Probing internal states detects confident hallucinations in financial QA

Researchers demonstrate that linear probes on the residual stream of large language models can reliably detect "confidently wrong" answers in financial question answering, outperforming behavioral confidence baselines. The study evaluates this approach on the FinQA and TAT-QA benchmarks using Qwen3-8B, Llama-3.1-8B, and Gemma-2-9B.

• Among confident answers where eight resampled outputs agreed, 15-23% were incorrect on FinQA.
• Probes achieved an AUROC of 0.68-0.77 for detecting these hallucinations, significantly higher than the 0.55-0.63 AUROC of baselines like token log-probabilities and self-assessment.
• The method leverages internal activations to identify errors that behavioral confidence metrics fail to catch.

The authors suggest that probing offers a cost-effective triage mechanism for routing LLM answers to human review in high-stakes financial applications.