Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Improving the Reliability of LLMs: Combining CoT, RAG, Self-Consistency, and Self-Verification

T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read Self-verification is the most reliable fix for LLM hallucination, the paper claims.

desk verdict Self-verification's reported advantage is an artifact of putting the ground-truth answer in the verification prompt. read the letter →

arxiv 2505.09031 v1 pith:CLX7I4MY submitted 2025-05-13 cs.AI cs.CL

classification cs.AIcs.CL
keywords LLMhallucinationchain-of-thoughtpromptingretrieval-augmentedgenerationself-consistencyself-verificationfactualaccuracyTruthfulQAFEVER
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

This paper asks which of four candidate fixes—step-by-step reasoning, retrieval of external evidence, sampling many answers, or having the model check its own answer—actually reduces LLM hallucination. Testing GPT-3.5-Turbo, LLaMA-2-7b, and DeepSeek-R1 on HaluEval, FEVER, and TruthfulQA, it finds that every intervention beats the unmodified baseline model. Retrieval-augmented reasoning and self-verification perform best, and self-verification is the strongest overall method, reaching roughly 90 percent accuracy on FEVER and 80 percent on TruthfulQA. The paper concludes that post-hoc verification grounded in retrieved evidence is the most effective way to keep model outputs truthful.

What carries the argument

The mechanism that carries the argument is the verification step in the self-verification pipeline: a second prompt receives the original query, the model's own generated answer, the ground-truth answer from the dataset, and the top five retrieved document chunks, and the model must decide whether the answer is hallucinated or correct. That step outperforms the paper's other interventions because it combines external evidence with an explicit judgement. The other load-bearing component is the self-consistency algorithm, which samples nine responses at varied temperatures, scores each against the reference answer with cosine similarity, applies a 0.5 threshold, and takes a majority vote on whether the output is hallucinated. The RAG pipeline, which embeds document chunks, retrieves the five most similar to the query, and feeds them to the model, supplies the external grounding for both retrieval-based and verification-based methods.

What would settle it

Run the same self-verification setup without the ground-truth answer in the prompt, keeping the query, generated answer, and retrieved documents. If FEVER accuracy and TruthfulQA MC2 drop to the level of RAG-plus-CoT or lower, the claimed advantage of self-verification depends on answer leakage rather than on independent checking.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that self-verification—a second pass in which the model receives its first answer, the ground-truth answer, and retrieved supporting documents and is asked to judge whether the answer is factually correct—is the most effective hallucination-mitigation strategy across the tested settings. The claim is comparative: self-verification outperforms baseline generation, chain-of-thought prompting, retrieval-augmented generation, their combination, and self-consistency voting, with LLaMA-2-7b slightly ahead of GPT-3.5-Turbo. On HaluEval, the best result is a RAG-plus-CoT setup reaching an 11 percent hallucination rate; on FEVER and TruthfulQA, self-verification tops the leaderboard. The paper presents these results as evidence that external grounding plus reflective checking, rather than reasoning alone, is what makes LLM outputs trustworthy.

Load-bearing premise

The load-bearing premise is that giving the model the correct answer inside the verification prompt measures its ability to verify; if the model is merely echoing that answer, the method's best performance is an experimental artifact.

Editorial extensions

If this is right

  • Teams deploying LLMs for fact-oriented tasks should add a self-verification pass with retrieved evidence rather than relying on chain-of-thought prompting alone.
  • Retrieval plus step-by-step reasoning is a close second and gives the lowest hallucination rate on HaluEval, so external grounding matters even when verification is not used.
  • All five strategies beat the unmodified model, meaning even lightweight prompting is better than no mitigation at all.
  • Self-verification is not model-specific: the pattern holds across GPT-3.5-Turbo, LLaMA-2-7b, and DeepSeek-R1, with LLaMA-2-7b slightly ahead.
  • Because self-consistency requires nine samples per query and self-verification needs retrieval plus a second prompt, the accuracy gains come with added inference cost.

