Pith. sign in

REVIEW 4 major objections 6 minor 43 references

ImageRef-VL: Enabling Contextual Image Referencing in Vision-Language Models

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that supervised fine-tuning on roughly 7,600 curated samples teaches a vision-language model when and where to insert relevant images into answers, letting an 8B model outscore GPT-4o, Claude-3.5-Sonnet, and open-source…

desk verdict A useful new task and benchmark for contextual image referencing in RAG chatbots, but the evaluation protocol needs tightening before the headline numbers can be taken at face value. read the letter →

arxiv 2501.12418 v1 pith:JFIYFLMH submitted 2025-01-20 cs.CV cs.AI

Jingwei Yi , Junhao Yin , Ju Xu , Peng Bao , Yongliang Wang , Wei Fan , Hao Wang This is my paper · ORCID
classification cs.CVcs.AI
keywords contextualimagereferencingvision-languagemodelsretrieval-augmentedgenerationinterleavedimage-textsupervisedfine-tuningplacementevaluationCIR-TestLLM-as-judge
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

Retrieval-augmented chatbots can cite text sources but rarely choose an image from the retrieved documents and place it where it helps the user. This paper defines that behavior as contextual image referencing and asks whether a vision-language model can be taught it. To answer, the authors build CIR-Test, a 456-sample test set with insertion-point labels scored 0-3, and CIR-Interleave/CIR-Caption, a training collection generated by LLMs and VLMs and then manually filtered. Fine-tuning InternVL2-8B and InternVL2-26B on this data gives ImageRef-VL, which the paper reports surpasses GPT-4o and Claude-3.5-Sonnet on human overall evaluation and improves 88% over the best open-source VLM, with human overall scores of 3.52 versus 1.87 for InternVL2-26B. If the result is right, a targeted supervised fine-tuning stage can give open models a capability that proprietary models currently score lower on.

What carries the argument

The mechanism that carries the argument is a three-stage data-construction pipeline followed by standard supervised fine-tuning. First, a text-only response is generated from the retrieved document text without images. Second, each image receives a context-aware caption in two stages: a purely visual description, then a supplemental pass that adds surrounding context. Third, an LLM inserts image references at chosen positions in the response, with manual filtering. The resulting samples form CIR-Interleave (7,645 samples with 73,833 images) and CIR-Caption (5,633 samples), mixed with InternVL2's own SFT data before fine-tuning InternVL2-8B and InternVL2-26B with a next-token generative loss. On the evaluation side, CIR-Test labels every potential insertion point with a 0-3 relevance score, from unrelated to perfectly matching, and defines precision, recall3, F1, plus human 5-point Likert ratings of text, image, and overall quality.

What would settle it

Build a fresh CIR-Test whose insertion-point labels are assigned by independent human annotators who never see the training data or the generation code, then re-run ImageRef-VL-8B and InternVL2-26B; if the human-overall gap of 3.52 versus 1.87 shrinks below noise or reverses, the reported capability is not separable from label style.

Watch

Extended reading notes

Core claim

The paper's central claim is that the missing ingredient in current VLMs is training signal for when and where to reference an image, not visual understanding or retrieval itself. The authors report that a relatively small volume of manually curated interleaved image-text responses transfers this ability: ImageRef-VL-8B reaches a human overall score of 3.52 on a 5-point scale, above GPT-4o (3.17), Claude-3.5-Sonnet (3.32), and InternVL2-26B (1.87), with a human image-quality score of 4.05. They also claim their evaluation framework is valid: the automated F1 for image placement correlates with the human image score at 0.9854, and the automated text score correlates with the human text score at 0.9033. The intended upshot is that the capability and a measurement for it have been established together.

Load-bearing premise

The headline numbers assume that CIR-Test's 0-3 insertion-point labels and the human 5-point ratings measure contextually appropriate image use, not just resemblance to the authors' own data-construction pipeline.

Editorial extensions

If this is right

  • Open-source VLM developers can add contextual image referencing without changing architecture or retrieval: a curated interleaved SFT set on the order of ten thousand samples is claimed to be sufficient.
  • End-to-end ImageRef-VL inference is cheaper than the three-stage pipeline, with the paper reporting 1.83 seconds per sample for the 26B model versus 9.77 seconds for the three-stage variant.
  • CIR-Test and its F1 metric provide an automatic screening signal for future work, because the paper reports a 0.9854 correlation between F1 and human image scores.
  • The authors state that starting from a well-pretrained VLM and including the collected dataset in supervised fine-tuning might yield better results, and that bad cases remain, so the method is a mechanism proof rather than a complete package.

