{
  "@context": {
    "@vocab": "http://schema.org/",
    "schema": "http://schema.org/",
    "hasPart": { "@id": "schema:hasPart", "@type": "@id" },
    "mainEntity": { "@id": "schema:mainEntity", "@type": "@id" },
    "about": { "@id": "schema:about", "@type": "@id" },
    "author": { "@id": "schema:author", "@type": "@id" },
    "publisher": { "@id": "schema:publisher", "@type": "@id" },
    "creator": { "@id": "schema:creator", "@type": "@id" },
    "contributor": { "@id": "schema:contributor", "@type": "@id" },
    "citation": { "@id": "schema:citation", "@type": "@id" },
    "isPartOf": { "@id": "schema:isPartOf", "@type": "@id" },
    "datePublished": { "@id": "schema:datePublished", "@type": "xsd:date" },
    "dateModified": { "@id": "schema:dateModified", "@type": "xsd:date" }
  },
  "@type": "ScholarlyArticle",
  "@id": "ark-knowledge-graph-exploration",
  "headline": "Autonomous Knowledge Graph Exploration with Adaptive Breadth-Depth Retrieval",
  "author": [
    {
      "@type": "Person",
      "name": "Joaquín Polonuer",
      "affiliation": [
        { "@type": "Organization", "name": "Department of Biomedical Informatics, Harvard Medical School" },
        { "@type": "Organization", "name": "Departamento de Computación, FCEyN, Universidad de Buenos Aires" }
      ],
      "contributorRole": "Equal contribution"
    },
    {
      "@type": "Person",
      "name": "Lucas Vittor",
      "affiliation": { "@type": "Organization", "name": "Department of Biomedical Informatics, Harvard Medical School" },
      "contributorRole": "Equal contribution"
    },
    {
      "@type": "Person",
      "name": "Iñaki Arango",
      "affiliation": { "@type": "Organization", "name": "Department of Biomedical Informatics, Harvard Medical School" },
      "contributorRole": "Equal contribution"
    },
    {
      "@type": "Person",
      "name": "Ayush Noori",
      "affiliation": [
        { "@type": "Organization", "name": "Department of Biomedical Informatics, Harvard Medical School" },
        { "@type": "Organization", "name": "Departamento de Computación, FCEyN, Universidad de Buenos Aires" }
      ],
      "contributorRole": "Equal contribution"
    },
    {
      "@type": "Person",
      "name": "David A. Clifton",
      "affiliation": [
        { "@type": "Organization", "name": "Department of Engineering Science, University of Oxford" },
        { "@type": "Organization", "name": "Oxford Suzhou Centre for Advanced Research, University of Oxford" }
      ]
    },
    {
      "@type": "Person",
      "name": "Luciano Del Corro",
      "affiliation": [
        { "@type": "Organization", "name": "ELIAS Lab, Departamento de Ingeniería, Universidad de San Andrés" },
        { "@type": "Organization", "name": "Kempner Institute for the Study of Natural and Artificial Intelligence" }
      ],
      "contactPoint": {
        "@type": "ContactPoint",
        "email": "delcorrol@udesa.edu.ar"
      },
      "contributorRole": "Co-senior author"
    },
    {
      "@type": "Person",
      "name": "Marinka Zitnik",
      "affiliation": [
        { "@type": "Organization", "name": "Department of Biomedical Informatics, Harvard Medical School" },
        { "@type": "Organization", "name": "Kempner Institute for the Study of Natural and Artificial Intelligence" },
        { "@type": "Organization", "name": "Broad Institute of MIT and Harvard" },
        { "@type": "Organization", "name": "Harvard Data Science Initiative" }
      ],
      "contactPoint": {
        "@type": "ContactPoint",
        "email": "marinka@hms.harvard.edu"
      },
      "contributorRole": "Co-senior author"
    }
  ],
  "datePublished": "2026-04-28",
  "abstract": "ARK (Adaptive Retriever of Knowledge) is a training-free, tool-using knowledge graph retriever that balances breadth and depth in retrieval by alternating between global lexical search and neighborhood exploration. It adapts retrieval strategies to query types without requiring seed selection or pre-set hop depth, achieving state-of-the-art performance on the STaRK benchmark and enabling efficient distillation into smaller models.",
  "articleBody": "ARK enables adaptive retrieval from knowledge graphs by combining global lexical search and neighborhood exploration, controlled by a language model agent. It addresses the breadth-depth tradeoff in multi-hop retrieval without task-specific training or fragile seed selection. ARK achieves strong performance on heterogeneous graphs and can be distilled into smaller models for efficiency.",
  "hasPart": [
    {
      "@type": "DefinedTermSet",
      "@id": "ark-defined-terms",
      "name": "Defined Terms in ARK Knowledge Graph Retrieval",
      "description": "Key terms used in the ARK framework and knowledge graph retrieval context.",
      "hasDefinedTerm": [
        {
          "@type": "DefinedTerm",
          "name": "Knowledge Graph (KG)",
          "description": "A data representation organizing evidence around entities and typed edges, supporting relational constraints and reuse across queries."