Reading between the lines

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

  • A stricter test would remove the ground-truth answer from the verification prompt; if accuracy collapses, the reported edge of self-verification is leakage rather than checking.
  • The comparison likely understates what a separate, stronger verifier model could do, since the paper uses the same LLM for generation and verification.
  • The dependence on a hand-tuned cosine-similarity threshold of 0.5 and on top-5 retrieval suggests results could shift with different embeddings or retrieval granularity.
  • A domain outside Wikipedia-style fact checking, such as medical or legal reasoning, would test whether verification still holds when retrieved documents are less authoritative.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper empirically compares hallucination-mitigation strategies for LLMs: baseline, chain-of-thought (CoT), retrieval-augmented generation (RAG), CoT+RAG, self-consistency, and self-verification. It evaluates GPT-3.5-Turbo, LLaMA-2-7b, and DeepSeek-R1 on 500 samples each from HaluEval, FEVER, and TruthfulQA, using hallucination rate, label accuracy, and a similarity-based MC2 proxy. The paper reports that self-verification achieves the best overall performance (e.g., roughly 90% accuracy on FEVER and 80% MC2 on TruthfulQA) and concludes in Section 6.2 that self-verification is the most robust approach. The experimental protocols are described in Sections 4.4-4.7, evaluation settings in Section 5, and results in Section 6.

Significance. If the comparison were valid, the paper would provide practically useful evidence about which mitigation stack to adopt, and the multi-model, multi-dataset setup is a reasonable starting point. The manuscript clearly describes its pipeline and identifies several engineering choices (e.g., retrieval counts, temperature, token limits) that matter in practice. However, the central empirical claim is currently not supported: the self-verification protocol gives the model the ground-truth answer, so the reported advantage of self-verification is an artifact of answer leakage rather than a measured property of the method. In addition, the lack of held-out tuning, small sample sizes, and absence of statistical uncertainty make the ranking unreliable. As it stands, the paper does not establish which method is most effective for reducing hallucination.

major comments (4)
  1. [Section 4.7, Figure 4] The verification prompt for self-verification explicitly includes the ground-truth answer from the dataset, along with the query, the model-generated answer, and retrieved documents. A model given the gold answer can label a response as correct simply by detecting that it matches the provided gold answer, so the FEVER accuracy around 90% and TruthfulQA MC2 around 80% in Section 6.1 do not measure an independent self-verification capability. The conclusion in Section 6.2 that 'Self-Verification had the best performance' is therefore an artifact of answer leakage rather than a property of the method. The authors should rerun the verification experiments with the gold answer removed from the verification prompt (e.g., verifying only against retrieved evidence) and report those numbers as the self-verification results.
  2. [Algorithm 1, Section 4.6; Section 5.1] The self-consistency decision in Algorithm 1 uses the ground-truth answer y in the cosine-similarity comparison, and the threshold tau=0.5 is said to be chosen 'through trial and error.' Section 5.1 also describes tuning the CoT prompt, the number of retrieved documents (2, 5, and 10), the temperature, and the max-token limit without specifying any train/validation split. If these choices were made on the same 500-sample evaluation sets used for the final numbers, the reported improvements in Section 6 may be optimistically biased and the method ranking may not generalize. Please state which hyperparameters were selected on held-out data and re-evaluate the final configurations on unseen test samples.
  3. [Section 4.2; Section 6.1] Only 500 samples per dataset are evaluated, and no error bars, confidence intervals, or statistical significance tests are reported. Because LLM outputs are stochastic, as the temperature sampling in Section 4.6 acknowledges, the small differences highlighted in Section 6.2 (e.g., 'LLaMA-2 slightly outperforming GPT-3.5-Turbo') may be sampling noise. The authors should report variability across independent runs or bootstrap confidence intervals, and run paired statistical tests for the central comparisons.
  4. [Section 5.3.3] The 'MC2' metric used for TruthfulQA is not the standard TruthfulQA MC2; it assigns each response a label by taking the highest cosine similarity to the sets of correct and incorrect reference answers and then computes accuracy. The claimed 'MC2 score of around 80%' in Section 6.1 should either be renamed (e.g., 'similarity-based truthfulness accuracy') or be computed with the official TruthfulQA evaluation script; otherwise the number is not comparable to MC2 results in the literature.
minor comments (6)
  1. [Title] The title as submitted ('Improving the Reliability of LLMs: Combining Chain-of-Thought Reasoning and Retrieval-Augmented Generation') does not match the arXiv title that includes self-consistency and self-verification; please make the titles consistent.
  2. [Section 4.3] The phrase 'for all dataset' should be 'for all datasets.'
  3. [Algorithm 1] The majority-vote rule does not specify a tie-breaking behavior when count_factual equals count_hallucinated.
  4. [Figure 5 caption] Figure 5 refers to 'our website' without providing a URL; for reproducibility, please include a publicly accessible repository or at least a stable URL.
  5. [References] Several references have incomplete author lists (e.g., Kojima et al., Wei et al., Liu et al., Singh and Kapoor, and Zhou et al.); please provide full author information or use standard bibliography formatting.
  6. [Section 4.2] The paper states that 500 samples per dataset were evaluated but does not describe the sampling procedure or whether a fixed random seed was used; please state how the subset was drawn.

