Pith. sign in

REVIEW 4 major objections 6 minor 59 references

Implicit Priors Editing in Stable Diffusion via Targeted Token Adjustment

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Editing only the word-token embedding of a target object—768 parameters in Stable Diffusion v1.4—changes the model's implicit assumptions (e.g., rose→blue rose) with no effect on unrelated concepts, beating prior methods on efficacy…

desk verdict Clean, parameter-efficient token-embedding edit, but the paper's own specificity numbers contradict its no-side-effect guarantee—major revision needed. read the letter →

arxiv 2412.03400 v1 pith:SQJWL2CW submitted 2024-12-04 cs.CV

Feng He , Chao Zhang , Zhixue Zhao This is my paper · ORCID
classification cs.CV
keywords implicitpriorseditingwordtokenembeddingsStableDiffusiontext-to-imagemodelsmodelsequentialgenderbiasmitigationparameterefficiency
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

People who prompt image models often rely on unspoken defaults: a 'rose' comes out red, a 'bear' brown. This paper argues that such implicit priors live, to a surprising degree, in a single layer of the model—the word-token embedding (WTE) of the object word—and that rewriting just that one vector can redirect the model's assumption. The proposed method, EMBEDIT, fine-tunes only the WTE of the target token so that the text encoder's final hidden state for the source prompt matches the hidden state for the explicit destination prompt. Because every other weight, and every other token's embedding, stays frozen, prompts that do not contain the edited object produce exactly the same images as before. The authors report consistent gains over the previous state of the art (efficacy from 87.17% to 93.18% on SD 1.4 single edits), and—unlike the prior method—stable performance when many edits are applied one after another.

What carries the argument

The central object is the word-token embedding (WTE) vector of the target object in the CLIP text encoder—a 768-dimensional vector in Stable Diffusion v1.4 and 2048 in XL. The mechanism precomputes the last hidden state of the text encoder for the destination prompt, then runs gradient descent on the target WTE alone to minimize the MSE between the hidden states of the source and destination prompts, stopping when the loss shrinks to a fraction $\lambda$ of its initial value. The frozen diffusion UNet is never touched. The paper's probe experiment—a logistic regression on WTEs to predict object color—is the empirical justification that implicit priors are present at this layer.

What would settle it

A concrete check: run prompts that contain the edited token in contexts that normally override the attribute—e.g., 'a black apple' after editing 'apple' to 'red apple'—and look for the edited attribute leaking; or take a compound like 'ice hockey' after the paper's own 'ice cream'→'strawberry ice cream' edit and measure whether 'ice hockey' images turn pink. If either reproduced, the claim of token-localized, side-effect-free editing is falsified. A more mechanistic test is to probe not just the WTE but also intermediate and final text-encoder layers for color/attribute information; 90% WTE accuracy would be undercut if deep layers carry comparable or stronger signal.

Watch

Extended reading notes

Core claim

EMBEDIT's central claim is that the implicit assumptions a Stable Diffusion model attaches to an underspecified word—color, shape, category, even gender—are encoded in the word-token embedding of that word, and that minimizing the mean squared error between the CLIP text encoder's last hidden states for the source prompt (e.g., 'bear') and the destination prompt (e.g., 'polar bear') is sufficient to transfer the attribute to generated images. The edit is a single gradient update loop over the 768- or 2048-dimensional embedding vector, stopped when the loss reaches a predefined fraction of its initial value. Since the diffusion weights and all non-target embeddings are untouched, the method is side-effect-free by construction for prompts that omit the target word, trivially reversible, and numerically stable under sequential editing—where the previous cross-attention-based method (TIME) collapses into noise. The paper also demonstrates the same mechanism mitigates profession-gender stereotypes, reducing the average deviation from balanced gender from 0.598 to 0.121, and shows the prior is visible at the embedding level via a logistic-regression probe that predicts object color (red vs. yellow) from the WTE with about 90% accuracy.

Load-bearing premise

The load-bearing premise is that a target concept's implicit prior lives mostly in that word's token embedding, and that matching the text encoder's final hidden state to the destination prompt is enough to move it—if the prior is spread across deeper text-encoder layers or cross-attention, the edit could fail or leak into other tokens.

