On Fairness of Unified Multimodal Large Language Model for Image Generation

Bhiksha Raj Ramakrishnan; Hao Chen; Jindong Wang; Liwen Wang; Ming Liu; Wensheng Zhang

REVIEW 6 major objections 5 minor 1 cited by

On Fairness of Unified Multimodal Large Language Model for Image Generation

T0 review · 6 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The paper claims that demographic bias in unified multimodal LLM image generation originates mostly in the language model's token predictions, and that a balanced preference loss reduces VILA-U gender bias from 0.89 to 0.25 while…

desk verdict Useful first bias benchmark for U-MLLMs, but the localization and mitigation claims rest on unvalidated labels and a flawed understanding protocol. read the letter →

arxiv 2502.03429 v1 pith:SK2CPNOF submitted 2025-02-05 cs.CL cs.AI

Ming Liu , Hao Chen , Jindong Wang , Liwen Wang , Bhiksha Raj Ramakrishnan , Wensheng Zhang This is my paper · ORCID

classification cs.CLcs.AI

keywords multimodallargelanguagemodelsdemographicbiasimagegenerationfairnessbalancedpreferenceoptimizationgenderracelocalizationtext-to-image

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 asks whether unified multimodal large language models—systems that understand and generate images through a single autoregressive token stream—inherit or amplify demographic stereotypes when asked to create pictures from neutral text prompts. Benchmarking six recent U-MLLMs, it finds that gender, race, and gender-race intersection biases are pervasive, often exceeding those of diffusion-only models such as Stable Diffusion. The paper argues, via a locate-then-fix audit, that the bias enters at the language model's image-token prediction stage rather than in the vision encoder or decoder. It then introduces a balanced preference loss that makes the model's preference between demographic groups symmetric, and reports that on VILA-U this cuts gender bias from 0.89 to 0.25—a 71.91% reduction—while keeping or slightly improving semantic fidelity. If the claim holds, it gives a practical target for debiasing future unified generators.

What carries the argument

Two mechanisms carry the argument. The first is the bias-localization audit: using Jensen-Shannon divergence to compare the empirical distribution over image tokens sampled for a neutral prompt with the distribution for demographic-augmented prompts; the close match between neutral and majority-demographic distributions is the evidence that bias lives in token generation. The second is the balanced preference loss, $L_{\mathrm{bal}}(\theta)=\log\left[1+\left(\sigma\left(\log \mathrm{OR}_\theta(y_{d_i},y_{d_j})\right)-\frac12\right)^2\right]$ summed over demographic pairs, where $\mathrm{OR}_\theta$ is the odds ratio of the model's probability of generating one demographic variant over another. This loss penalizes any pair whose preference probability deviates from $1/2$, so it drives the policy toward equal likelihood across demographics; it is applied after a supervised finetuning stage, with LoRA for parameter-efficient updates.

What would settle it

Generate 100 images from FLUX.1-dev for prompts like "a photo of an Asian nurse" and ask human raters to label the gender and ethnicity actually depicted; if the diffusion model rarely renders the requested attribute, the training pairs are mislabeled and the balanced preference loss is optimizing on the wrong demographic groups. A second check is to replace the pretrained attribute classifier in the evaluation with human-annotated demographic labels for the same generated images and recompute representation disparity; if the disparity scores diverge sharply, the reported reductions may be classifier artifacts rather than true demographic balance.

Watch

Extended reading notes

Core claim

The central claim on the paper's own terms is that demographic bias in U-MLLM text-to-image generation originates primarily in the language model's autoregressive distribution over image tokens, not in the vision encoder or decoder that turns images into and out of those tokens. The supporting analysis compares empirical token distributions: for a neutral prompt such as "a photo of a firefighter", the distribution over image tokens is far closer to that of "a photo of a male firefighter" than to "a photo of a female firefighter", matching the majority demographic actually generated. The paper further claims that its balanced preference loss reduces that skew: VILA-U's gender representation disparity falls from 0.89 to 0.25 and its inception score rises to 2.10, with similar but smaller gains for race and intersectional bias while CLIP-based semantic alignment remains essentially flat. It also reports a "partial alignment" phenomenon: models with low or guarded understanding bias can still generate heavily skewed images, and finetuning for understanding does not reduce generation bias.

