---
kind: concept
slug: knowledge-graph-reasoning
title: Knowledge graph reasoning
summary: >-
  Knowledge graph reasoning infers a relation between two unlinked entities. GraIL scores unseen entities from subgraph shape (author-reported best inductive AUC-PR on 12 splits). CronKGQA answers when-questions over timed facts (author-reported Hits@1 0.647, but 0.392 on complex questions). A 2020 multi-hop QA baseline scores 5.41 joint F1, so pick a method by task.
tags: [knowledge-graphs, reasoning, link-prediction, question-answering, temporal-reasoning]
papers: ["1911.06962", "2011.01060", "2104.06378", "2106.01515", "2303.13948", "2401.06853", "2412.10390", "2505.15062"]
related: [graph-neural-network, knowledge-graph-completion, knowledge-graph-embedding, knowledge-base-question-answering, multi-hop-reasoning, commonsense-knowledge-injection, dataset-curation, neural-symbolic-integration, reinforcement-learning-verifiable-rewards, tool-use, llm-agent]
status: draft
updated_at: 2026-11-01
---

## Overview

Knowledge graph reasoning asks whether a relation r holds between two entities
h and t that are both in the graph but not connected by an edge [arxiv:2303.13948 §5.5].
The survey names two task types: single-hop prediction, where one element of a
triple is predicted from the other two, and multi-hop reasoning over logical
queries [arxiv:2303.13948 §5.5]. It treats reasoning as the last of five
technical challenges for knowledge graphs, after embedding, acquisition,
completion and fusion [arxiv:2303.13948 §2.2]. A benchmark paper gives a
concrete multi-hop case: each question comes with the Wikidata triples that
lead to its answer [arxiv:2011.01060 §2.1]. A 2024 survey sorts the same field
by query type instead of by method family: one missing fact, a logical
combination of conditions, a question in natural language, or a multi-turn
conversation [arxiv:2412.10390 §1] [§2.8]. The variants below use both framings.

## Variants

The survey groups reasoning methods into three families: logic rule-based,
distributed-representation-based and neural-network-based [arxiv:2303.13948 §5.5].
It does not describe any single method in enough detail to implement, and it
gives no hyperparameters for any of them [arxiv:2303.13948 §5.5]. The survey
card [Knowledge Graphs: Opportunities and Challenges](/papers/2303.13948) is the
source for the taxonomy. The benchmark card
[2WikiMultiHopQA](/papers/2011.01060) is the source for the rule-based question generator and the evaluation numbers.
Four method cards add concrete methods that the survey does not describe in
enough detail to implement: [GraIL](/papers/1911.06962) for inductive link
prediction, [QA-GNN](/papers/2104.06378) for question answering,
[CronKGQA](/papers/2106.01515) for questions about when a fact held, and
[TG-LLM](/papers/2401.06853) for reasoning over a timeline written from text.
The 2024 survey [Neural-Symbolic Reasoning over Knowledge Graphs](/papers/2412.10390)
is the source for the complex-query and natural-language families below. It
proposes no method and reports no experiments [arxiv:2412.10390 §1].
[SAKE](/papers/2505.15062) adds a different kind of method: a language model
trained with reinforcement learning to call graph lookup tools while it answers.

### Logic rule-based

Reasoning from rules over the graph. The survey lists this family without
describing a rule language, a rule-mining procedure or a result
[arxiv:2303.13948 §5.5]. The benchmark paper shows one concrete use of rules:
its inference questions are built from 28 logical rules over Wikidata
relations, checked by hand from rules mined with AMIE [arxiv:2011.01060 §3.2]
[app. A.3]. Only relations with a single value per entity are used, so each
question has one answer [arxiv:2011.01060 app. A.3]. The paper uses the rules to
generate questions and does not evaluate rule-based reasoning as a method. It
also notes that several rules can be false in some cases, and checks each new
triple against Wikidata before use [arxiv:2011.01060 app. A.3].

