Pith. sign in

REVIEW 4 major objections 5 minor 53 references

LPOI: Listwise Preference Optimization for Vision Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Training a vision-language model to rank progressively masked images by object visibility lowers hallucination rates below pairwise preference methods, the paper claims, and needs no extra annotations.

desk verdict Genuinely new VLM preference optimization with plausible empirical gains; the core monotonicity assumption needs a sanity check, but this deserves peer review. read the letter →

arxiv 2505.21061 v1 pith:LIAPXWD2 submitted 2025-05-27 cs.CV cs.AI

classification cs.CVcs.AI
keywords vision-languagemodelshallucinationreductionlistwisepreferenceoptimizationobjectmaskinghardnegativesamplingdirectvisualgroundingmultimodalalignment
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

The paper tries to establish that a vision-language model hallucinates less when it is trained to order images by how clearly the object named in its answer is visible, rather than merely to prefer one image over another. LPOI creates that order automatically: it detects and masks a critical object, then interpolates the masked fraction from the full image to a fully masked hard negative, producing a ranked list at no extra annotation cost. A listwise preference loss then pushes the model's likelihood of the chosen text to fall as visibility falls. Experiments on MMHalBench, AMBER, and Object HalBench report lower hallucination rates and better scores than DPO, mDPO, and other preference-optimization baselines across three base VLMs. If the monotonicity assumption behind the rank is accepted, the result is that dense visual-ordering supervision can serve as a practical antidote to visual grounding failures.

What carries the argument

The load-bearing mechanism is the automatically generated ranked image list together with the listwise loss in Eq. (1). A zero-shot open-vocabulary object detector finds an object named or implied by the chosen answer; its bounding box is progressively masked from the image edge inward, and the masked region is circled so the model attends to it. A filtering step keeps the list only when the fully masked image makes a reference VLM hallucinate. The listwise loss is the negative log-likelihood of the desired ordering of normalized log-likelihood scores, so the model is trained to make the chosen text more probable exactly when the object is more visible. This machinery carries the argument because it converts standard pairwise preference data into dense, graded visual supervision without listwise human annotations.

What would settle it

On a held-out set, compute the trained model's log-likelihood of the chosen answer for the $L$ progressively masked images and have annotators judge whether that answer is still true at each mask level; if likelihood does not decrease monotonically with masking on a substantial share of items, or if the answer is already false on the original image, the listwise training signal is enforcing a false premise. A complementary control is to train on the same masked images as unordered pairs; if the hallucination gains survive, the ranking order is not the active ingredient.

Watch

Extended reading notes

Core claim

LPOI's central claim is that a VLM can be taught to stop hallucinating a mentioned object by optimizing the model to rank images by that object's visibility. The paper claims this is the first listwise preference optimization applied to vision-language models. Given an image, a question, a chosen answer, and a rejected answer, the method detects the critical object, builds a hard negative by masking it, and creates $L$ images whose masks cover $\frac{k-1}{L-1}\times100\%$ of the bounding box for $k=1,\dots,L$, with a visual prompt on the masked region. The training objective is $L_{\mathrm{total}}=L_{\mathrm{DPO}}+L_{\mathrm{Anchor}}+L_{\mathrm{Listwise}}$, where the listwise term is the negative log-likelihood of the permutation of scores $S_k=\beta\log\frac{\pi_\theta(w\mid x_k,q)}{\pi_{\mathrm{ref}}(w\mid x_k,q)}$, pushing $S_1>S_2>\cdots>S_L$. The paper reports that this ordering signal lowers hallucination rates relative to pairwise methods on MMHalBench, AMBER, and Object HalBench across three base models, and that human annotators prefer its responses over mDPO- and DPO-trained ones.

Load-bearing premise

The load-bearing premise is that for every generated list the original chosen answer stays the correct response and the model's likelihood of producing it should fall smoothly as more of the object is covered, with no independent check that this holds at intermediate or full masking.