Load-bearing premise

The whole mitigation chain assumes that the diffusion model used to create the synthetic training images actually depicts the demographic attribute written in the prompt, since each image is labeled by its prompt with no classifier or human check.

Editorial extensions

If this is right

For VILA-U, gender bias falls from 0.89 to 0.25 (a 71.91% reduction) while the inception score rises from 1.87 to 2.10, indicating that demographic balancing need not cost image quality on this model.
Race debiasing on VILA-U reduces race bias from 0.48 to 0.26 and improves the inception score to 2.31, while the combined gender-race bias drops to 0.15.
Finetuning the same balanced data in the image-to-text (understanding) direction does not reduce generation bias, so understanding-level debiasing does not carry over to generation.
The same two-stage protocol applied to TokenFlow lowers its gender bias from 0.84 to 0.55, suggesting the approach transfers to other U-MLLMs.
Because the bias is localized to image-token generation, model builders can target the language model's conditional distribution over those tokens rather than replacing the vision tokenizer.

Reading between the lines

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

If the localization result holds, decoding-time interventions that steer the language model's image-token distribution toward uniform demographic likelihood could debias generation without any finetuning; the paper does not test this.
The method's effectiveness is bounded by the demographic fidelity of the synthetic-data generator, so as diffusion models improve, the training signal should improve as well—and a diffusion generator with its own biases would cap the achievable fairness.
A testable extension is measuring whether the balanced policy generalizes to occupation prompts outside the 50 used for evaluation and to finer-grained groups such as age or religion, both listed by the paper as future work.
Because the evaluation labels come from a pretrained classifier, human-annotation studies of the debiased model's outputs would decide how much of the reported reduction is real demographic balance rather than classifier agreement.

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.

Desk Editor's Note

read the letter

Two things to know. First, this is the first systematic benchmark of demographic bias in unified multimodal LLMs for image generation, and the headline finding — six current U-MLLMs all skew toward majority demographics, with gender bias between 0.83 and 0.90 — is plausible and consistent with the diffusion-model fairness literature. Second, don't take the localization and mitigation claims at face value yet; the support is thinner than the abstract suggests.

What's genuinely new and useful: the benchmark covers VILA-U, Show-o, Janus, Janus-Pro, TokenFlow, and Emu3 using the occupation-prompt set from Shen et al., so the numbers are comparable to prior diffusion-era fairness work. The balanced preference loss (BPO) is a reasonable ORPO-style objective — it penalizes the deviation of the sigmoid of the log odds ratio from 0.5, which is a clean way to express 'prefer neither demographic.' The two-stage SFT+BPO recipe is clearly specified. The observation that text-to-image finetuning reduces generation bias more than image-to-text finetuning is a useful data point, if the synthetic labels are correct.

Now the soft spots, in order of severity. The training data in Sec. 4.1 is labeled by the demographic written in the FLUX prompt, with no verification that the generated image actually depicts it. The evaluation classifier from Shen et al. is used without reporting its accuracy on U-MLLM outputs. If either is wrong, the BPO bias reductions in Table 1 could be artifacts — and that table is the only evidence the method works. This is load-bearing. Second, the 'partial alignment' claim in Sec. 5.3 uses an empty image as input; that measures the text prior, not visual understanding. Calling it understanding bias is misleading. Third, the localization of bias to the language model rests on a qualitative JSD observation with no numbers or tests, and the linear probing actually shows the vision encoder preserves demographic information, which undercuts the claim that the encoder is not a source. Fourth, Table 1 shows the same method (e.g., Finetune(I→T)) with different values in the Gender, Race, and G×R blocks, with no error bars or seeds. It looks like either separate training runs for each target or a typo; the reader can't tell.

