Self-Reflective Planning with Knowledge Graphs: Enhancing LLM Reasoning Reliability for Question Answering

Jiajun Zhu; Kai Zhang; Meikai Bao; Qi Liu; Yanghai Zhang; Ye Liu

REVIEW 5 major objections 6 minor 14 references

Self-Reflective Planning with Knowledge Graphs: Enhancing LLM Reasoning Reliability for Question Answering

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that letting an LLM judge and edit its own knowledge-graph reasoning paths until the answer is found makes question answering more accurate and more factually grounded.

desk verdict A plausible incremental KGQA pipeline whose accuracy gains are credible, but the reliability headline rests on an unmeasured stopping rule and a vague grounding metric. read the letter →

arxiv 2505.19410 v1 pith:F4APEI3P submitted 2025-05-26 cs.CL

Jiajun Zhu , Ye Liu , Meikai Bao , Kai Zhang , Yanghai Zhang , Qi Liu This is my paper

classification cs.CL

keywords knowledgegraphquestionansweringself-reflectionself-reflectiveplanningLLMreasoningpathmulti-hopretrieval-augmentedgenerationself-correction

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Large language models asked to answer questions over a knowledge graph often generate reasoning paths that look plausible but miss the relation the question actually needs. This paper claims that adding a self-reflection loop—where the model judges the triples it retrieved, prunes them, and edits its path when the answer is absent—makes the reasoning reliable enough to beat strong baselines on three benchmark datasets. The proposed framework, Self-Reflective Planning (SRP), also checks the topic entity's one-hop relations before planning and pulls in similar solved examples as references. If the claim holds, a training-free prompting method can rival or exceed fine-tuned KGQA systems while keeping answers traceable to graph triples.

What carries the argument

The load-bearing mechanism is the reflection loop: a sequence judge that inspects the retrieved triplet sequence, prunes irrelevant triples, and emits a 'have answer' or 'no answer' verdict, paired with a path editor that rewrites the reasoning path when the verdict is negative. The loop is seeded by two planning aids: a reference search that retrieves similar solved questions with their reasoning paths, and a relation check that scores the topic entity's one-hop relations so the initial path starts from relations that actually exist in the graph. The loop stops only when the judge says the answer is present, and then the model answers from the pruned sequence.

What would settle it

Measure the sequence judge's precision on the three datasets: run SRP on a sample of questions, record each 'have answer' verdict, and check whether the pruned triples actually contain the gold answer. If a substantial share of 'have answer' stops occur on wrong or incomplete triples, the reliability gain over Readi would shrink or disappear when the loop is forced to continue.

Watch

Extended reading notes

Core claim

SRP's central claim is that reliability in KGQA comes from treating the reasoning path as editable state, not a one-shot prediction. The model first retrieves reference cases, scores the 1-hop relations of the topic entity to seed the path, generates a path, and instantiates it into triplet sequences via semantic retrieval. A sequence judge then decides whether the answer is present; if not, a path editor replaces the offending relation using candidate relations from the graph, and the loop repeats. On WebQSP and CWQ, SRP with GPT-4.1-mini reports Hits@1 of 83.6 and 69.0, respectively, surpassing the Readi baseline by 2.7 and 8.8 points, and on GrailQA it reports 78.8 overall, second only to ToG-GPT4 among the compared methods. The paper also reports higher searching success rates and reliable answering rates than Readi, arguing that SRP's correct answers are more often grounded in actual KG triples.

Load-bearing premise

The whole method depends on the model's own 'have answer' verdict being reliable enough to stop the reflection loop; the paper does not measure how often the judge stops early on a path that lacks the true answer.

Editorial extensions

If this is right

If SRP's reported numbers hold, a prompting-only LLM can beat fine-tuned KGQA models such as TIARA and Flexkbqa on WebQSP and CWQ without retraining.
The relation check and the reference search are load-bearing: removing them drops WebQSP accuracy from 78.5 to 76.9 and 75.7, respectively.
The reflection loop is especially valuable on complex multi-hop questions: on CWQ, removing self-reflection lowers Hits@1 from 58.7 to 55.5 for GPT-3.5.
SRP's reliable answering rate on WebQSP reaches 92.2% with GPT-4.1-mini, versus 70.5% for Readi, implying correct answers are more frequently supported by graph triples.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