Editorial extensions

If this is right

  • On the three benchmarks, LPOI consistently lowers hallucination metrics relative to DPO and mDPO across LLaVA-v1.5-7B, LLaVA-v1.5-13B, and Idefics2-8B.
  • Under a fixed 20-hour GPU budget on Idefics2-8B, LPOI with list size 5 still outperforms DPO and mDPO on MMHalBench score and hallucination rate.
  • Larger list sizes improve hallucination metrics on Object HalBench, indicating that the graded list itself, not just a single hard negative, contributes to the gain.
  • Removing the text DPO loss or the anchor loss degrades performance, so the listwise image signal complements rather than replaces textual preference learning.

Reading between the lines

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

  • If the monotonicity link between object visibility and answer likelihood holds broadly, the same interpolation idea should transfer to other evidence-damaging transforms such as blur, occlusion by other objects, or lighting changes; the paper's mechanism predicts each would reduce hallucination if applied as a ranked list.
  • Because the ranked list is generated from a single chosen text, LPOI implicitly teaches the model to modulate its confidence with visible evidence, which could generalize to calibration under partial occlusion beyond the three hallucination benchmarks, though that transfer is not established here.
  • A control the paper does not run would train on the same masked images as unordered pairs; if the hallucination gains persist without the ordering, the active ingredient would be extra masked images rather than listwise ranking.
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 / 5 minor

Summary. The paper proposes LPOI, a listwise preference optimization method for vision-language models (VLMs) aimed at reducing object hallucination. The method identifies a critical object in the image, masks it progressively, and interpolates between the original and fully masked images to build a ranked list of images. The model is trained with a listwise loss (Eq. 1) so that the normalized likelihood of the chosen text decreases as the object is more heavily masked, in addition to the standard DPO and an anchor loss. Empirical evaluation on MMHalBench, AMBER, and Object HalBench across three backbones (LLaVA-v1.5-7B/13B, Idefics2-8B) reports consistent hallucination reduction compared to DPO, mDPO, and several published baselines, with additional ablations, a human evaluation, and a same-budget comparison.

Significance. If the reported gains are robust, LPOI is a meaningful contribution: it is the first listwise preference optimization for VLMs, it requires no additional annotations beyond standard pairwise preference data, and it shows consistent improvements across multiple models and benchmarks. The paper includes useful strengths: code release, several ablations (list size, visual prompting, loss components, verification module), a human evaluation with agreement scores, and a same-GPU-budget comparison. The central idea—using progressively masked images as a ranked list to ground responses in object visibility—is intuitive and could influence subsequent work on hallucination mitigation.

