Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

EmoFeedback^2 claims that an LVLM feedback loop — scoring generated images on valence and arousal, rewarding the generator, and rewriting the prompt — yields continuous emotional control that beats current text-to-image methods.

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

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

T0 review · deepseek-v4-flash

2026-08-03 20:21 UTC pith:JC7G3HBZ

load-bearing objection The framework is a reasonable novel assembly, but the central SOTA claim collapses because the same LVLM trained on synthetic labels is both the reward model and the evaluation metric, and the paper's own human check shows errors jump 50-84%. the 4 major comments →

arxiv 2511.19982 v3 pith:JC7G3HBZ submitted 2025-11-25 cs.CV cs.AI

EmoFeedback²: Reinforcement of Continuous Emotional Image Generation via LVLM-based Reward and Textual Feedback

Kai Shu , Jingyang Jia , Gang Yang , Long Xing , Xun Chen , Aiping Liu This is my paper
classification cs.CV cs.AI
keywords continuous emotional image generationvalence-arousal modelemotion-aware reward feedbacklarge vision-language modeltextual prompt feedbackreinforcement fine-tuningemotional fidelitytext-to-image generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Continuous emotional image generation usually works by injecting a valence-arousal (V-A) target into the text prompt and hoping the diffusion model obeys it. EmoFeedback^2 claims that the missing piece is feedback: an emotion-understanding vision-language model (LVLM) looks at each generated image, scores its V-A values, uses the gap from the target as a reward to fine-tune the generator, and at inference rewrites the prompt based on what the image is still missing. The authors argue this generation-understanding-feedback loop gives users continuous control over emotional expression — shifting lighting, color, objects, and atmosphere along the valence and arousal axes — while preserving semantic content. They report that the loop beats existing emotion-conditioned and general text-to-image methods on V-A error, image-text alignment, perceptual quality, and user preference, including on an unseen cross-domain dataset. If true, this turns emotional image generation from a categorical trick into a continuously tunable, self-correcting process.

Core claim

The paper claims that the missing ingredient in continuous emotional image generation is feedback: previous methods inject valence-arousal (V-A) values into the prompt and hope the generator obeys them, but never look at the emitted image. EmoFeedback^2 closes that loop. A large vision-language model is first fine-tuned to predict V-A scores and emotion categories from images; it is then used as a reward model to fine-tune the text-to-image generator through group-relative policy optimization, so the generator is directly pushed toward images whose predicted emotion matches the target. At inference, the same LVLM compares the best and worst generated samples against the target emotion, produ

What carries the argument

The load-bearing mechanism is the emotion-understanding LVLM: a vision-language model fine-tuned by group-relative policy optimization under three reward terms — a format reward, a step-thresholded valence-arousal regression reward, and an eight-class emotion classification reward. This LVLM does double duty. During training, its V-A score discrepancy with the target becomes a normalized advantage signal that reinforcement-fine-tunes the diffusion generator, alongside a human-preference-based quality reward that suppresses reward hacking. At inference, it powers self-promotion textual feedback: the emotional gap is treated as a 'loss,' the LVLM's natural-language suggestions as a 'gradient,'

Load-bearing premise

The whole pipeline takes the emotion-understanding LVLM's valence-arousal predictions as the measure of emotional fidelity, but that model was trained on synthetic V-A labels randomly sampled from category-level lexicon statistics rather than on per-image human emotion ratings — so if those predictions do not track what people actually feel, both the reward and the evaluation lose their ground truth.

What would settle it

Ask independent human raters to score valence and arousal (1–9) for images generated by EmoFeedback^2 and by each baseline against the same target V-A values. If the mean absolute errors computed from those human ratings are not lower for EmoFeedback^2, the claimed emotional-fidelity gain is not real. A faster check: measure the correlation between the emotion-understanding LVLM's V-A predictions and per-image human ratings on a held-out set; low correlation would mean the reward signal is not tracking human emotion perception.

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

If this is right

  • Continuous valence-arousal control becomes usable in practice: users can give only a mood or V-A numbers and receive images whose lighting, color, atmosphere, and objects shift as the target emotion shifts.
  • Training-time reward feedback and test-time textual feedback each contribute separately: reward feedback anchors the initial emotional content, textual feedback refines details, and the combination outperforms either alone.
  • Jointly optimizing V-A regression and emotion classification in the LVLM improves both, and the resulting model transfers to an unseen domain, lowering V-A error on the public benchmark relative to baselines.
  • Emotional fidelity can be raised without sacrificing semantic fidelity: the method also improves image-text alignment and perceptual quality, and the human-preference reward term blocks the reward-hacking failure mode where content quality degrades.
  • The design is deployable at variable compute budgets: a single generation round costs roughly the same as a lightweight baseline, and users can choose to iterate for higher emotional fidelity.

