Pith. sign in

REVIEW 4 major objections 5 minor 27 references

A prompt-only pipeline for LLM self-verification claims to improve factual accuracy and citations—but the paper itself calls its results fabricated.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-05 05:03 UTC pith:GJLGOEKP

load-bearing objection The paper's own admission that its experimental results are fabricated is the load-bearing flaw; no amount of prompt-engineering scaffolding can survive that. the 4 major comments →

arxiv 2509.05741 v1 pith:GJLGOEKP submitted 2025-09-06 cs.CL

Enhancing Factual Accuracy and Citation Generation in LLMs via Multi-Stage Self-Verification

Fernando Gabriela Garc\'ia , Qiyang Shi , Zilin Feng This is my paper
classification cs.CL
keywords large language modelshallucinationchain-of-thoughtself-verificationcitation generationprompt engineeringfactual accuracyretrieval-augmented generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper proposes VeriFact-CoT, a four-stage, prompt-engineering-only method that asks an LLM to generate an initial chain-of-thought, decompose it into factual claims, simulate a fact-checking and evidence-retrieval step using the model's own internal knowledge, and then rewrite the answer with corrections and per-claim citations. The intended contribution is a fine-tuning-free way to reduce hallucination and add traceable sources. The paper's own text, however, explicitly labels its experimental results as 'fabricated,' so the reported numbers—for example, factual accuracy rising from 72 percent to 83 percent and hallucination falling from 25 percent to 12 percent on complex QA—are not supported by any real evaluation. What remains is the architecture of the idea and the unresolved question of whether a model can verify itself without external grounding.

Core claim

The central claim is that a multi-stage 'fact verification–reflection–citation integration' loop embeds fact-checking and source attribution into LLM reasoning. Each stage is a generative function: initial CoT generation, claim extraction with verification queries, simulated verification and evidence generation, and final refinement with citation integration. The paper argues this closed loop lets the model correct its own factual errors and produce verifiable outputs. The paper also states, without irony, that the accompanying experimental results demonstrating this superiority are 'fabricated,' which removes the empirical basis of the claim.

What carries the argument

The carrying object is the four-stage prompt sequence, formalized as equations (1)–(4): (C0,A0)=G_InitialCoT(Q); (F,V)=G_ClaimExtract(C0,A0); E=G_VerifySimulate(V); (Cf,Af)=G_RefineIntegrate(C0,A0,E). The third stage is the crucial one: the LLM invents 'evidence' and 'plausible formatted citation sources' from its parametric knowledge, acting as a self-contained RAG. That internal simulation is what makes the method fine-tuning-free, and it is also the load-bearing assumption that the model's own verification can be trusted.

Load-bearing premise

The method assumes an LLM can reliably verify its own factual claims using only internal parametric knowledge and then generate truthful citations; if that assumption fails, the verification stage simply adds new hallucinations instead of removing them.

What would settle it

Run the four-stage pipeline on a standard benchmark such as HotpotQA or Natural Questions with real ground truth, then check every generated citation against an actual database: if factual accuracy does not improve over standard CoT or if citations are non-existent or irrelevant, the central claim is refuted.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the loop works, LLM outputs in fact-sensitive domains could carry inline citations without external retrieval infrastructure.
  • The method would be a drop-in prompt template usable on any instruction-following model, with no training or architecture change.
  • Citation quality would become a first-class output of LLM generation, enabling automated traceability checks.
  • The ablation logic implies that each stage contributes: removing claim extraction, verification simulation, or refinement degrades accuracy and citation scores.
  • The reported gains would make self-verification competitive with basic RAG while keeping the system self-contained.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the self-admission of fabrication means the only testable residue is whether internal self-verification can improve factuality at all; a clean experiment would need to check generated citations against real sources, because plausible-looking but invented references are a known failure mode.
  • Editorial inference: the method's success hinges on the model's inability to confidently assert falsehoods during verification; if the model is uniformly overconfident, the refinement stage has no error signal to act on.
  • Editorial inference: a natural extension is to replace the simulated verification with real retrieval (search API or knowledge graph) while keeping the claim-extraction and refinement stages; the paper itself lists this as future work.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes VeriFact-CoT, a multi-stage prompt-engineering framework for LLMs that interleaves initial chain-of-thought generation, claim identification, simulated internal fact-checking, and refinement with citation integration. The method is formalized in Equations (1)-(4). The paper claims that VeriFact-CoT outperforms standard CoT and basic RAG-augmented CoT on factual accuracy, hallucination rate, and citation quality, across multiple tasks and LLM backbones. However, the manuscript explicitly states in the abstract and Section 1 that its experimental results are 'fabricated,' and the verification stage is itself an LLM-generated simulation that can produce invented citation sources. No code, data, prompt templates, or evaluation protocol are provided.