### Distributed-representation-based

Entities and relations are embedded as vectors and a score for each candidate
triple is computed from them [arxiv:2303.13948 §5.1]. The survey's scoring
functions for this family are the ones listed under knowledge graph embeddings
[arxiv:2303.13948 §5.1]. Single-hop prediction in this family is the same
operation as completion: rank candidates for the missing element of a triple
[arxiv:2303.13948 §5.3] [arxiv:2303.13948 §5.5].

### Neural-network-based

Reasoning with a neural network over the graph structure. The survey lists this
family for reasoning [arxiv:2303.13948 §5.5] but describes its models, R-GCN and
ConvKB among them, under knowledge graph embeddings [arxiv:2303.13948 §5.1.3].
Their reported link prediction results on FB15K and WN18RR are in
[arxiv:2303.13948 tab. 3]. The R-GCN equation printed in the survey does not
show a relation-dependent weight, so check the original paper before
implementing [eq. 4].

### Inductive subgraph scoring (GraIL)

Scores a candidate triple (u, r, v) from the subgraph around u and v. It uses
no entity embeddings and no node features, so it can score entities that were
not in training [arxiv:1911.06962 §3.1]. The method assumes the local
neighbourhood of the candidate contains the evidence needed to decide it
[arxiv:1911.06962 §3.1]. Each node is labelled with its distances to u and to v,
and a relation-aware GNN with edge attention scores the subgraph
[arxiv:1911.06962 §3.1]. The inductive splits were built for this test: the
training and test graphs share the relation set but no entities
[arxiv:1911.06962 §4.1] [app. F]. Author-reported: best inductive AUC-PR in all
12 columns against RuleN, Neural-LP and DRUM [tab. 1], and best Hits@10 in 10
of 12 [tab. 2]. Hits@10 ranks each positive against 50 sampled negatives
[arxiv:1911.06962 §4.1], so the number depends on the sampler. Using the union
of the two neighbourhoods instead of their intersection lowers FB15k-237 v3 from
91.68 to 84.25 AUC-PR [arxiv:1911.06962 tab. 7]. Most of the cost is in
extracting subgraphs, not in the network [arxiv:1911.06962 §3.4]. In the
transductive setting, ensembles with embedding models give small gains on two of
three datasets [arxiv:1911.06962 §4.2] [tab. 3] [tab. 4] [tab. 5]; the paper
reports no significance tests [arxiv:1911.06962 §4.2].

### Multi-hop reasoning over paths

Answers that need a chain of facts, for example a question that links two
hops through an intermediate entity. The survey cites multi-hop relation
reasoning as the knowledge-graph technique behind two question-answering
systems, MHPGM and EmbedKGQA [arxiv:2303.13948 tab. 1]. In recommenders, path
generation from user-item interactions is listed for KPRN [arxiv:2303.13948 tab. 1].
The survey reports that most multi-hop reasoning work runs on graphs of about
63K entities and 592K relations [arxiv:2303.13948 §5.5], which is small next to
the large academic and industrial graphs it lists [arxiv:2303.13948 §4.2].

### Evaluating multi-hop chains (2WikiMultiHopQA)

The benchmark has about 193k English questions in four types: comparison,
inference, compositional and bridge-comparison [arxiv:2011.01060 §2.2]
[tab. 2]. Inference questions follow the logic-rule pattern above; the
inference type has 7,478 examples [arxiv:2011.01060 tab. 2]. Each sample has
a question, a context of 10 paragraphs, supporting sentences, an answer and
evidence triples [arxiv:2011.01060 §2.1] [§3.2].

