Pith. sign in

REVIEW 4 major objections 6 minor 87 references

This paper claims that a panel of vision-language models can use one another's five-axis Likert scores as preference data and improve their average on fifteen benchmarks from 48% to 57% with no human labels.

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 ↗

A panel of LVLMs that generate, evaluate, and learn from each other's outputs improves average benchmark scores by 9 points across 15 tasks.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Useful synthesis of panel-judges and self-rewarding for LVLMs, with strong ablations; the headline gain is likely inflated by benchmark overlap and the reward is never calibrated, but it deserves review. the 4 major comments →

arxiv 2509.01610 v1 pith:2N7PRQ3K submitted 2025-09-01 cs.CV

Improving Large Vision and Language Models by Learning from a Panel of Peers

Jefferson Hernandez , Jing Shi , Simon Jenni , Vicente Ordonez , Kushal Kafle This is my paper
classification cs.CV
keywords large vision-language modelspeer evaluationpreference optimizationself-improvementreward ensemblingknowledge transfermultimodal alignmentSimPO
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.

The reading

The paper tries to establish that a group of similar-capability vision-language models can improve themselves purely from each other's judgments. Each model answers unlabeled image-question prompts, and the whole panel grades every answer on five quality axes: helpfulness, correctness, coherence, complexity, and verbosity, with the averaged grades acting as rewards. After rejection sampling and pairing best with worst answers, the preference data fine-tunes all panel members. On fifteen vision-language benchmarks the average rises from 48% to 57% over three iterations, and a purposely weak-OCR member learns OCR from peers. The authors argue this is a scalable alternative to human-curated preference data, which is expensive, and to self-supervised preference data, which tends to hallucinate.

Core claim

The central claim is that peer evaluation—models acting simultaneously as answer generators and as judges—supplies a sufficient training signal for improving LVLMs without ground-truth answers or human preference labels. The paper builds a panel of three open 7B/8B vision-language models, samples 15 candidate answers per prompt from unlabeled image-question pairs, has every model score every answer on a five-axis rubric, and averages the scores into a reward. It keeps answers scoring at least 0.85, pairs the highest- and lowest-scoring answers as preferred and dispreferred responses, and fine-tunes all panel members with a reference-free preference objective (SimPO). Three iterations move th

What carries the argument

The machinery is the 'Panel-of-Peers': M autoregressive LVLMs used as both answer generators and judges. The judge prompt asks each model to rate another model's answer from 1 to 5 on Helpfulness, Correctness, Coherence, Complexity, and Verbosity, and the panel's mean vote becomes the reward. That reward is used to reject low-quality samples, to keep the best and worst remaining answers as chosen/dispreferred pairs, and to define the SimPO loss used to update every member. Repeating generation, evaluation, and fine-tuning is what lets a weak member inherit capabilities, such as OCR, that exist elsewhere in the panel.

Load-bearing premise

The method rests on the assumption that a model's self-assessed Likert score for another model's answer—helpfulness, correctness, coherence, complexity, verbosity—actually tracks whether that answer is right, with no calibration against human judgments or ground truth; if the panel reliably rewards fluent but false answers, the loop would amplify the panel's biases.

What would settle it

Take a random sample of several hundred of the preference pairs PoP constructs and have human annotators decide which of the two answers is actually correct, without seeing which the panel preferred. If the panel-chosen answer is correct no more often than the rejected one, the peer reward is not tracking correctness; equivalently, rerunning PoP with randomly shuffled rewards should erase the 9-point gain if the reward signal is doing the work.

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

If this is right

  • If correct, LVLM alignment no longer requires expensive human preference labels: a diverse panel of existing models can generate its own training signal from unlabeled image-question pairs.
  • Every panel member improves individually, so the method costs no extra inference capacity at deployment—one model can be used alone afterward.
  • A model that lacks a specific skill (the paper demonstrates OCR) can acquire it from peers that have it, so panels can repair individual weaknesses.
  • The reported gains plateau around the third iteration, suggesting diminishing returns and a practical stopping point for the self-improvement loop.
  • Because the panel as a zero-shot evaluator beats its average member at every scale tested, simply ensembling peer judgments can improve answer selection even without fine-tuning.