Where Pith is reading between the lines

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

  • Editorial: because the same LVLM both produces the training reward and scores the final evaluation, its own bias can be self-consistent; an independent human-rated benchmark is the only way to verify that the fidelity gain is real rather than a closed-loop artifact.
  • Editorial: the 'textual gradient descent' framing is generic — the same loss-analyze-rewrite cycle could steer style, humor, or narrative tone by swapping the LVLM's evaluation instruction.
  • Editorial: a small set of per-image human V-A ratings could calibrate the LVLM's predictions before they enter the reward; the gap between lexicon-based and human-annotated errors reported in Table 6 suggests such calibration would make the gains more convincing.
  • Editorial: the test-time loop is a form of inference-time optimization; tracking how much V-A error drops per iteration and how prompt length and complexity grow would reveal when additional rounds stop paying off.

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 proposes EmoFeedback^2, a generation-understanding-feedback reinforcement paradigm for continuous emotional image generation (C-EICG). A Qwen2.5-VL-7B model is fine-tuned with GRPO on a custom dataset built from EmoSet-118K to perform emotion understanding (V-A regression and emotion classification). This fine-tuned LVLM is then used as (i) the reward model for reinforcement fine-tuning of SD3.5-Medium and (ii) the scoring model for the headline V-Error/A-Error metrics. A second, inference-time component iteratively uses the LVLM to compare best/worst generated images and rewrite prompts for better emotional fidelity. The paper reports state-of-the-art results against EmotiCrafter, EmoEdit, SD3.5-Large, and FLUX on V-Error/A-Error and image-quality metrics, plus ablations and a user study.

Significance. If the central claim were established, the paper would make a useful contribution to continuous emotion control in text-to-image generation by showing that an LVLM-based reward and an iterative prompt-refinement mechanism can improve emotional fidelity. The proposed pipeline is coherent, and the paper includes useful ablations (e.g., the PickScore reward to mitigate reward hacking, multi-task vs. single-task emotion training). However, the central quantitative claim is currently not supported: the model used as the reward signal is the same model used as the evaluation metric, and both are trained on synthetic V-A labels sampled from category-level Gaussians. The paper's own human-annotation check shows a large degradation in V-Error/A-Error and does not evaluate baselines under human ground truth. As a result, the claimed SOTA outperformance is not established.

major comments (4)
  1. [§4.1, §3.2, Tables 1–2] The evaluation loop is circular. The emotion-understanding LVLM is fine-tuned on V-A labels that are randomly sampled from per-category Gaussian distributions built from Warriner et al. lexicon statistics (§4.1), not from human ratings of the images. The same LVLM is then used both as the reward model in Eq. (1) and as the scorer for V-Error/A-Error in Tables 1–2. The generator is therefore optimized to satisfy the same model that measures the headline outcome, and the custom test targets come from the same synthetic sampling process. Improvements in V-Error/A-Error may simply reflect the generator learning the judge's category-conditioned mapping rather than improved human-perceived emotional fidelity. This is a load-bearing issue for the central outperformance claim.
  2. [§4.4.3, Table 6] The paper's own validity check does not resolve the circularity and instead weakens the claim. Under per-image human annotation by eight experts, V-Error rises from 0.521 to 0.781 and A-Error from 0.710 to 1.310, i.e., relative increases of roughly 50% and 84%. The text calls these 'comparable,' but no statistical test or per-image agreement measure is provided, and Figure 6 only compares sorted distributions, which can look similar even when per-image predictions are poor. Critically, none of the baseline methods are evaluated under human-annotated ground truth, so the claimed SOTA advantage may disappear or reverse when emotional fidelity is measured by humans rather than by the model being optimized.
  3. [§4.1, Tables 1–2] The custom dataset's V-A targets are not a valid ground truth for continuous emotional values. For each image, V-A values are sampled independently from a Gaussian distribution determined only by the image's discrete emotion category. This makes the targets synthetic and ignores image-specific content. Since the same synthetic targets are used for training the reward model and for evaluating all methods, the numerical V-Error/A-Error values in Tables 1–2 should not be interpreted as emotional-fidelity measurements. At minimum, the paper must re-evaluate with human-annotated targets for all methods, not only for the proposed method.
  4. [§4.3.3, Table 3] The user study provides some human evidence for qualitative preference, but it does not compensate for the flawed V-Error/A-Error evaluation. Participants were asked which image 'best conveys the target emotion' among images from different methods, but the target emotion is derived from the same synthetic V-A labels, and the study does not ask participants to rate V/A values. Preference of 53.83% may reflect stylistic differences rather than accurate continuous emotion control. The quantitative SOTA claim therefore rests entirely on the circular metric.
minor comments (6)
  1. [Abstract / §4.3.2] The arXiv abstract claims outperformance on 'both our custom dataset and public dataset,' but the full-text abstract only mentions the custom dataset. The EMOTIC experiment is presented as an extra validation, but the main abstract should be consistent.
  2. [Appendix B.2 / §3.1] Hyperparameters α1, α2, and threshold ε are mentioned in the appendix but not defined in the main text. The threshold τ in Figure 2 is never given a numeric value, and its relation to ε is unclear. Please define all reward parameters where the rewards are introduced.
  3. [Appendix C, Table 7] The reported per-sample inference latency of 6.4s is hard to reconcile with the statement that generating 8 images takes 11s and LVLM evaluation/refinement takes 9s. If one round of 8 images plus feedback takes about 20s, the per-sample latency should be about 2.5s, not 6.4s. Please clarify what is included in the latency.
  4. [Figure 2] The reward formulas contain typographical issues (e.g., '＞' instead of '>', and the ternary logic is not fully formal). Please rewrite the reward definitions in proper mathematical notation.
  5. [§4.4.1, Table 4] The variant 'Qwen-7B-C' is described as using a continuous reward function, but the exact functional form is not given. Without this, the ablation is not reproducible.
  6. [§4.1] The MLLM used to generate neutral and emotional prompts for the dataset is not identified. Please specify the model and provide examples or validation of the prompt-generation process.

