SIGMA: Search-Augmented On-Demand Knowledge Integration for Agentic Mathematical Reasoning

A new multi-agent framework called SIGMA has posted a 7.4% absolute performance gain on mathematical reasoning benchmarks, outperforming both open- and closed-source systems, according to a paper on arXiv [1]. The framework, formally named Search-Augmented On-Demand Knowledge Integration for AGentic Mathematical reAsoning, orchestrates specialized agents that independently reason, perform targeted searches, and synthesize findings through a moderator mechanism [1]. Each agent generates hypothetical passages to optimize retrieval for its own analytic perspective, which the authors describe as a method to keep knowledge integration both context-sensitive and computation-efficient [1]. The paper was submitted to arXiv, the open-access e-print repository that hosts preprints across mathematics, computer science, and other fields without formal peer review [7]. The first version appeared on 31 October 2025, and a revised version followed on 18 June 2026 [1]. SIGMA was evaluated on three challenging benchmarks: MATH500, AIME, and the PhD-level science QA dataset GPQA [1]. Across those tests, the framework consistently outperformed existing systems, recording the 7.4% absolute improvement [1]. The authors argue that current retrieval-augmented models often rely on a single perspective and inflexible search strategies, limiting their ability to combine information from multiple sources [1]. By assigning distinct agents to different analytic viewpoints, SIGMA aims to overcome that bottleneck [1]. The paper appears on arXiv, which as of late 2024 was receiving roughly 24,000 submissions per month and had surpassed two million total articles by the end of 2021 [7]. The repository is moderated but does not conduct peer review, meaning the SIGMA results have not yet been vetted through a journal process [7]. The authors state they will release the code upon publication [1]. The lead author is listed as Ali Asgarov [1].