An extension the paper does not run is to measure the sequence judge's precision and recall directly; if the judge often says 'have answer' before the true answer is present, the loop's stopping rule is the weakest point.
The same judge-edit loop could transfer to other structured reasoning environments such as SQL databases or code execution, where a path is a query or a program and the judge checks intermediate results.
Because the reference base is drawn from each dataset's training set, domain-shift robustness is an open question; on out-of-distribution questions the benefit of reference search may shrink or require adaptive reference selection.
SRP's own stated limitation—higher LLM call counts from iterative reflection—points to a testable cost-benefit question: whether the accuracy gain over Readi persists when measured per API call or per unit of latency.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

First thing you should know: the accuracy numbers are probably real, but the 'reliability' claim is softer than it looks. The SRP pipeline—KNN reference retrieval, relation check, iterative path reflection—does beat Readi on WebQSP and CWQ with both GPT-3.5 and GPT-4.1-mini, and the ablations show each piece pulls its weight. That is a solid incremental contribution.

What's new is the specific combination. None of the components is novel by itself: reference retrieval resembles in-context learning, relation check is reminiscent of ToG's relation pruning, and self-reflection is Readi/Reflexion territory. But the paper shows the combination works across two backbones and three datasets, and the code is available. The case studies are actually informative.

The soft spots are the ones the reader flagged. The sequence judge in Section 4.5 is load-bearing: if it says 'have answer' when it shouldn't, reflection stops and the model answers from a pruned sequence that lacks the answer. The paper never measures judge precision or recall, doesn't specify a maximum number of reflection rounds, and the answering prompt in Table 10 explicitly lets the model fall back on its own parametric knowledge. That makes the 'reliable answering rate' in Section 5.4 hard to interpret—it is never precisely defined, and it could count answers that came from the model's memory rather than from the triples. The ablation shows removing self-reflection costs 1.6–3.2 points, so the mechanism matters; without judge-reliability data, the reliability advantage over Readi is not fully established.

Other issues are minor but worth noting: no error bars or repeated runs, GrailQA uses ToG's test subset while some baselines come from earlier papers, and there's no same-backbone ToG comparison with GPT-4.1-mini. These don't sink the paper; they just bound how strongly you can read the comparisons.

Overall, this is a competent empirical paper with a modest but useful contribution. I'd send it to review, but ask for three things: a definition and measurement of the judge's accuracy, a clearer grounding metric that separates triple-grounded answers from parametric ones, and repeated runs or confidence intervals. If those come back okay, the reliability claim will actually be supported.

Referee Report

5 major / 6 minor

Summary. The paper proposes Self-Reflective Planning (SRP), a prompting pipeline for KGQA over Freebase. Given a question and a topic entity, SRP retrieves similar reference cases consisting of questions, reasoning paths, and answers from the training set; checks the candidate 1-hop relations of the topic entity; generates an initial multi-hop reasoning path; instantiates it as a triplet sequence via BM25/Contriever retrieval; and then iteratively judges and edits the path until the sequence judge emits a "have answer" verdict, after which a chain-of-thought answer is generated from the retained triples. Experiments on WebQSP, CWQ, and a GrailQA subset compare SRP with fine-tuned, LLM-only, and prompting baselines; the paper reports state-of-the-art Hits@1 on WebQSP and CWQ, second-best overall on GrailQA, and higher searching-success and reliable-answering rates than Readi. Ablation studies show that removing relation check, self-reflection, or references lowers accuracy, and a case study illustrates path editing behavior.

Significance. If the reported results are robust, SRP is a useful empirical contribution to KGQA prompting. The relation-check step and the reflection loop are clearly motivated, and the ablation in Table 3 shows that each component contributes non-trivially, so the framework is not merely a wrapper around the baseline. Comparing Readi on the same GPT-3.5 and GPT-4.1-mini backbones is a fairer comparison than many prior works. The paper is empirical rather than theoretical: it provides no machine-checked proofs and the code is only an anonymous repository link, so the value rests on the reliability of the measured gains and on the operational definitions of the proposed reliability metrics. The central claims are plausible but currently rest on several unquantified design choices, most notably the sequence judge's stopping rule and the definition of the reliable answering rate.

major comments (5)