Circularity Check

2 steps flagged

The custom-dataset emotional-fidelity claim reduces to the same LVLM/synthetic-label loop used for both reward and metric.

specific steps
  1. fitted input called prediction [Sec. 4.1 (Dataset Construction)]
    "we extract the mean and standard deviation of the V–A values for each emotion category. Based on these statistics, we build independent Gaussian distributions for each emotion label and randomly sample a pair of the V–A values from the corresponding distribution for each image."

    The V-A values created by this random Gaussian sampling are the ground-truth labels used to GRPO-fine-tune the emotion LVLM (Sec 3.1) and are the target values against which V-Error/A-Error are computed in Tables 1-2. Since the same paragraph creates a 1,000-sample test split by the same procedure, both the judge and the target come from the same fitted distribution. Reporting low V-Error/A-Error on this custom dataset therefore measures agreement with a synthetic, category-conditioned label generator, not with per-image human emotional ratings.

  2. self definitional [Sec. 3.2 (Eq. 1); Sec. 4.2 evaluation metrics]
    "We design an emotion-aware reward feedback strategy for the reinforcement fine-tuning of SD3.5-M by employing the emotion understanding model to assess the generated images. ... Next, the emotion understanding model predicts V–A scores of each image and computes reward values as reward feedback."

    The RL objective in Eqs. (1)-(2) trains the generator using advantages derived from the fine-tuned LVLM's V-A predictions against the synthetic targets. Tables 1-2 define V-Error/A-Error as the absolute error between predicted and target V-A values, i.e., the same comparison that constitutes the reward. Thus the generator is explicitly optimized to reduce the headline evaluation metric; outperforming baselines that were not trained against this same LVLM judge is by construction, not independent evidence of better human-perceived emotion.

full rationale

The central custom-dataset outperformance claim is circular by construction. Sec 4.1 generates per-image V-A ground truth by randomly sampling from per-category Gaussian distributions built from lexicon statistics; these synthetic labels are used to fine-tune the emotion LVLM (Sec 3.1) and also serve as targets for V-Error/A-Error in Tables 1-2. The same fine-tuned LVLM is then employed both as the RL reward model for SD3.5-M (Sec 3.2, Eq. 1) and as the scorer behind the headline V-Error/A-Error metrics (Sec 4.2). The generator is therefore trained to satisfy the very model that measures the outcome, making the reported improvement a fit to the judge rather than an independent measure of emotional fidelity. The paper's own human re-annotation check (Sec 4.4.3, Table 6) does not break the loop: V-Error rises from 0.521 to 0.781 and A-Error from 0.710 to 1.310, roughly 50% and 84% relative increases, yet the text calls these rates 'comparable,' and no baseline is re-evaluated with human ground truth. Some independent evidence exists (the 20-subject user study in Table 3 and the EMOTIC cross-domain results in Table 2), so the paper is not wholly vacuous, but the principal quantitative SOTA claim on the custom dataset reduces to optimizing and then measuring with the same fitted judge; the conclusion's own caveat that 'emotion understanding is highly subjective' further undercuts the unified synthetic-label metric. Score 8 reflects that the central claim is forced by the definition of the reward/evaluation loop, while not reaching 10 because a few independent checks remain.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claim rests on a small set of modeling choices. The most fragile are the synthetic V-A label generator and the use of the same trained LVLM as both reward and evaluator. No genuinely new conceptual entities are introduced.