Significance. If the method worked as claimed, a fine-tuning-free, purely prompt-based pipeline for reducing hallucinations and generating citations would be practically useful, especially for deployment settings where external retrieval is unavailable. The paper should also be credited for laying out a four-stage pipeline in a clear, formal way and for acknowledging computational overhead and the risk of simulated hallucinations in Section 4.9. However, these strengths do not compensate for the absence of any real empirical evidence. The explicit admission that the experimental results are fabricated voids every quantitative claim, and the internal-simulation design cannot support the concept of 'verifiable' citations. No code, data, or prompt templates are shipped, so the claimed gains are not reproducible. The significance of the paper as currently submitted is therefore nil; at most it is a proposal sketch.

major comments (4)
  1. [Abstract and Section 1] The paper states: 'Our fabricated experimental results demonstrate that VeriFact-CoT consistently outperforms...' This is a direct admission that the reported numbers are invented. Every quantitative claim in the manuscript—Table 1's accuracy/hallucination/citation figures, Table 2's ablation, Figure 2's cross-model results, and Figure 1's human evaluation—rests on this fabricated basis. No code, data, prompt templates, or evaluation protocol are provided that could convert these numbers into measured outcomes. This is a load-bearing failure that invalidates the central claim.
  2. [Section 3.3, Equation (3)] The verification stage instructs the LLM to 'generate' evidence and 'produce a plausible, formatted citation source' from its own pretrained knowledge, with no external lookup. Section 4.9 concedes that this stage can 'hallucinate verification results or fabricate citation sources.' Consequently, the evidence used to correct the initial output and the citations inserted in the final answer are generated by the same model being verified. The reported citation-quality F1, including its 'verifiability' component (Section 4.3.3), cannot measure traceability to real sources; it measures only self-plausibility. This circularity undermines the method's core claim of objective factual self-correction.
  3. [Section 4.1.2 and References] The literature foundation is inaccurate for central works. The reference for Chain-of-Thought, [9], is actually 'System Prompt Optimization with Meta-Learning'; the reference for RAG, [10], is a mathematics paper on weighted spherical means; and the reference for Natural Questions, [12], is a computability-theory preprint. These are not the standard CoT, RAG, or Natural Questions citations, and the baselines 'Standard CoT' and 'CoT + Basic RAG' are therefore not properly grounded in prior work.
  4. [Section 4] Even setting aside the explicit 'fabricated' admission, the experimental section lacks the basic elements needed for evaluation: dataset splits and sizes, full prompt templates, inference hyperparameters, number of human annotators, inter-annotator agreement, and statistical significance tests. Tables 1-2 and Figures 1-2 report effect sizes with no variance or sample-size information. Thus, no independent artifact is provided that could substantiate the empirical claims.
minor comments (5)
  1. [Abstract] The sentence 'while reducing hallucination rate to 12% from 25% and 18' appears to be truncated; the intended comparison is incomplete.
  2. [Table 2] The table header contains the typo 'Citation Qualitys'; it should read 'Citation Quality'.
  3. [Section 4.3.3] The definition of Citation Quality as an F1 score is under-specified: it is unclear how precision and recall are computed for 'relevance' and 'verifiability.' The paper should provide exact formulas.
  4. [Section 4.10] The qualitative example labels all sources as 'Simulated Source,' confirming that no real citation is verified. This should be explicitly connected to the limitation stated in Section 4.9.
  5. [References] Several references are misformatted or non-standard, and some arXiv identifiers contain version tags that are not standard citations. A thorough cleanup is needed if the paper is revised.

