PolyGnosis 2.0: Enhancing LLM Reasoning via Agentic Harness Engineering for Polymarket and OSINT Insight Extraction

Problem
This paper addresses the gap in predictive intelligence extraction from financial markets, specifically focusing on the integration of Polymarket anomaly signals with Open Source Intelligence (OSINT) streams. The authors introduce the concept of “Perspective Mismatches,” which refers to the divergence between Polymarket sentiment and global media narratives. This work is presented as a preprint and has not yet undergone peer review.

Method
The core technical contribution is the PolyGnosis 2.0 architecture, a multi-agent system that employs various “Harness Engineering” techniques to enhance reasoning capabilities in high-noise environments. Key methodologies include:

• Reflection Loops: Mechanisms for agents to iteratively refine their outputs.
• Tool-Calling: Integration of external tools to augment decision-making.
• Divide-and-Conquer (D&C) Partitioning: A strategy to break down complex tasks into manageable sub-tasks, facilitating multi-dimensional alignment.
• Chain-of-Thought (CoT): A structured reasoning approach that allows agents to articulate their thought processes.

The authors rigorously quantify the effectiveness of these techniques, particularly noting that while structural partitioning is essential, excessive reflection can lead to logical drift. The training compute specifics are not disclosed, but the architecture is designed to optimize for both analytical precision and operational efficiency.

Results
The empirical evaluation of PolyGnosis 2.0 against human-expert benchmarks demonstrates significant improvements in predictive accuracy. The authors report that their optimal configuration achieves a professional-grade analytical precision, outperforming baseline models in terms of both latency and token efficiency. Specific metrics are not detailed in the abstract, but the results indicate a clear advantage over existing methodologies in the context of financial narrative reasoning.

Limitations
The authors acknowledge several limitations, including the presence of “consensus bias” across all agent configurations, which can skew narrative reasoning outcomes. They also note that while their techniques improve performance, the potential for logical drift due to unconstrained terminal reflection remains a concern. Additionally, the study’s reliance on specific datasets (Polymarket and GDELT) may limit generalizability to other domains or data sources.

Why it matters
The implications of this work are significant for the fields of financial forecasting and automated intelligence systems. By providing a robust framework for integrating diverse data streams and enhancing reasoning capabilities, PolyGnosis 2.0 sets a precedent for future research in predictive analytics. The findings suggest that harnessing multi-agent architectures with advanced reasoning techniques can lead to more accurate and timely insights in volatile markets, paving the way for improved decision-making in trading and investment strategies.

Authors: Daren Wang, Hong Xu, Jiawen Xian
Source: arXiv:2605.25958
URL: https://arxiv.org/abs/2605.25958v1