Reading between the lines

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

  • A testable extension the paper does not run: apply the same data-construction recipe to a VLM without InternVL2's interleaved-image pretraining, to see whether the SFT data or the base model is the active ingredient.
  • The 0-3 insertion-point labels could be converted into a reward model, letting follow-up work use reinforcement learning to optimize the same image-placement objective and reduce the bad cases the authors acknowledge.
  • Since the training data comes from LLM/VLM generation plus human filtering, the pipeline could be bootstrapped: ImageRef-VL's own outputs, after filtering, could seed the next round of training data and lower the manual curation burden.
  • An independent-annotation version of CIR-Test, with labels written by people blind to the generation pipeline, would test whether the measured gap reflects contextual appropriateness rather than stylistic mimicry of the training-data generation process.
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.

Referee Report

4 major / 6 minor

Summary. The paper introduces Contextual Image Reference, a new capability for vision-language models (VLMs) to insert contextually relevant images from retrieved documents into their responses, and proposes ImageRef-VL, a supervised fine-tuning recipe built on InternVL2. The authors construct two training datasets (CIR-Interleave with 7,645 conversations and CIR-Caption with 5,633 samples) via LLM/VLM generation followed by manual filtering, mix these with the InternVL2-SFT data, and fine-tune InternVL2-8B and 26B. They also construct a test set (CIR-Test) with 456 samples and propose automated metrics (text evaluation, image position precision/recall/F1) plus a 5-point Likert human evaluation. The main claim is that ImageRef-VL-8B achieves an 88% improvement over state-of-the-art open-source VLMs in human overall score (3.52 vs 1.87 for InternVL2-26B), and that it outperforms proprietary closed-source models (GPT-4o, Claude-3.5-Sonnet) on the same human metrics.

Significance. If the claimed results are valid, this is a useful contribution to multimodal conversational AI and retrieval-augmented generation. The paper identifies a real and previously unaddressed gap—contextually appropriate image referencing—and provides a practical SFT method that appears to transfer this skill to open-source VLMs with a large human-evaluation margin. The evaluation framework, including the labeled insertion-point scores and the metric-validation correlations, is a useful starting point for future work. The paper also evaluates computational efficiency, showing an advantage over a three-stage generation pipeline. However, the significance is conditional on the evaluation protocol being trustworthy; the current manuscript lacks sufficient detail to rule out test-set contamination and selection-on-test bias, and the human evaluation is under-reported.

major comments (4)
  1. [Sections 5.1, 6.1, Table 1] The construction of CIR-Test is not described anywhere. The paper states that CIR-Test is constructed 'as outlined in Section 5' (Section 6.1), but Section 5 only defines the metric formulas and the scoring rubric; it does not describe how the 456 samples, 3,767 images, prompts, or retrieved documents were sampled, nor who assigned the 0-3 position labels. The paper also never reports any overlap or deduplication analysis between CIR-Test and the training sets CIR-Interleave (7,645 conversations, 73,833 images) and CIR-Caption (5,633 samples, 29,558 images). Because both training and test data appear to come from the same retrieved-document conversation format, the headline 88% improvement on human overall evaluation (Table 2) could be inflated by memorized page-image associations rather than by generalized contextual referencing. Please provide a full construction protocol and an overlap analysis (e.g., exact or near-duplicate text and image detection) to support the generalization claim.
  2. [Section 6.4, Figure 4, Section 6.1] The final hyperparameters (data mixture ratio 1:4 for the 8B model and 1:1 for the 26B model; training steps 950 and 550, respectively) are selected by evaluating the text evaluation score and image F1 on CIR-Test. Since CIR-Test is the same benchmark on which the main comparison in Table 2 is reported, the reported numbers are the best of a small hyperparameter sweep rather than a fixed-protocol prediction, which introduces selection-on-test bias. Please introduce a validation split for hyperparameter tuning, or report results across all sweep points so that the reported values are not cherry-picked from the test set.
  3. [Section 5.2, Table 2, Figure 3] The human evaluation reports mean 5-point Likert scores (e.g., 3.52 for ImageRef-VL-8B versus 1.87 for InternVL2-26B) without specifying the number of annotators, the annotation instructions, whether annotators were blinded to model identity, or any inter-annotator agreement measure. Without this information and without error bars or significance tests, the 88% improvement claim is not statistically substantiated. Please add these experimental details and report per-item variance or bootstrap confidence intervals for the reported means.
  4. [Section 6.2, Table 2] The image position evaluation is not applied uniformly across methods. Open-source baselines are evaluated with 'controlled sampling' to insert image references, ImageRef-VL is evaluated end-to-end, and closed-source models are not evaluated on this metric at all. The paper itself acknowledges that 'a fair comparison is not feasible' for the three-stage approaches, yet it still reports these numbers and uses them for hyperparameter selection (Section 6.4) and for the metric-validity Pearson correlation (0.9854). This makes the automated-metric validation and the position scores non-comparable across systems. Please either apply a single evaluation protocol for all models or report the position metric only for end-to-end models and explicitly justify its use in metric validation.
minor comments (6)
  1. [Sections 4.1, 6.1, Figure 2] The naming is inconsistent: 'IMI-interleave', 'MI-Interleave', 'IMI-Caption', and 'IMI-VL' are used interchangeably; please unify the terminology (e.g., CIR-Interleave, CIR-Caption, ImageRef-VL).
  2. [Equation (3), Section 5.1] In the Recall3 definition, the indicator 'I(sp = 1)' should read 'I(sp = 3)' to match the text description that follows.
  3. [Section 6.2, Figure 3] The text mentions 'ImageRef-VL-8B and -32B' when discussing Figure 3, but the experiments only use the 8B and 26B models; please correct the typo.
  4. [Section 5.1, Equation (2)] The symbol N is used both for the number of test samples in Eq. (1) and for the total number of inserted images in Eq. (2); please use distinct symbols for clarity.
  5. [Section 6.2] The 'controlled sampling' procedure used for open-source baselines is not described; please provide the prompt template and sampling mechanism so the evaluation is reproducible.
  6. [Abstract and Section 6.2] The '88% performance improvement' should be qualified as a relative gain on the human overall score over InternVL2-26B on the proposed test set, since it is not an aggregate across all reported metrics.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical SFT recipe whose reported gains are not forced by construction or by a self-citation chain.