Circularity Check

2 steps flagged · score 8.0 of 10

Self-verification's reported advantage is not identifiable because the verification prompt includes the ground-truth answer, allowing label leakage.

  1. self definitional [Section 4.7, Self Verification (Figure 4); claimed in Section 6.2]
    "This second prompt is given the original query, the model's generated answer, the ground truth answer, and some supporting documents retrieved from a vector database similar to how it was done for RAG. The model used for generation is then asked to decide whether the generated answer is factually correct or hallucinated, based on the given evidence."

    The verifier receives the gold answer as part of its input, so the 'self-verification' decision can be made by matching the generated answer to the provided correct answer. The reported FEVER accuracy ~90%, TruthfulQA MC2 ~80%, and the Section 6.2 conclusion that 'Self-Verification had the best performance' therefore measure a label-leaking classification task, not an independent verification capability. The best-method result is forced by the protocol rather than by any property of self-verification.

  2. other [Section 4.6, paragraph after Algorithm 1]
    "We came up with a threshold tau = 0.5 through trial and error, if S(y_i_hat, y) > tau, the output is considered factually consistent; otherwise, it is classified as hallucinated."

    The factual-versus-hallucinated label for each sampled answer is assigned by cosine similarity to ground truth y, and the threshold is tuned on the data. The reported HaluEval hallucination rate for self-consistency is thus a ground-truth-supervised classification of samples, not a measure of consistency among the model's own generations. Giving the method access to the answer key at inference time makes its apparent improvement over baselines an artifact of construction.

full rationale

The central claim of the paper is the Section 6.2 ranking that 'Self-Verification had the best performance.' That ranking is not identifiable from the reported experiments because Section 4.7 states that the verification prompt contains 'the ground truth answer' alongside the query, the model's generated answer, and retrieved documents. A verifier with the gold answer in its input can label the generated answer correct simply by matching the provided answer, so the FEVER and TruthfulQA scores for self-verification are confounded with label leakage. The same structural leak appears in Section 4.6, where the self-consistency criterion labels each response as factually consistent or hallucinated by thresholding cosine similarity to the ground truth. The resulting HaluEval rates are therefore not independent measurements of the mitigation strategy. No load-bearing self-citation chain is present; the circularity is in the experimental definitions themselves. Although self-verification may in principle be a useful technique, the evidence presented cannot distinguish its intrinsic effect from the effect of supplying the answer key, so the paper's headline comparison is forced by construction. Score 8 reflects that the best-method conclusion reduces to the leak in the protocol, not that every number in the paper is vacuous.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new theory; its empirical claims rest on several unvalidated modeling choices. Most importantly, the self-verification prompt contains the ground-truth answer, and the similarity threshold is tuned on the same data used for final numbers.

