Duanyu Feng

Jimin Huang, Mengxi Xiao, Dong Li et al.

Financial LLMs hold promise for advancing financial tasks and domain-specific applications. However, they are limited by scarce corpora, weak multimodal capabilities, and narrow evaluations, making them less suited for real-world application. To address this, we introduce \textit{Open-FinLLMs}, the first open-source multimodal financial LLMs designed to handle diverse tasks across text, tabular, time-series, and chart data, excelling in zero-shot, few-shot, and fine-tuning settings. The suite includes FinLLaMA, pre-trained on a comprehensive 52-billion-token corpus; FinLLaMA-Instruct, fine-tuned with 573K financial instructions; and FinLLaVA, enhanced with 1.43M multimodal tuning pairs for strong cross-modal reasoning. We comprehensively evaluate Open-FinLLMs across 14 financial tasks, 30 datasets, and 4 multimodal tasks in zero-shot, few-shot, and supervised fine-tuning settings, introducing two new multimodal evaluation datasets. Our results show that Open-FinLLMs outperforms afvanced financial and general LLMs such as GPT-4, across financial NLP, decision-making, and multi-modal tasks, highlighting their potential to tackle real-world challenges. To foster innovation and collaboration across academia and industry, we release all codes (https://anonymous.4open.science/r/PIXIU2-0D70/B1D7/LICENSE) and models under OSI-approved licenses.

Duanyu Feng, Yongfu Dai, Jimin Huang et al.

In the financial industry, credit scoring is a fundamental element, shaping access to credit and determining the terms of loans for individuals and businesses alike. Traditional credit scoring methods, however, often grapple with challenges such as narrow knowledge scope and isolated evaluation of credit tasks. Our work posits that Large Language Models (LLMs) have great potential for credit scoring tasks, with strong generalization ability across multiple tasks. To systematically explore LLMs for credit scoring, we propose the first open-source comprehensive framework. We curate a novel benchmark covering 9 datasets with 14K samples, tailored for credit assessment and a critical examination of potential biases within LLMs, and the novel instruction tuning data with over 45k samples. We then propose the first Credit and Risk Assessment Large Language Model (CALM) by instruction tuning, tailored to the nuanced demands of various financial risk assessment tasks. We evaluate CALM, existing state-of-art (SOTA) methods, open source and closed source LLMs on the build benchmark. Our empirical results illuminate the capability of LLMs to not only match but surpass conventional models, pointing towards a future where credit scoring can be more inclusive, comprehensive, and unbiased. We contribute to the industry's transformation by sharing our pioneering instruction-tuning datasets, credit and risk assessment LLM, and benchmarks with the research community and the financial industry.

Yongfu Dai, Duanyu Feng, Jimin Huang et al.

General and legal domain LLMs have demonstrated strong performance in various tasks of LegalAI. However, the current evaluations of these LLMs in LegalAI are defined by the experts of computer science, lacking consistency with the logic of legal practice, making it difficult to judge their practical capabilities. To address this challenge, we are the first to build the Chinese legal LLMs benchmark LAiW, based on the logic of legal practice. To align with the thinking process of legal experts and legal practice (syllogism), we divide the legal capabilities of LLMs from easy to difficult into three levels: basic information retrieval, legal foundation inference, and complex legal application. Each level contains multiple tasks to ensure a comprehensive evaluation. Through automated evaluation of current general and legal domain LLMs on our benchmark, we indicate that these LLMs may not align with the logic of legal practice. LLMs seem to be able to directly acquire complex legal application capabilities but perform poorly in some basic tasks, which may pose obstacles to their practical application and acceptance by legal experts. To further confirm the complex legal application capabilities of current LLMs in legal application scenarios, we also incorporate human evaluation with legal experts. The results indicate that while LLMs may demonstrate strong performance, they still require reinforcement of legal logic.

Yuxin Wang, Duanyu Feng, Yongfu Dai et al.