full rationale

This paper presents an empirical method: training data for contextual image referencing is generated with existing LLMs/VLMs and manual filtering (Section 4.1), a VLM is fine-tuned with standard SFT (Section 4.2), and performance is measured with automated metrics and human Likert ratings (Section 5). There is no claimed derivation, uniqueness theorem, or fitted parameter that is renamed as a prediction. The 88% improvement over open-source VLMs is an arithmetic comparison of reported human-evaluation scores in Table 2, not a quantity defined into existence by the method. The main concerns noted by a skeptical reader are about evaluation validity — how CIR-Test was constructed, whether it overlaps with the training data, and the lack of annotator agreement or a fixed validation protocol before the hyperparameter sweep in Section 6.4. These are legitimate scientific-validity concerns, but they are not circularity as defined here: no quoted passage shows a test quantity or headline result being reduced by construction to a training input, and the paper does not invoke a load-bearing self-citation or an author-imported uniqueness result. Accordingly, the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper's contribution is an empirical recipe; its central claim rests on the quality of generated training data, the validity of the evaluation protocol, and the reliability of human ratings. The paper leaves these as assumptions rather than demonstrated facts.

free parameters (3)
  • data mixture ratio (CIR data : InternVL2-SFT) = 1:1 (26B), 1:4 (8B)
    Selected in Section 6.4 based on CIR-Test text eval and F1; this is a hand-tuned hyperparameter chosen using the test set.
  • training steps = 550 (26B), 950 (8B)
    Selected in Section 6.4 based on convergence of test metrics.
  • image position score thresholds (0/1/2/3) = categories 3, 2, 1, 0
    Definition of label categories is a design choice in Section 5.1 that determines Precision/Recall3/F1; no justification or validation of these thresholds is provided.
assumptions (4)
  • domain assumption The CIR training data, generated by LLMs/VLMs and manually filtered, is sufficiently high quality to teach appropriate image referencing.
    Section 4.1 relies on this for SFT to transfer the capability; no quality audit or inter-annotator statistics are reported.
  • domain assumption CIR-Test and its insertion-point labels reflect the true task of contextual image referencing.
    Section 5 defines the test format but does not describe construction or labeling protocol; validity of the metric depends on this.
  • domain assumption The LLM-as-judge text score is a valid proxy for response quality even though it does not see the images.
    Section 5.1 reports Pearson 0.9033 with human text scores but provides no details on the judge prompt, variance, or failure cases.
  • domain assumption Human Likert ratings are reliable without reported annotator agreement.
    Section 5.2 uses 5-point Likert scores but does not report number of annotators, instructions, or inter-annotator agreement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ImageRef-VL: Enabling Contextual Image Referencing in Vision-Language Models." pith.science (2026). https://pith.science/paper/JFIYFLMH

