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

2025-09-09 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences, Robotics & Autonomous Systems · Depth: Expert, long

Summary

SIGMA (Search-Augmented On-Demand Knowledge Integration for Agentic Mathematical Reasoning) is a unified framework designed to solve complex mathematical reasoning problems by orchestrating specialized agents. It addresses limitations of current retrieval-augmented models, which often rely on single perspectives and struggle with multi-source information. SIGMA employs four agents (Factual, Logical, Computational, Completeness) that independently reason, perform targeted searches using hypothetical passages, and synthesize findings via a moderator. This framework achieves an absolute performance improvement of 7.4% on benchmarks like MATH500, AIME, and PhD-level science QA GPQA. SIGMA variants at 1.5B, 3B, and 7B parameters outperform larger closed-source models, including GPT-4o by 8.1% on MATH500, Claude-3.5-Haiku by 1.4%, and show strong gains on AMC23 (5.0%) and AIME24 (3.3%).

Key takeaway

For AI Scientists and Machine Learning Engineers developing advanced reasoning systems, SIGMA offers a robust blueprint for improving performance on knowledge-intensive tasks. You should consider implementing a multi-agent architecture with on-demand, perspective-specific search and a moderator for synthesizing diverse reasoning paths. This approach can significantly boost accuracy and efficiency, especially for complex mathematical or scientific problem-solving, potentially outperforming larger monolithic models.

Key insights

SIGMA uses specialized agents and on-demand search with hypothetical documents to enhance complex mathematical reasoning accuracy and efficiency.

Principles

• Decompose complex tasks across specialized agents.
• Generate hypothetical passages for targeted retrieval.
• Synthesize multi-agent outputs via a moderator.

Method

SIGMA orchestrates Factual, Logical, Computational, and Completeness agents. Each agent performs reasoning-search cycles, generating hypothetical documents for queries, and a moderator synthesizes their outputs.

In practice

• Apply multi-agent systems for complex problem decomposition.
• Use hypothetical document generation to refine search queries.
• Implement a heuristic moderator for output synthesis.

Topics

• Multi-Agent Systems
• Mathematical Reasoning
• Retrieval-Augmented Generation
• Hypothetical Document Enhancement
• Large Language Models
• Benchmarking

Code references

• tongyx361/AwesomeLLM4Math

Best for: AI Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• New Live Session: Designing Multi-Agent Deep Search Systems — Multi-agent deep search systems enhance retrieval by planning, validating, and merging information from diverse sources.
• LEAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks — Agentic frameworks like LEAP enable LLMs to achieve state-of-the-art formal theorem proving by bridging informal reasoning with verifiable proofs.
• MathNet: a Global Multimodal Benchmark for Mathematical Reasoning and Retrieval — MathNet is a new multimodal, multilingual benchmark for advanced mathematical reasoning and retrieval, highlighting current model limitations.
• Integrating Graphs, Large Language Models, and Agents: Reasoning and Retrieval — Graph-LLM integration enhances reasoning and retrieval, with optimal strategies depending on task and data.
• Accelerating Mathematical and Scientific Discovery with Gemini Deep Think — Gemini Deep Think, with agentic reasoning, achieves gold-medal performance in math/programming and solves professional research problems.
• The Illusion of Multi-Agent Advantage — Automated Multi-Agent Systems often incur high costs for superficial complexity, failing to outperform simpler Single-Agent Systems.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.