Where Pith is reading between the lines

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

  • One open consequence the paper leaves implicit: if peer rewards favor style over factual content, the loop could converge to confident, fluent but wrong answers; a human-calibration study on the preference pairs would reveal how far this risk is realized.
  • Several benchmarks are flagged as overlapping with the unlabeled training mixture, so part of the gain may be exposure to benchmark-like questions; held-out-only evaluation would isolate true generalization beyond the training distribution.
  • The method should scale to larger and more diverse panels, including frontier models, as long as members remain comparable in quality; a natural test is whether the same loop improves a panel of state-of-the-art models without any human feedback.
  • Best-of-N from a single model underperforms a diverse panel of one sample each, suggesting that diversity of generators, not just number of candidates, drives the gains; one could test this by varying both dimensions independently.
Share X Bluesky LinkedIn Reddit HN

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 / 6 minor

Summary. The paper introduces Panel-of-Peers (PoP), a self-improvement method for large vision-language models (LVLMs). A panel of three LLaVA-1.5 variants (Vicuna-7B, Mistral-7B, Llama3-8B) generates candidate answers to unlabeled image-question pairs sampled from Cambrian-7M. Each panel member scores every candidate on a five-axis Likert prompt (Helpfulness, Correctness, Coherence, Complexity, Verbosity); the mean score is used as a pseudo-reward, followed by rejection sampling and SimPO fine-tuning of all panel members, iterated three times. The central claim is that this peer-feedback loop raises the average score on 15 VLMEvalKit benchmarks from 48% to 57%, and that it can transfer abilities such as OCR from stronger to weaker peers. The paper also reports zero-shot ensemble evaluation and ablations over reward aggregation, scoring type, panel composition, and SFT-vs-SimPO objectives.

Significance. If the result holds, PoP is a practically attractive alignment method: it requires no human preference labels, no external reward model, and no extra ground-truth answers, and it improves every panel member. The paper has real strengths: a large and diverse evaluation suite, a transparent footnote marking four benchmarks whose training sets overlap the Cambrian-7M mixture, an SFT-on-ground-truth control, and a dedicated OCR-knowledge-transfer experiment. The ablations on reward aggregation and scoring type are thoughtful. However, the central claim rests on an unvalidated peer-generated reward signal, and the headline number aggregates benchmarks that may be contaminated by the same data used to construct the training signal. These issues are fixable, but they are load-bearing for the paper's main conclusion.

major comments (4)
  1. [§3.1, Eq. (2)] The load-bearing assumption is that R_μ(y_j) = mean_i π_i(·|p∪x∪y_j) is a reliable correctness signal. This reward is an uncalibrated average of five Likert axes scored by the same models that generate the candidates. The paper provides no correlation of R_μ with ground-truth correctness or human judgments, and no per-axis analysis. Axes such as Complexity (higher = more expert) and Verbosity (higher = more verbose) can reward style over factuality. Because every preference pair and the SimPO update in Eq. (2) are built on this reward, the reported 9-point gain could reflect amplification of panel biases rather than improved reasoning. The SFT-on-ground-truth control in §5.5 (54.0 vs 57.0) is a useful sanity check, but it covers only Vicuna and does not validate the reward as a correctness signal. Please add a calibration study on a labeled subset (e.g., Cambrian-7M ground-truth answers
  2. [Table 2] The headline '48% to 57%' aggregates 15 benchmarks, including four (†AI2D, †ScienceQA, †ChartQA, †TextVQA) whose training sets are observed in the Cambrian-7M mixture used to construct the self-improvement data (§4). Because the model is fine-tuned on preference pairs for the same image/question prompts that later appear in evaluation, gains on those benchmarks cannot be attributed to generalization. The † disclosure is to the authors' credit, but it does not repair the headline. Please report the primary average excluding all overlapping benchmarks (a quick calculation from Table 2 suggests the gain shrinks from about 9 to about 7 points), and verify that no unmarked benchmark split overlaps the Cambrian-7M mixture.
  3. [Tables 1, 2] No error bars, confidence intervals, or multiple runs are reported for any evaluation. Many headline differences between iterations and baselines are small (for example, several 1–2 point changes in Table 2), and the claim of consistent outperformance in §5.1 needs variance information. At minimum, report evaluation noise from VLMEvalKit (e.g., bootstrap CIs or multiple evaluation seeds) or multiple fine-tuning seeds for the central result. Without this, the precision of '48% to 57%' is not justified.
  4. [§3.3, Eq. (2)] The preference data construction selects the highest- and lowest-reward responses after rejection sampling, but the paper does not report how often the dispreferred response is actually incorrect rather than merely stylistically disfavored. Since SimPO maximizes the margin between preferred and dispreferred responses, noisy dispreferred labels can distort generation. Please include a human or rule-based audit of a sample of preference pairs, and report the fraction in which the preferred answer is factually correct (or at least strictly better) according to an independent judge. This is a concrete way to test the self-referential loop concern.
minor comments (6)
  1. [Eq. (2)] Equation (2) writes πθt(y(n)) and πθt(y(1)) without the logarithm; SimPO's loss uses log-likelihoods normalized by length. As written the expression is dimensionally implausible. Likely a typo, but it should be fixed for reproducibility.
  2. [Table 1] The text says 'For fairness ... comparing only to PoP performed for only one iteration,' but Table 1 also reports PoP-iter3 (900k samples) against methods trained on 80k or fewer samples. Please clarify which comparison is the controlled one and label data sizes consistently.
  3. [Table 2 caption] The table caption refers to colored symbols for PoP-Mistral, PoP-Vicuna, and PoP-LLaMA3, but the symbols are not visible in the text. Please ensure the column ordering is explicit and readable.
  4. [Related Work / Baselines] The baseline list contains 'Self-Rewarding [80], CSR [80]' — both point to the same reference. Self-Rewarding should likely cite [74] or a distinct entry. Please correct.
  5. [§4] The rejection threshold 0.85 and margin 0.75 are introduced as 'we found that this helps,' but no sensitivity analysis is shown. Given that these are free parameters of the method, please provide at least a small ablation or a plot.
  6. [§5.5] The SFT-on-ground-truth control reports 54.0 for one model. Please specify the exact hyperparameters (epochs, learning rate, data size) and whether the same 900K samples were used, so the comparison with PoP-Vicuna's 57.0 is interpretable.

