Reproduce “Expect the Unexpected: FailSafe Long Context QA for Finance” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kiran Kamble, Melisa Russak, Dmytro Mozolevskyi, Muayad Ali, Mateusz Russak, Waseem AlShikh
This paper introduces FailSafeQA, a new long-context financial benchmark, to evaluate Large Language Models' robustness and context grounding under realistic conditions, including user errors and degraded data. The benchmark reveals that achieving balanced "fail-safe" behavior—accurately answering valid queries under imperfect conditions while reliably refusing to hallucinate when information is insufficient—remains a significant challenge, with top models reaching a compliance score of 0.81.
Key insights
- All evaluated LLMs experienced performance degradation under robustness perturbations, with OCRed context and out-of-domain queries proving to be the most challenging input types.
- A significant trade-off exists between LLM Robustness (ability to answer despite perturbations) and Context Grounding (ability to refuse when context is insufficient), with some robust models fabricating information in over 40% of unanswerable cases.
- Text generation tasks were found to be more susceptible to context alterations and hallucination than question-answering tasks, highlighting a higher risk for generative use cases in finance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.