The authors are candid about some limitations (Appendix F mentions resource constraints and limited model coverage), which is good. But the current paper overclaims. It's a strong benchmark paper in the making, not a finished one. I'd bring it to a reading group to talk about the methodology; the benchmark is worth citing with a caveat, but the mitigation numbers shouldn't be used until the synthetic labels and classifier are validated.

For peer review: yes, a serious editor should send this out. The topic is timely, the benchmark is new, and every problem I've listed is addressable with additional experiments and transparency. I'd ask for validation of the FLUX labels, classifier accuracy, quantitative JSD with significance, a corrected understanding-bias protocol, and a reconciled Table 1. After that, it could be a solid contribution.

Referee Report

6 major / 5 minor

Summary. This paper studies demographic bias (gender and race) in unified multimodal large language models (U-MLLMs) for text-to-image generation. It benchmarks six recent U-MLLMs using a classifier-based representation disparity metric, proposes a 'locate-then-fix' strategy in which bias is attributed to the language model via linear probing of vision-encoder embeddings and Jensen-Shannon divergence between image-token distributions, and introduces a balanced preference optimization (BPO) loss that encourages equal likelihood of images from different demographic groups during fine-tuning on synthetic FLUX.1-dev data. Experiments on VILA-U and TokenFlow report substantial bias reductions (e.g., VILA-U gender bias from 0.89 to 0.25) with modest effects on CLIP scores and inception score.

Significance. If the results are reproducible, the paper provides one of the first systematic fairness audits of U-MLLMs and a practical debiasing method for autoregressive token-based generation. The benchmark across six models is a useful contribution, and the balanced-preference-loss idea is simple and clearly described in Algorithm 1. The paper also explicitly acknowledges limitations in Appendix F, including the focus on overt demographic categories and the limited number of open-source models tested. However, the wide adoption of this method is contingent on the evaluation being robust to label noise and sampling variability, which the current manuscript does not demonstrate.

major comments (6)

[Section 3.2] The JSD-based localization argument is not well-defined as written. The empirical token distribution is defined in Eq. (3) as 1/M times a sum of delta functions over full token sequences; for long image token sequences, two independently sampled sets will almost surely have disjoint support, in which case the JSD equals log 2 for every comparison and the claim that 'the JSD ... is significantly smaller' when the prompt matches the predicted demographic is vacuous. The authors need to specify a meaningful distribution (e.g., over individual token positions or a smoothed n-gram distribution) and report actual JSD values with variance or a significance test.
[Section 4.1 and Section 5.1] The entire debiasing evaluation rests on two unvalidated labeling steps: the synthetic training images are labeled with the demographic attribute inserted into the FLUX.1-dev prompt, with no check that the image actually depicts that attribute, and the demographic classifier from Shen et al. (2024) is applied to generated images without reporting its accuracy on U-MLLM outputs or on the synthetic images. If FLUX misrenders modifiers or the classifier is biased on debiased outputs, the bias reductions in Table 1 could be artifacts of label noise. Please add a validation of the synthetic labels (human or classifier) and report the classifier's accuracy on the evaluated model outputs.
[Section 5.2, Table 1] Table 1 reports single point estimates of RD with no confidence intervals, standard errors, or significance tests. Given that each prompt is sampled with N=160 generations, the variability across the 50 prompts is likely to be substantial; for instance, comparisons such as Finetune(T→I) gender bias 0.27 vs. BPO 0.25 are within plausible sampling noise. The authors should report bias over seeds or bootstrap intervals, and test whether the differences between BPO and SFT baselines are statistically significant.
[Section 5.2, Table 1] The claim that the approach preserves semantic fidelity is only partially supported. For BPO in the Gender block, CLIP-S drops from 28.24 (VILA-U) to 27.74 and CLIP-IQA from 0.84 to 0.77, while IS increases. The paper should discuss this trade-off explicitly rather than asserting that quality is preserved, and the choice of which quality metric is most appropriate for token-based generation should be justified.
[Section 5.3] The understanding-bias protocol is not a valid measure of visual understanding: it feeds an 'empty image' and asks the model to answer 'What is the gender of {occupation}?' This measures the language model's prior, not the model's multimodal understanding. The 'partial alignment' conclusion (e.g., Janus shows low understanding bias but high generation bias) is therefore not established. Either use a proper image-understanding task (e.g., caption or VQA on images with known demographic attributes) or frame the result as a comparison of language-prior bias versus generation bias.
[Section 5.1] It is not stated whether the 50 evaluation prompts are disjoint from the 1,000 training prompts. If they overlap, the finetuning results could reflect memorization of the test prompts. The authors should state the proportion of overlap or make the split explicit.