@misc{pith2026250112418,
  author       = {Pith},
  title        = {Pith review of: ImageRef-VL: Enabling Contextual Image Referencing in Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JFIYFLMH}},
  note         = {Machine review of arXiv:2501.12418}
}
read the original abstract

Vision-Language Models (VLMs) have demonstrated remarkable capabilities in understanding multimodal inputs and have been widely integrated into Retrieval-Augmented Generation (RAG) based conversational systems. While current VLM-powered chatbots can provide textual source references in their responses, they exhibit significant limitations in referencing contextually relevant images during conversations. In this paper, we introduce Contextual Image Reference -- the ability to appropriately reference relevant images from retrieval documents based on conversation context -- and systematically investigate VLMs' capability in this aspect. We conduct the first evaluation for contextual image referencing, comprising a dedicated testing dataset and evaluation metrics. Furthermore, we propose ImageRef-VL, a method that significantly enhances open-source VLMs' image referencing capabilities through instruction fine-tuning on a large-scale, manually curated multimodal conversation dataset. Experimental results demonstrate that ImageRef-VL not only outperforms proprietary models but also achieves an 88% performance improvement over state-of-the-art open-source VLMs in contextual image referencing tasks. Our code is available at https://github.com/bytedance/ImageRef-VL.

Figures

Figures reproduced from arXiv: 2501.12418 by the authors.

Figure 1
Figure 1. An example of contextual image reference, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The training strategy of the proposed IMI-VL model. Stage 1: Training dataset construction involves generating textual responses and image descriptions through a language model and a vision-language model. These are combined into interleaved responses using image contexts and captions. Stage 2: Supervised fine-tuning refines the model with a vision encoder, adapter, and language model, optimizing through generative … view at source ↗
Figure 3
Figure 3. Human evaluation score distribution of four methods. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Impact of the hyper-parameters on our ImageRef-VL-8B and ImageRef-VL-26B. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Computational cost comparison between our [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 21 canonical work pages

  1. [1]

    Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. 2024. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219

    arXiv 2024

  2. [2]

    Anthropic. 2024. https://www.anthropic.com/news/claude-3-5-sonnet Claude 3.5 sonnet

    2024

  3. [3]

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. 2024. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185--24198

    work page 2024

  4. [4]

    Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. 2024. A survey on rag meeting llms: Towards retrieval-augmented large language models. In KDD, pages 6491--6501

    work page 2024

  5. [5]

    Constanza Fierro, Reinald Kim Amplayo, Fantine Huot, Nicola De Cao, Joshua Maynez, Shashi Narayan, and Mirella Lapata. 2024. Learning to plan and generate text with citations. arXiv preprint arXiv:2404.03381

    arXiv 2024

  6. [6]

    Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. 2024. https://arxiv.org/abs/2306.13394 Mme: A comprehensive evaluation benchmark for multimodal large language models . Preprint, arXiv:2306.13394

    arXiv 2024

  7. [7]

    Tianyu Gao, Howard Yen, Jiatong Yu, and Danqi Chen. 2023 a . Enabling large language models to generate text with citations. arXiv preprint arXiv:2305.14627

    arXiv 2023

  8. [8]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023 b . Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997

    arXiv 2023

Show all 43 references
  1. [9]

    Shailja Gupta, Rajesh Ranjan, and Surya Narayan Singh. 2024. https://arxiv.org/abs/2410.12837 A comprehensive survey of retrieval-augmented generation (rag): Evolution, current landscape and future directions . Preprint, arXiv:2410.12837

    2024 arXiv

  2. [10]

    Lucas Torroba Hennigen, Shannon Shen, Aniruddha Nrusimha, Bernhard Gapp, David Sontag, and Yoon Kim. 2023. Towards verifiable text generation with symbolic references. arXiv preprint arXiv:2311.09188

    2023 arXiv

  3. [11]

    Chengyu Huang, Zeqiu Wu, Yushi Hu, and Wenya Wang. 2024. Training language models to generate text with citations via fine-grained rewards. arXiv preprint arXiv:2402.04315

    2024 arXiv

  4. [12]

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. 2023. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. TOIS

    2023

  5. [13]

    Drew A Hudson and Christopher D Manning. 2019. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, pages 6700--6709

    2019

  6. [14]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

    2024 arXiv

  7. [15]

    Ankur Joshi, Saket Kale, Satish Chandel, and D Kumar Pal. 2015. Likert scale: Explored and explained. British journal of applied science & technology, 7(4):396--403