        },
        {
          "@type": "DefinedTerm",
          "name": "Global Lexical Search",
          "description": "A retrieval tool that performs broad search over node descriptors using lexical matching (e.g., BM25)."
        },
        {
          "@type": "DefinedTerm",
          "name": "Neighborhood Exploration",
          "description": "A retrieval tool that explores one-hop neighbors of a node, filtered by node and edge types and optionally ranked by query relevance."
        },
        {
          "@type": "DefinedTerm",
          "name": "Breadth-Depth Tradeoff",
          "description": "The balance between broad search across many entities and deep multi-hop traversal along relational paths in knowledge graph retrieval."
        },
        {
          "@type": "DefinedTerm",
          "name": "Hit@1",
          "description": "A retrieval metric measuring the fraction of queries for which the top-ranked retrieved node is relevant."
        },
        {
          "@type": "DefinedTerm",
          "name": "Mean Reciprocal Rank (MRR)",
          "description": "A retrieval metric that averages the reciprocal rank of the first relevant retrieved node across queries."
        },
        {
          "@type": "DefinedTerm",
          "name": "Trajectory",
          "description": "A sequence of agent states, tool invocations, and observations during interactive retrieval."
        },
        {
          "@type": "DefinedTerm",
          "name": "Distillation",
          "description": "The process of training a smaller student model to imitate the tool-use trajectories of a larger teacher model."
        },
        {
          "@type": "DefinedTerm",
          "name": "Relevance Function",
          "description": "A scoring function (e.g., BM25) used to rank candidate nodes based on textual subqueries."
        },
        {
          "@type": "DefinedTerm",
          "name": "Self-Consistency Voting",
          "description": "An aggregation method combining multiple agent retrieval lists by frequency and earliest occurrence to improve robustness."
        }
      ]
    },
    {
      "@type": "Question",
      "@id": "q1",
      "name": "What is the main challenge in retrieving evidence from knowledge graphs?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "The main challenge is balancing the breadth-depth tradeoff: covering broad graph regions for multi-entity queries while performing deep multi-hop traversal for relational evidence."
      }
    },
    {
      "@type": "Question",
      "@id": "q2",
      "name": "How does ARK differ from existing retrieval methods?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "ARK combines global lexical search and neighborhood exploration interactively without requiring seed selection, pre-set hop depth, or task-specific training, adapting retrieval strategy to query needs."
      }
    },
    {
      "@type": "Question",
      "@id": "q3",
      "name": "What tools does ARK use for knowledge graph retrieval?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "ARK uses two tools: Global Search for broad lexical retrieval and Neighborhood Exploration for one-hop relational expansion."
      }
    },
    {
      "@type": "Question",
      "@id": "q4",
      "name": "What is the role of the relevance function in ARK?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "The relevance function ranks candidate nodes returned by tools based on textual subqueries, implemented using BM25 lexical scoring for stable and fast retrieval."
      }
    },
    {
      "@type": "Question",
      "@id": "q5",
      "name": "How does ARK handle multi-hop retrieval?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "ARK performs multi-hop retrieval by alternating between global search and neighborhood exploration, dynamically deciding when to expand or stop based on retrieved evidence."
      }
    },
    {
      "@type": "Question",
      "@id": "q6",
      "name": "What datasets were used to evaluate ARK?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "ARK was evaluated on the STaRK benchmark, which includes three heterogeneous knowledge graphs: AMAZON (e-commerce), MAG (scholarly), and PRIME (biomedical)."
      }
    },
    {
      "@type": "Question",
      "@id": "q7",
      "name": "How does ARK perform compared to baselines?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "ARK achieves the best average retrieval performance across all STaRK graphs, outperforming both training-free and trained retrieval-based and agent-based baselines."
      }
    },
    {
      "@type": "Question",
      "@id": "q8",
      "name": "What is the benefit of distilling ARK into smaller models?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Distillation enables efficient inference by training smaller student models to imitate ARK's tool-use trajectories without requiring ground-truth labels, preserving most retrieval quality."
      }
    },
    {
      "@type": "Question",
      "@id": "q9",
      "name": "How does ARK adapt its retrieval strategy to query types?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "ARK uses global search predominantly for text-heavy queries and neighborhood exploration for relation-heavy queries, autonomously shifting tool use to match query demands."