minor comments (5)

[Abstract and Section 1] There are several typos and grammatical errors, such as 'we audit show how' in the abstract, 'obatin' in Section 4.1, and 'Qualitive' in Section 5.3; these should be corrected.
[Algorithm 1, line 11] The notation L^{(d_{k-1}, d_k)}_{bal} is inconsistent with the pairwise definition in Eq. (10); it should be written as a sum over all pairs (d_k, d_l) with k < l.
[Section 5.1] The bias(P) formula uses a pairwise average of absolute frequency differences; please cite the exact metric from Shen et al. (2024) and clarify whether this is their 'representation disparity' metric.
[Section 4.1] The use of ChatGPT-4o for paraphrasing is not reproducible without a list of paraphrases or a fixed seed; provide the paraphrased prompts or include them in a supplement.
[Figure 4] The y-axis label 'Bias Score' and the legend are present, but no numbers are given; the finding that 'most models exhibit a pronounced understanding bias' is not supported by the figure as printed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark, bias localization, and mitigation results are empirical and do not reduce to their own inputs by construction.

full rationale

The paper's three load-bearing claims are each supported by independent empirical procedures rather than by definitional equivalence or self-citation. First, the demographic bias benchmark is measured with an external attribute classifier from Shen et al. (2024) and external U-MLLMs, so the reported bias scores are observations, not derived quantities. Second, the claim that bias originates primarily from the language model is supported by a Jensen-Shannon divergence comparison between token distributions for neutral and demographic-augmented prompts; this is an observational localization step, and while the choice of grouping by the classifier's predicted label makes the 'majority group has smallest JSD' finding partly descriptive, it is not a fitted identity or a renamed input. Third, the balanced preference loss (Eq. 10) is designed to equalize demographic likelihoods, and the evaluation metric measures demographic frequency, so the training objective is intentionally aligned with the evaluation goal; however, the evaluation is performed on held-out prompts and compares against multiple baselines, so the reported bias reduction is an empirical generalization result rather than a quantity forced by construction. The paper contains a self-citation (Chen et al., 2024) used as background for the existence of bias in training data, but that citation is not load-bearing for any derivation. The unvalidated demographic labeling of synthetic images and the lack of reported classifier accuracy are validity concerns for correctness, not circularity, because the paper never derives a prediction from those labels in a way that makes the conclusion equivalent to the input. Overall, no circular step with quoted text and a specific reduction can be exhibited.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The central claims rest on five domain assumptions: the external demographic classifier's accuracy, the uniform-distribution definition of fairness, the fidelity of FLUX.1-dev synthetic labels, the interpretation of linear probing as excluding encoder bias, and the empty-image protocol as a measure of understanding bias. The only free parameter listed in the algorithm, λ, is never used in the reported loss, leaving the training objective incompletely specified. No invented entities are introduced.

free parameters (1)

