Selected work

Representative Papers

FinSphere: A Conversational Stock Analysis Agent Equipped with Quantitative Tools based on Real-Time Database

Jan 08, 2025arXiv.org

Existing financial large language models (FinLLMs) face two critical bottlenecks in stock analysis: the absence of objective, quantitative evaluation metrics for report quality and insufficient analytical depth to generate professional-grade insights. To address these limitations, we propose the first end-to-end conversational AI agent framework specifically designed for stock analysis. Our method integrates a real-time financial database, a quantitative computation module, and an instruction-tuned LLM to enable multi-step reasoning and structured output generation. Key contributions include: (1) the Stocksis dataset—curated and expert-annotated by finance professionals; (2) AnalyScore—a novel, interpretable, multi-dimensional evaluation metric for analytical report quality; and (3) a modular agent architecture supporting domain-specific reasoning. Experimental results demonstrate that our agent significantly outperforms both general-purpose and financial-domain LLMs, as well as existing agent systems, on professional stock analysis tasks—yielding substantial improvements in report quality, interpretability, and operational applicability.

1 citationsRead paper

Enhancing LLMs' Reasoning-Intensive Multimedia Search Capabilities through Fine-Tuning and Reinforcement Learning

May 24, 2025

Existing LLM-based search agents rely on prompt engineering to decompose queries, limiting their capability on complex reasoning tasks; moreover, Python-based search planning and insufficient multimodal (text-image) integration lead to high token overhead. Method: We propose a natural-language-based search plan representation and design a multimodal search agent supporting visual input and output. We introduce a novel joint training paradigm integrating supervised fine-tuning (SFT) with search-feedback-driven reinforcement learning (RLSF), and establish an automated pipeline for constructing multimodal training data and benchmarks. Contribution/Results: Our approach achieves +36.60% and +54.54% performance gains over Perplexity Pro on FinSearchBench-24 and our newly curated SearchExpertBench-25, respectively. Human evaluation confirms substantial improvements in response readability and coherence.

0 citationsRead paper

Recent publications

Latest Papers

Enhancing LLMs' Reasoning-Intensive Multimedia Search Capabilities through Fine-Tuning and Reinforcement Learning

May 24, 2025

Existing LLM-based search agents rely on prompt engineering to decompose queries, limiting their capability on complex reasoning tasks; moreover, Python-based search planning and insufficient multimodal (text-image) integration lead to high token overhead. Method: We propose a natural-language-based search plan representation and design a multimodal search agent supporting visual input and output. We introduce a novel joint training paradigm integrating supervised fine-tuning (SFT) with search-feedback-driven reinforcement learning (RLSF), and establish an automated pipeline for constructing multimodal training data and benchmarks. Contribution/Results: Our approach achieves +36.60% and +54.54% performance gains over Perplexity Pro on FinSearchBench-24 and our newly curated SearchExpertBench-25, respectively. Human evaluation confirms substantial improvements in response readability and coherence.

0 citationsRead paper

FinSphere: A Conversational Stock Analysis Agent Equipped with Quantitative Tools based on Real-Time Database

Jan 08, 2025arXiv.org

Existing financial large language models (FinLLMs) face two critical bottlenecks in stock analysis: the absence of objective, quantitative evaluation metrics for report quality and insufficient analytical depth to generate professional-grade insights. To address these limitations, we propose the first end-to-end conversational AI agent framework specifically designed for stock analysis. Our method integrates a real-time financial database, a quantitative computation module, and an instruction-tuned LLM to enable multi-step reasoning and structured output generation. Key contributions include: (1) the Stocksis dataset—curated and expert-annotated by finance professionals; (2) AnalyScore—a novel, interpretable, multi-dimensional evaluation metric for analytical report quality; and (3) a modular agent architecture supporting domain-specific reasoning. Experimental results demonstrate that our agent significantly outperforms both general-purpose and financial-domain LLMs, as well as existing agent systems, on professional stock analysis tasks—yielding substantial improvements in report quality, interpretability, and operational applicability.

1 citationsRead paper