Graph-Based Evidence Curation

• Graph-based evidence curation is a method that transforms scattered evidence into structured graphs, facilitating multi-hop reasoning and accurate verification.
• It employs precise node and edge construction with scoring and pruning algorithms to model relationships and reveal bridge evidence in complex data domains.
• Empirical evaluations demonstrate significant improvements in fact verification, biomedical knowledge extraction, and multi-hop question answering using this approach.

Graph-based evidence curation refers to the process of organizing, prioritizing, and synthesizing distributed pieces of evidence into a structured graph, facilitating robust reasoning, accurate prediction, and transparent verification in multi-hop question answering, fact verification, biomedical knowledge extraction, and other complex data domains. Unlike flat ranked lists produced by traditional retrieval pipelines, graph-based curation explicitly models relationships among evidence, enabling multi-step reasoning, bridge discovery, redundancy pruning, and chain-of-thought construction. This paradigm is essential for scalable, zero-shot inference and has demonstrated substantial empirical gains across diverse tasks.

1. Principles and Motivations

The central goal of graph-based evidence curation is to overcome the limitations of list-based retrieval, where individually relevant text fragments may miss critical interconnections, and retrieval noise can obscure reasoning chains. Graph representations encode evidence pieces as nodes and their lexical, semantic, or logical relationships as edges. Nodes often represent passages, serialized table rows, entities, functions, or claim-evidence pairs; edges model shared terms, predicate-argument relations, logical dependencies, or co-mention statistics (Sharafath et al., 10 Jan 2026). The weighted graph structure enables explicit detection of bridge documents that connect reasoning steps, a capability absent from flat ranking approaches.

Curation is not merely retrieval: it includes node and edge construction, scoring/pruning routines to remove redundancy or noise, and aggregation algorithms that synthesize connected subgraphs into human-readable narratives. Graph-based organization is indispensable for multi-hop reasoning, where chains of evidence must be assembled across heterogeneous and noisy sources (Sharafath et al., 10 Jan 2026).

2. Evidence Graph Construction Methodologies

Graph construction methodologies vary by task and domain. In table-text QA, nodes typically represent either passage or “Table:…|Row:…” units, edges are generated when documents share at least one lexical term, and weights are computed as TF-IDF overlap to reduce semantic drift (Sharafath et al., 10 Jan 2026). In logic-level fact verification, heterogeneous graphs contain entity nodes (table cells, statement spans) and function nodes (symbolic operators), with edges modeling either parent-child (program tree) connections or same-entity glue between programs (Shi et al., 2021).

Multi-evidence fact verification frameworks, such as GET, form graphs over tokenized claims and evidence snippets, merging identical word forms into single nodes and connecting tokens via co-occurrence windows or sliding deep attention (Xu et al., 2022). Biomedical knowledge graphs, e.g., SimpleGermKG, link extracted and normalized entities (genes/diseases) with article nodes via part-whole relations and attribute semantics (Gonzalez et al., 2023).

Construction is often preceded by entity extraction, normalization, and schema mapping (e.g., PICO for medical evidence), and may use deterministic, rule-based routines or ensemble relation extraction (triplet agreement among LLMs for clinical reasoning) (Mu et al., 15 Dec 2025). In all cases, design choices around node granularity, edge semantics, and weighting have significant downstream impact on reasoning and interpretability.

3. Scoring, Pruning, and Bridge Selection Algorithms

Effective graph-based curation hinges on principled node and edge ranking, pruning, and bridge selection mechanisms. The N2N-GQA framework introduces GraphRank, combining normalized semantic relevance (retrieval score from ColBERTv2) and weighted degree centrality, yielding a hybrid score for node selection: $$\text{Score}_{GR}(v) = S_{sem}(v) \times [1 + (1 - \alpha) S_{struct}(v)]$$ with $\alpha$ controlling the amplification effect (Sharafath et al., 10 Jan 2026). Top-$N$ nodes by Score$_{GR}$ are retained, substantially reducing context length while preserving reasoning chains.

Bridge-aware selectors further prioritize key connectors between evidence types (e.g., passages and tables), boosting scores of contextually overlapping nodes (Sharafath et al., 10 Jan 2026). In confidential GNNs (CO-GAT), node confidences $s_p$ are soft-gated to attenuate information flow from noisy evidence, using convex combinations with blank nodes to prevent spurious propagation (Lan et al., 2024).

Other frameworks employ redundancy-scoring subnetworks for node dropout (GET uses GGNN-based scores, discarding high-redundancy nodes iteratively) (Xu et al., 2022), or auxiliary sparsification via relevance prediction and residual aggregation (LERGV applies probabilistic node pruning after each GAT layer) (Shi et al., 2021). Uncertainty-aware fusion models such as EFGNN compute cumulative belief fusion analytically, guaranteeing monotonicity and uncertainty reduction in multi-hop evidence aggregation (Chen et al., 16 Jun 2025).