Data serves as the fundamental foundation for advancing deep learning, particularly tabular data presented in a structured format, which is highly conducive to modeling. However, even in the era of LLM, obtaining tabular data from sensitive domains remains a challenge due to privacy or copyright concerns. Hence, exploring how to effectively use models like LLMs to generate realistic and privacy-preserving synthetic tabular data is urgent. In this paper, we take a step forward to explore LLMs for tabular data synthesis and privacy protection, by introducing a new framework HARMONIC for tabular data generation and evaluation. In the tabular data generation of our framework, unlike previous small-scale LLM-based methods that rely on continued pre-training, we explore the larger-scale LLMs with fine-tuning to generate tabular data and enhance privacy. Based on idea of the k-nearest neighbors algorithm, an instruction fine-tuning dataset is constructed to inspire LLMs to discover inter-row relationships. Then, with fine-tuning, LLMs are trained to remember the format and connections of the data rather than the data itself, which reduces the risk of privacy leakage. In the evaluation part of our framework, we develop specific privacy risk metrics DLT for LLM synthetic data generation, as well as performance evaluation metrics LLE for downstream LLM tasks. Our experiments find that this tabular data generation framework achieves equivalent performance to existing methods with better privacy, which also demonstrates our evaluation framework for the effectiveness of synthetic data and privacy risks in LLM scenarios.

Qianqian Xie, Weiguang Han, Zhengyu Chen et al.

LLMs have transformed NLP and shown promise in various fields, yet their potential in finance is underexplored due to a lack of comprehensive evaluation benchmarks, the rapid development of LLMs, and the complexity of financial tasks. In this paper, we introduce FinBen, the first extensive open-source evaluation benchmark, including 36 datasets spanning 24 financial tasks, covering seven critical aspects: information extraction (IE), textual analysis, question answering (QA), text generation, risk management, forecasting, and decision-making. FinBen offers several key innovations: a broader range of tasks and datasets, the first evaluation of stock trading, novel agent and Retrieval-Augmented Generation (RAG) evaluation, and three novel open-source evaluation datasets for text summarization, question answering, and stock trading. Our evaluation of 15 representative LLMs, including GPT-4, ChatGPT, and the latest Gemini, reveals several key findings: While LLMs excel in IE and textual analysis, they struggle with advanced reasoning and complex tasks like text generation and forecasting. GPT-4 excels in IE and stock trading, while Gemini is better at text generation and forecasting. Instruction-tuned LLMs improve textual analysis but offer limited benefits for complex tasks such as QA. FinBen has been used to host the first financial LLMs shared task at the FinNLP-AgentScen workshop during IJCAI-2024, attracting 12 teams. Their novel solutions outperformed GPT-4, showcasing FinBen's potential to drive innovation in financial LLMs. All datasets, results, and codes are released for the research community: https://github.com/The-FinAI/PIXIU.

Duanyu Feng, Li Ding, Hongru Liang et al.

Edge classification, a crucial task for graph applications, remains relatively under-explored compared to link prediction. Current methods often overlook the potential causal influences of node features on edge features, leading to a loss of relevant prior information. In this work, we present an empirical exploration using the Causal Edge Classification Framework (CECF). Unlike conventional causal inference methods, CECF is the first framework to apply causal inference principles to the edge classification task and to explore modeling edge features as a high-dimensional treatment within a causal framework. Based on the node embedding of Graph Neural Network (GNN), CECF seeks to learn a balanced representation of high-dimensional edge features by mitigating the potential influence of node features. Then, a cross-attention network captures the complex dependencies between node and edge features for final edge classification.Extensive experiments demonstrate that CECF not only achieves superior performance but also serves as a flexible, plug-and-play enhancement for existing methods.We also provide empirical analyses, offering insights into when and how this high-dimensional causal modeling framework works for the edge classification.

Li Ding, Duanyu Feng, Chen Huang et al.

Protein-Text Question Answering (QA) is crucial for interpreting biological sequences through natural language. The integration of Large Language Models (LLMs) with Retrieval-Augmented Generation (RAG) that efficiently leverages biological databases and facilitates reasoning offers a potent approach for it. However, constrained by the standard RAG pipeline, these models often rely on curated, static datasets instead of expert-proven biological workflows, lacking the fine-grained information processing and struggling to generalize to novel (OOD) proteins. To bridge this gap, we propose 2D-ProteinRAG, a novel framework that empowers LLMs to operate within the gold-standard biological research workflow (BLAST). To further extract high-quality information from noisy retrieval contexts, we introduce a dual-dimensional (2D) filtering strategy following the expert analytical paradigms. Horizontal Fine-grained Attribute Alignment utilizes a lightweight, intent-aware discriminative filter to prune irrelevant metadata and align database entries with specific user queries. Vertical Homology-based Semantic Denoising resolves functional contradictions and redundancy across multiple homologs via hierarchical clustering. Extensive evaluations on both In-Distribution and diverse biological OOD benchmarks demonstrate that 2D-ProteinRAG consistently achieves state-of-the-art performance, outperforming fine-tuned baselines and other RAG methods. Our results validate the framework's robustness and scalability, providing a practical solution for interpreting protein functions in real-world scientific scenarios.