The benchmark scores the answer, the supporting sentences and the evidence
triples, and a joint score that requires all three [arxiv:2011.01060 §2.1]
[eq. 1]. Its baseline is a HotpotQA-style reading model with one added block that
predicts evidence triples [arxiv:2011.01060 §5.2]. On the test split it reaches 43.93 answer
F1, 65.26 supporting-fact F1, 14.94 evidence F1 and 5.41 joint F1
[arxiv:2011.01060 tab. 5]. On 100 test samples, human annotators reach 75.63
evidence F1 on average, against 16.74 for the model [arxiv:2011.01060 tab. 7].
The same multi-hop model scores lower on this benchmark than on HotpotQA:
26.76 joint F1 against 40.52 [arxiv:2011.01060 tab. 4].

The evidence scores show that a correct answer does not mean the chain of
facts was found. On the test split the baseline's evidence F1 is 2.85 for
inference questions and 17.65 for compositional ones [arxiv:2011.01060 tab. 6].

### Language model with a knowledge-graph subgraph (QA-GNN)

For multiple-choice questions, a pretrained language model (RoBERTa-large, or
SapBERT for medical questions) is combined with a ConceptNet subgraph. The
subgraph holds the question's topic entities and their 2-hop neighbours, cut to
the top 200 nodes by relevance [arxiv:2104.06378 §3] [§4.2]. The language model
scores how relevant each node is to the question and answer text
[arxiv:2104.06378 §3.2] [eq. 1]. The question and answer are added as one extra
node, linked to the question and answer entities, and five attention GNN layers
pass messages over the graph [arxiv:2104.06378 §3.1] [§3.3]. Author-reported:
73.41% on the CommonsenseQA in-house test, against 68.69% for RoBERTa-large alone
and 71.11% for MHGRN [arxiv:2104.06378 tab. 3]. On OpenBookQA test with
RoBERTa-large it reports 67.80% against 64.80% [arxiv:2104.06378 tab. 5]. On
CommonsenseQA questions with negation it reports 58.8% against 54.2% for RoBERTa
[arxiv:2104.06378 tab. 10]. The cost is about 20 hours on two RTX 2080 Ti for
CommonsenseQA, with a model of about 360M parameters [arxiv:2104.06378 §4.3]
[§4.5]. The released code computes the relevance scores differently from eq. 1
and uses other settings the paper leaves open, so reproducing the numbers needs
the card's [Paper vs code](/papers/2104.06378) section.

### Temporal question answering over a knowledge graph (CronKGQA)

Adds the time at which a fact held. A pretrained language model encodes the
question, and two projections of that encoding give an entity query and a time
query [arxiv:2106.01515 §5] [fig. 1]. Every entity is scored from the question's
subject and time, and every year from the subject and object, with a temporal
embedding model (TComplEx) [arxiv:2106.01515 §4.2] [eq. 4] [eq. 5]. The
entities and times in the question must be given as input; entity linking is
assumed to be solved [arxiv:2106.01515 §3.2]. The authors built CronQuestions,
about 410k question-answer pairs from templates and paraphrases over a timed
Wikidata graph, split 350,000 / 30,000 / 30,000 [arxiv:2106.01515 §3.2] [tab. 4].
Author-reported: Hits@1 0.647 overall, against 0.288 for the best baseline
[arxiv:2106.01515 tab. 5]. Simple questions reach 0.987 and complex ones 0.392
[tab. 5]. For time answers, TComplEx embeddings give 0.549 Hits@1 and ComplEx
embeddings give 0.057 [arxiv:2106.01515 tab. 7]. The paper does not say which
split Table 5 reports [arxiv:2106.01515 tab. 5].

