Selected work

Representative Papers

DFM: Difference Feature Modeling with Text-Guided Gated Contrastive Loss for Remote Sensing Image Change Captioning

Jun 25, 2026

This work addresses the limitations of existing remote sensing image change captioning methods, which rely on a single autoregressive generation paradigm and tend to favor frequent vocabulary while overlooking discriminative differences. To overcome this, the authors propose a difference-aware feature modeling framework that leverages a text-guided gated contrastive loss to steer the visual encoder—via linguistic cues—toward salient change regions. The approach integrates a pretrained change detection model with a multi-scale joint feature modeling module to comprehensively capture spatiotemporal discrepancies between multi-temporal images. By transcending the constraints of conventional generative paradigms, the method achieves significant improvements in both caption accuracy and discriminative expression across multiple remote sensing change captioning benchmarks, demonstrating its effectiveness and novelty.

0 citationsRead paper

Fin-RATE: A Real-world Financial Analytics and Tracking Evaluation Benchmark for LLMs on SEC Filings

Feb 07, 2026

This work addresses the limitations of existing financial large language model (LLM) evaluation benchmarks, which fail to capture the complex reasoning capabilities required of professional analysts across multiple documents, entities, and time dimensions, and lack fine-grained attribution of error sources. To bridge this gap, we introduce Fin-RATE—the first benchmark grounded in U.S. SEC filings that simulates real-world financial analysis workflows. It encompasses three task types: fine-grained single-document reasoning, cross-entity thematic comparison, and longitudinal firm tracking. We systematically evaluate 17 prominent models under both retrieved and given-context settings and, for the first time, categorize and quantify errors arising from retrieval, generation, financial reasoning, and contextual understanding. Experiments reveal accuracy drops of 18.60% and 14.35% on longitudinal and cross-entity tasks, respectively, primarily due to comparative hallucinations, temporal misalignment, and entity mismatches, thereby filling a critical void in evaluating complex financial reasoning.

0 citationsRead paper

Adaptation of Embedding Models to Financial Filings via LLM Distillation

Dec 08, 2025

To address the limitations of generic embedding models in financial retrieval—including insufficient domain expertise, reliance on manual annotations, and trade-offs between efficiency and accuracy—this paper proposes an unsupervised knowledge distillation framework. It employs a large language model (LLM) as a discriminator to automatically mine hard negative examples from financial filings, and iteratively refines a dual-encoder student model via teacher-student interaction. The method requires no human annotation and achieves effective domain adaptation. Experiments on 21,800 query-document pairs demonstrate a 27.7% improvement in MRR@5 and a 44.6% gain in mean DCG@5; NDCG also improves significantly across three of four document categories in FinanceBench. Our core contribution is the first LLM-guided, unsupervised embedding distillation paradigm tailored for financial text—balancing domain specificity, scalability, and deployment efficiency.

0 citationsRead paper

Recent publications

Latest Papers

DFM: Difference Feature Modeling with Text-Guided Gated Contrastive Loss for Remote Sensing Image Change Captioning

Jun 25, 2026

This work addresses the limitations of existing remote sensing image change captioning methods, which rely on a single autoregressive generation paradigm and tend to favor frequent vocabulary while overlooking discriminative differences. To overcome this, the authors propose a difference-aware feature modeling framework that leverages a text-guided gated contrastive loss to steer the visual encoder—via linguistic cues—toward salient change regions. The approach integrates a pretrained change detection model with a multi-scale joint feature modeling module to comprehensively capture spatiotemporal discrepancies between multi-temporal images. By transcending the constraints of conventional generative paradigms, the method achieves significant improvements in both caption accuracy and discriminative expression across multiple remote sensing change captioning benchmarks, demonstrating its effectiveness and novelty.

0 citationsRead paper

Fin-RATE: A Real-world Financial Analytics and Tracking Evaluation Benchmark for LLMs on SEC Filings

Feb 07, 2026

This work addresses the limitations of existing financial large language model (LLM) evaluation benchmarks, which fail to capture the complex reasoning capabilities required of professional analysts across multiple documents, entities, and time dimensions, and lack fine-grained attribution of error sources. To bridge this gap, we introduce Fin-RATE—the first benchmark grounded in U.S. SEC filings that simulates real-world financial analysis workflows. It encompasses three task types: fine-grained single-document reasoning, cross-entity thematic comparison, and longitudinal firm tracking. We systematically evaluate 17 prominent models under both retrieved and given-context settings and, for the first time, categorize and quantify errors arising from retrieval, generation, financial reasoning, and contextual understanding. Experiments reveal accuracy drops of 18.60% and 14.35% on longitudinal and cross-entity tasks, respectively, primarily due to comparative hallucinations, temporal misalignment, and entity mismatches, thereby filling a critical void in evaluating complex financial reasoning.

0 citationsRead paper

Adaptation of Embedding Models to Financial Filings via LLM Distillation

Dec 08, 2025

To address the limitations of generic embedding models in financial retrieval—including insufficient domain expertise, reliance on manual annotations, and trade-offs between efficiency and accuracy—this paper proposes an unsupervised knowledge distillation framework. It employs a large language model (LLM) as a discriminator to automatically mine hard negative examples from financial filings, and iteratively refines a dual-encoder student model via teacher-student interaction. The method requires no human annotation and achieves effective domain adaptation. Experiments on 21,800 query-document pairs demonstrate a 27.7% improvement in MRR@5 and a 44.6% gain in mean DCG@5; NDCG also improves significantly across three of four document categories in FinanceBench. Our core contribution is the first LLM-guided, unsupervised embedding distillation paradigm tailored for financial text—balancing domain specificity, scalability, and deployment efficiency.

0 citationsRead paper