major comments (4)
  1. [Section 3.2, Eq. (1)] The load-bearing assumption that the likelihood of the chosen text w decreases monotonically with the mask fraction is not verified. For chosen responses that describe multiple objects, attributes, or scene context, masking 20–40% of one bounding box may leave the full response well-supported, so the sequence-level likelihood S_k can be flat or non-monotone. The paper should measure the ordering of S_k on a held-out sample before training and report the frequency with which monotonicity holds, and also report the absolute values of S_L to demonstrate that the fully masked image is actually suppressed. Without such a check, the training signal cannot be attributed to object visibility rather than to arbitrary ordering constraints.
  2. [Algorithm 1, Line 10 vs Section 3.1] The verification criterion is contradictory. Section 3.1 states that the masked image is kept as a hard negative when the filtering model hallucinates (i.e., still outputs the chosen text w), and that another object is selected only if the model does not hallucinate. Algorithm 1, Line 10 instead says to retry with another object when the filtering model 'answers (x_L,q,w) to be positive answer,' which is precisely the hallucinating case. This internal inconsistency must be resolved, as it changes which examples are included in training.
  3. [Section 4, Tables 1–4, 8–10] The main results are single runs without error bars or statistical significance tests. Only Figure 5 reports variance, and only for one model and one benchmark. Given that some improvements over mDPO are small (e.g., Idefics2-8B on AMBER HalRate 0.36 vs 0.40 in Table 1), the claim of consistent improvement needs repeated runs or significance testing across the reported metrics.
  4. [Section 4.1, Baselines] HALVA, HA-DPO, and V-DPO are taken from their original papers rather than re-run under the same training dataset and budget, while DPO and mDPO are reproduced under the same setting. The paper acknowledges this 'for reference,' but the abstract's claim of outperforming 'state-of-the-art VLM preference optimization approaches' rests in part on these incomparable numbers. A fair comparison would retrain these baselines under the same protocol, or the scope of the claim should be narrowed to the reproduced baselines.
minor comments (5)
  1. [Eq. (1) and Algorithm 1] The listwise loss uses z as the list length in Eq. (1), while Algorithm 1 and the text use L; unify the notation.
  2. [Section 3.2] Typo: 'postive' should be 'positive.'
  3. [Throughout] The paper consistently renders the model name as 'LLaV A-v1.5' (with a space), which should be 'LLaVA-v1.5'; also the Figure 3 caption refers to 'Idefics-2B' instead of 'Idefics2-8B,' and Table 3 uses 'V .P.' with spaces.
  4. [Appendix A, Table 6] Hyperparameters beta and delta (used in the losses) are not listed in Table 6; the temperature and anchor margin should be specified for reproducibility.
  5. [Figure 5] The legend uses 'MDPO' in capitals, inconsistent with the 'mDPO' notation used elsewhere.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LPOI is an empirical training method evaluated on external hallucination benchmarks, with no fitted parameter renamed as a prediction.

full rationale

The paper's central claim is that listwise preference optimization over progressively masked images reduces VLM hallucination. This claim is supported by training three base VLMs with a fixed objective (Eq. 1) and measuring performance on MMHalBench, AMBER, and Object HalBench, which are external to the training data and are not constructed from the method's own outputs. The listwise loss enforces S_1 > S_2 > ... > S_L by construction, but that is a training objective, not a prediction derived from fitted constants; the paper never claims to predict a benchmark number from the same data used to fit a parameter. Reusing the chosen text w as the target for masked images is a modeling assumption, and the unverified monotonicity of S_k with mask fraction is a genuine correctness risk, but neither reduces the benchmark results to the training objective by definition. The hard-negative filtering step uses Idefics2-8B's hallucination behavior to select training examples, which is a data-selection heuristic rather than a fitted parameter later called a prediction. There is no load-bearing self-citation chain, no imported uniqueness theorem, and no renamed empirical pattern presented as derivation. The apparent discrepancy between Algorithm 1's filtering condition and Section 3.1's description is a wording issue, not a circular step. The paper's derivation is therefore self-contained relative to its external evaluations.

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

LPOI introduces no new physical or conceptual entities; its machinery is composed of existing components (object detection, masking, visual prompting, listwise loss). The ledger entries above capture the hand-chosen hyperparameters and the domain assumptions that are load-bearing for the method's success.

free parameters (4)
  • beta (DPO/listwise temperature)
    Scales the log-probability ratios in LDPO, LAnchor, and LListwise. Standard DPO hyperparameter, value not stated in main text; affects the strength of the preference signal.
  • delta (anchor margin)
    Margin in LAnchor, borrowed from mDPO; chosen by hand to push the chosen text probability above a threshold on the original image.
  • List size L = 5 (default; 3,4,5 tested)
    Number of images in the list. Table 4 shows performance varies with L; larger L generally improves results, so the method's efficacy depends on this choice.
  • Masking ratio schedule = (k-1)/(L-1) x 100%, from edge
    Linear interpolation of the masked fraction of the bounding box. This specific schedule is an ad hoc modeling choice; no principled derivation is given.
assumptions (4)
  • domain assumption Grounding-DINO-Tiny localizes the text-relevant object accurately enough for the masks to remove the object from the model's visual input.
    Section 3.1 uses Grounding-DINO to select objects; if detection fails, the masked images do not actually hide the critical object and the ranking signal is corrupted.
  • ad hoc to paper The chosen text w remains the correct response for all partially masked images, with likelihood monotonically decreasing in masking ratio.
    Section 3.2 and Eq. (1) use the same w for every list position with a ranking target; no ground truth for intermediate masks is available, and an explicit 'object absent' text is never used.
  • domain assumption The base model's hallucination on a fully masked image is a valid indicator that the masked image is a useful hard negative.
    Algorithm 1 Lines 10-12 retain only samples where the verification model (Idefics2-8B) hallucinates the object; this assumes the failure transfers to other target models.
  • standard math Listwise ranking loss from Cao et al. (2007) and LiPO is a valid objective for optimizing the desired order.
    The loss in Eq. (1) is taken from the learning-to-rank literature without new proof; accepted as standard.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LPOI: Listwise Preference Optimization for Vision Language Models." pith.science (2026). https://pith.science/paper/LIAPXWD2