free parameters (5)
  • Similarity threshold tau = 0.5
    Set 'through trial and error' in Algorithm 1 and Section 5.1; directly determines HaluEval hallucination rate.
  • Number of retrieved documents k = 5
    Tuned by comparing 2, 5, and 10; chosen because 2 gave incomplete context and 10 introduced noise. This is a per-task hyperparameter.
  • Sampling temperature = 0.4
    Chosen after trying 0.3 to 0.7; balances determinism and diversity. Used across all datasets.
  • Max token limit = 150
    Chosen to allow complete TruthfulQA answers; part of generation settings.
  • Number of self-consistency samples n = 9
    Chosen to reduce variance; mentioned in Algorithm 1.
assumptions (5)
  • domain assumption Cosine similarity between sentence embeddings of the generated answer and the reference answer is a valid measure of factual consistency.
    Used in Algorithm 1 to classify outputs as hallucinated or non-hallucinated; no validation that similarity correlates with factuality.
  • domain assumption Self-verification by the same model, given ground truth and retrieved documents, faithfully reflects the model's ability to detect its own hallucinations.
    Section 4.7: the model is asked to judge its answer with the correct answer in the prompt; this assumes no leakage.
  • domain assumption The ground truth labels in HaluEval, FEVER, and TruthfulQA are correct and the evaluation metrics (hallucination rate, label accuracy, MC2 via cosine similarity) capture the intended quality.
    Section 5.3; the MC2 approximation may not match official TruthfulQA scoring.
  • domain assumption Wikipedia retrieval provides adequate grounding for TruthfulQA questions, even though the dataset's source URLs span many websites.
    Section 4.4.3: the authors infer domains and use Wikipedia instead of the provided sources; this changes the task.
  • domain assumption The 500-sample subset is representative of each full dataset.
    Section 4.2: 'Due to limited hardware we evaluated on 500 samples per dataset.'

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving the Reliability of LLMs: Combining CoT, RAG, Self-Consistency, and Self-Verification." pith.science (2026). https://pith.science/paper/CLX7I4MY

@misc{pith2026250509031,
  author       = {Pith},
  title        = {Pith review of: Improving the Reliability of LLMs: Combining CoT, RAG, Self-Consistency, and Self-Verification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CLX7I4MY}},
  note         = {Machine review of arXiv:2505.09031}
}
read the original abstract

Hallucination, where large language models (LLMs) generate confident but incorrect or irrelevant information, remains a key limitation in their application to complex, open-ended tasks. Chain-of-thought (CoT) prompting has emerged as a promising method for improving multistep reasoning by guiding models through intermediate steps. However, CoT alone does not fully address the hallucination problem. In this work, we investigate how combining CoT with retrieval-augmented generation (RAG), as well as applying self-consistency and self-verification strategies, can reduce hallucinations and improve factual accuracy. By incorporating external knowledge sources during reasoning and enabling models to verify or revise their own outputs, we aim to generate more accurate and coherent responses. We present a comparative evaluation of baseline LLMs against CoT, CoT+RAG, self-consistency, and self-verification techniques. Our results highlight the effectiveness of each method and identify the most robust approach for minimizing hallucinations while preserving fluency and reasoning depth.

Figures

Figures reproduced from arXiv: 2505.09031 by the authors.