### Temporal graph reasoning with an LLM (TG-LLM)

The graph here is not a knowledge graph. A fine-tuned Llama 2 13B first writes
an ordered timeline of the events in a story, with the start and end of each
event as separate entries [arxiv:2401.06853 §3.1]. A second adapter reads that
timeline, some computed time gaps and the question, then writes a chain of
thought and an answer [arxiv:2401.06853 §3] [fig. 1]. The reasoning adapter is
trained on chains of thought chosen by a contrastive score, and on graph
perturbations: renamed entities, shifted years, dropped irrelevant facts and
swapped relation names [arxiv:2401.06853 §3.2.1] [§3.2.2]. The training data,
TGQA, is synthetic: stories written by GPT-3.5 from subgraphs of YAGO11k
[arxiv:2401.06853 §2]. Author-reported: on TGQA, exact match rises from 0.646
for the chain-of-thought baseline to 0.797 for the full model [arxiv:2401.06853
tab. 5], against 0.821 for GPT-4 with chain of thought and 0.772 for GPT-4
with a standard prompt [tab. 4]. Human raters find fewer temporal-graph errors
after training: 0.05 for the full model against 0.10 for in-context learning
[arxiv:2401.06853 tab. 3].

### Complex logical queries (tree-form and EFO-1)

Answers a query that combines conditions with conjunction, disjunction, negation
and existential variables, not a single fact or a chain of lookups. The 2024
survey splits these queries into two families, tree-form and EFO-1, and notes
that neither family contains the other [arxiv:2412.10390 §4.1]. Tree-form
queries are built from a projection, complement, intersection and union, and
their neural solvers embed each term and implement the operators as differentiable
functions [arxiv:2412.10390 §4.2.1]. Boxes make intersection a box intersection,
while union needs a rewrite into disjunctive normal form [arxiv:2412.10390 §4.2.1].
EFO-1 queries are solved as a graph, for example by message passing over the
query graph [arxiv:2412.10390 §4.2.2].

Neural-symbolic solvers search for variable assignments that satisfy the query.
A link predictor turns each triple into a fuzzy truth value between 0 and 1, and
the search runs as beam search on acyclic query graphs, backtracking on acyclic
queries, or a method that also covers cyclic graphs [arxiv:2412.10390 §4.3]. Some
of these solvers relax the search into embedding space for gradient-based
optimisation, and one learns its link predictor from search feedback
[arxiv:2412.10390 §4.3]. The survey gives no results for any of them
[arxiv:2412.10390 §1].

### Natural-language questions and LLM combinations

Answers a question written in English by retrieving a subgraph around candidate
answers, ranking answers by embedding similarity, or ranking candidate query
graphs and generating the correct one [arxiv:2412.10390 §5.1]. In multi-turn
dialogue, the user's reformulations serve as feedback for reinforcement-learning
agents that walk the graph [arxiv:2412.10390 §5.2]. The survey's question-answering
side is covered on the [knowledge base question answering](/concepts/knowledge-base-question-answering)
page.

For language models, the survey lists three ways to combine them with a graph.
The graph can supply facts to the model, as in QA-GNN, GreaseLM and retrieval
with a knowledge graph [arxiv:2412.10390 §6.1]. The model can walk the graph from
topic entities, scoring neighbours and moving on when the current node lacks the
answer [arxiv:2412.10390 §6.2]. Or the two can be trained together with a shared
objective [arxiv:2412.10390 §6.3]. The survey gives one example for each and does
not compare them [arxiv:2412.10390 §6].

### Agentic reinforcement learning with graph tools (SAKE)

The language model is the reasoner, and the graph is reached through two
deterministic tools rather than a learned scorer. In a three-turn rollout the
model names entities in the question, the first tool returns for each entity
a group of the entity and its most similar graph concepts, and the model keeps
the groups it needs. The second tool returns the graph triplets that link two
kept groups, and the model then reasons by analogy over those triplets and
writes an answer [arxiv:2505.15062 §3.1] [§3.2] [eq. 6] [eq. 7]. The tools have
no trainable parameters, and tool output is excluded from the training loss
[arxiv:2505.15062 §3.1] [alg. 1].