@misc{pith2026250521061,
  author       = {Pith},
  title        = {Pith review of: LPOI: Listwise Preference Optimization for Vision Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LIAPXWD2}},
  note         = {Machine review of arXiv:2505.21061}
}
read the original abstract

Aligning large VLMs with human preferences is a challenging task, as methods like RLHF and DPO often overfit to textual information or exacerbate hallucinations. Although augmenting negative image samples partially addresses these pitfalls, no prior work has employed listwise preference optimization for VLMs, due to the complexity and cost of constructing listwise image samples. In this work, we propose LPOI, the first object-aware listwise preference optimization developed for reducing hallucinations in VLMs. LPOI identifies and masks a critical object in the image, and then interpolates the masked region between the positive and negative images to form a sequence of incrementally more complete images. The model is trained to rank these images in ascending order of object visibility, effectively reducing hallucinations while retaining visual fidelity. LPOI requires no extra annotations beyond standard pairwise preference data, as it automatically constructs the ranked lists through object masking and interpolation. Comprehensive experiments on MMHalBench, AMBER, and Object HalBench confirm that LPOI outperforms existing preference optimization methods in reducing hallucinations and enhancing VLM performance. We make the code available at https://github.com/fatemehpesaran310/lpoi.

Figures

Figures reproduced from arXiv: 2505.21061 by the authors.

Figure 1
Figure 1. Comparison of preference optimization (PO) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the LPOI framework. (1) Given an input image, prompt and corresponding set of chosen and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Human evaluation results on a subset of the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: MMHalBench results for different preference [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results of the base model (LLaVA-v1.5-7B) its finetuned versions with DPO, mDPO, and [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Comparison of saliency maps with or without visual prompting (highlighted in red circle). [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Qualitative results of the base model (LLaVA-v1.5-7B) its variants optimized with DPO, mDPO, and [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: User interface and instruction for human evaluation. [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 3 canonical work pages

  1. [1]

    Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, ...

  2. [2]

    Hritik Bansal, Ashima Suvarna, Gantavya Bhatt, Nanyun Peng, Kai-Wei Chang, and Aditya Grover. 2024. https://arxiv.org/abs/2404.00530 Comparing bad apples to good oranges: Aligning large language models via joint preference optimization . Preprint, arXiv:2404.00530

  3. [3]

    Meyer, Yuning Chai, and Yong Jae Lee

    Mu Cai, Haotian Liu, Dennis Park, Siva Karthik Mustikovela, Gregory P. Meyer, Yuning Chai, and Yong Jae Lee. 2024. https://arxiv.org/abs/2312.00784 Vip-llava: Making large multimodal models understand arbitrary visual prompts . Preprint, arXiv:2312.00784

  4. [4]

    Zhe Cao, Tao Qin, Tie-Yan Liu, Ming-Feng Tsai, and Hang Li. 2007 a . https://doi.org/10.1145/1273496.1273513 Learning to rank: From pairwise approach to listwise approach . volume 227, pages 129--136

  5. [5]