Circularity Check

0 steps flagged

No significant circularity: the central 48→57 benchmark claim is checked on external evaluation suites, and no equation reduces the reported gain to the peer-reward definition by construction.

full rationale

The paper's load-bearing claim is the 9-point average improvement on fifteen benchmarks. That improvement is measured with VLMEvalKit on external benchmark suites, not with the panel's own reward signal. The training loop is self-referential in the broad sense that the same models generate and score candidate responses, but the evaluation is independent, so the loop is broken at the benchmark stage. No fitted parameter is renamed as a prediction: the rejection thresholds (0.85 reward and 0.75 margin) are hand-set, and the paper does not claim the benchmark scores are derived from them. The paper explicitly marks AI2D, ScienceQA, ChartQA, and TextVQA with a dagger because they overlap the Cambrian-7M prompt mixture, which is a contamination risk and weakens the 'unseen ability' interpretation, but it is not a case where an equation reduces the prediction to its input. The only plausible author-overlap citation is Cambrian-1 [55], used as a source of unlabeled prompts and images, not as a justification of the method's correctness; it is not load-bearing. No uniqueness theorem, ansatz-smuggling via self-citation, or renaming of a known result was found. The unvalidated peer reward is the weakest assumption, but that is an empirical correctness/alignment risk, not a circularity step under the definitions used here.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 0 invented entities

The method introduces no new physical entities. The central assumptions are that model self-evaluation is reliable, that the five scoring axes are valid, and that the Cambrian-7M samples do not overlap with test benchmarks. The free parameters are hyperparameters chosen by hand, not fitted to benchmark outcomes.

free parameters (4)
  • rejection_threshold = 0.85
    Responses with reward below 0.85 are discarded before preference construction; chosen by the authors without a systematic search.
  • reward_margin = 0.75
    Minimum reward gap between preferred and dispreferred answers; described as 'as we found that this helps preference optimization.'
  • num_candidates_per_model = 15
    Each panel member samples 15 responses for best-of-N selection; single-try variant st-PoP is the control.
  • reward_aggregation = mean voting
    Averaging panel member scores chosen for simplicity, despite uncertainty-weighted being slightly better in ablations.
axioms (3)
  • domain assumption LVLMs can reliably evaluate peer outputs without ground truth using a 5-axis Likert prompt
    Section 3.1 defines rewards purely from the judge's internal knowledge; no calibration against human judgments is provided. This is the load-bearing premise for constructing preference data.
  • domain assumption The five evaluation criteria (Helpfulness, Correctness, Coherence, Complexity, Verbosity) are valid measures of answer quality
    The prompt template in Figure A.1 assumes these dimensions capture quality; Complexity and Verbosity may reward longer, more complex answers regardless of correctness.
  • domain assumption Panel members are similar-capacity peers and their collective judgment is more reliable than individual judgment
    Stated in the introduction; relies on prior work PoLL for panels of judges, but not re-validated on the specific LVLM setting.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Large Vision and Language Models by Learning from a Panel of Peers." pith.science (2026). https://pith.science/paper/2N7PRQ3K