4. Graph-based Reasoning and Answer Synthesis

Once curated, evidence graphs can be exploited for structured reasoning and answer synthesis. Multi-hop QA frameworks decompose queries into sequential hops, retrieve and organize context per hop, and assemble final subgraphs that trace explicit reasoning paths. N2N-GQA's noise-to-narrative module presents the LLM with a pruned, highly connected sequence of evidence, interleaved with hop decompositions, resulting in concise, verifiable synthesis (Sharafath et al., 10 Jan 2026).

Graph neural networks, typically GAT or GGNN/GCN variants, enable information propagation across nodes, leveraging centrality, attention, or aggregation schemes. LERGV deploys multi-head GATs over logic graphs, using pooling and auxiliary textual features for entailment classification (Shi et al., 2021). GEAR employs message-passing over fully connected evidence graphs, followed by attention-weighted aggregation conditioned on claim embeddings (Zhou et al., 2019).

Critical Evidence Graphs (CEGs) in MedCEG represent clinically valid reasoning pathways, distilled via backward traversal and transitive reduction, and are used not only for answer synthesis but as direct reward signals for RL-enhanced model alignment, enforcing pathway integrity (Mu et al., 15 Dec 2025). In biomedical settings, knowledge graphs encode co-occurrence, part-whole, and direct relations, supporting provenance-traceable query responses (Gonzalez et al., 2023, Belluomo et al., 2024).

5. Empirical Evaluations and Performance Metrics

Graph-based evidence curation frameworks have consistently demonstrated substantial empirical gains. N2N-GQA yields a +19.9 absolute EM improvement over strong list-based retrieval baselines, essentially matching fine-tuned retrievers without task-specific training (48.8 EM vs. 49.0 EM for CORE, and approaching COS at 56.9 EM) on OTT-QA (Sharafath et al., 10 Jan 2026). Ablation studies confirm that graph curation and pruning contribute independently to performance.

In fact verification, GEAR achieves 67.10% FEVER score, outperforming BERT-only baselines especially on difficult claims requiring multi-evidence composition (Zhou et al., 2019). CO-GAT surpasses GAT-only models and is robust to under-masking, with entropy analyses revealing sharper attention on relevant nodes (Lan et al., 2024). MedCEG attains 58.6% in-distribution (vs. 48.3% for Huatuo-o1-8B) and 64.1% out-of-distribution reasoning, and improves process quality by 9.5% in five expert-aligned criteria (Mu et al., 15 Dec 2025).

Structural similarity metrics, node/edge precision/recall, coverage, faithfulness, semantic similarity, and specialized metrics (chain completeness, structural correctness, reasoning reward) are commonly reported (Zhang et al., 1 Jan 2026, Spranger et al., 2016, Mu et al., 15 Dec 2025). Limitations in evaluation often stem from dataset domain, annotation granularity, and normalization challenges.

6. Applications and Domain-specific Extensions

Graph-based curation supports a broad spectrum of applications:

• Hybrid table-text QA, multi-hop reasoning, and bridge document discovery (Sharafath et al., 10 Jan 2026).
• Logic program verification over semi-structured tables and symbolic reasoning (Shi et al., 2021).
• Clinical reasoning trace alignment and reinforcement learning via graph-based rewards (Mu et al., 15 Dec 2025).
• Biomedical literature mining, gene-disease curation, and provenance-traceable knowledge graph construction (Gonzalez et al., 2023, Belluomo et al., 2024).
• Robust fact verification, fake news detection, and semantic redundancy pruning (Xu et al., 2022, Zhou et al., 2019).
• Small LM distillation and hallucination mitigation via graph and evidence alignment (Chen et al., 2 Jun 2025).
• Large-scale literature graph search, citation network analysis, and subgraph-based evidence retrieval (Ammar et al., 2018).

Methodological extensions include adaptive node/edge weighting, learned and analytical fusion, combinatorial optimization for motif finding in precision medicine, and hierarchical graph compression for scalable retrieval (Belluomo et al., 2024, Zhang et al., 1 Jan 2026).

7. Limitations, Challenges, and Research Directions

Despite their strengths, graph-based curation frameworks face challenges including imperfect entity normalization, context-sensitive ambiguity, missing complex relations, and lack of global optimization in subgraph assembly (Spranger et al., 2016). Many biomedical graphs rely on co-occurrence at the article level rather than predicate-argument relations, impacting precision (Gonzalez et al., 2023). Robustness to unseen domains, dynamic composition, and scalable learning are ongoing research problems.

Future directions entail context-aware species normalization, complex formation detection, compositional graph algorithms for merging subgraphs, graph-theoretic reward design for RL integration, and end-to-end GNN embedding for predictive link discovery. Evaluating against held-out gold standards, implementing motif-based reasoning, and extending frameworks to policy analysis, law, and scientific argumentation remain active avenues (Sharafath et al., 10 Jan 2026, Mu et al., 15 Dec 2025, Belluomo et al., 2024).

---

Graph-based evidence curation transforms noisy, list-based retrieval into structured, high-connectivity graphs that enable explicit, multi-hop reasoning and process-transparent answer synthesis. This paradigm, underlying state-of-the-art QA, fact verification, and biomedical knowledge extraction, is essential for scalable, interpretable, and robust inference across diverse domains.