Figure 1
Figure 1. Illustration of Retrieval-Augmented Genera [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. CoT Outputs with different prompts [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Self Verification Output vs.Base Output For each input, we sample n = 9 responses from the model to reduce variance and avoid ambiguous cases, such as a 50-50 split. To promote diverse reasoning trajectories, we vary the temperature pa￾rameter T during decoding, where higher T values increase output randomness, and lower T values make the generation more deterministic. Each output yˆi is then evaluated against the g… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Self Verification Architecture answer from the dataset. To check if the model’s generated answer is trustworthy, we ask the model to verify its response. This second prompt is given the original query, the model’s generated answer, [PITH_FULL_IMAGE:figures/full_fig_p0…
Figure 5
Figure 5. Figure 5: Sample Working of Self Consistency on HaluEval Dataset on our website [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: HaluEval Results The following are the finding that we can see: • All the methods performed better than base￾line models [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: FEVER Results [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: TruthfulQA Results • Every method CoT, RAG, RAG + CoT, Self￾Consistency, and Self-Verification shows im￾provements over the base model across all datasets. This validates that hallucination mit￾igation strategies, whether through reasoning, retrieval, or verification, …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. RoT: Enhancing Table Reasoning with Iterative Row-Wise Traversals

    cs.CL 2025-05 conditional novelty 5.0 of 10

    A training-free row-wise traversal prompt improves table question answering accuracy over long-CoT reasoning models on WikiTableQuestions and TableBench.

Reference graph

Works this paper leans on

13 extracted references · 13 linked inside Pith · cited by 1 Pith paper

  1. [2]

    Retrieval-augmented gen- eration for large language models: A survey.arXiv preprint arXiv:2312.10997. T. Kojima, Y . Tsuchiya, and A. Ogawa

  2. [7]

    Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian

    Mitigating hal- lucination in retrieval-augmented chain-of-thought reasoning.arXiv preprint, arXiv:2303.08896. Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian

  3. [8]

    arXiv preprint arXiv:2307.06435

    A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435. S. Singh and A. Kapoor

  4. [9]

    James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal

    Hallucinations in open-domain question answering: Solutions through chain-of-thought reasoning.arXiv preprint, arXiv:2407.07071. James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal

  5. [11]

    Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171. J. Wei, P. Wang, D. Schuurmans, M. Bosma, and D. Chen

  6. [12]

    Yixuan Weng, Minjun Zhu, Fei Xia, Bin Li, Shizhu He, Shengping Liu, Bin Sun, Kang Liu, and Jun Zhao

    Chain-of-thought prompting elicits reasoning in large language models.arXiv preprint, arXiv:2201.11903. Yixuan Weng, Minjun Zhu, Fei Xia, Bin Li, Shizhu He, Shengping Liu, Bin Sun, Kang Liu, and Jun Zhao

  7. [13]

    Large language models are better reasoners with self-verification.arXiv preprint arXiv:2212.09561. W. Zhou, H. Wang, and Z. Yu

  8. [14]

    Reduc- ing hallucinations in retrieval-augmented generation with chain-of-thought prompting.arXiv preprint, arXiv:2305.13534

Show all 13 references
  1. [2018]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou

    Fever: a large-scale dataset for fact extraction and verification.arXiv preprint arXiv:1803.05355. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou

  2. [2021]

    Truthfulqa: Measuring how models mimic human falsehoods.arXiv preprint arXiv:2109.07958. Y . Liu, W. Chen, and J. Gao

  3. [2023]

    Stephanie Lin, Jacob Hilton, and Owain Evans

    Halueval: A large- scale hallucination evaluation benchmark for large language models.arXiv preprint arXiv:2305.11747. Stephanie Lin, Jacob Hilton, and Owain Evans

  4. [2024]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Meng Wang, and Haofen Wang

    Factchd: Benchmarking fact-conflicting hallucina- tion detection.arXiv preprint arXiv:2310.12086. Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Meng Wang, and Haofen Wang

  5. [2025]

    Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen

    Cot-rag: Integrating chain of thought and retrieval-augmented generation to en- hance reasoning in large language models.arXiv preprint arXiv:2504.13534. Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.