@misc{pith2026250901610,
  author       = {Pith},
  title        = {Pith review of: Improving Large Vision and Language Models by Learning from a Panel of Peers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2N7PRQ3K}},
  note         = {Machine review of arXiv:2509.01610}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Traditional alignment methods for Large Vision and Language Models (LVLMs) primarily rely on human-curated preference data. Human-generated preference data is costly; machine-generated preference data is limited in quality; and self-supervised preference data often introduces hallucinations. To overcome these limitations, we propose a novel Panel-of-Peers learning framework inspired by collaborative learning among humans. This approach leverages a panel of LVLMs, each evaluating and learning from their collective outputs through an iterative self-improvement process. By simulating a peer review system, our models generate, assess, and refine outputs in response to a curated set of prompts, mimicking a classroom learning environment. We demonstrate that this methodology enhances model performance without requiring extensive human-labeled datasets. Our experiments show significant improvement across multiple benchmarks, demonstrating the potential of peer evaluations as a scalable alternative to self-supervised alignment. Notably, we show that Panel-of-Peers increases the average score on fifteen benchmarks from 48% to 57%

Figures

Figures reproduced from arXiv: 2509.01610 by Jefferson Hernandez, Jing Shi, Kushal Kafle, Simon Jenni, Vicente Ordonez.

Figure 1
Figure 1. Figure 1: A Panel-of-Peers (PoP) generates candidate responses from multiple LVLMs. The panel’s scoring of these responses is used to build a preference set, which is used to tune one or all the members of the panel, improving their accuracy individually. PoP significantly outperforms other forms of reaching consensus across many benchmarks. The training process for LVLMs has largely converged into three stages: 1) … view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the overall learning from feedback from peers approach. Our post-alignment strategy involves rejection sampling, supervised finetuning, and preference optimization methods to learn from peers. See text for details. the alignment and reliability of LVLMs. Collecting high-quality multimodal data can be one of the most straightforward solutions but doing so on a large scale is often expensive,… view at source ↗
Figure 5
Figure 5. Figure 5: Learning a new skill from peers We start with a model with a limited knowl￾edge of OCR (≈0% - 100%) and use PoP to teach the model OCR knowledge. and vision-centric tasks suggests that enhanced OCR capa￾bilities contribute to better general alignment in multimodal understanding. This experiment shows how Peer-to-Peer Learning, with varying levels of specialized knowledge, can iteratively improve a model’s … view at source ↗
Figure 6
Figure 6. Figure 6: Peer-to-Peer Learning ablation experiments We use the LLaVA-1.5, configuration using the Vicuna-7B language model and CLIP/L-14 vision model except for the first ablation which is evaluation only. We evaluate on 15 selected benchmarks. Absolute scoring can be aggregated in a similar way. As seen in the figure, Absolute scoring with averaging produced the highest average score, followed closely by Absolute … 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

87 extracted references · 28 canonical work pages · 1 internal anchor

  1. [1]

    Phi-3 techni- cal report: A highly capable language model locally on your phone

    Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadal- lah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 techni- cal report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 17

    Pith/arXiv arXiv 2024

  2. [2]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2

    Pith/arXiv arXiv 2023

  3. [3]

    The llama 3 herd of models

    AI@Meta. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 4, 13, 17

    Pith/arXiv arXiv 2024

  4. [4]

    Theoretical guarantees on the best-of-n alignment policy

    Ahmad Beirami, Alekh Agarwal, Jonathan Berant, Alexan- der D’Amour, Jacob Eisenstein, Chirag Nagpal, and Ananda Theertha Suresh. Theoretical guarantees on the best-of-n alignment policy. arXiv preprint arXiv:2401.01879,

    Pith/arXiv arXiv

  5. [5]

    Paligemma: A versatile 3b vlm for transfer

    Lucas Beyer, Andreas Steiner, Andr´e Susano Pinto, Alexan- der Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arXiv preprint arXiv:2407.07726, 2024. 17

    Pith/arXiv arXiv 2024

  6. [6]

    Are we on the right way for evaluating large vision- language models? arXiv preprint arXiv:2403.20330, 2024

    Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision- language models? arXiv preprint arXiv:2403.20330, 2024. 5, 6, 14, 17

    Pith/arXiv arXiv 2024

  7. [7]

    Self-play fine-tuning converts weak lan- guage models to strong language models

    Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak lan- guage models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 3

    Pith/arXiv arXiv 2024

  8. [8]

    How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites

    Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 7, 15, 17

    Pith/arXiv arXiv 2024

  9. [9]

    Gonzalez, Ion Stoica, and Eric P

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, 2023. 1, 4, 13

    2023

  10. [10]

    Self-Improving Robust Preference Optimization

    Eugene Choi, Arash Ahmadian, Matthieu Geist, Oil- vier Pietquin, and Mohammad Gheshlaghi Azar. Self- improving robust preference optimization. arXiv preprint arXiv:2406.01660, 2024. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  11. [11]

    Training verifiers to solve math word problems

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 3

    Pith/arXiv arXiv 2021

  12. [12]

    Reward model ensembles help mitigate overop- timization

    Thomas Coste, Usman Anwar, Robert Kirk, and David Krueger. Reward model ensembles help mitigate overop- timization. In The Twelfth International Conference on Learn- ing Representations, 2024. 2

    2024

  13. [13]

    Nvlm: Open frontier-class multimodal llms

    Wenliang Dai, Nayeon Lee, Boxin Wang, Zhuolin Yang, Zi- han Liu, Jon Barker, Tuomas Rintamaki, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nvlm: Open frontier-class multimodal llms. arXiv preprint, 2024. 15

    2024

  14. [14]

    Flashattention: Fast and memory-efficient exact attention with io-awareness

    Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christo- pher R ´e. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344–16359, 2022. 13

    2022

  15. [15]

    Molmo and pixmo: Open weights and open data for state-of-the-art multi- modal models

    Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and pixmo: Open weights and open data for state-of-the-art multi- modal models. arXiv preprint arXiv:2409.17146, 2024. 15

    Pith/arXiv arXiv 2024

  16. [16]

    Enhancing large vision language models with self-training on image comprehension

    Yihe Deng, Pan Lu, Fan Yin, Ziniu Hu, Sheng Shen, James Zou, Kai-Wei Chang, and Wei Wang. Enhancing large vision language models with self-training on image comprehension. arXiv preprint arXiv:2405.19716, 2024. 3, 5, 6

    Pith/arXiv arXiv 2024

  17. [17]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 13

    Pith/arXiv arXiv 2010

  18. [18]

    Vlmevalkit: An open-source toolkit for evaluating large multi-modality models

    Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmevalkit: An open-source toolkit for evaluating large multi-modality models. arXiv preprint arXiv:2407.11691, 2024. 5, 13, 15

    Pith/arXiv arXiv 2024

  19. [19]

    vila2: Vila augmented vila

    Yunhao Fang, Ligeng Zhu, Yao Lu, Yan Wang, Pavlo Molchanov, Jang Hyun Cho, Marco Pavone, Song Han, and Hongxu Yin. vila2: Vila augmented vila. arXiv preprint arXiv:2407.17453, 2024. 3

    Pith/arXiv arXiv 2024

  20. [20]

    Hallusionbench: an advanced diagnostic suite for entangled language hallucination and vi- sual illusion in large vision-language models

    T Guan, F Liu, X Wu, R Xian, Z Li, X Liu, X Wang, L Chen, F Huang, Y Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and vi- sual illusion in large vision-language models. arxiv. 10.48550. arXiv preprint arXiv.2310.14566, 2023. 5, 6, 14, 17

    Pith/arXiv arXiv 2023

  21. [21]

    Cogvlm2: Visual language models for image and video understanding

    Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qing- song Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500,

    Pith/arXiv arXiv

  22. [22]

    Mistral 7b (2023)

    AQ Jiang, A Sablayrolles, A Mensch, C Bamford, DS Chaplot, D de las Casas, F Bressand, G Lengyel, G Lample, L Saulnier, et al. Mistral 7b (2023). arXiv preprint arXiv:2310.06825,

    Pith/arXiv arXiv 2023

  23. [23]

    A diagram is worth a dozen images

    Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, Octo- ber 11–14, 2016, Proceedings, Part IV 14 , pages 235–251. Springer, 2016. 5, 6, 14, 17

    2016

  24. [24]

    What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024

    Hugo Laurenc ¸on, L´eo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 17

    Pith/arXiv arXiv 2024

  25. [25]