    2015

  8. [16]

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. NIPS, 35:22199--22213

    2022

  9. [17]

    Dongyub Lee, Taesun Whang, Chanhee Lee, and Heuiseok Lim. 2023. Towards reliable and fluent large language models: Incorporating feedback learning loops in qa systems. arXiv preprint arXiv:2309.06384

    2023 arXiv

  10. [18]

    Dongfang Li, Zetian Sun, Baotian Hu, Zhenyu Liu, Xinshuo Hu, Xuebo Liu, and Min Zhang. 2024. Improving attributed text generation of large language models via preference learning. arXiv preprint arXiv:2403.18381

    2024 arXiv

  11. [19]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730--19742. PMLR

    2023

  12. [20]

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888--12900

    2022

  13. [21]

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2024 a . Improved baselines with visual instruction tuning. In CVPR, pages 26296--26306

    2024

  14. [22]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2024 b . Visual instruction tuning. NIPS, 36

    2024

  15. [23]

    Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. 2024. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525

    2024 arXiv

  16. [24]

    Meta. 2024. https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/ Llama 3.2

    2024

  17. [25]

    Seyed Mahed Mousavi, Simone Alghisi, and Giuseppe Riccardi. 2024. Is your llm outdated? benchmarking llms & alignment algorithms for time-sensitive knowledge. arXiv preprint arXiv:2404.08700

    2024 arXiv

  18. [26]

    Masayasu Muraoka, Ryosuke Kohita, and Etsuko Ishii. 2020. Image position prediction in multimodal documents. In LREC, pages 4265--4274

    2020

  19. [27]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. NIPS, 35:27730--27744

    2022

  20. [28]

    Gabriel Poesia, Oleksandr Polozov, Vu Le, Ashish Tiwari, Gustavo Soares, Christopher Meek, and Sumit Gulwani. 2022. Synchromesh: Reliable code generation from pre-trained language models. arXiv preprint arXiv:2201.11227

    2022 arXiv

  21. [29]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In ICML, pages 8748--8763

    2021

  22. [30]

    Jiajun Shen, Tong Zhou, Suifeng Zhao, Yubo Chen, and Kang Liu. 2024. Citekit: A modular toolkit for large language model citation generation. arXiv preprint arXiv:2408.04662

    2024 arXiv

  23. [31]

    Marzieh Tahaei, Aref Jafari, Ahmad Rashid, David Alfonso-Hermelo, Khalil Bibi, Yimeng Wu, Ali Ghodsi, Boxing Chen, and Mehdi Rezagholizadeh. 2024. Efficient citer: Tuning large language models for enhanced answer quality and verification. In Findings of NAACL, pages 4443--4450

    2024

  24. [32]

    SM Tonmoy, SM Zaman, Vinija Jain, Anku Rani, Vipula Rawte, Aman Chadha, and Amitava Das. 2024. A comprehensive survey of hallucination mitigation techniques in large language models. arXiv preprint arXiv:2401.01313

    2024 arXiv

  25. [33]

    A Vaswani. 2017. Attention is all you need. NIPS

    2017

  26. [34]

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. 2024. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191

    2024 arXiv

  27. [35]

    Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. 2024. Hallucination is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817

    2024 arXiv

  28. [36]

    Xi Ye, Ruoxi Sun, Sercan Arik, and Tomas Pfister. 2024. Effective large language model adaptation for improved grounding and citation generation. In NAACL, pages 6237--6251

    2024

  29. [37]

    Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. 2023. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490

    2023 arXiv

  30. [38]

    Biao Zhang, Barry Haddow, and Alexandra Birch. 2023. Prompting large language model for machine translation: A case study. In ICML, pages 41092--41110

    2023

  31. [39]

    Jiajie Zhang, Yushi Bai, Xin Lv, Wanjun Gu, Danqing Liu, Minhao Zou, Shulin Cao, Lei Hou, Yuxiao Dong, Ling Feng, et al. 2024 a . Longcite: Enabling llms to generate fine-grained citations in long-context qa. arXiv preprint arXiv:2409.02897

    2024 arXiv

  32. [40]

    Jingyu Zhang, Marc Marone, Tianjian Li, Benjamin Van Durme, and Daniel Khashabi. 2024 b . Verifiable by design: Aligning language models to quote from pre-training data. arXiv preprint arXiv:2404.03862

    2024 arXiv

  33. [41]

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. NIPS, 36:46595--46623

    2023

  34. [42]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  35. [43]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.