[§5.4, Figure 3] The "reliable answering rate" is not defined precisely enough to support the reliability claim. The text says it is "the proportion of correct answers supported by factual triples from the KG," but it does not specify how "supported" was determined, who annotated it, or how the values in Figure 3 were computed. This matters because the answering prompt in Appendix C (Table 10) explicitly instructs the model to use its own knowledge when the triples are insufficient; a correct answer can therefore be produced without any grounding in the retrieved triples. Please provide an operational definition, an annotation procedure, and preferably per-case examples and error analysis showing that the reported reliable answering rate excludes answers generated from parametric memory.
[§4.5, Reflection and Reasoning] The sequence judge's "have answer" verdict stops reflection with no recovery path, and the accuracy of this judge is never measured. A false "have answer" on a pruned sequence that lacks the true answer terminates the loop prematurely and can lead to a wrong or incomplete final answer. The Table 3 ablation shows that removing self-reflection costs about 1.6–3.2 points on WebQSP and CWQ and 2.5 points on GrailQA, so the mechanism is material to the reported gains. The paper should report judge precision and recall against an oracle label (e.g., whether the pruned triplet sequence entails the gold answer), the distribution of reflection rounds, and a maximum-round policy, or otherwise demonstrate that premature stopping is not a dominant failure mode.
[§4.2, Reference Searching and Appendix C prompts] The reference base is constructed from the training set, and each reference includes the question, the reasoning path, and the correct answer. Because these references are injected into the relation-check, sequence-judge, and path-edit prompts, the inference-time guidance includes gold-answer information for similar training questions. The paper does not quantify how much of the gain comes from this answer leakage rather than from the proposed planning and reflection mechanisms. Please ablate references with the answers removed (paths only) or with the answers withheld, and clarify whether this constitutes a form of supervision at inference time rather than a purely training-free prompting method.
[§5.1, Table 2 and Appendix B] GrailQA results are reported on "the same test samples established in ToG," but several baseline numbers in Table 2, such as TIARA and Flexkbqa, are cited from their original publications and may use the full GrailQA test set or a different subset. If the baseline values are not all computed on the same 1,000-question subset, the claim of "second-highest overall performance on GrailQA" is not supported. Please either recompute all baselines on the same subset or clearly flag which comparisons are not directly comparable.
[§5.2, Table 2] All main results appear to come from single runs at temperature 0.3, with no standard errors, repeated seeds, or significance tests. The reported gains over Readi on WebQSP and CWQ are 2.7–8.8 points, and without variance information it is hard to judge whether the state-of-the-art claim is stable across sampling variability. Please report multiple runs or statistical significance measures, and disclose the API call dates and model versions used.

minor comments (6)

[Table 2, Table 3, Figure 3] There are several typos: "dataets" in the Table 2 header, "GraliQA" in Table 3 and Figure 3, and "theree datasets" in the Figure 3 caption.
[§4.4 and §5.1] "Contriver" should be "Contriever" in both the Knowledge Retrieval description and the Implementation Details paragraph.
[Appendix C, Table 8] The expected output tag appears as "<HA VE_ANSWER>" with a space in the prompt text. If the actual prompt contains this token, the sequence judge may never emit the intended string; please correct this typo.
[§5.1] The sentence "we extract 100 questions and corresponding references from each dataset" should specify how the 100 questions are selected from the clusters and should explicitly confirm that they come from the training split and are disjoint from the test questions.
[Limitations] The statement "Because we use data from training set, this limitation doesn't influence the performance of SRP" is too strong; using training-set-derived references can still overfit to the benchmark distribution. Please soften this and discuss how the method would behave under domain shift.
[§5.4, Table 4] The "searching success rate" is described as "how efficiently each approach retrieves answer of question from the KG," but the computation is not defined. Please specify whether success requires exact match of the gold answer entity, containment in the retrieved sequence, or some other criterion.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: SRP is an empirical KGQA pipeline evaluated on external benchmarks; training-set references are an inference-time supervision choice, not a constructional reduction.

full rationale

The paper does not present a formal derivation whose conclusion is equivalent to its assumptions. The reported results are measured Hits@1 numbers on the external WebQSP, CWQ, and GrailQA benchmarks, compared against cited baselines. The reference base in Section 4.2 is built from training-set questions with gold reasoning paths and answers and is used as in-context guidance, but this is not a fitted parameter renamed as a prediction; the ablation in Table 3 shows that removing or randomizing references changes performance, so the outcome is not forced by construction. The under-specified 'reliable answering rate' in Section 5.4 and the unmeasured accuracy of the sequence judge in Section 4.5 are empirical validity concerns, not circularity. There are no load-bearing self-citations, no imported uniqueness theorems, and no equation that defines a predicted quantity in terms of itself.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method's success is empirical and rests on domain assumptions about KG completeness and LLM judgment, plus several hand-chosen hyperparameters. It introduces no new entities, forces, or physical postulates.

