Scaling Test-Time Inference with Parallel Graph-Retrieval-Augmented Reasoning Chains

Analysis

Why This Paper Matters

This paper addresses a critical limitation in current large reasoning models (LRMs) applied to medical question-answering: error accumulation along single, linear reasoning chains. While chain-of-thought prompting and retrieval-augmented generation (RAG) have improved test-time scaling, they often incorporate unstructured text in a flat, context-agnostic manner, leading to cascading errors. MIRAGE tackles this by leveraging structured medical knowledge graphs and parallel inference paths, which not only reduces error propagation but also enhances traceability—a key requirement in clinical settings.

The significance is twofold. First, it demonstrates that test-time scaling can be improved through architectural choices (multi-path inference) rather than solely increasing model size or compute. Second, by grounding reasoning in explicit graph paths, MIRAGE provides interpretable chains that can be verified by domain experts, addressing the black-box nature of many LRMs. This makes the framework particularly relevant for high-stakes domains like healthcare, where both accuracy and explainability are paramount.

Technical Contributions

• Query Decomposition: Complex medical queries are broken into entity-grounded sub-questions, enabling focused retrieval and reasoning.
• Parallel Inference Paths: Multiple independent reasoning chains are executed simultaneously, reducing the risk of error accumulation seen in single-path approaches.
• Adaptive Evidence Retrieval: The framework dynamically expands neighbor nodes and performs multi-hop traversal over the knowledge graph, retrieving contextually relevant evidence without relying on flat text.
• Cross-Path Verification: Answers from parallel paths are integrated through a verification mechanism that resolves contradictions, improving final answer reliability.
• Interpretability: Each factual claim is traced to concrete paths in the knowledge graph, providing explicit reasoning chains for auditability.

Results

MIRAGE was evaluated on three medical QA benchmarks: GenMedGPT-5k, CMCQA, and ExplainCPE. It consistently outperformed GPT-4o, Tree-of-Thought variants, and other retrieval-augmented baselines in both automatic metrics (e.g., accuracy, F1) and human evaluations. The paper reports that MIRAGE achieves superior performance without requiring additional model fine-tuning, highlighting its efficiency as a test-time scaling method. The explicit reasoning chains also received higher interpretability scores from human evaluators.

Significance

MIRAGE represents a step forward in making LRMs more reliable for critical applications. By combining structured knowledge graphs with multi-path inference, it offers a blueprint for reducing error accumulation in complex reasoning tasks. The framework's emphasis on traceability aligns with regulatory needs in healthcare and other high-stakes fields. Future work could explore extending MIRAGE to other domains (e.g., legal, scientific) and optimizing computational costs of parallel inference.