    Zhe Cao, Tao Qin, Tie-Yan Liu, Ming-Feng Tsai, and Hang Li. 2007 b . https://api.semanticscholar.org/CorpusID:207163577 Learning to rank: from pairwise approach to listwise approach . In International Conference on Machine Learning

  6. [6]

    Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. 2024. https://arxiv.org/abs/2401.17270 Yolo-world: Real-time open-vocabulary object detection . Preprint, arXiv:2401.17270

  7. [7]

    Brown, Miljan Martic, Shane Legg, and Dario Amodei

    Paul Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. 2023. https://arxiv.org/abs/1706.03741 Deep reinforcement learning from human preferences . Preprint, arXiv:1706.03741

  8. [8]

    Yihe Deng, Pan Lu, Fan Yin, Ziniu Hu, Sheng Shen, Quanquan Gu, James Zou, Kai-Wei Chang, and Wei Wang. 2024. https://arxiv.org/abs/2405.19716 Enhancing large vision language models with self-training on image comprehension . Preprint, arXiv:2405.19716

Show all 53 references
  1. [9]

    Leo Gao, John Schulman, and Jacob Hilton. 2022. https://arxiv.org/abs/2210.10760 Scaling laws for reward model overoptimization . Preprint, arXiv:2210.10760

  2. [10]

    Amelia Glaese, Nat McAleese, Maja Trębacz, John Aslanides, Vlad Firoiu, Timo Ewalds, Maribeth Rauh, Laura Weidinger, Martin Chadwick, Phoebe Thacker, Lucy Campbell-Gillingham, Jonathan Uesato, Po-Sen Huang, Ramona Comanescu, Fan Yang, Abigail See, Sumanth Dathathri, Rory Greig...

  3. [11]

    Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. 2024. https://arxiv.org/abs/2310.14566 Hallusionbench: An advanced diagnostic suite for entangled language hallucinat...

  4. [12]

    Hadsell, S

    R. Hadsell, S. Chopra, and Y. LeCun. 2006. https://doi.org/10.1109/CVPR.2006.100 Dimensionality reduction by learning an invariant mapping . In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 2, pages 1735--1742

  5. [13]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2015. https://arxiv.org/abs/1512.03385 Deep residual learning for image recognition . Preprint, arXiv:1512.03385

  6. [14]

    Jiwoo Hong, Noah Lee, and James Thorne. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.626 ORPO : Monolithic preference optimization without reference model . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 11170--11189, Miam...

  7. [15]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. https://arxiv.org/abs/2106.09685 Lora: Low-rank adaptation of large language models . Preprint, arXiv:2106.09685

  8. [16]

    Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. 2024. https://arxiv.org/abs/2405.02246 What matters when building vision-language models? Preprint, arXiv:2405.02246

  9. [17]

    Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. 2023 a . https://arxiv.org/abs/2312.10665 Silkie: Preference distillation for large visual language models . Preprint, arXiv:2312.10665

  10. [18]

    Zheng Li, Caili Guo, Xin Wang, Zerun Feng, and Yanjun Wang. 2023 b . https://arxiv.org/abs/2305.16566 Integrating listwise ranking into pairwise-based image-text retrieval . Preprint, arXiv:2305.16566

  11. [19]

    Weifeng Lin, Xinyu Wei, Ruichuan An, Peng Gao, Bocheng Zou, Yulin Luo, Siyuan Huang, Shanghang Zhang, and Hongsheng Li. 2024. https://arxiv.org/abs/2403.20271 Draw-and-understand: Leveraging visual prompts to enable mllms to comprehend what you want . Preprint, arXiv:2403.20271

  12. [20]

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2024 a . https://arxiv.org/abs/2310.03744 Improved baselines with visual instruction tuning . Preprint, arXiv:2310.03744

  13. [21]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. https://arxiv.org/abs/2304.08485 Visual instruction tuning . Preprint, arXiv:2304.08485

  14. [22]

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. 2024 b . https://arxiv.org/abs/2303.05499 Grounding dino: Marrying dino with grounded pre-training for open-set object detection . Prep...