Circularity Check

2 steps flagged

Paper admits 'fabricated experimental results' and closes the verification loop inside the same LLM, so the reported gains reduce to self-generated evidence with no external anchor.

specific steps
  1. self definitional [Section 3.3, Equations 2-4; Section 4.9]
    "our method employs a novel prompt engineering strategy to guide the LLM to simulate this verification process internally. The LLM leverages its extensive pre-trained knowledge to 'judge' each query ... it is prompted to 'generate' a verification result or piece of evidence e_i ... The LLM is also instructed to 'produce' a plausible, formatted citation source s_i."

    The correction loop is closed inside the same model: (C0,A0) -> F,V -> E -> (Cf,Af). Equation 3 makes E the LLM's own simulated evidence, and Equation 4 uses that same E to revise C0,A0. There is no external fact-check; the 'verification' is the model generating plausible-sounding evidence for claims derived from its own answer. Consequently the reported improvements in factual accuracy and citation quality measure internal self-consistency/plausibility, not objective verification. The paper itself concedes in §4.9 that the GVerifySimulate stage can 'hallucinate verification results or fabricate citation sources.'

  2. other [Abstract and Section 1]
    "Our fabricated experimental results demonstrate that VeriFact-CoT consistently outperforms traditional CoT and basic RAG-enhanced CoT methods across all evaluated tasks and metrics."

    All quantitative claims in the paper—Table 1 (83% vs 72/78% factual accuracy, 12% vs 25/18% hallucination, citation F1 0.75), Table 2 ablation, Figure 2 cross-model results, and Figure 1 human evaluation—are explicitly attributed to experiments the manuscript itself labels 'fabricated.' No code, data, or prompt templates are provided. Thus the central claim's evidence is not merely self-referential; it is nonexistent, and the stated performance gains reduce to the paper's own unsupported assertion.

full rationale

The paper's derivation chain has two fatal problems. First, the method is self-verifying by construction: the 'evidence' E used to correct the initial answer is generated by the same LLM from queries derived from that same initial answer, with simulated sources rather than real retrieval. This makes the fact-checking loop entirely internal—there is no independent ground truth in the loop, and the paper admits the verification stage can itself hallucinate. Second, the paper explicitly labels its experimental results as 'fabricated,' which voids the evidentiary basis for every reported number. Even if one were to accept the method as a prompt-engineering proposal, the measured gains cannot be distinguished from the model's own self-confirmation bias or from pure invention. This is not a minor self-citation issue; the central claim's support reduces to self-generated evidence and self-admitted fabrication, warranting a score of 9.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 1 invented entities

The paper introduces no quantitative model, so there are no genuine fitted parameters. Its axiomatic load comes from assumptions about the reliability of self-verification and simulated citations, all lacking independent evidence. The only invented entity is the simulated citation itself, fabricated without external grounding.

axioms (3)
  • domain assumption An LLM can accurately verify its own factual claims using internal parametric knowledge alone.
    Section 3.3 simulates verification without external retrieval; this unproven assumption is contradicted by the paper's own acknowledgment of simulated hallucinations in verification.
  • domain assumption Model-generated simulated citations (hypothetical URLs, report titles, institution names) are acceptable evidence of verifiability.
    Sections 3.3 and 3.4 instruct the model to produce plausible citations; the paper treats citation quality as an improvement metric despite these being fabricated.
  • domain assumption The four-stage prompt sequence does not introduce error propagation beyond what the model can correct.
    The method is implemented purely via prompts (Section 3.5); the paper provides no analysis of stage-by-stage error accumulation.
invented entities (1)
  • Simulated citation sources (hypothetical URLs, authoritative report titles, institution names) no independent evidence
    purpose: To provide verifiability and source attribution for the LLM's claims
    Generated by the LLM itself in Section 3.3, with no external check; the paper calls them simulated, plausible, and hypothetical.

