https://serhii.net/tags/finance/ Recent content in Finance on serhii.net Hugo en-us Serhii Hamotskyi / serhii.net Sat, 16 May 2026 00:00:00 +0200 https://serhii.net/dtb/260516-1107-fnp-2026-at-lrec-paper-notes/ Sat, 16 May 2026 00:00:00 +0200 https://serhii.net/dtb/260516-1107-fnp-2026-at-lrec-paper-notes/ <ul> <li><a href="https://wp.lancs.ac.uk/cfie/fnp2026/">FNP – FNP 2026</a></li> <li>FNP 2026 <a href="http://lrec-conf.org/proceedings/lrec2026/workshops/fnp/2026.fnp-1.0.pdf">proceedings</a> PDF.</li> </ul> <h3 id="when-tables-go-crazy-evaluating-multimodal-models-on-french-financial-documents">When Tables Go Crazy: Evaluating Multimodal Models on French Financial Documents</h3> <ul> <li>Links <ul> <li><a href="https://arxiv.org/abs/2602.10384">[2602.10384] When Tables Go Crazy: Evaluating Multimodal Models on French Financial Documents</a></li> <li><a href="https://github.com/dseddah/Scribe_finance">Github</a> contains both dataset and scripts!</li> </ul> </li> <li>Benchmark French visual finance docs, includes evaluation on multiple VLMs <ul> <li>Chart checkboxes graphs and weird edge cases specifically present</li> </ul> </li> <li>Bits <ul> <li>LLMs when asked to generate questions usually generate only questions they have answers to</li> </ul> </li> </ul> <h3 id="cfqa-a-chinese-financial-question-answering-benchmark-from-corporate-annual-reports">CFQA: A Chinese Financial Question Answering Benchmark from Corporate Annual Reports</h3> <ul> <li>Links: <ul> <li><a href="https://huggingface.co/datasets/ZackZhu00/CFQA_Chinese_Finance_Question_Answering">ZackZhu00/CFQA_Chinese_Finance_Question_Answering · Datasets at Hugging Face</a></li> <li><a href="https://github.com/zhutianning/Hallucination-detection-for-RAG">zhutianning/Hallucination-detection-for-RAG: This project aims to design and implement a hallucination detection and evaluation pipeline for Retrieval- Augmented Generation (RAG) systems processing multi-modal financial reports (text, tables, images/charts).</a></li> <li>Unclear relation: <ul> <li><a href="https://github.com/ygan/CFQA">Github</a> (?) / <a href="https://link.springer.com/article/10.1140/epjds/s13688-025-00601-6#Sec17">Addressing investor concerns: a Chinese financial question-answering benchmark with LLM-based evaluation | EPJ Data Science | Springer Nature Link</a> (??)</li> </ul> </li> </ul> </li> <li>Takeaway <ul> <li>RAG decreases hallucinations and improves fact extraction scores but not clear for other tasks involving reasoning, and decrease scores w/ some model[s</li> </ul> </li> </ul> <h3 id="-verifiable-financial-enterprise-question-answering-via-inference-time-grounding-and-traceability">(!) Verifiable Financial Enterprise Question Answering via Inference-Time Grounding and Traceability</h3> <ul> <li>LLMs grounding verifiable citations</li> <li>Framework, modular, real-time</li> <li>Verify citations and fix citation drift by looking at ovelaps/presence w/ source documents. <ul> <li><strong>Sentence-level citations</strong> allegedly lead to better groundedness</li> </ul> </li> <li><strong>TODO</strong> Many many interesting citations to parse <img src="https://serhii.net/assets/pasted-image-20260516114801.png" alt="Pasted image 20260516114801.png"> </li> <li>Related paper I just found: <a href="https://arxiv.org/abs/2604.23588">[2604.23588] FinGround: Detecting and Grounding Financial Hallucinations via Atomic Claim Verification</a></li> </ul>