Fin-R1: A Large Language Model for Financial Reasoning through Reinforcement Learning

Reasoning large language models are rapidly evolving across various domains. However, their capabilities in handling complex financial tasks still require in-depth exploration. In this paper, we introduce Fin-R1, a reasoning large language model specifically designed for the financial sector. Fin-R1 is built using a two-stage architecture, leveraging a financial reasoning dataset distilled and processed based on DeepSeek-R1. Through supervised fine-tuning (SFT) and reinforcement learning (RL) training, it demonstrates performance close to DeepSeek-R1 with a parameter size of 7 billion across a range of financial reasoning tasks. It achieves the state-of-the-art (SOTA) in the FinQA and ConvFinQA tasks between those LLMs in our evaluation, surpassing larger models in other tasks as well. Fin-R1 showcases strong reasoning and decision-making capabilities, providing solutions to various problems encountered in the financial domain. Our code is available at https://github.com/SUFE-AIFLM-Lab/Fin-R1.

What Makes a Reward Model a Good Teacher? An Optimization Perspective

The success of Reinforcement Learning from Human Feedback (RLHF) critically depends on the quality of the reward model. While this quality is primarily evaluated through accuracy, it remains unclear whether accuracy fully captures what makes a reward model an effective teacher. We address this question from an optimization perspective. First, we prove that regardless of how accurate a reward model is, if it induces low reward variance, then the RLHF objective suffers from a flat landscape. Consequently, even a perfectly accurate reward model can lead to extremely slow optimization, underperforming less accurate models that induce higher reward variance. We additionally show that a reward model that works well for one language model can induce low reward variance, and thus a flat objective landscape, for another. These results establish a fundamental limitation of evaluating reward models solely based on accuracy or independently of the language model they guide. Experiments using models of up to 8B parameters corroborate our theory, demonstrating the interplay between reward variance, accuracy, and reward maximization rate. Overall, our findings highlight that beyond accuracy, a reward model needs to induce sufficient variance for efficient optimization.

ReelWave: A Multi-Agent Framework Toward Professional Movie Sound Generation

Film production is an important application for generative audio, where richer context is provided through multiple scenes. In ReelWave, we propose a multi-agent framework for audio generation inspired by the professional movie production process. We first capture semantic and temporal synchronized "on-screen" sound by training a prediction model that predicts three interpretable time-varying audio control signals comprising loudness, pitch, and timbre. These three parameters are subsequently specified as conditions by a cross-attention module. Then, our framework infers "off-screen" sound to complement the generation through cooperative interaction between communicative agents. Each agent takes up specific roles similar to the movie production team and is supervised by an agent called the director. Besides, we investigate when the conditional video consists of multiple scenes, a case frequently seen in videos extracted from movies of considerable length. Consequently, our framework can capture a richer context of audio generation conditioned on video clips extracted from movies.

Dyads: Artist-Centric, AI-Generated Dance Duets

Existing AI-generated dance methods primarily train on motion capture data from solo dance performances, but a critical feature of dance in nearly any genre is the interaction of two or more bodies in space. Moreover, many works at the intersection of AI and dance fail to incorporate the ideas and needs of the artists themselves into their development process, yielding models that produce far more useful insights for the AI community than for the dance community. This work addresses both needs of the field by proposing an AI method to model the complex interactions between pairs of dancers and detailing how the technical methodology can be shaped by ongoing co-creation with the artistic stakeholders who curated the movement data. Our model is a probability-and-attention-based Variational Autoencoder that generates a choreographic partner conditioned on an input dance sequence. We construct a custom loss function to enhance the smoothness and coherence of the generated choreography. Our code is open-source, and we also document strategies for other interdisciplinary research teams to facilitate collaboration and strong communication between artists and technologists.

Invisible Strings: Revealing Latent Dancer-to-Dancer Interactions with Graph Neural Networks

Dancing in a duet often requires a heightened attunement to one's partner: their orientation in space, their momentum, and the forces they exert on you. Dance artists who work in partnered settings might have a strong embodied understanding in the moment of how their movements relate to their partner's, but typical documentation of dance fails to capture these varied and subtle relationships. Working closely with dance artists interested in deepening their understanding of partnering, we leverage Graph Neural Networks (GNNs) to highlight and interpret the intricate connections shared by two dancers. Using a video-to-3D-pose extraction pipeline, we extract 3D movements from curated videos of contemporary dance duets, apply a dedicated pre-processing to improve the reconstruction, and train a GNN to predict weighted connections between the dancers. By visualizing and interpreting the predicted relationships between the two movers, we demonstrate the potential for graph-based methods to construct alternate models of the collaborative dynamics of duets. Finally, we offer some example strategies for how to use these insights to inform a generative and co-creative studio practice.

