VERA-8B is an end-to-end audit reasoning system designed to identify financial risks in SEC filings before enforcement actions occur. Developed by Menghan Liu and Elynn Chen, the model addresses the "grounding gap" in standard financial language models, which often prioritize fluent, plausible-sounding answers over verifiable evidence. VERA-8B aims to provide auditors with reliable, evidence-backed findings that are ready for professional review.
Addressing the Audit Grounding Gap
Standard language models often struggle with audit tasks because they may cite non-existent text, misattribute quotes, or treat boilerplate language as substantive evidence. The authors note that simply identifying "red-flag" words in a filing is insufficient for audit work, as the timing and context of a disclosure determine its actual evidential weight. VERA-8B is built to solve this by requiring that every audit risk identified be directly supported by specific, verified evidence from the filing.
How VERA-8B Works
The system uses a two-stage framework to ensure audit reliability:
Stage 1 (Supervision Construction): The researchers created an "executable rulebook" that converts public audit knowledge—such as PCAOB standards and historical enforcement mechanisms—into a versioned JSON contract. This contract acts as an evidence gate, ensuring the model only learns from filing spans that meet strict criteria for quotation, context, and relevance.
Stage 2 (Evidence-Closed Alignment): The model undergoes post-training using a combination of Supervised Fine-Tuning (SFT) and Group-Relative Policy Optimization (GRPO). A shared verifier enforces the evidence contract throughout the process. If the model cannot find sufficient evidence to support a claim, it is trained to abstain from making a prediction rather than providing an unsupported answer.
AuditBridge and Reliability
To make the model’s output practical for auditors, the authors designed "AuditBridge." This component transforms raw SEC filings into verified JSON records and renders them as structured, reviewer-ready reports. The system also incorporates uncertainty qualification, which routes ambiguous or evidence-incomplete cases to human reviewers. This ensures that only findings that pass the model’s internal evidence-verification gates are automated.
Performance and Results
In evaluations against general-purpose and finance-specific language models, VERA-8B demonstrated significant improvements in audit-specific metrics. The model achieved a 95% binary F1 score and a 94% verified-positive recall. Notably, the authors report that VERA-8B reduced unsupported claims by 81% compared to the strongest baseline models evaluated. The system is designed to be reusable, with the authors suggesting that its framework for evidence-constrained reasoning can be applied to other tasks where accuracy and verifiability are critical.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!