  15. [23]

    Liu, and Xuanhui Wang

    Tianqi Liu, Zhen Qin, Junru Wu, Jiaming Shen, Misha Khalman, Rishabh Joshi, Yao Zhao, Mohammad Saleh, Simon Baumgartner, Jialu Liu, Peter J. Liu, and Xuanhui Wang. 2024 c . https://arxiv.org/abs/2402.01878 Lipo: Listwise preference optimization through learning-to-rank . Prepr...

  16. [24]

    Yu Meng, Mengzhou Xia, and Danqi Chen. 2024. https://arxiv.org/abs/2405.14734 Simpo: Simple preference optimization with a reference-free reward . Preprint, arXiv:2405.14734

  17. [25]

    Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. 2024. https://arxiv.org/abs/2306.09683 Scaling open-vocabulary object detection . Preprint, arXiv:2306.09683

  18. [26]

    OpenAI. 2024. https://arxiv.org/abs/2410.21276 Gpt-4o system card . Preprint, arXiv:2410.21276

  19. [27]

    Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and...

  20. [28]

    Renjie Pi, Tianyang Han, Wei Xiong, Jipeng Zhang, Runtao Liu, Rui Pan, and Tong Zhang. 2024. https://arxiv.org/abs/2403.08730 Strengthening multimodal large language model with bootstrapped preference optimization . Preprint, arXiv:2403.08730

  21. [29]

    Manning, and Chelsea Finn

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2024. https://arxiv.org/abs/2305.18290 Direct preference optimization: Your language model is secretly a reward model . Preprint, arXiv:2305.18290

  22. [30]

    Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. 2021. https://arxiv.org/abs/2010.04592 Contrastive learning with hard negative samples . Preprint, arXiv:2010.04592

  23. [31]

    Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. 2019. https://arxiv.org/abs/1809.02156 Object hallucination in image captioning . Preprint, arXiv:1809.02156

  24. [32]

    Arık, and Tomas Pfister

    Pritam Sarkar, Sayna Ebrahimi, Ali Etemad, Ahmad Beirami, Sercan Ö. Arık, and Tomas Pfister. 2024. https://arxiv.org/abs/2405.18654 Data-augmented phrase-level alignment for mitigating object hallucination . Preprint, arXiv:2405.18654

  25. [33]

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. 2015. https://doi.org/10.1109/cvpr.2015.7298682 Facenet: A unified embedding for face recognition and clustering . In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), page 815–823. IEEE

  26. [34]

    Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. 2023. https://arxiv.org/abs/2304.06712 What does clip know about a red circle? visual prompt engineering for vlms . Preprint, arXiv:2304.06712

  27. [35]

    Feifan Song, Bowen Yu, Minghao Li, Haiyang Yu, Fei Huang, Yongbin Li, and Houfeng Wang. 2024. https://arxiv.org/abs/2306.17492 Preference ranking optimization for human alignment . Preprint, arXiv:2306.17492

  28. [36]

    Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul Christiano

    Nisan Stiennon, Long Ouyang, Jeff Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul Christiano. 2022. https://arxiv.org/abs/2009.01325 Learning to summarize from human feedback . Preprint, arXiv:2009.01325

  29. [37]

    Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, Kurt Keutzer, and Trevor Darrell. 2023. https://arxiv.org/abs/2309.14525 Aligning large multimodal models with factually augmented rlhf . Prepri...

  30. [38]

    Huang, Nan Xu, Sheng Zhang, Hoifung Poon, and Muhao Chen

    Fei Wang, Wenxuan Zhou, James Y. Huang, Nan Xu, Sheng Zhang, Hoifung Poon, and Muhao Chen. 2024 a . https://arxiv.org/abs/2406.11839 mdpo: Conditional preference optimization for multimodal large language models . Preprint, arXiv:2406.11839

  31. [39]