The policy is trained end-to-end with GRPO from an instruction-tuned base model,
with no supervised stage [arxiv:2505.15062 §3.4] [§3.5]. The reward adds a
format term, which checks that the closing tags are present, to an exact-match
term on the answer [arxiv:2505.15062 §3.5] [eq. 9] [eq. 10] [eq. 11]. The
paper does not state the clipping range or the KL weight [arxiv:2505.15062 §3.4]
[eq. 8]. Author-reported: a Qwen2.5-7B model reaches a weighted average of 75.4
on UMLS question answering, against 70.1 for GPT-3.5-Turbo with GIVE, and 81.3
on CommonsenseQA [arxiv:2505.15062 tab. 1]. The paper's token counts per
question are 639 to 1,156 for SAKE, against 7,010 to 19,461 for ToG and GIVE
[arxiv:2505.15062 tab. 2]. The UMLS knowledge graph has 135 nodes and 5,877
triplets, which is small next to the ConceptNet graph used for CommonsenseQA
[arxiv:2505.15062 §4.1]. The UMLS test set is 224 questions, so one question
moves the weighted average by about 0.45 points [arxiv:2505.15062 §4.1]
[tab. 1]. The released code differs from the paper in several places, including
the group construction and the tool-call trigger; see the card's
[Paper vs code](/papers/2505.15062) section before reproducing numbers.

## When to use which

The survey does not compare reasoning methods on a common benchmark, and it
gives no latency, memory or data-volume figures for reasoning. The table
therefore chooses by task. Its entries are what the cited sources state, not
results measured here.

| Situation | Use | Why |
|---|---|---|
| One fact is missing and the other two elements are known | Single-hop prediction, typically a distributed-representation scorer | The survey defines this as predicting one element from the other two [arxiv:2303.13948 §5.5]. |
| Answer requires chaining several facts (multi-hop question answering) | Multi-hop reasoning over paths, e.g. MHPGM or EmbedKGQA | The survey lists multi-hop relation reasoning and KG-embedding multi-hop QA as the technique for this case [arxiv:2303.13948 tab. 1]. |
| Need to test whether a system finds the chain of facts, not only the answer | Evaluate with evidence triples, e.g. 2WikiMultiHopQA | The benchmark gives evidence triples per question and reports a separate evidence score [arxiv:2011.01060 §2.1] [tab. 5]. |
| Need a dataset of multi-hop questions with gold reasoning paths | 2WikiMultiHopQA, after checking the released files | The counts in the paper's tables are from its arXiv v2 and were not checked against the current release [arxiv:2011.01060 tab. 1] [tab. 2]. |
| Recommendations should come with an explanation | Path-based reasoning, e.g. KPRN | The survey claims explainable recommendation paths as an advantage, but does not measure it [arxiv:2303.13948 §3.1.2]. |
| Need a rule you can read and audit | Logic rule-based reasoning | The survey names the family [arxiv:2303.13948 §5.5]. The benchmark's rule list [arxiv:2011.01060 tab. 10] is one concrete rule set, but the paper does not evaluate it as a reasoner. |
| Test entities may be absent from training, and retraining per entity is not an option | Inductive subgraph scoring, e.g. GraIL | It uses no entity embeddings, and its inductive splits are built for unseen entities [arxiv:1911.06962 §3.1] [§4.1]. |
| Transductive link prediction, all entities known, and an ensemble is acceptable | Embedding scorer plus GraIL in an ensemble | The ensemble gains are small and on two of three datasets [arxiv:1911.06962 §4.2] [tab. 3] [tab. 4] [tab. 5]. |
| Multiple-choice commonsense QA with a knowledge graph and a pretrained LM | LM plus subgraph GNN, e.g. QA-GNN | Gains over the LM alone are reported on CommonsenseQA and OpenBookQA [arxiv:2104.06378 tab. 3] [tab. 5]. Training cost is about 20 hours on two RTX 2080 Ti for CommonsenseQA [arxiv:2104.06378 §4.3]. |
| Questions ask when a fact held, over a graph with timed facts | CronKGQA with temporal (TComplEx) embeddings | Time-answer Hits@1 is 0.549 with TComplEx and 0.057 with ComplEx on the same model [arxiv:2106.01515 tab. 7]. |
| Complex temporal questions (before/after, first/last, time join) | No cited method is strong on them yet; CronKGQA is the best reported, with 0.392 Hits@1 on complex questions | The authors name complex questions as the weak point [arxiv:2106.01515 §6.3] [tab. 5]. |
| Ordering and duration questions about a story, with no timeline annotated | TG-LLM-style two-step pipeline | The authors report 0.797 EM on synthetic TGQA [arxiv:2401.06853 tab. 4]. Building the data needs GPT-3.5 and GPT-4 [arxiv:2401.06853 §4.2], and the data is described as for research use only [arxiv:2401.06853 Ethics Statement]. |
| Question combines several conditions with AND, OR or NOT over the graph | Tree-form query embedding, e.g. Query2Box, or a neural-symbolic solver, e.g. CQD | Both are described for complex logical queries; the survey gives no results to compare them [arxiv:2412.10390 §4.1] [§4.2.1] [§4.3]. |
| Question is in English and the answer is found by searching the graph | Single-turn retrieval or path search, e.g. PullNet or an RL walker | The survey lists these for single-turn questions [arxiv:2412.10390 §5.1]. Multi-turn follow-ups are covered in [§5.2]. |
| A language model must answer with facts from a knowledge graph | Give the graph to the model as input (QA-GNN, GreaseLM), or let the model walk the graph (Think-on-Graph) | Both are listed as ways to combine the two [arxiv:2412.10390 §6.1] [§6.2]. The survey does not compare them [§6]. |
| Small biomedical or domain graph, a question needs facts that the model lacks, and a token budget per question | Agentic RL with graph tools, e.g. SAKE, if you can train a 3B or 7B model | The authors report 639 to 1,156 tokens per question against several thousand for ToG and GIVE, on a 135-node UMLS graph [arxiv:2505.15062 tab. 2] [§4.1]. Training used two A6000 GPUs [arxiv:2505.15062 app. B]. |
| Choosing a method by reported accuracy | Not possible from these cards | The survey's results table mixes datasets and metrics and is not a fair comparison [arxiv:2303.13948 tab. 3]. |

