Google Paper: Financial AI Agents Excel at the Past, Fail at the Future
rohanpaul_ai · x · 2026-08-04
A new Google paper reveals that financial deep-research agents are much better at reconstructing the past than anticipating the future.
- Low Overall Accuracy: Across 17 baselines tested on 400 expert-annotated questions, every model scored below 40% overall.
- Workflow Matters: Using the same Qwen3.6-27B backbone, upgrading from a simple search loop to a full finance-oriented tool stack raised the score from 25.3% to 32.4%.
- Marginal Training Gains: Extra training via Group Relative Policy Optimization (GRPO) barely changed the result, adding only 0.4 percentage points.
The researchers conclude that the main bottleneck is no longer search or report structure, but the need for better causal and scenario reasoning to improve forward-looking judgment.
More from coding & agent
- ADHD Skill for Claude Code: 6 Isolated Threads to Break AI Tunnel Vision — techNmak · 2026-08-04
- Kuaishou's RecHarness: Automating Recommender Optimization with Bandit-Routed LLM Agents — kuaishou · 2026-08-04
- Survey of 396 Fintechs: 64% of 'AI Agents' Are Just Copilots — LolaCronje · 2026-08-04
- BitBox: Non-engineers successfully audit codebase using AI agents — RSync25 · 2026-08-04
- The Vibe-Coding Curse: AI Impresses Initially but Degrades with Every Detail Tweak — arjunrajlab · 2026-08-04
- Is Traditional RAG Dead? Developers Discuss Six Alternative Architectures — SwordfishWest6860 · 2026-08-04