FinanceHarness: Autonomous Financial Deep Research Framework
创建于 更新于
摘要
FinanceHarness introduces a layered, expert-guided harness and a point-in-time financial deep research benchmark (FinanceGym) to address the gap between generic deep-research systems and analyst-style financial reports. The PIT sandbox enables per-question cutoffs and forward-looking evaluation, while FinanceGym provides 400 expert-annotated questions with 2,464 rubric items across topics, sectors, and reasoning types, split into pre-cutoff evidence retrieval and post-cutoff outcome anticipation [page::2][page::4][page::5]. The framework demonstrates that even strong open-weight models struggle to bridge retrieval with forward-looking synthesis, with FinanceHarness achieving a 32.4% overall rubric score on a 27B backbone—close to larger backbones and well above baseline open-weight systems—and GRPO trainer adds a further ~0.4 points, underscoring the value of a shared, temporally controlled research environment [page::12][page::9][page::11]. The paper details the architecture (orchestration, tools, runtime, and skills), dataset pipeline, evaluation contract, and illustrative case studies indicating the harness's capacity to produce citation-grounded analyst-style reports within a reproducible PIT framework [page::6][page::7][page::8].
