high-quality training data. 2) no open-source foundation LLMs. To alleviate these problems, we formally model interpretable mental health analysis as a text generation task, and build the first multi-task and multi-source interpretable mental health instruction (IMHI) dataset with 105K data samples to support LLM instruction tuning and evaluation. The raw social media data are collected from 10 existing sources covering 8 mental health analysis tasks. We prompt ChatGPT with expert-designed few-shot prompts to obtain explanations. To ensure the reliability of the explanations, we perform strict automatic and human evaluations on the correctness, consistency, and quality of generated data. Based on the IMHI dataset and LLaMA2 foundation models, we train MentaLLaMA, the first open-source instruction-following LLM series for interpretable mental health analysis on social media. We evaluate MentaLLaMA and other advanced methods on the IMHI benchmark, the first holistic evaluation benchmark for interpretable mental health analysis. The results show that MentaLLaMA approaches state-of-the-art discriminative methods in correctness and generates human-level explanations. MentaLLaMA models also show strong generalizability to unseen tasks. The project is available at https://github.com/SteveKGYang/MentaLLaMA.

Temporal reasoning is a crucial natural language processing (NLP) task, providing a nuanced understanding of time-sensitive contexts within textual data. Although recent advancements in Large Language Models (LLMs) have demonstrated their potential in temporal reasoning, the predominant focus has been on tasks such as temporal expression detection, normalization, and temporal relation extraction. These tasks are primarily designed for the extraction of direct and past temporal cues from given contexts and to engage in simple reasoning processes. A significant gap remains when considering complex reasoning tasks such as event forecasting, which requires multi-step temporal reasoning on events and prediction on the future timestamp. Another notable limitation of existing methods is their incapability to illustrate their reasoning process for explaining their prediction, hindering explainability. In this paper, we introduce the first task of explainable temporal reasoning, to predict an event's occurrence at a future timestamp based on context which requires multiple reasoning over multiple events, and subsequently provide a clear explanation for their prediction. Our task offers a comprehensive evaluation of both the LLMs' complex temporal reasoning ability, the future event prediction ability, and explainability-a critical attribute for AI applications. To support this task, we present the first instruction-tuning dataset of explainable temporal reasoning (ExpTime) with 26k derived from the temporal knowledge graph datasets, using a novel knowledge-graph-instructed-generation strategy. Based on the dataset, we propose the first open-source LLM series TimeLlaMA based on the foundation LLM LlaMA2, with the ability of instruction following for explainable temporal reasoning. We compare the performance of our method and a variety of LLMs, where our method achieves the state-of-the-art performance of temporal prediction and explanation generation. We also explore the impact of instruction tuning and different training sizes of instruction-tuning data, highlighting LLM's capabilities and limitations in complex temporal prediction and explanation generation.

Due to the overload of published scientific articles, citation recommendation has long been a critical research problem for automatically recommending the most relevant citations of given articles. Relational topic models (RTMs) have shown promise on citation prediction via joint modeling of document contents and citations. However, existing RTMs can only capture pairwise or direct (first-order) citation relationships among documents. The indirect (high-order) citation links have been explored in graph neural network–based methods, but these methods suffer from the well-known explainability problem. In this article, we propose a model called Graph Neural Collaborative Topic Model that takes advantage of both relational topic models and graph neural networks to capture high-order citation relationships and to have higher explainability due to the latent topic semantic structure. Experiments on three real-world citation datasets show that our model outperforms several competitive baseline methods on citation recommendation. In addition, we show that our approach can learn better topics than the existing approaches. The recommendation results can be well explained by the underlying topics.

Graph Neural Networks (GNNs) such as GCN can effectively learn document representations via the semantic relation graph among documents and words. However, despite a few exceptions, most of the previous work in this line of research does not consider the underlying topical semantics inherited in document contents and the relation graph, making the representations less effective and hard to interpret. In a few recent studies trying to incorporate latent topics into GNNs, the topics have been learned independently from the relation graph modeling. Intuitively, topic extraction can benefit much from the information propagation of the relation graph structure - directly and indirectly connected documents and words have similar topics. In this paper, we propose a novel Graph Topic Neural Network (GTNN) model to mine latent topic semantics for interpretable document representation learning, taking into account the document-document, document-word, and word-word relationships in the graph. We also show that our model can be viewed as semi-amortized inference for relational topic model based on Poisson distribution, with high order correlations. We test our model in several settings: unsupervised, semi-supervised, and supervised representation learning, for both connected and unconnected documents. In all the cases, our model outperforms the state-of-the-art models for these tasks.

Deep generative models (DGMs) have shown strong performance in semi-supervised learning (SSL), which incorporate discrete class information into the learning process. Yet existing methods generally overfit to the given labeled data, for only considering the conditional probability of labels. In this paper, we propose a novel discriminative regularized deep generative method for SSL, which fully exploits the discriminative and geometric information of data to address the aforementioned issue. Our method introduces the cluster and manifold assumption that maximizes the classification margin between clusters and simultaneously smooths the predictions of the data which is close in the sub-manifold of each cluster, to regularize the learning of the classifier in DGMs. To derive the regularization based on introduced assumptions, we adopt the generated data of DGMs along with labelled and unlabelled data, to model the data manifold and yield clusters based on the Gumbel-softmax distribution. Experimental results on both text and image datasets demonstrate the effectiveness and flexibility of our method, and prove that two introduced assumptions are complementary in guiding the classification boundary, thus improving the discriminative ability of the classifier.