Analytica: Soft Propositional Reasoning for Robust and Scalable LLM-Driven Analysis
New architecture slashes variance and cost while improving accuracy on complex analysis.
A new paper from Junyan Cheng, Kyle Richardson, and Peter Chin (ICLR 2026) introduces Analytica, an LLM agent architecture designed to overcome the stochastic instability and lack of verifiable structure that plagues current reasoning systems. The core innovation is Soft Propositional Reasoning (SPR), which reframes analysis as estimating the soft truth values of outcome propositions, allowing formal modeling and minimization of bias and variance. Analytica operationalizes this through a parallel divide-and-conquer framework: it decomposes problems into a tree of sub-propositions, uses tool-equipped LLM grounder agents (including a novel Jupyter Notebook agent for data-driven analysis) to validate facts and reduce bias, and recursively synthesizes results using robust linear models that average out noise.
On economic, financial, and political forecasting benchmarks, Analytica improves accuracy by 15.84% on average over diverse base models, reaching 71.06% accuracy with the lowest variance of 6.02% when using a Deep Research grounder. The Jupyter Notebook grounder is particularly cost-effective, achieving 70.11% accuracy with 90.35% less cost and 52.85% less time. The system also shows noise-resilient performance scaling with near-linear time complexity, good adaptivity to open-weight LLMs, and strong results in scientific domains. This makes Analytica a promising approach for robust, scalable, and interactive "what-if" analysis in high-stakes fields like finance and science.
- Analytica uses Soft Propositional Reasoning (SPR) to formally model and minimize estimation bias and variance in LLM-driven analysis.
- Achieves 71.06% accuracy with 6.02% variance on forecasting tasks, improving 15.84% over base models.
- Novel Jupyter Notebook grounder agent cuts costs by 90.35% and time by 52.85% while maintaining 70.11% accuracy.
Why It Matters
Analytica makes LLM agents more reliable and cost-effective for high-stakes analysis, like financial forecasting and scientific discovery.