Transformers Learn to Implement Multi-step Gradient Descent with Chain of Thought

Chain of Thought (CoT) prompting has been shown to significantly improve the performance of large language models (LLMs), particularly in arithmetic and reasoning tasks, by instructing the model to produce intermediate reasoning steps. Despite the remarkable empirical success of CoT and its theoretical advantages in enhancing expressivity, the mechanisms underlying CoT training remain largely unexplored. In this paper, we study the training dynamics of transformers over a CoT objective on an in-context weight prediction task for linear regression. We prove that while a one-layer linear transformer without CoT can only implement a single step of gradient descent (GD) and fails to recover the ground-truth weight vector, a transformer with CoT prompting can learn to perform multi-step GD autoregressively, achieving near-exact recovery. Furthermore, we show that the trained transformer effectively generalizes on the unseen data. With our technique, we also show that looped transformers significantly improve final performance compared to transformers without looping in the in-context learning of linear regression. Empirically, we demonstrate that CoT prompting yields substantial performance improvements.

ZiGong 1.0: A Large Language Model for Financial Credit

Large Language Models (LLMs) have demonstrated strong performance across various general Natural Language Processing (NLP) tasks. However, their effectiveness in financial credit assessment applications remains suboptimal, primarily due to the specialized financial expertise required for these tasks. To address this limitation, we propose ZiGong, a Mistral-based model enhanced through multi-task supervised fine-tuning. To specifically combat model hallucination in financial contexts, we introduce a novel data pruning methodology. Our approach utilizes a proxy model to score training samples, subsequently combining filtered data with original datasets for model training. This data refinement strategy effectively reduces hallucinations in LLMs while maintaining reliability in downstream financial applications. Experimental results show our method significantly enhances model robustness and prediction accuracy in real-world financial scenarios.

Humanity's Last Exam

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 3,000 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Class Incremental Fault Diagnosis under Limited Fault Data via Supervised Contrastive Knowledge Distillation

Class-incremental fault diagnosis requires a model to adapt to new fault classes while retaining previous knowledge. However, limited research exists for imbalanced and long-tailed data. Extracting discriminative features from few-shot fault data is challenging, and adding new fault classes often demands costly model retraining. Moreover, incremental training of existing methods risks catastrophic forgetting, and severe class imbalance can bias the model's decisions toward normal classes. To tackle these issues, we introduce a Supervised Contrastive knowledge distiLlation for class Incremental Fault Diagnosis (SCLIFD) framework proposing supervised contrastive knowledge distillation for improved representation learning capability and less forgetting, a novel prioritized exemplar selection method for sample replay to alleviate catastrophic forgetting, and the Random Forest Classifier to address the class imbalance. Extensive experimentation on simulated and real-world industrial datasets across various imbalance ratios demonstrates the superiority of SCLIFD over existing approaches. Our code can be found at https://github.com/Zhang-Henry/SCLIFD_TII.

SeMi: When Imbalanced Semi-Supervised Learning Meets Mining Hard Examples

Semi-Supervised Learning (SSL) can leverage abundant unlabeled data to boost model performance. However, the class-imbalanced data distribution in real-world scenarios poses great challenges to SSL, resulting in performance degradation. Existing class-imbalanced semi-supervised learning (CISSL) methods mainly focus on rebalancing datasets but ignore the potential of using hard examples to enhance performance, making it difficult to fully harness the power of unlabeled data even with sophisticated algorithms. To address this issue, we propose a method that enhances the performance of Imbalanced Semi-Supervised Learning by Mining Hard Examples (SeMi). This method distinguishes the entropy differences among logits of hard and easy examples, thereby identifying hard examples and increasing the utility of unlabeled data, better addressing the imbalance problem in CISSL. In addition, we maintain a class-balanced memory bank with confidence decay for storing high-confidence embeddings to enhance the pseudo-labels' reliability. Although our method is simple, it is effective and seamlessly integrates with existing approaches. We perform comprehensive experiments on standard CISSL benchmarks and experimentally demonstrate that our proposed SeMi outperforms existing state-of-the-art methods on multiple benchmarks, especially in reversed scenarios, where our best result shows approximately a 54.8\% improvement over the baseline methods.