Qinyi Zhang, Duanyu Feng, Ronghui Han et al.

Simulating microstructure evolution (MicroEvo) is vital for materials design but demands high numerical accuracy, efficiency, and physical fidelity. Although recent studies on deep learning (DL) offer a promising alternative to traditional solvers, the field lacks standardized benchmarks. Existing studies are flawed due to a lack of comparing specialized MicroEvo DL models with state-of-the-art spatio-temporal architectures, an overemphasis on numerical accuracy over physical fidelity, and a failure to analyze error propagation over time. To address these gaps, we introduce MicroEvoEval, the first comprehensive benchmark for image-based microstructure evolution prediction. We evaluate 14 models, encompassing both domain-specific and general-purpose architectures, across four representative MicroEvo tasks with datasets specifically structured for both short- and long-term assessment. Our multi-faceted evaluation framework goes beyond numerical accuracy and computational cost, incorporating a curated set of structure-preserving metrics to assess physical fidelity. Our extensive evaluations yield several key insights. Notably, we find that modern architectures (e.g., VMamba), not only achieve superior long-term stability and physical fidelity but also operate with an order-of-magnitude greater computational efficiency. The results highlight the necessity of holistic evaluation and identify these modern architectures as a highly promising direction for developing efficient and reliable surrogate models in data-driven materials science.

Shuai Li, Chen Huang, Duanyu Feng et al.

To augment Large Language Models (LLMs) for multi-hop question answering, a mainstream solution within Graph Retrieval Augmented Generation (GraphRAG) leverages lightweight retrievers to efficiently extract information from a given Knowledge Graph (KG). However, existing methods often overlook the inherent challenge of sparse semantic information in graphs. Specifically, our experiments reveal that these methods produce biased retrieval Semantic Shortcut Bias and Long-Tail Path Bias, leading to inadequate semantic modeling and limited GraphRAG effectiveness. To address these issues, we propose STAR, a semantic-tuned and tail-adaptive retriever for GraphRAG. STAR integrates two key learning paradigms: token-level interaction learning and path-weighted contrastive learning. The former employs a cross-attention architecture and a hard path mining mechanism to jointly model the query and path, thereby mitigating the Semantic Shortcut Bias. The latter introduces a tailored contrastive learning objective that utilizes tail-adaptive path weighting, designed to optimize the training process and ease the Long-Tail Path Bias. Extensive experiments demonstrate that STAR consistently outperforms baselines, achieving average retrieval performance gains of 1.8\% and LLM QA performance improvements of 2.2\% across all benchmark datasets. Our code is available at https://anonymous.4open.science/r/STAR-C583.

Jimin Huang, Duanyu Feng, Nuo Chen et al.

Federated learning (FL) enables training large language models (LLMs) without sharing raw data, but adapting LLMs under strict data isolation and non-IID client distributions remains challenging in practice. Synthetic data offers a natural privacy-preserving surrogate for local training, yet existing federated pipelines typically treat synthetic generation as static or loosely coupled with downstream optimization, leading to rapidly diminishing utility under heterogeneous clients. We study federated adaptation of LLMs on tabular tasks where raw records and validation data cannot be shared, and local training must rely entirely on synthetic tables. We propose Concordia, a tri-level optimization framework that aligns synthetic data generation with federated validation utility despite these constraints. At the client level, models are adapted via parameter-efficient LoRA training on synthetic tables. Clients additionally learn lightweight utility scorers from private validation feedback to reweight synthetic samples during local training. At the outer level, each client refines its own synthetic table generator using group-relative policy optimization (GRPO), guided by an ensemble of heterogeneous scorers shared across clients, without aggregating generator parameters or exposing validation data. Experiments on privacy-sensitive tabular benchmarks from finance and healthcare demonstrate that Concordia consistently improves federated performance, cross-client stability, and robustness to distribution shift compared to static and decoupled synthetic-data baselines.

Youcheng Huang, Jingkun Tang, Duanyu Feng et al.