Trade-off parameter lambda (Algorithm 1) = Not reported
Algorithm 1 takes a trade-off parameter λ as input, but the Stage 2 loss update in the pseudocode uses only Lbal without λ, and no λ value is reported in Section 5.1. The final training objective is therefore incompletely specified.

assumptions (5)

domain assumption The demographic classifier from Shen et al. (2024) correctly labels generated images.
All bias metrics and the evaluation labels in Table 1 and Figure 4 depend on this classifier (Section 5.1). No accuracy or agreement is reported for U-MLLM outputs.
domain assumption An unbiased model should generate a uniform distribution of demographic groups.
The RD metric and the BPO loss define fairness as equal frequency across groups (Section 2 and Eq. 9), which is a value choice rather than a derived property.
domain assumption FLUX.1-dev renders the demographic attribute specified in the augmented prompt.
Section 4.1 creates synthetic images and labels them by the prompt's demographic without validation; the debiasing dataset relies on this for the six demographic groups.
domain assumption High linear-probing accuracy on vision embeddings implies the vision encoder is not a source of generation bias.
Section 3.1 argues the encoder is not the bias source because a linear classifier can predict demographics from embeddings, but this only shows the attributes are encoded, not that the encoder fails to introduce or amplify bias.
domain assumption Empty-image responses to demographic questions measure understanding bias.
Section 5.3 feeds an empty image with the question 'What is the gender of {occupation}?' and interprets the answer as visual understanding bias; this actually measures the language prior, not multimodal understanding.

how reviews work

0 comments

Cite this review

Pith. "Pith review of On Fairness of Unified Multimodal Large Language Model for Image Generation." pith.science (2026). https://pith.science/paper/SK2CPNOF

@misc{pith2026250203429,
  author       = {Pith},
  title        = {Pith review of: On Fairness of Unified Multimodal Large Language Model for Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SK2CPNOF}},
  note         = {Machine review of arXiv:2502.03429}
}

read the original abstract

Unified multimodal large language models (U-MLLMs) have demonstrated impressive performance in visual understanding and generation in an end-to-end pipeline. Compared with generation-only models (e.g., Stable Diffusion), U-MLLMs may raise new questions about bias in their outputs, which can be affected by their unified capabilities. This gap is particularly concerning given the under-explored risk of propagating harmful stereotypes. In this paper, we benchmark the latest U-MLLMs and find that most exhibit significant demographic biases, such as gender and race bias. To better understand and mitigate this issue, we propose a locate-then-fix strategy, where we audit and show how the individual model component is affected by bias. Our analysis shows that bias originates primarily from the language model. More interestingly, we observe a "partial alignment" phenomenon in U-MLLMs, where understanding bias appears minimal, but generation bias remains substantial. Thus, we propose a novel balanced preference model to balance the demographic distribution with synthetic data. Experiments demonstrate that our approach reduces demographic bias while preserving semantic fidelity. We hope our findings underscore the need for more holistic interpretation and debiasing strategies of U-MLLMs in the future.

Figures

Figures reproduced from arXiv: 2502.03429 by the authors.

**Figure 1.** U-MLLMs are capable of generating images with high quality, but the generation lacks diversity. The model has a bias for some occupations. In this example, given the prompt “construction worker”, the model generate most of images with demographic attribute as “male” and “white”. tioning on user prompts in natural language. However, these U-MLLMs with unified capabilities may inadvertently reproduce or amplify biase… view at source ↗

**Figure 2.** The optimization objective is to minimize the deviation of preference between different demographic group. {Asian, Black, Indian, White} under neutral prompts, rather than overwhelmingly generate one demographic group. This shift in perspective means we want the absolute difference in preference across demographics to be minimized, rather than pushing a single demographic to be more or less likely in isolation. To e… view at source ↗