pith-pipeline@v1.3.0-alltime-deepseek · 12564 in / 9888 out tokens · 93741 ms · 2026-08-05T05:03:57.752561+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Enhancing Factual Accuracy and Citation Generation in LLMs via Multi-Stage Self-Verification." pith.science (2026). https://pith.science/paper/GJLGOEKP

@misc{pith2026250905741,
  author       = {Pith},
  title        = {Pith review of: Enhancing Factual Accuracy and Citation Generation in LLMs via Multi-Stage Self-Verification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GJLGOEKP}},
  note         = {Machine review of arXiv:2509.05741}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This research introduces VeriFact-CoT (Verified Factual Chain-of-Thought), a novel method designed to address the pervasive issues of hallucination and the absence of credible citation sources in Large Language Models (LLMs) when generating complex, fact-sensitive content. By incorporating a multi-stage mechanism of 'fact verification-reflection-citation integration,' VeriFact-CoT empowers LLMs to critically self-examine and revise their intermediate reasoning steps and final answers. This process significantly enhances the objective accuracy, trustworthiness, and traceability of the generated outputs, making LLMs more reliable for applications demanding high fidelity such as scientific research, news reporting, and legal consultation.

Figures

Figures reproduced from arXiv: 2509.05741 by Fernando Gabriela Garc\'ia, Qiyang Shi, Zilin Feng.

Figure 1
Figure 1. Figure 1: Human Evaluation Results on Output Quality and Trustworthiness [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: VeriFact-CoT Performance Across Different LLM Backbones on Complex Factual QA [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 13 canonical work pages · 6 internal anchors

  1. [1]

    Cedille: A large autoregressive French language model

    Martin Müller and Florian Laurent. Cedille: A large autoregressive french language model.arXiv preprint arXiv:2202.03371v1, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  2. [2]

    Dpl-slam: enhancing dynamic point-line slam through dense semantic methods.IEEE Sensors Journal, 24(9):14596–14607, 2024

    Zhihao Lin, Qi Zhang, Zhen Tian, Peizhuo Yu, and Jianglin Lan. Dpl-slam: enhancing dynamic point-line slam through dense semantic methods.IEEE Sensors Journal, 24(9):14596–14607, 2024

    work page 2024

  3. [3]

    Enhanced visual slam for collision-free driving with lightweight autonomous cars.Sensors, 24(19):6258, 2024

    Zhihao Lin, Zhen Tian, Qi Zhang, Hanyang Zhuang, and Jianglin Lan. Enhanced visual slam for collision-free driving with lightweight autonomous cars.Sensors, 24(19):6258, 2024

    work page 2024

  4. [4]

    Slam2: Simultaneous localization and multimode mapping for indoor dynamic environments.Pattern Recognition, 158:111054, 2025

    Zhihao Lin, Qi Zhang, Zhen Tian, Peizhuo Yu, Ziyang Ye, Hanyang Zhuang, and Jianglin Lan. Slam2: Simultaneous localization and multimode mapping for indoor dynamic environments.Pattern Recognition, 158:111054, 2025

    work page 2025

  5. [5]

    Score: Story coherence and retrieval enhancement for ai narratives.arXiv preprint arXiv:2503.23512, 2025

    Qiang Yi, Yangfan He, Jianhui Wang, Xinyuan Song, Shiyao Qian, Xinhang Yuan, Miao Zhang, Li Sun, Keqin Li, Kuan Lu, et al. Score: Story coherence and retrieval enhancement for ai narratives.arXiv preprint arXiv:2503.23512, 2025

    arXiv 2025

  6. [6]