People tell lies when seeking rewards. Large language models (LLMs) are aligned to human values with reinforcement learning where they get rewards if they satisfy human preference. We find that this also induces dishonesty in helpful and harmless alignment where LLMs tell lies in generating harmless responses. Using the latest interpreting tools, we detect dishonesty, show how LLMs can be harmful if their honesty is increased, and analyze such conflicts at the parameter-level. Given these preliminaries and the hypothesis that reward-seeking stimulates dishonesty, we theoretically show that the dishonesty can in-turn decrease the alignment performances and augment reward-seeking alignment with representation regularization. Extensive results, including GPT-4 annotated win-rates, perplexities, and cases studies demonstrate that we can train more honest, helpful, and harmless LLMs. We will make all our codes and results be open-sourced upon this paper's acceptance.

Ruibo Deng, Duanyu Feng, Wenqiang Lei

Offline preference optimization offers a simpler and more stable alternative to RLHF for aligning language models. However, their effectiveness is critically dependent on ranking accuracy, a metric where further gains are highly impactful. This limitation arises from a fundamental problem that we identify and formalize as the Overfitting-Underfitting Dilemma: current margin designs cause models to apply excessive, wasteful gradients to correctly ranked samples (overfitting) while providing insufficient corrective signals for misranked ones (underfitting). To resolve this dilemma, we propose Adaptive Margin-attached Preference Optimization (AMaPO), a simple yet principled algorithm. AMaPO employs an instance-wise adaptive margin, refined by Z-normalization and exponential scaling, which dynamically reallocates learning effort by amplifying gradients for misranked samples and suppressing them for correct ones. Extensive experiments on widely used benchmarks demonstrate that AMaPO not only achieves better ranking accuracy and superior downstream alignment performance, but targeted analysis also confirms that it successfully mitigates the core overfitting and underfitting issues.

Duanyu Feng, Bowen Qin, Chen Huang et al.

Direct Preference Optimization (DPO), which derives reward signals directly from pairwise preference data, has shown its effectiveness on aligning Large Language Models (LLMs) with human preferences. Despite its widespread use across various tasks, DPO has been criticized for its sensitivity to the SFT's effectiveness and its hindrance to the learning capacity towards human-preferred responses, leading to less satisfactory performance. To overcome those limitations, the theoretical understanding of DPO are indispensable but still lacking. To this end, we take a step towards theoretically analyzing and understanding the limitations of DPO. Specifically, we provide an analytical framework using the field theory to analyze the optimization process of DPO. By analyzing the gradient vector field of the DPO loss function, we find that the DPO loss function decreases the probability of producing human dispreferred data at a faster rate than it increases the probability of producing preferred data. This provides theoretical insights for understanding the limitations of DPO discovered in the related research experiments, thereby setting the foundation for its improvement.

Xiao Zhang, Ruoyu Xiang, Chenhan Yuan et al.

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 reveal a significant multilingual performance gap and bias in existing LLMs. FinMA-ES models surpass SOTA LLMs such as GPT-4 in Spanish financial tasks, due to strategic instruction tuning and leveraging data from diverse linguistic resources, highlighting the positive impact of cross-linguistic transfer. All our datasets, models, and benchmarks have been released.

Bowen Qin, Duanyu Feng, Xi Yang

Reinforcement Learning from Human Feedback (RLHF) is a widely used framework for the training of language models. However, the process of using RLHF to develop a language model that is well-aligned presents challenges, especially when it comes to optimizing the reward model. Our research has found that existing reward models, when trained using the traditional ranking objective based on human preference data, often struggle to effectively distinguish between responses that are more or less favorable in real-world scenarios. To bridge this gap, our study introduces a novel method to estimate the preference differences without the need for detailed, exhaustive labels from human annotators. Our experimental results provide empirical evidence that incorporating margin values into the training process significantly improves the effectiveness of reward models. This comparative analysis not only demonstrates the superiority of our approach in terms of reward prediction accuracy but also highlights its effectiveness in practical applications.

Youcheng Huang, Chen Huang, Duanyu Feng et al.

Understanding the inner workings of Large Language Models (LLMs) is a critical research frontier. Prior research has shown that a single LLM's concept representations can be captured as steering vectors (SVs), enabling the control of LLM behavior (e.g., towards generating harmful content). Our work takes a novel approach by exploring the intricate relationships between concept representations across different LLMs, drawing an intriguing parallel to Plato's Allegory of the Cave. In particular, we introduce a linear transformation method to bridge these representations and present three key findings: 1) Concept representations across different LLMs can be effectively aligned using simple linear transformations, enabling efficient cross-model transfer and behavioral control via SVs. 2) This linear transformation generalizes across concepts, facilitating alignment and control of SVs representing different concepts across LLMs. 3) A weak-to-strong transferability exists between LLM concept representations, whereby SVs extracted from smaller LLMs can effectively control the behavior of larger LLMs.