**Figure 3.** Given the test prompt: “a photo of the face of a electrical and electronics repairer, a person”. Images generated by: VILA-U (bottom-left), our method reduced bias in gender(bottom-right), race(top-left), gender-race(top-right). For each image, the first color-coded bar represents the predicted gender: male or female. The second bar represents the predicted race: While, Asian, Black, or Indian. The height of each ba… view at source ↗

Figures from the paper (10 more)

**Figure 4.** Figure 4: Compare the understanding and generation bias. generation capabilities. Compared to early MLLMs, which focus purely on understanding, such as Llava series (Liu et al., 2024; 2023), these more recent works, represented by VILA-U (Wu et al., 2024b), Show-o (Xie et al., 2…

**Figure 5.** Figure 5: Overview of framework’s multi-modal training and inference process(Wu et al., 2024b) Visual inputs are converted into discrete tokens and merged with textual tokens to create a unified multi-modal token sequence. This sequence is used in next-token prediction process, …

**Figure 6.** Figure 6: Overview of unified foundation vision tower(Wu et al., 2024b) Input images are processed by the vision encoder, where features are extracted and discretized using residual quantization. These discrete vision features are then utilized in two ways: they are fed into the…

**Figure 7.** Figure 7: Detecting bias in LM(top), Vision encoder(bottom); C. BPO Algorithm Algorithm 1 Balanced Preference Optimization Input: U-MLLM with parameters θ0; SFT dataset DSFT = {(xi , zi)} of prompts xi and image tokens zi ; Balanced dataset Dbal = {(xj , yj1 , . . . , yjK )}, ea…

**Figure 8.** Figure 8: Given the test prompt: “a photo of the face of a electrical and electronics repairer, a person”, images generated by VILA-U. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

**Figure 9.** Figure 9: Given the test prompt: “a photo of the face of a electrical and electronics repairer, a person”, images generated by Janus 15 [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

**Figure 10.** Figure 10: Given the test prompt: “a photo of the face of a electrical and electronics repairer, a person”, images generated by Janus-Pro 16 [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]

**Figure 11.** Figure 11: Given the test prompt: “a photo of the face of a electrical and electronics repairer, a person”, images generated by Show-o [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]

**Figure 12.** Figure 12: Given the test prompt: “a photo of the face of a electrical and electronics repairer, a person”, images generated by TokenFlow 17 [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]

**Figure 13.** Figure 13: Given the test prompt: “a photo of the face of a electrical and electronics repairer, a person”, images generated by Emu3 18 [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

Bias Analysis in Unconditional Image Generative Models
cs.CV 2025-06 conditional novelty 6.0 of 10

In unconditional image generators, measured attribute bias shifts are small and are strongly influenced by whether the attribute classifier's decision boundary falls in a dense or sparse region of the attribute's dist...

Reference graph

Works this paper leans on

18 extracted references · 1 canonical work pages · cited by 1 Pith paper

[1]

Direct pref- erence optimization with an offset

Amini, A., Vieira, T., and Cotterell, R. Direct pref- erence optimization with an offset. arXiv preprint arXiv:2402.10571,

arXiv
[5]

Elazar, Y ., Bhagia, A., Magnusson, I., Ravichander, A., Schwenk, D., Suhr, A., Walsh, P., Groeneveld, D., Sol- daini, L., Singh, S., et al

URL https://arxiv.org/abs/ 2302.00070. Elazar, Y ., Bhagia, A., Magnusson, I., Ravichander, A., Schwenk, D., Suhr, A., Walsh, P., Groeneveld, D., Sol- daini, L., Singh, S., et al. What’s in my big data? arXiv preprint arXiv:2310.20707,

arXiv
[7]

Hong, J., Lee, N., and Thorne, J

URL https://arxiv.org/abs/2308.14761. Hong, J., Lee, N., and Thorne, J. Orpo: Monolithic prefer- ence optimization without reference model,

arXiv
[8]