Editorial extensions

If this is right

  • Each edit changes only 0.002% of the model's parameters (SD 1.4), making thousands of sequential edits feasible without touching the diffusion weights.
  • Prompts that do not contain the edited token produce outputs identical to the unedited model, so specificity is preserved by construction, not by tuning.
  • Sequential editing, where TIME collapses into salt-and-pepper noise, stays stable for EMBEDIT (e.g., 96.59% efficacy on SD 1.4 sequential vs. 93.18% single).
  • Beyond object attributes, the same WTE-only procedure moves profession-gender bias toward balance (average deviation $\Delta$ from 50% female from 0.598 to 0.121, outperforming TIME's 0.308).
  • The method generalizes to Stable Diffusion XL with the same hyperparameters, and specificity improves on the larger model.

Reading between the lines

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

  • If implicit priors really are concentrated in the WTE, then the same recipe could transfer to other CLIP-conditioned diffusion models—and possibly to other conditional generators that share a frozen text encoder—though the paper only tests Stable Diffusion.
  • The known failure on multi-token words like 'ice cream' suggests the method's locality guarantee is per token, not per concept: editing 'ice' leaks 'strawberry' into 'ice hockey'. A testable remedy would be editing a concept-level direction in embedding space rather than the raw token vector.
  • The probe shows color signal in the WTE, but does not show the signal is absent from deeper layers; a causal test (ablating mid/late text-encoder layers during editing) would tell whether last-hidden-state MSE is necessary or just convenient.
  • Because the stopping threshold $\lambda$ controls edit strength, the paper implicitly offers a dial for how much bias to remove; a principled, per-profession automatic choice of $\lambda$ (the paper's automatic variant underperforms manual tuning) remains an open extension.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes EMBEDIT, a method for editing implicit priors in Stable Diffusion by fine-tuning only the word-token embedding (WTE) of a target object. The optimization minimizes the MSE between the last hidden states of the CLIP text encoder for a source prompt and a destination prompt, with an early-stopping threshold controlled by a free parameter λ. The authors report improvements over TIME on the TIMED benchmark for single and sequential editing across SD 1.4 and SD XL, and for gender-bias mitigation on six professions. The paper also includes a probing experiment intended to show that WTE encodes object-color priors.

Significance. If the main claims held, the method would be an attractive baseline for concept editing in text-to-image models: it is parameter-efficient (768 or 2048 parameters), reversible, and sequential editing appears stable where TIME collapses. The paper also provides a useful comparison against TIME on two model sizes and includes ablations and a supplementary comparison to ReFACT. However, the central advertised property—that prompts not containing the edited token are completely unaffected—is contradicted by the paper's own specificity numbers in Table 3. Because this property is the main differentiator from prior work, the contribution is currently not established at the level claimed.