## Implementation sketch

The survey gives no algorithm for reasoning. The first sketch covers only the
single-hop case, which the survey defines as ranking candidates for a missing
element [arxiv:2303.13948 §5.5]. Scoring is left as a parameter because the
survey does not fix one scoring function for this use [arxiv:2303.13948 §5.1].
The second sketch is the input step of the inductive variant. It builds the
labelled subgraph for one candidate; the scoring GNN is in the GraIL card.
Multi-hop reasoning is not specified here.

```pseudocode
# Rank every entity as the tail of the query (head, relation, ?)
def rank_tails(head, relation, entities, score):
    # score(h, r, t): plausibility of the triple, from a trained model
    scored = [(score(head, relation, t), t) for t in entities]
    # highest plausibility first
    return [t for _, t in sorted(scored, reverse=True)]
```

```pseudocode
# Labelled enclosing subgraph for a candidate (u, r, v)  [arxiv:1911.06962 §3.1]
def enclosing_subgraph(triples, u, v, k):
    # k-hop undirected neighbourhood of each target  [§3.1]
    near_u = within_hops(triples, u, k)
    near_v = within_hops(triples, v, k)
    # keep nodes close to both targets (intersection)  [§3.1]
    nodes = (near_u & near_v) | {u, v}
    # the two targets have fixed labels  [§3.1]
    labels = {u: (0, 1), v: (1, 0)}
    for n in nodes - {u, v}:
        # distance to each target, not passing through the other  [§3.1]
        d_u = dist_avoiding(triples, n, u, avoid=v)
        d_v = dist_avoiding(triples, n, v, avoid=u)
        # drop nodes farther than k from either target  [§3.1]
        if d_u <= k and d_v <= k:
            labels[n] = (d_u, d_v)
    return labels
```