      }
    },
    {
      "@type": "Question",
      "@id": "q10",
      "name": "What are the limitations of ARK?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "ARK incurs higher latency due to multiple LLM calls, relies on large proprietary LLMs for best performance, assumes informative node descriptors for lexical search, and may underperform on graphs with limited text."
      }
    },
    {
      "@type": "HowTo",
      "@id": "howto1",
      "name": "How to perform adaptive knowledge graph retrieval with ARK",
      "description": "Step-by-step process for using ARK to retrieve evidence from knowledge graphs.",
      "step": [
        {
          "@type": "HowToStep",
          "position": 1,
          "name": "Issue Global Search",
          "text": "Start retrieval by issuing a global lexical search over node descriptors using a textual subquery."
        },
        {
          "@type": "HowToStep",
          "position": 2,
          "name": "Select Candidate Nodes",
          "text": "Select nodes retrieved by global search as candidates for further exploration."
        },
        {
          "@type": "HowToStep",
          "position": 3,
          "name": "Perform Neighborhood Exploration",
          "text": "Expand selected nodes by retrieving their one-hop neighbors filtered by node and edge types, optionally ranked by query relevance."
        },
        {
          "@type": "HowToStep",
          "position": 4,
          "name": "Alternate Between Tools",
          "text": "Alternate between global search and neighborhood exploration based on query needs and retrieved evidence."
        },
        {
          "@type": "HowToStep",
          "position": 5,
          "name": "Terminate Retrieval",
          "text": "Stop the retrieval process when sufficient evidence is found or a maximum step budget is reached."
        }
      ]
    },
    {
      "@type": "HowTo",
      "@id": "howto2",
      "name": "How to distill ARK into a smaller model",
      "description": "Process for training a compact student model to imitate ARK's retrieval behavior.",
      "step": [
        {
          "@type": "HowToStep",
          "position": 1,
          "name": "Generate Teacher Trajectories",
          "text": "Run the ARK teacher agent on training queries to collect tool-use trajectories including tool calls and observations."
        },
        {
          "@type": "HowToStep",
          "position": 2,
          "name": "Prepare Training Data",
          "text": "Mask user messages and tool outputs, keeping only assistant-generated tokens for next-token prediction."
        },
        {
          "@type": "HowToStep",
          "position": 3,
          "name": "Fine-tune Student Model",
          "text": "Train the student model with supervised fine-tuning on the collected trajectories using LoRA adapters."
        }
      ]
    },
    {
      "@type": "HowTo",
      "@id": "howto3",
      "name": "How to aggregate multiple agent retrieval outputs",
      "description": "Method to combine retrieval results from multiple parallel ARK agents.",
      "step": [
        {
          "@type": "HowToStep",
          "position": 1,
          "name": "Collect Ordered Lists",
          "text": "Obtain ordered lists of retrieved nodes from each independent agent."
        },
        {
          "@type": "HowToStep",
          "position": 2,
          "name": "Concatenate Lists",
          "text": "Concatenate the lists in agent order to form a combined list."
        },
        {
          "@type": "HowToStep",
          "position": 3,
          "name": "Rank by Frequency and Position",
          "text": "Rank nodes by decreasing frequency across lists, breaking ties by earliest occurrence."
        }
      ]
    }
  ],
  "publisher": {
    "@type": "Organization",
    "name": "Harvard Medical School",
    "location": {
      "@type": "Place",
      "address": {
        "@type": "PostalAddress",
        "addressLocality": "Boston",
        "addressRegion": "MA",
        "addressCountry": "USA"
      }
    }
  },
  "citation": [
    {
      "@type": "ScholarlyArticle",
      "name": "STaRK: benchmarking LLM retrieval on textual and relational knowledge bases",
      "author": "Wu et al.",
      "datePublished": "2024"
    },
    {
      "@type": "ScholarlyArticle",
      "name": "Retrieval-augmented generation for knowledge-intensive NLP tasks",
      "author": "Lewis et al.",
      "datePublished": "2020"
    },
    {
      "@type": "ScholarlyArticle",
      "name": "Distilling the Knowledge in a Neural Network",
      "author": "Hinton et al.",
      "datePublished": "2015"
    }
  ],
  "keywords": [
    "Knowledge Graph Retrieval",
    "Adaptive Retrieval",
    "Multi-hop Retrieval",
    "Global Lexical Search",
    "Neighborhood Exploration",
    "Large Language Models",
    "Retrieval-Augmented Generation",
    "Distillation",
    "STaRK Benchmark",
    "Label-free Imitation"
  ],
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "mainEntity": [
    "q1",
    "q2",
    "q3",
    "q4",
    "q5",
    "q6",
    "q7",
    "q8",
    "q9",
    "q10",
    "howto1",
    "howto2",
    "howto3",
    "ark-defined-terms"
  ]
}