free parameters (6)

reference_base_size = 100
The authors extract 100 questions and corresponding references from each dataset; this number is chosen by hand and not swept.
num_references_k = 4
The number of retrieved references for each test question is set to k=4 in Section 5.1.
relation_check_top_k = variable
The LLM scores candidate relations and the top-K are selected, but K is not fixed across examples; the prompt asks for 3 scored relations while examples show variable selection.
temperature = 0.3
Temperature is set to 0.3 for all LLM modules; no sensitivity analysis is reported.
retrieval_top_k = 5
For each relation in the reasoning path, the hybrid searcher retrieves top-5 similar relations on Freebase.
demonstration_shots = 3-5 per prompt/dataset
The number of few-shot demonstrations varies by prompt and dataset, chosen by hand and not justified by a sweep.

assumptions (4)

domain assumption Freebase contains complete and correct facts for all test questions.
The retrieval and answering modules assume that the relevant triples exist in Freebase; this enters in Sections 3 and 5.1.
domain assumption The reference base from the training set is representative of test questions and embedding similarity retrieves useful references.
Section 4.2 builds the reference base from training questions and uses MiniLM similarity; if domain shift breaks this, planning and reflection lose guidance.
domain assumption The LLM's relevance scoring of 1-hop relations correlates with correctness of KG paths.
Section 4.3 uses LLM scores to select initial relations; this assumes the LLM can map natural language questions to Freebase relation names.
domain assumption The sequence judge can accurately determine whether an answer is present in a triplet sequence.
Section 4.5 relies on the judge's 'have answer' verdict to stop reflection; no accuracy measurement for this judge is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Self-Reflective Planning with Knowledge Graphs: Enhancing LLM Reasoning Reliability for Question Answering." pith.science (2026). https://pith.science/paper/F4APEI3P

@misc{pith2026250519410,
  author       = {Pith},
  title        = {Pith review of: Self-Reflective Planning with Knowledge Graphs: Enhancing LLM Reasoning Reliability for Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F4APEI3P}},
  note         = {Machine review of arXiv:2505.19410}
}

read the original abstract

Recently, large language models (LLMs) have demonstrated remarkable capabilities in natural language processing tasks, yet they remain prone to hallucinations when reasoning with insufficient internal knowledge. While integrating LLMs with knowledge graphs (KGs) provides access to structured, verifiable information, existing approaches often generate incomplete or factually inconsistent reasoning paths. To this end, we propose Self-Reflective Planning (SRP), a framework that synergizes LLMs with KGs through iterative, reference-guided reasoning. Specifically, given a question and topic entities, SRP first searches for references to guide planning and reflection. In the planning process, it checks initial relations and generates a reasoning path. After retrieving knowledge from KGs through a reasoning path, it implements iterative reflection by judging the retrieval result and editing the reasoning path until the answer is correctly retrieved. Extensive experiments on three public datasets demonstrate that SRP surpasses various strong baselines and further underscore its reliable reasoning ability.

Figures

Figures reproduced from arXiv: 2505.19410 by the authors.

**Figure 2.** The framework of our SRP method. It consists of four main parts: (1) Reference Searching, (2) Path [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Reliable answering rate of Readi and SRP on theree datasets. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Case study of Readi and SRP. Question 1 comes from WebQSP, and question 2 comes from CWQ. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 10 canonical work pages

[1]

(’language.human_language.main_country’, 0.4): This relation is highly relevant as it directly relates to the country whose president is being asked for, and the main country where Brahui language is spoken in 1980

work page 1980
[2]

(’language.human_language.countries_spoken_in’, 0.3): This relation is also relevant as it provides information on the countries where Brahui language is spoken, which could help narrow down the search for the president

work page
[3]

Table 6: Prompts of relation check

(’base.rosetta.languoid.parent’, 0.2): This relation is less relevant but still provides some context on the language family to which Brahui belongs, which could be useful in understanding the linguistic and cultural background of the country in question. Table 6: Prompts of relation check. Instruction You are tasked with generating relation paths to help...

work page
[4]

Wen-tau Yih, Matthew Richardson, Chris Meek, Ming- Wei Chang, and Jina Suh

Tree of thoughts: Deliberate problem solving with large language models.Advances in neural information processing systems, 36:11809–11822. Wen-tau Yih, Matthew Richardson, Chris Meek, Ming- Wei Chang, and Jina Suh. 2016. The value of se- mantic parse labeling for knowledge base question answering. InProceedings of the 54th Annual Meet- ing of the Associat...

arXiv 2016
[8]

One or more topic entity that is central to the question

work page
[9]

princess leia

A set of valuable relations associated with the topic entity. Your goal is to generate relation paths that start with the topic entity and follow a sequence of relations to help answer the question. Demonstration Example Question: who played princess leia in star wars movies? Topic Entity: princess leia Valuable Relations: {“princess leia”: [’film.film_ch...