    Enhancing code llms with reinforcement learning in code generation

    Junqiao Wang, Zeng Zhang, Yangfan He, Yuyang Song, Tianyu Shi, Yuchen Li, Hengyuan Xu, Kunyu Wu, Guangwu Qian, Qiuwu Chen, et al. Enhancing code llms with reinforcement learning in code generation. arXiv preprint arXiv:2412.20367, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    Weak to strong generalization for large language models with multi-capabilities

    Yucheng Zhou, Jianbing Shen, and Yu Cheng. Weak to strong generalization for large language models with multi-capabilities. InThe Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  8. [8]

    Large language models as software components: A taxonomy for llm-integrated applications

    Irene Weber. Large language models as software components: A taxonomy for llm-integrated applications. arXiv preprint arXiv:2406.10300v1, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    System prompt optimization with meta-learning.arXiv preprint arXiv:2505.09666v1, 2025

    Yumin Choi, Jinheon Baek, and Sung Ju Hwang. System prompt optimization with meta-learning.arXiv preprint arXiv:2505.09666v1, 2025

    arXiv 2025

  10. [10]

    Weighted spherical means generated by generalized translation and general Euler-Poisson-Darboux equation

    Elina Shishkina. Weighted spherical means generated by generalized translation and general euler-poisson- darboux equation.arXiv preprint arXiv:1703.06340v1, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  11. [11]

    Answeringcomplexopen-domainquestions with multi-hop dense retrieval.arXiv preprint arXiv:2009.12756v2, 2020

    WenhanXiong, XiangLorraineLi, SriniIyer, JingfeiDu, PatrickLewis, WilliamYangWang, YasharMehdad, WentauYih, SebastianRiedel, DouweKiela, andBarlasOğuz. Answeringcomplexopen-domainquestions with multi-hop dense retrieval.arXiv preprint arXiv:2009.12756v2, 2020

    Pith/arXiv arXiv 2009

  12. [12]

    The combinatorial equivalence of a computability theoretic question

    Lu Liu. The combinatorial equivalence of a computability theoretic question.arXiv preprint arXiv:2012.13588v1, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2012

  13. [13]

    ResNetVLLM-2: Addressing ResNetVLLM's Multi-Modal Hallucinations

    Ahmad Khalil, Mahmoud Khalil, and Alioune Ngom. Resnetvllm-2: Addressing resnetvllm’s multi-modal hallucinations.arXiv preprint arXiv:2504.14429v1, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  14. [14]

    Uncertainty-aware fusion: An ensemble framework for mitigating hallucinations in large language models.arXiv preprint arXiv:2503.05757v1, 2025

    Prasenjit Dey, Srujana Merugu, and Sivaramakrishnan Kaveri. Uncertainty-aware fusion: An ensemble framework for mitigating hallucinations in large language models.arXiv preprint arXiv:2503.05757v1, 2025

    Pith/arXiv arXiv 2025

  15. [15]

    Hyper-rag: Combating llm hallucinations using hypergraph-driven retrieval-augmented generation.arXiv preprint arXiv:2504.08758v1, 2025

    Yifan Feng, Hao Hu, Xingliang Hou, Shiquan Liu, Shihui Ying, Shaoyi Du, Han Hu, and Yue Gao. Hyper-rag: Combating llm hallucinations using hypergraph-driven retrieval-augmented generation.arXiv preprint arXiv:2504.08758v1, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Zero-knowledge llm hallucination detection and mitigation through fine-grained cross-model consistency.arXiv preprint arXiv:2508.14314v1, 2025

    Aman Goel, Daniel Schwartz, and Yanjun Qi. Zero-knowledge llm hallucination detection and mitigation through fine-grained cross-model consistency.arXiv preprint arXiv:2508.14314v1, 2025

    work page arXiv 2025

  17. [17]

    Improving medical large vision-language models with abnormal-aware feedback.arXiv preprint arXiv:2501.01377, 2025