URL https://arxiv.org/abs/2403.07691. Hu, E. J., Shen, Y ., Wallis, P., Allen-Zhu, Z., Li, Y ., Wang, S., Wang, L., and Chen, W. Lora: Low-rank adaptation of large language models,

arXiv
[11]

Li, C., Chen, M., Wang, J., Sitaram, S., and Xie, X

URL https://arxiv.org/ abs/2411.16318. Li, C., Chen, M., Wang, J., Sitaram, S., and Xie, X. Cul- turellm: Incorporating cultural differences into large lan- guage models. arXiv preprint arXiv:2402.10946, 2024a. Li, C., Teney, D., Yang, L., Wen, Q., Xie, X., and Wang, J. Culturepark: Boosting cross-cultural understanding in large language models. arXiv pre...

arXiv
[12]

Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C

URL https://arxiv.org/abs/2405.14734. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P., Leike, J., and Lowe, R. Training language models to follow instructions with human feedback,

arXiv
[13]

Park, R., Rafailov, R., Ermon, S., and Finn, C

URL https: //arxiv.org/abs/2203.02155. Park, R., Rafailov, R., Ermon, S., and Finn, C. Disentan- gling length from quality in direct preference optimiza- tion,

arXiv
[14]

URL https://arxiv.org/abs/2403. 19159. Qu, L., Zhang, H., Liu, Y ., Wang, X., Jiang, Y ., Gao, Y ., Ye, H., Du, D. K., Yuan, Z., and Wu, X. Tokenflow: Unified image tokenizer for multimodal understanding and generation. arXiv preprint arXiv:2412.03069,

arXiv

Show all 18 references

[15]

Tong, S., Fan, D., Zhu, J., Xiong, Y ., Chen, X., Sinha, K., Rabbat, M., LeCun, Y ., Xie, S., and Liu, Z

URL https://arxiv.org/abs/ 2311.07604. Tong, S., Fan, D., Zhu, J., Xiong, Y ., Chen, X., Sinha, K., Rabbat, M., LeCun, Y ., Xie, S., and Liu, Z. Metamorph: Multimodal understanding and generation via instruc- tion tuning,

arXiv
[16]

Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., Fan, Y ., Dang, K., Du, M., Ren, X., Men, R., Liu, D., Zhou, C., Zhou, J., and Lin, J

URL https://arxiv.org/abs/ 2412.14164. Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., Fan, Y ., Dang, K., Du, M., Ren, X., Men, R., Liu, D., Zhou, C., Zhou, J., and Lin, J. Qwen2-vl: Enhancing vision-language model’s perception of ...

arXiv
[17]

Transfusion: Predict the next token and dif- fuse images with one multi-modal model

Zhou, C., Yu, L., Babu, A., Tirumala, K., Yasunaga, M., Shamis, L., Kahn, J., Ma, X., Zettlemoyer, L., and Levy, O. Transfusion: Predict the next token and dif- fuse images with one multi-modal model. arXiv preprint arXiv:2408.11039,

arXiv
[18]

Since then, numerous new loss functions have been proposed (Meng et al., 2024; Park et al., 2024; Hong et al., 2024; Ethayarajh et al., 2024; Azar et al., 2023)

has emerged as a promising technique to address biases in machine learning models, especially LLM. Since then, numerous new loss functions have been proposed (Meng et al., 2024; Park et al., 2024; Hong et al., 2024; Ethayarajh et al., 2024; Azar et al., 2023). Recent advances ...