## Contradictions

None known between papers. The cards measure different things on different
benchmarks: inductive splits [arxiv:1911.06962 tab. 1], CommonsenseQA and
OpenBookQA [arxiv:2104.06378 tab. 3], and 2WikiMultiHopQA [arxiv:2011.01060
tab. 5]. Add CronQuestions [arxiv:2106.01515 tab. 5] and TGQA [arxiv:2401.06853 tab. 4]
to that list. Their numbers are not comparable. The survey's advantage claims for
knowledge-graph systems are argued qualitatively and not measured
[arxiv:2303.13948 §3]. The benchmark measures the evidence chain of one text
baseline at 14.94 evidence F1 [arxiv:2011.01060 tab. 5], which is a measurement
of one system, not of knowledge-graph explanations in general. The 2024 survey reports no
numbers, so it adds no entries to these comparisons [arxiv:2412.10390 §1].

SAKE's results are on UMLS question answering [arxiv:2505.15062 tab. 1] and on a
246-question CommonsenseQA subset over ConceptNet [arxiv:2505.15062 tab. 3].
Its CommonsenseQA score of 81.3 and QA-GNN's 73.41 are from different test sets
and backbones, so they are not comparable [arxiv:2104.06378 tab. 3]. SAKE's
numbers are also not comparable with the inductive scores above.

Two papers disagree with their own tables. TG-LLM's text says the 13B model
is comparable to or better than GPT-4 on all datasets [arxiv:2401.06853 §4.3].
Its Table 4 shows GPT-4 with a standard prompt ahead on TimeQA easy and on both
TempReason levels, and GPT-4 with chain of thought ahead on TGQA [arxiv:2401.06853
tab. 4]. CronKGQA's text says every model except EmbedKGQA does best on time-join
questions [arxiv:2106.01515 §6.3], but its Table 6 shows the two
Entities-as-Experts variants scoring higher on before/after and first/last
questions than on time join [arxiv:2106.01515 tab. 6].

SAKE's text says it improves every benchmark on both model sizes without
exception [arxiv:2505.15062 §4.1]. Its Table 1 shows 80.0 on ProcessBank for
SAKE at 7B, against 82.5 for GraphRAG and ToG [arxiv:2505.15062 tab. 1].

## Papers

- [Inductive Relation Prediction by Subgraph Reasoning](/papers/1911.06962): GraIL, a 2019 inductive link predictor that scores a candidate from its labelled enclosing subgraph; the source for the inductive variant and its numbers.
- [QA-GNN: Reasoning with Language Models and Knowledge Graphs for Question Answering](/papers/2104.06378): a 2021 method that joins a pretrained LM with a ConceptNet subgraph for multiple-choice QA; the source for the LM-plus-subgraph variant.
- [2WikiMultiHopQA](/papers/2011.01060): a 2020 multi-hop QA benchmark with evidence triples from Wikidata; the source for the rule-based question generator and the evaluation numbers on this page.
- [Knowledge Graphs: Opportunities and Challenges](/papers/2303.13948): a 2023 survey whose §5.5 gives the reasoning taxonomy and task types used on this page.
- [Question Answering Over Temporal Knowledge Graphs](/papers/2106.01515): a 2021 method and dataset for questions about when facts held; the source for the CronKGQA variant.
- [Large Language Models Can Learn Temporal Reasoning](/papers/2401.06853): a 2024 two-step LoRA pipeline that writes a timeline from a story and reasons over it; the source for the TG-LLM variant.
- [Neural-Symbolic Reasoning over Knowledge Graphs: A Survey from a Query Perspective](/papers/2412.10390): a 2024 survey that organises reasoning by query type and has no experiments; the source for the complex-query and natural-language sections.
- [SAKE: Structured Agentic Knowledge Extrapolation for Complex LLM Reasoning via Reinforcement Learning](/papers/2505.15062): a 2025 method that trains a Qwen2.5 model with GRPO to call two knowledge-graph tools; the source for the agentic variant and its numbers.