Fine-Grained Inconsistency Classification in Financial Disclosures
A recent study posted on arXiv addresses the challenge of identifying subtle discrepancies in financial disclosures. Researchers analyzed a dataset of 5,940 instances from the SBID-FD benchmark, utilizing 11 distinct inconsistency labels and associated evidence. The study examined various models, including frozen embedding classifiers and fine-tuned encoders, all under a standardized evaluation system. Notably, a fine-tuned 300M encoder achieved an accuracy rate of 61.9%, while a LoRA-adapted Qwen3.5-9B model followed closely with 61.5%. The paper underscores that distinguishing a conflict is only the initial step; precise classification is essential for thorough validation. The work is cataloged under arXiv ID 2607.26368.
Key facts
- Study addresses fine-grained inconsistency classification in financial disclosures.
- Uses 5,940-instance snapshot of SBID-FD benchmark.
- Benchmark includes 11 inconsistency labels and paired reference evidence spans.
- Compared frozen embedding classifiers, fine-tuned encoders, evidence-augmented classifiers, prompted LLMs, and LoRA-adapted generative models.
- Fine-tuned 300M encoder achieved 61.9% accuracy.
- LoRA-adapted Qwen3.5-9B model achieved 61.5% accuracy.
- Another model achieved 61.3% accuracy.
- Inconsistency types include numerical, temporal, referential, factual, and normative.
- Published on arXiv with ID 2607.26368.
Entities
Institutions
- arXiv