    Junyang Wang, Yuhang Wang, Guohai Xu, Jing Zhang, Yukai Gu, Haitao Jia, Jiaqi Wang, Haiyang Xu, Ming Yan, Ji Zhang, and Jitao Sang. 2024 b . https://arxiv.org/abs/2311.07397 Amber: An llm-free multi-dimensional benchmark for mllms hallucination evaluation . Preprint, arXiv:2311.07397

  32. [40]

    Rossi, Ruiyi Zhang, Subrata Mitra, Dimitris N

    Junda Wu, Zhehao Zhang, Yu Xia, Xintong Li, Zhaoyang Xia, Aaron Chang, Tong Yu, Sungchul Kim, Ryan A. Rossi, Ruiyi Zhang, Subrata Mitra, Dimitris N. Metaxas, Lina Yao, Jingbo Shang, and Julian McAuley. 2024. https://arxiv.org/abs/2409.15310 Visual prompting in multimodal large...

  33. [41]

    Liwei Wu, Cho-Jui Hsieh, and James Sharpnack. 2019. https://arxiv.org/abs/1803.00114 Sql-rank: A listwise approach to collaborative ranking . Preprint, arXiv:1803.00114

  34. [42]

    Wenyi Xiao, Ziwei Huang, Leilei Gan, Wanggui He, Haoyuan Li, Zhelun Yu, Hao Jiang, Fei Wu, and Linchao Zhu. 2024. https://arxiv.org/abs/2404.14233 Detecting and mitigating hallucination in large vision language models via fine-grained ai feedback . Preprint, arXiv:2404.14233

  35. [43]

    Yuxi Xie, Guanzhen Li, Xiao Xu, and Min-Yen Kan. 2024. https://arxiv.org/abs/2411.02712 V-dpo: Mitigating hallucination in large vision language models via vision-guided direct preference optimization . Preprint, arXiv:2411.02712

  36. [44]

    Haoran Xu, Amr Sharaf, Yunmo Chen, Weiting Tan, Lingfeng Shen, Benjamin Van Durme, Kenton Murray, and Young Jin Kim. 2024 a . https://arxiv.org/abs/2401.08417 Contrastive preference optimization: Pushing the boundaries of llm performance in machine translation . Preprint, arXi...

  37. [45]

    Jing Xu, Andrew Lee, Sainbayar Sukhbaatar, and Jason Weston. 2024 b . https://arxiv.org/abs/2312.16682 Some things are more cringe than others: Iterative preference optimization with the pairwise cringe loss . Preprint, arXiv:2312.16682

  38. [46]

    Zheng Yuan, Hongyi Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Fei Huang. 2023. https://arxiv.org/abs/2304.05302 Rrhf: Rank responses to align language models with human feedback without tears . Preprint, arXiv:2304.05302

  39. [47]

    Zihao Yue, Liang Zhang, and Qin Jin. 2024. https://arxiv.org/abs/2402.14545 Less is more: Mitigating multimodal hallucination from an eos decision perspective . Preprint, arXiv:2402.14545

  40. [48]

    Fatemeh Pesaran Zadeh, Juyeon Kim, Jin-Hwa Kim, and Gunhee Kim. 2024. https://arxiv.org/abs/2410.04064 Text2chart31: Instruction tuning for chart generation with automatic feedback . Preprint, arXiv:2410.04064

  41. [49]

    Zhiyuan Zhao, Bin Wang, Linke Ouyang, Xiaoyi Dong, Jiaqi Wang, and Conghui He. 2024. https://arxiv.org/abs/2311.16839 Beyond hallucinations: Enhancing lvlms through hallucination-aware direct preference optimization . Preprint, arXiv:2311.16839

  42. [50]

    Yiyang Zhou, Chenhang Cui, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. 2024. https://arxiv.org/abs/2402.11411 Aligning modalities in vision large language models via preference fine-tuning . Preprint, arXiv:2402.11411

  43. [51]

    Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B

    Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. 2020. https://arxiv.org/abs/1909.08593 Fine-tuning language models from human preferences . Preprint, arXiv:1909.08593

  44. [52]

    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...

  45. [53]

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