free parameters (4)
  • V-A discrepancy threshold tau = 0.70
    Controls when the V-A regression reward is 1, 0.5, or 0; set in Appendix B.2 and directly shapes the reward used to train both the judge and the generator.
  • Reward weights alpha1, alpha2 = 0.25, 0.75
    Weights for the reward terms in GRPO training of the emotion understanding model; specified in Appendix B.2 but not derived from any principle.
  • PickScore reward weight = unspecified
    PickScore is added as an extra reward to prevent reward hacking, but the weighted combination is only described qualitatively in Section E.1.1; the actual weight is not reported.
  • Per-category Gaussian V-A target sampler = means/std from Warriner lexicon; per-image random draws
    Target V-A labels are randomly sampled from category-level Gaussians rather than human-per-image ratings (Sec 4.1). These sampled values become the ground truth for training the LVLM and the targets for the main V-Error/A-Error evaluation.
axioms (5)
  • domain assumption The Warriner lexicon V-A norms for words transfer to whole-image emotion categories.
    Sec 4.1 builds per-emotion Gaussian label distributions from lexicon statistics; if this transfer is invalid, the synthetic labels are meaningless.
  • domain assumption An image's emotional content can be summarized by one (V,A) pair and one of eight discrete categories.
    Central task definition inherited from EmotiCrafter; used throughout the rewards and the evaluation metrics.
  • domain assumption The fine-tuned Qwen2.5-VL LVLM produces valid emotion judgments after GRPO training on synthetic labels.
    Sec 3.1; this model is the sole reward source and the sole evaluator for emotional fidelity in the main tables.
  • domain assumption PickScore is a valid proxy for semantic consistency and human preference and offsets reward hacking.
    Sec 3.2 and E.1.1 rely on PickScore to prevent content distortion; support is qualitative and no quantitative validation is provided.
  • standard math The GRPO/Flow-GRPO objectives are valid for fine-tuning the diffusion/flow-matching generator.
    Equations (1) and (2) are imported from Shao et al. (2024) and Liu et al. (2025); accepted as standard in the RL fine-tuning literature.

pith-pipeline@v1.3.0-alltime-deepseek · 15784 in / 10422 out tokens · 104435 ms · 2026-08-03T20:21:22.554098+00:00 · methodology

0 comments
read the original abstract

Continuous emotional image content generation (C-EICG) is emerging rapidly due to its ability to produce images aligned with both user descriptions and continuous emotional values. However, existing approaches lack emotional feedback from generated images, limiting the control of emotional continuity. Additionally, their simple emotion-text alignment fails to adaptively adjust emotional prompts according to image content, leading to insufficient emotional fidelity. To address these concerns, we propose a novel generation-understanding-feedback reinforcement paradigm (EmoFeedback$^2$) for C-EICG, which exploits the reasoning capability of the fine-tuned large vision-language model (LVLM) to provide reward and textual feedback for generating high-quality images with continuous emotions. Specifically, we introduce an emotion-aware reward feedback strategy, where the LVLM evaluates the emotional values of generated images and computes the reward against target emotions, guiding the reinforcement fine-tuning of the generative model and enhancing the emotional continuity of images. Furthermore, we design a self-promotion textual feedback framework, in which the LVLM iteratively analyzes the emotional content of generated images and adaptively produces refinement suggestions for the next-round prompt, improving the emotional fidelity with fine-grained content. Extensive experimental results demonstrate that our approach effectively generates high-quality images with the desired emotions, outperforming existing state-of-the-art methods on both our custom dataset and public dataset.

Figures

Figures reproduced from arXiv: 2511.19982 by Aiping Liu, Gang Yang, Jingyang Jia, Kai Shu, Long Xing, Xun Chen.