    Prometheusvision: Vision-language model as a judge for fine-grained evaluation

    Seongyun Lee, Seungone Kim, Sue Hyun Park, Geewook Kim, and Minjoon Seo. Prometheusvision: Vision-language model as a judge for fine-grained evaluation. arXiv preprint arXiv:2401.06591, 2024. 3

    Pith/arXiv arXiv 2024

  26. [26]

    Seed-bench: Benchmark- ing multimodal large language models

    Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmark- ing multimodal large language models. In Proceedings of 9 the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13299–13308, 2024. 5, 6, 14, 17

    2024

  27. [27]

    Llava-onevision: Easy visual task transfer

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chun- yuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 15

    Pith/arXiv arXiv 2024

  28. [28]

    Blip: Bootstrapping language-image pre-training for unified vision- language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision- language understanding and generation. In International conference on machine learning, pages 12888–12900. PMLR,

  29. [29]

    Silkie: Preference distillation for large visual language models

    Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models. arXiv preprint arXiv:2312.10665, 2023. 5

    Pith/arXiv arXiv 2023

  30. [30]

    Multi-modal pref- erence alignment remedies degradation of visual instruction tuning on language models

    Shengzhi Li, Rongyu Lin, and Shichao Pei. Multi-modal pref- erence alignment remedies degradation of visual instruction tuning on language models. arXiv preprint arXiv:2402.10884,

    Pith/arXiv arXiv

  31. [31]

    Evaluating object hallucination in large vision- language models

    Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision- language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 5, 6, 14, 17

    work page 2023

  32. [32]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296–26306, 2024. 1, 3, 4, 5, 13

    work page 2024

  33. [33]

    Llava-next: Improved reason- ing, ocr, and world knowledge, 2024

    Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-next: Improved reason- ing, ocr, and world knowledge, 2024. 1, 7, 15, 17

    work page 2024

  34. [34]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 3, 4, 13

    work page 2024

  35. [35]

    Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023

    Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 2, 5, 6, 14, 17

    Pith/arXiv arXiv 2023

  36. [36]

    On the hidden mystery of ocr in large multimodal models

    Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery of ocr in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 5, 6, 14, 17

    Pith/arXiv arXiv 2023

  37. [37]

    Learn to explain: Multimodal reasoning via thought chains for science question answering

    Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507–2521, 2022. 5, 6, 14, 17

    work page 2022

  38. [38]

    Mathvista: Evaluating mathemat- ical reasoning of foundation models in visual contexts

    Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathemat- ical reasoning of foundation models in visual contexts. In The Twelfth International Conference on Learning Represen- tations, 2024. 5, 6, 14, 17

    work page 2024

  39. [39]

    Wildvision: Evaluating vision-language models in the wild with human preferences

    Yujie Lu, Dongfu Jiang, Wenhu Chen, William Yang Wang, Yejin Choi, and Bill Yuchen Lin. Wildvision: Evaluating vision-language models in the wild with human preferences. arXiv preprint arXiv:2406.11069, 2024. 2

    Pith/arXiv arXiv 2024

  40. [40]

    Chartqa: A benchmark for question answering about charts with visual and logical reasoning

    Ahmed Masry, Xuan Long Do, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2263–2279, 2022. 5, 6, 14, 17

    work page 2022

  41. [41]

    Simpo: Simple preference optimization with a reference-free reward

    Yu Meng, Mengzhou Xia, and Danqi Chen. Simpo: Simple preference optimization with a reference-free reward. arXiv preprint arXiv:2405.14734, 2024. 4, 7

    Pith/arXiv arXiv 2024

  42. [42]

    Ocr-vqa: Visual question answering by reading text in images

    Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947–952. IEEE, 2019. 5, 6, 14, 17

    work page 2019

  43. [43]

    Cognitive perspectives on peer learning

    Angela M O’donnell and Alison King. Cognitive perspectives on peer learning. Routledge, 2014. 2

    work page 2014

  44. [44]

    Gpt-4o system card, 2024

    OpenAI. Gpt-4o system card, 2024. Accessed: 2024-09-30. 15

    work page 2024

  45. [45]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth´ee Darcet, Th´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1

    Pith/arXiv arXiv 2023

  46. [46]

    Im2text: Describing images using 1 million captioned photographs

    Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. Advances in neural information processing systems, 24, 2011. 1, 13

    work page 2011

  47. [47]

    Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744,

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Car- roll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744,

