FinEvolveKG introduces block-constrained prompt optimization for financial knowledge graph construction
A new method separates reusable general rules from persona-specific policies, confining prompt evolution to avoid fragmenting shared grounding while improving extraction quality.
1 source · cross-referenced
- FinEvolveKG constrains prompt optimization to persona-specific blocks to preserve shared grounding rules in financial knowledge graph construction.
- Evaluated on ticker-disjoint SEC 10-K splits, the block-constrained workflow (v4) outperformed a frozen prompt baseline by 19.4–30.9% and an unconstrained full-prompt workflow (v5) by 2.2–6.4%.
- A blinded two-SME audit favored v4 over v5 in 70.1% of contrast-enriched judgments, supporting the method’s empirical and auditability benefits.
Financial institutions build knowledge graphs (KGs) from SEC 10-K filings to support risk analysis, portfolio management, and investment research, but these tasks require different extraction rules. Hand-authored persona prompts and fine-tuning are costly, while unconstrained prompt optimization risks fragmenting shared grounding, typing, and temporal rules across tasks.
The authors introduce FinEvolveKG, a KG-construction framework that separates reusable general blocks from persona-specific policy blocks and confines a reflective prompt optimizer (GEPA) to the designated policy block. During persona adaptation, the general block remains byte-stable, and every optimizer edit is localized to a separately versioned persona policy, improving auditability.
The evaluation uses five mutation scopes on ticker-disjoint 10-K splits with three seeds across three personas. The complete two-stage block-constrained workflow (v4) improves over a frozen prompt baseline (v1) by 19.4–30.9% and over an unconstrained full-prompt workflow (v5) by 2.2–6.4%, leading v5 on both objective components and winning all nine paired-seed comparisons with the lowest seed variance in every persona.
Without further optimization, v4 also exceeds v5 across all evaluated out-of-distribution shifts. A subject-matter expert (SME) audit supports using the generated triplets as ground truth, and a blinded two-SME audit favors v4 over v5 in 70.1% of contrast-enriched judgments.
- Nov 11, 2026 · The Verge — AI
Google DeepMind unveils AlphaGenome 2 to predict genetic variant effects on cell types
Trust79 - Nov 11, 2026 · Google DeepMind — Blog
Google DeepMind unveils AlphaGenome 2 to predict DNA variant effects in specific cell types
Trust84 - Nov 11, 2026 · arXiv cs.AI
Researchers demonstrate auditable human-agent teams for autonomous synchrotron X-ray experiments
Trust79