Chen Huang, Duanyu Feng, Wenqiang Lei et al.

Deploying dense retrieval models efficiently is becoming increasingly important across various industries. This is especially true for enterprise search services, where customizing search engines to meet the time demands of different enterprises in different domains is crucial. Motivated by this, we develop a time-efficient approach called DREditor to edit the matching rule of an off-the-shelf dense retrieval model to suit a specific domain. This is achieved by directly calibrating the output embeddings of the model using an efficient and effective linear mapping. This mapping is powered by an edit operator that is obtained by solving a specially constructed least squares problem. Compared to implicit rule modification via long-time finetuning, our experimental results show that DREditor provides significant advantages on different domain-specific datasets, dataset sources, retrieval models, and computing devices. It consistently enhances time efficiency by 100-300 times while maintaining comparable or even superior retrieval performance. In a broader context, we take the first step to introduce a novel embedding calibration approach for the retrieval task, filling the technical blank in the current field of embedding calibration. This approach also paves the way for building domain-specific dense retrieval models efficiently and inexpensively.

Youcheng Huang, Bowen Qin, Chen Huang et al.

Large Reasoning Models (LRMs) have demonstrated remarkable problem-solving abilities in mathematics, as evaluated by existing benchmarks exclusively on well-defined problems. However, such evaluation setup constitutes a critical gap, since a genuine intelligent agent should not only solve problems (as a math quiz solver), but also be able~to ask for information when the problems lack sufficient information, enabling proactivity in responding users' requests. To bridge such gap, we proposes a new dataset consisting of two types of incomplete problems with diverse contexts. Based on the dataset, our systematical evaluation of LRMs reveals their inability in proactively asking for information. In addition, we uncover the behaviors related to overthinking and hallucination of LRMs, and highlight the potential and challenges of supervised fine-tuning in learning such ability. We hope to provide new insights in developing LRMs with genuine intelligence, rather than just solving problems.

Ronghui Han, Duanyu Feng, Hongyu Du et al.

Accurate time series forecasting, predicting future values based on past data, is crucial for diverse industries. Many current time series methods decompose time series into multiple sub-series, applying different model architectures and training with an end-to-end overall loss for forecasting. However, this raises a question: does this overall loss prioritize the importance of critical sub-series within the decomposition for the better performance? To investigate this, we conduct a study on the impact of overall loss on existing time series methods with sequence decomposition. Our findings reveal that overall loss may introduce bias in model learning, hindering the learning of the prioritization of more significant sub-series and limiting the forecasting performance. To address this, we propose a hybrid loss framework combining the global and component losses. This framework introduces component losses for each sub-series alongside the original overall loss. It employs a dual min-max algorithm to dynamically adjust weights between the overall loss and component losses, and within component losses. This enables the model to achieve better performance of current time series methods by focusing on more critical sub-series while still maintaining a low overall loss. We integrate our loss framework into several time series methods and evaluate the performance on multiple datasets. Results show an average improvement of 0.5-2% over existing methods without any modifications to the model architectures.

Duanyu Feng, Bowen Qin, Chen Huang et al.

The success of the reward model in distinguishing between responses with subtle safety differences depends critically on the high-quality preference dataset, which should capture the fine-grained nuances of harmful and harmless responses. This motivates the need to develop a dataset involving preference margins, which accurately quantify how harmless one response is compared to another. In this paper, we take the first step to propose an effective and cost-efficient framework to promote the margin-enhanced preference dataset development. Our framework, Legend, Leverages representation engineering to annotate preference datasets. It constructs the specific direction within the LLM's embedding space that represents safety. By leveraging this safety direction, Legend can then leverage the semantic distances of paired responses along this direction to annotate margins automatically. We experimentally demonstrate our effectiveness in both reward modeling and harmless alignment for LLMs. Legend also stands out for its efficiency, requiring only the inference time rather than additional training. This efficiency allows for easier implementation and scalability, making Legend particularly valuable for practical applications in aligning LLMs with safe conversations.