2024
[1952]

Chen, H., Raj, B., Xie, X., and Wang, J

URL https: //api.semanticscholar.org/CorpusID: 125209808. Chen, H., Raj, B., Xie, X., and Wang, J. On catastrophic inheritance of large foundation models. arXiv preprint arXiv:2402.01909,

arXiv
[2019]

URL https: //arxiv.org/abs/1908.04913. Labs, B. F. Flux. https://github.com/ black-forest-labs/flux,

1908 arXiv
[2021]

org/abs/2106.09685

URL https://arxiv. org/abs/2106.09685. Kim, Y ., Na, B., Park, M., Jang, J., Kim, D., Kang, W., and Moon, I.-C. Training unbiased diffusion models from biased dataset. In The Twelfth International Conference on Learning Representations,

arXiv
[2022]

Bradley, R

URL https://arxiv.org/abs/2210.15230. Bradley, R. A. and Terry, M. E. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39:324,

arXiv
[2023]

Bachmann, R., Kar, O

URL https://arxiv.org/abs/ 2310.12036. Bachmann, R., Kar, O. F., Mizrahi, D., Garjani, A., Gao, M., Griffiths, D., Hu, J., Dehghan, A., and Zamir, A. 4m-21: An any-to-any vision model for tens of tasks and modalities. arXiv preprint arXiv:2406.09406,

arXiv
[2024]

Gandikota, R., Orgad, H., Belinkov, Y ., Materzy´nska, J., and Bau, D

URL https://arxiv.org/abs/ 2402.01306. Gandikota, R., Orgad, H., Belinkov, Y ., Materzy´nska, J., and Bau, D. Unified concept editing in diffusion models,

arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

Direct pref- erence optimization with an offset

Amini, A., Vieira, T., and Cotterell, R. Direct pref- erence optimization with an offset. arXiv preprint arXiv:2402.10571,

arXiv

[2] [5]

Elazar, Y ., Bhagia, A., Magnusson, I., Ravichander, A., Schwenk, D., Suhr, A., Walsh, P., Groeneveld, D., Sol- daini, L., Singh, S., et al

URL https://arxiv.org/abs/ 2302.00070. Elazar, Y ., Bhagia, A., Magnusson, I., Ravichander, A., Schwenk, D., Suhr, A., Walsh, P., Groeneveld, D., Sol- daini, L., Singh, S., et al. What’s in my big data? arXiv preprint arXiv:2310.20707,

arXiv

[3] [7]

Hong, J., Lee, N., and Thorne, J

URL https://arxiv.org/abs/2308.14761. Hong, J., Lee, N., and Thorne, J. Orpo: Monolithic prefer- ence optimization without reference model,

arXiv

[4] [8]

URL https://arxiv.org/abs/2403.07691. Hu, E. J., Shen, Y ., Wallis, P., Allen-Zhu, Z., Li, Y ., Wang, S., Wang, L., and Chen, W. Lora: Low-rank adaptation of large language models,

arXiv

[5] [11]

Li, C., Chen, M., Wang, J., Sitaram, S., and Xie, X

URL https://arxiv.org/ abs/2411.16318. Li, C., Chen, M., Wang, J., Sitaram, S., and Xie, X. Cul- turellm: Incorporating cultural differences into large lan- guage models. arXiv preprint arXiv:2402.10946, 2024a. Li, C., Teney, D., Yang, L., Wen, Q., Xie, X., and Wang, J. Culturepark: Boosting cross-cultural understanding in large language models. arXiv pre...

arXiv

[6] [12]

Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C

URL https://arxiv.org/abs/2405.14734. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P., Leike, J., and Lowe, R. Training language models to follow instructions with human feedback,

arXiv

[7] [13]

Park, R., Rafailov, R., Ermon, S., and Finn, C

URL https: //arxiv.org/abs/2203.02155. Park, R., Rafailov, R., Ermon, S., and Finn, C. Disentan- gling length from quality in direct preference optimiza- tion,

arXiv

[8] [14]

URL https://arxiv.org/abs/2403. 19159. Qu, L., Zhang, H., Liu, Y ., Wang, X., Jiang, Y ., Gao, Y ., Ye, H., Du, D. K., Yuan, Z., and Wu, X. Tokenflow: Unified image tokenizer for multimodal understanding and generation. arXiv preprint arXiv:2412.03069,

arXiv