  48. [48]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, pages 8748–8763. PMLR, 2021. 1, 13

    work page 2021

  49. [49]

    Direct prefer- ence optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct prefer- ence optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 1, 3

    work page 2024

  50. [50]

    Laion-400m: Open dataset of clip-filtered 400 million image-text pairs

    Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 13

    Pith/arXiv arXiv 2021

  51. [51]

    Laion-5b: An open large-scale dataset for training next gen- eration image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next gen- eration image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 1

    work page 2022

  52. [52]

    10 Towards vqa models that can read

    Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xin- lei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. 10 Towards vqa models that can read. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5, 6, 14, 17

    work page 2019

  53. [53]

    Aligning large multi- modal models with factually augmented rlhf

    Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chun- yuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu- Xiong Wang, Yiming Yang, et al. Aligning large multi- modal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 2, 3, 5

    Pith/arXiv arXiv 2023

  54. [54]

    Gemini: a family of highly capable multimodal models

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2, 3, 15

    Pith/arXiv arXiv 2023

  55. [55]

    Cambrian-1: A fully open, vision-centric exploration of multimodal llms

    Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 4, 13, 17

    Pith/arXiv arXiv 2024

  56. [56]

    Replacing judges with juries: Evaluating llm generations with a panel of diverse models

    Pat Verga, Sebastian Hofstatter, Sophia Althammer, Yixuan Su, Aleksandra Piktus, Arkady Arkhangorodsky, Minjie Xu, Naomi White, and Patrick Lewis. Replacing judges with juries: Evaluating llm generations with a panel of diverse models. arXiv preprint arXiv:2404.18796, 2024. 3, 4

    Pith/arXiv arXiv 2024

  57. [57]

    To see is to believe: Prompting gpt-4v for better visual instruction tuning

    Junke Wang, Lingchen Meng, Zejia Weng, Bo He, Zuxuan Wu, and Yu-Gang Jiang. To see is to believe: Prompting gpt-4v for better visual instruction tuning. arXiv preprint arXiv:2311.07574, 2023. 2

    Pith/arXiv arXiv 2023

  58. [58]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191,

    Pith/arXiv arXiv

  59. [59]

    Self-taught evalu- ators

    Tianlu Wang, Ilia Kulikov, Olga Golovneva, Ping Yu, Weizhe Yuan, Jane Dwivedi-Yu, Richard Yuanzhe Pang, Maryam Fazel-Zarandi, Jason Weston, and Xian Li. Self-taught evalu- ators. arXiv preprint arXiv:2408.02666, 2024. 3

    Pith/arXiv arXiv 2024

  60. [60]

    Enhancing visual- language modality alignment in large vision language mod- els via self-improvement

    Xiyao Wang, Jiuhai Chen, Zhaoyang Wang, Yuhang Zhou, Yiyang Zhou, Huaxiu Yao, Tianyi Zhou, Tom Goldstein, Parminder Bhatia, Furong Huang, et al. Enhancing visual- language modality alignment in large vision language mod- els via self-improvement. arXiv preprint arXiv:2405.15973,

    Pith/arXiv arXiv

  61. [61]

    Helpsteer2: Open-source dataset for training top-performing reward models

    Zhilin Wang, Yi Dong, Olivier Delalleau, Jiaqi Zeng, Ger- ald Shen, Daniel Egert, Jimmy J Zhang, Makesh Narsimhan Sreedhar, and Oleksii Kuchaiev. Helpsteer2: Open-source dataset for training top-performing reward models. arXiv preprint arXiv:2406.08673, 2024. 3, 4

    Pith/arXiv arXiv 2024

  62. [62]

    Helpsteer: Multi-attribute helpfulness dataset for steerlm

    Zhilin Wang, Yi Dong, Jiaqi Zeng, Virginia Adams, Makesh Narsimhan Sreedhar, Daniel Egert, Olivier Delalleau, Jane Scowcroft, Neel Kant, Aidan Swope, et al. Helpsteer: Multi-attribute helpfulness dataset for steerlm. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technolo...

    work page 2024

  63. [63]

    Pytorch image models

    Ross Wightman. Pytorch image models. https : / / github . com / rwightman / pytorch - image - models, 2019. 13

    work page 2019

  64. [64]

    Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation

    Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22227–22238, 2024. 2

    work page 2024

  65. [65]