Figure 1
Figure 1. Figure 1: The framework of the EmoFeedback2 . During training, given a neutral prompt, an emo￾tional prompt, the V and A scores, the generative model produces a set of images. The emotion understanding model then evaluates the images to provide reward feedback. During testing, the emotional prompt is omitted due to the users’ affective gap, and the model instead iteratively gener￾ates textual feedback to refine the … view at source ↗
Figure 2
Figure 2. Figure 2: The Emotion Understanding Model Training Process. The training image is input into the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparisons with baselines under specific emotional states. Our approach [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparisons with baselines under varying emotional values. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ablation study on the reward and textual feedback. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Additional Ablation Study on the valence distribution difference between human annota [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Additional Ablation Study results under reward hacking. [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Additional Qualitative results under specific emotional states. [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Additional Qualitative results under specific emotional states. [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Additional Qualitative results under specific emotional states. [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Additional Qualitative results under varying emotional states. The original neutral [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Additional Qualitative results under varying emotional states. The original neutral [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Additional Qualitative results under varying emotional states. The original neutral [PITH_FULL_IMAGE:figures/full_fig_p025_13.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

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

  1. What Can I Edit? Open-Ended Strategy Discovery and the Emotion Editability Landscape

    cs.CV 2026-07 conditional novelty 6.0

    An open-ended 'what can I edit' planner that discovers image-specific emotional affordances before editing produces edits humans strongly prefer over fixed-taxonomy emotional image editors, while standard emotion clas...

  2. CogBlender: Towards Continuous Cognitive Intervention in Text-to-Image Generation

    cs.CV 2026-03 unverdicted novelty 6.0

    CogBlender enables continuous multi-dimensional control over cognitive properties (valence, arousal, dominance, memorability) in text-to-image generation by creating discrete prompt variants and interpolating their ve...

Reference graph

Works this paper leans on

10 extracted references · 7 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923,

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923,

    Pith/arXiv arXiv

  2. [9]

    Msnet: A deep architecture using multi-sentiment semantics for sentiment-aware image style transfer

    Shikun Sun, Jia Jia, Haozhe Wu, Zijie Ye, and Junliang Xing. Msnet: A deep architecture using multi-sentiment semantics for sentiment-aware image style transfer. InICASSP 2023-2023 IEEE international conference on acoustics, speech and signal processing (ICASSP), pp. 1–5. IEEE,

    2023

  3. [10]

    differentiation

    Mert Yuksekgonul, Federico Bianchi, Joseph Boen, Sheng Liu, Zhi Huang, Carlos Guestrin, and James Zou. Textgrad: Automatic "differentiation" via text.arXiv preprint arXiv:2406.07496,

    Pith/arXiv arXiv

  4. [1980]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    Pith/arXiv arXiv

  5. [2015]

    Learning multi-level deep representations for image emotion classification.Neural processing letters, 51(3):2043–2061,

    Tianrong Rao, Xiaoxu Li, and Min Xu. Learning multi-level deep representations for image emotion classification.Neural processing letters, 51(3):2043–2061,

    2043

  6. [2018]

    Flow-grpo: Training flow matching models via online rl.arXiv preprint arXiv:2505.05470,

    Jie Liu, Gongye Liu, Jiajun Liang, Yangguang Li, Jiaheng Liu, Xintao Wang, Pengfei Wan, Di Zhang, and Wanli Ouyang. Flow-grpo: Training flow matching models via online rl.arXiv preprint arXiv:2505.05470,

    Pith/arXiv arXiv

  7. [2020]

    Emoticrafter: Text-to- emotional-image generation based on valence-arousal model.arXiv preprint arXiv:2501.05710,

    Shengqi Dang, Yi He, Long Ling, Ziqing Qian, Nanxuan Zhao, and Nan Cao. Emoticrafter: Text-to- emotional-image generation based on valence-arousal model.arXiv preprint arXiv:2501.05710,

    arXiv

  8. [2022]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    Pith/arXiv arXiv

  9. [2024]

    Q-insight: Understanding image quality via visual reinforcement learning.arXiv preprint arXiv:2503.22679, 2025a

    Weiqi Li, Xuanyu Zhang, Shijie Zhao, Yabin Zhang, Junlin Li, Li Zhang, and Jian Zhang. Q-insight: Understanding image quality via visual reinforcement learning.arXiv preprint arXiv:2503.22679, 2025a. Yafu Li, Xuyang Hu, Xiaoye Qu, Linjie Li, and Yu Cheng. Test-time preference optimization: On-the-fly alignment via iterative textual feedback. InForty-secon...

    Pith/arXiv arXiv

  10. [2025]

    Clipscore: A reference-free evaluation metric for image captioning.arXiv preprint arXiv:2104.08718,

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning.arXiv preprint arXiv:2104.08718,

    Pith/arXiv arXiv