    Yucheng Zhou, Lingran Song, and Jianbing Shen. Improving medical large vision-language models with abnormal-aware feedback.arXiv preprint arXiv:2501.01377, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Mitigating Large Language Model Hallucinations via Autonomous Knowledge Graph-based Retrofitting

    Xinyan Guan, Yanjiang Liu, Hongyu Lin, Yaojie Lu, Ben He, Xianpei Han, and Le Sun. Mitigating large language model hallucinations via autonomous knowledge graph-based retrofitting.arXiv preprint arXiv:2311.13314v1, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  19. [19]

    Evaluating the Factual Consistency of Large Language Models Through News Summarization

    Derek Tam, Anisha Mascarenhas, Shiyue Zhang, Sarah Kwan, Mohit Bansal, and Colin Raffel. Evaluating the factual consistency of large language models through news summarization.arXiv preprint arXiv:2211.08412v2, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  20. [20]

    The dawn after the dark: An empirical study on factuality hallucination in large language models.arXiv preprint arXiv:2401.03205v1, 2024

    Junyi Li, Jie Chen, Ruiyang Ren, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. The dawn after the dark: An empirical study on factuality hallucination in large language models.arXiv preprint arXiv:2401.03205v1, 2024

    Pith/arXiv arXiv 2024

  21. [21]

    Multihal: Multilingual dataset for knowledge-graph grounded evaluation of llm hallucinations.arXiv preprint arXiv:2505.14101v1, 2025

    Ernests Lavrinovics, Russa Biswas, Katja Hose, and Johannes Bjerva. Multihal: Multilingual dataset for knowledge-graph grounded evaluation of llm hallucinations.arXiv preprint arXiv:2505.14101v1, 2025

    work page arXiv 2025

  22. [22]

    Structured chain-of-thought prompting for code generation.arXiv preprint arXiv:2305.06599v3, 2023

    Jia Li, Ge Li, Yongmin Li, and Zhi Jin. Structured chain-of-thought prompting for code generation.arXiv preprint arXiv:2305.06599v3, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Automatic prompt augmentation and selection with chain- of-thought from labeled data.arXiv preprint arXiv:2302.12822v3, 2023

    KaShun Shum, Shizhe Diao, and Tong Zhang. Automatic prompt augmentation and selection with chain- of-thought from labeled data.arXiv preprint arXiv:2302.12822v3, 2023

    Pith/arXiv arXiv 2023

  24. [24]

    Prompt baking.arXiv preprint arXiv:2409.13697v1, 2024

    Aman Bhargava, Cameron Witkowski, Alexander Detkov, and Matt Thomson. Prompt baking.arXiv preprint arXiv:2409.13697v1, 2024

    Pith/arXiv arXiv 2024

  25. [25]

    Large language models have intrinsic self-correction ability

    DanchengLiu, AmirNassereldine, ZimingYang, ChenhuiXu, YutingHu, JiajieLi, UtkarshKumar, Changjae Lee, Ruiyang Qin, Yiyu Shi, and Jinjun Xiong. Large language models have intrinsic self-correction ability. arXiv preprint arXiv:2406.15673v2, 2024

    Pith/arXiv arXiv 2024

  26. [26]

    Do advanced language models eliminate the need for prompt engineering in software engineering? arXiv preprint arXiv:2411.02093v1, 2024

    Guoqing Wang, Zeyu Sun, Zhihao Gong, Sixiang Ye, Yizhou Chen, Yifan Zhao, Qingyuan Liang, and Dan Hao. Do advanced language models eliminate the need for prompt engineering in software engineering? arXiv preprint arXiv:2411.02093v1, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Visual in-context learning for large vision- language models

    Yucheng Zhou, Xiang Li, Qianning Wang, and Jianbing Shen. Visual in-context learning for large vision- language models. InFindings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 15890–15902. Association for Computational Linguistics, 2024

    work page 2024