    Self-play preference opti- mization for language model alignment

    Yue Wu, Zhiqing Sun, Huizhuo Yuan, Kaixuan Ji, Yim- ing Yang, and Quanquan Gu. Self-play preference opti- mization for language model alignment. arXiv preprint arXiv:2405.00675, 2024. 3

    Pith/arXiv arXiv 2024

  66. [66]

    Grok-1.5 vision preview, 2024

    x.ai Team. Grok-1.5 vision preview, 2024. 5, 6, 14, 17

    work page 2024

  67. [67]

    Llava- critic: Learning to evaluate multimodal models.arXiv preprint arXiv:2410.02712, 2024

    Tianyi Xiong, Xiyao Wang, Dong Guo, Qinghao Ye, Haoqi Fan, Quanquan Gu, Heng Huang, and Chunyuan Li. Llava- critic: Learning to evaluate multimodal models.arXiv preprint arXiv:2410.02712, 2024. 3, 5

    Pith/arXiv arXiv 2024

  68. [68]

    The perfect blend: Redefining rlhf with mixture of judges

    Tengyu Xu, Eryk Helenowski, Karthik Abinav Sankarara- man, Di Jin, Kaiyan Peng, Eric Han, Shaoliang Nie, Chen Zhu, Hejia Zhang, Wenxuan Zhou, et al. The perfect blend: Redefining rlhf with mixture of judges. arXiv preprint arXiv:2409.20370, 2024. 3

    Pith/arXiv arXiv 2024

  69. [69]

    xgen-mm (blip-3): A family of open large multimodal models

    Le Xue, Manli Shu, Anas Awadalla, Jun Wang, An Yan, Senthil Purushwalkam, Honglu Zhou, Viraj Prabhu, Yu- tong Dai, Michael S Ryoo, et al. xgen-mm (blip-3): A family of open large multimodal models. arXiv preprint arXiv:2408.08872, 2024. 17

    arXiv 2024

  70. [70]

    Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback

    Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, et al. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13807–13816, 2024. 5

    work page 2024

  71. [71]

    Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness

    Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, and Maosong Sun. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. arXiv preprint arXiv:2405.17220, 2024. 2

    arXiv 2024

  72. [72]

    Mm-vet: Evaluating large multimodal models for integrated capabilities

    Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. In Forty-first International Conference on Ma- chine Learning, 2024. 2, 5, 6, 14, 17

    work page 2024

  73. [73]

    Florence: A new foundation model for computer vision

    Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432,

    Pith/arXiv arXiv

  74. [74]

    Self-rewarding language models

    Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Sain- bayar Sukhbaatar, Jing Xu, and Jason Weston. Self-rewarding language models. arXiv preprint arXiv:2401.10020, 2024. 3

    Pith/arXiv arXiv 2024

  75. [75]

    Mmmu: A massive multi-discipline multi- modal understanding and reasoning benchmark for expert agi

    Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multi- modal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556–9567, 2024. 5, 6, 14, 17 11

    work page 2024

  76. [76]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975–11986, 2023. 1

    2023

  77. [77]

    Svit: Scaling up visual instruction tuning

    Bo Zhao, Boya Wu, and Tiejun Huang. Svit: Scaling up visual instruction tuning. arXiv preprint arXiv:2307.04087, 2023. 2

    Pith/arXiv arXiv 2023

  78. [78]

    Pytorch fsdp: experiences on scaling fully sharded data parallel

    Yanli Zhao, Andrew Gu, Rohan Varma, Liang Luo, Chien- Chin Huang, Min Xu, Less Wright, Hamid Shojanazeri, Myle Ott, Sam Shleifer, et al. Pytorch fsdp: experiences on scaling fully sharded data parallel. arXiv preprint arXiv:2304.11277,

    Pith/arXiv arXiv

  79. [79]

    Aligning modalities in vision large lan- guage models via preference fine-tuning

    Yiyang Zhou, Chenhang Cui, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Aligning modalities in vision large lan- guage models via preference fine-tuning. arXiv preprint arXiv:2402.11411, 2024. 3, 5

    Pith/arXiv arXiv 2024

  80. [80]

    Calibrated self-rewarding vision language models

    Yiyang Zhou, Zhiyuan Fan, Dongjie Cheng, Sihan Yang, Zhaorun Chen, Chenhang Cui, Xiyao Wang, Yun Li, Lin- jun Zhang, and Huaxiu Yao. Calibrated self-rewarding vision language models. arXiv preprint arXiv:2405.14622, 2024. 3, 5, 6

    Pith/arXiv arXiv 2024

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.