work page
[11]

Aviano Air Base

[(“Aviano Air Base”, “aviation.airport.serves”, “Aviano”)] Thinking Process: First, based on the triplet (“Aviano Air Base”, “location.location.containedby”, “Italy”), I can answer the question. So, I think these triplet sequences have enough information to answer the question. <HA VE_ANSWER> Retained sequences:

work page
[12]

Aviano Air Base

[(“Aviano Air Base”, “location.location.containedby”, “Italy”)]

work page

Show all 14 references

[13]

relation XXX not instantiated

[] Table 8: Prompts of sequence judge. Instruction Task: Given an Inital Path and some feedback information of a Question, please correct the Inital Path. Note: (1)When you receive Error Message, please edit the path based on Instantiate Paths. For example, if the Error Messag...
[14]

<cvt></cvt> in the end
[15]

place.religion.major_religions

relation “place.religion.major_religions” not instantiated. »» Instantiation Context Instantiate Paths: United Kingdom -> location.location.contains -> Heaton railway station United Kingdom -> location.statistical_region.religions -> <cvt></cvt> United Kingdom -> location.loca...
[2022]

Nils Reimers and Iryna Gurevych

Training language models to follow instruc- tions with human feedback.Advances in neural in- formation processing systems, 35:27730–27744. Nils Reimers and Iryna Gurevych. 2019. Sentence- BERT: Sentence embeddings using Siamese BERT- networks. InProceedings of the 2019 Confere...

2019 arXiv
[2023]

Liyi Chen, Panrong Tong, Zhongming Jin, Ying Sun, Jieping Ye, and Hui Xiong

Sparks of artificial general intelligence: Early experiments with gpt-4. Liyi Chen, Panrong Tong, Zhongming Jin, Ying Sun, Jieping Ye, and Hui Xiong. 2024. Plan-on- graph: Self-correcting adaptive planning of large lan- guage model on knowledge graphs.arXiv preprint arXiv:2410...

2024 arXiv
[2024]

In Proceedings of the AAAI conference on artificial in- telligence, volume 38, pages 18608–18616

Flexkbqa: A flexible llm-powered framework for few-shot knowledge base question answering. In Proceedings of the AAAI conference on artificial in- telligence, volume 38, pages 18608–18616. Linhao Luo, Yuan-Fang Li, Gholamreza Haffari, and Shirui Pan. 2023. Reasoning on graphs:...

2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

(’language.human_language.main_country’, 0.4): This relation is highly relevant as it directly relates to the country whose president is being asked for, and the main country where Brahui language is spoken in 1980

work page 1980

[2] [2]

(’language.human_language.countries_spoken_in’, 0.3): This relation is also relevant as it provides information on the countries where Brahui language is spoken, which could help narrow down the search for the president

work page

[3] [3]

Table 6: Prompts of relation check

(’base.rosetta.languoid.parent’, 0.2): This relation is less relevant but still provides some context on the language family to which Brahui belongs, which could be useful in understanding the linguistic and cultural background of the country in question. Table 6: Prompts of relation check. Instruction You are tasked with generating relation paths to help...

work page

[4] [4]

Wen-tau Yih, Matthew Richardson, Chris Meek, Ming- Wei Chang, and Jina Suh

Tree of thoughts: Deliberate problem solving with large language models.Advances in neural information processing systems, 36:11809–11822. Wen-tau Yih, Matthew Richardson, Chris Meek, Ming- Wei Chang, and Jina Suh. 2016. The value of se- mantic parse labeling for knowledge base question answering. InProceedings of the 54th Annual Meet- ing of the Associat...

arXiv 2016

[5] [8]

One or more topic entity that is central to the question

work page

[6] [9]

princess leia

A set of valuable relations associated with the topic entity. Your goal is to generate relation paths that start with the topic entity and follow a sequence of relations to help answer the question. Demonstration Example Question: who played princess leia in star wars movies? Topic Entity: princess leia Valuable Relations: {“princess leia”: [’film.film_ch...

work page

[7] [11]

Aviano Air Base

[(“Aviano Air Base”, “aviation.airport.serves”, “Aviano”)] Thinking Process: First, based on the triplet (“Aviano Air Base”, “location.location.containedby”, “Italy”), I can answer the question. So, I think these triplet sequences have enough information to answer the question. <HA VE_ANSWER> Retained sequences:

work page

[8] [12]

Aviano Air Base

[(“Aviano Air Base”, “location.location.containedby”, “Italy”)]

work page