major comments (4)
  1. [§4, Algorithm 1; §5.4, Table 3] The central invariance claim is contradicted by the paper's own specificity results. Since only the WTE of the target token is updated, any prompt whose token sequence does not contain that token should produce exactly the original text-encoder hidden states and, under a fixed seed, identical images. Table 3 reports single-edit Specificity of 77.09 versus the 88.56 baseline on SD 1.4, and 87.71 versus 94.28 on SD XL, with sequential Specificity of 69.92. These differences are not explainable by the stated mechanism. The manuscript itself documents leakage for multi-token edits ("ice cream" affects "a bucket of ice", §5.4, Fig. 7), so the invariant holds at best only for prompts avoiding every edited token, and no single-token-only breakdown is provided. Please report specificity separately for single-token versus multi-token edits, state the evaluation protocol precisely (including seeds and whether negative prompts ever contain the edited token), and reconcile Table 3 with the theoretical guarantee.
  2. [§5.2, Table B.1] The sequential-editing evaluation is run on a filtered subset of TIMED. The paper removes 27 of 104 entries, with the stated criterion that objects which are specificity negatives in single editing become positives in sequential editing, and Table B.1 lists additional removed objects deemed "unsuitable." This changes the task distribution and makes the single-versus-sequential comparison in Table 3 not matched on the same edit set. Because the filtering removes hard specificity cases, it can inflate the sequential specificity figure (69.92) and the claim that sequential editing causes no degradation. Please report sequential results on the full TIMED set as well as on the filtered set, or provide per-entry results so the effect of filtering can be assessed.
  3. [§6, §6.3, Supplementary B.4] The gender-bias mitigation results involve several manually tuned free parameters. Section 6 states that λ is set differently for professions with stronger versus weaker bias, and the automatic method in Sec. B.4 introduces α_min and α=10·Δ with α_min=2 chosen manually. With per-profession manual λ, the reported Δ=0.121 is not a fair comparison to a fixed-hyperparameter baseline, and the automatic method is evaluated with 10 images per prompt while the main gender experiment uses 24. Please either fix λ across all professions and report sensitivity, or list the manual settings for every profession; otherwise the advantage over TIME on this task is not established.
  4. [§3, §5.3] The probing experiment is too weak to support the localization premise that implicit priors reside mainly in the WTE. The probe is a logistic regression on 200 ChatGPT-generated objects (100 red, 100 yellow) with an 80/20 split; there are no control conditions on later-layer representations, random embeddings, word frequency, or token count, and only accuracy with standard deviation is reported. More importantly, 90% accuracy in classifying WTE color does not imply that modifying the WTE is sufficient to transfer an attribute in generation, which is the load-bearing assumption of the method. Please add control probes on other text-encoder layers and an analysis of which TIMED entries involve single-token objects, to support the locality claim.
minor comments (6)
  1. [Eq. (1)] The notation "(h_orig, h_new)^2" is undefined; the MSE should be written as the mean squared Euclidean distance, e.g., (1/d) Σ_i (h_orig,i − h_new,i)^2.
  2. [Algorithm 1] Line 11 says "Update original WTE via L.step"; this should specify that an optimizer step is taken and that the destination hidden state is kept fixed (no gradient flows through the destination prompt).
  3. [Abstract; §1] The paper repeatedly says "model weights remain unchanged" while updating the WTE, which is itself a model parameter. Clarify that all other parameters remain unchanged.
  4. [Abstract; §5.4] The claim of "at least a 6.01% improvement (from 87.17% to 93.18%)" is only the SD 1.4 single-edit efficacy number; several other metrics in Table 3 improve by less or even decrease (specificity). Qualify the summary claim to avoid overstating the result.
  5. [Supplementary B.4] The automatic gender method is compared in Table B.3 using 10 images per prompt, while the main gender evaluation in §6.2 uses 24 images per prompt; the comparison is not on equal footing and should be rerun or clearly labeled as a pilot.
  6. [Table 3] FID and CLIP Score are reported without standard deviations, and the FID protocol (which MS-COCO split, how many images, whether the same captions are used for all methods) is not described. Please specify the exact protocol or cite the reference implementation.

Circularity Check

2 steps flagged · score 6.0 of 10

EMBEDIT's side-effect-free guarantee is a tautology of Algorithm 1 and is contradicted by its own specificity numbers; gender-bias lambda is tuned per profession from the measured bias, making those results fitted rather than predicted.

  1. self definitional [Abstract; Sec. 1; Algorithm 1; Sec. 5.4, Table 3]
    "This targeted adjustment prevents unintended effects on other objects in the model's knowledge base, as the WTEs for unrelated objects and the model weights remain unchanged. Consequently, when a prompt does not contain the edited object, all representations, and the model outputs are identical to those of the original, unedited model."

    Algorithm 1 updates only wteorig at token index Iorig; all other WTEs and weights are frozen. Any prompt avoiding Iorig therefore produces bit-identical hidden states and, with a fixed seed, identical images. The side-effect-free claim is an analytic consequence of the implementation, not an empirical finding; using Specificity as a measured strength is circular because the metric is guaranteed to equal Baseline by construction. Table 3 contradicts the guarantee (EMBEDIT 77.09 vs Baseline 88.56 on SD1.4; 87.71 vs 94.28 on SDXL), which is impossible under the stated mechanism unless unstated evaluation components change. The invariance is an input assumption, not a derived prediction.

  2. fitted input called prediction [Sec. 6; Sec. 6.3; Supplementary B.4]
    "As practice in [34], we manually set a lower λ for professions with stronger biases and a slightly higher λ for those with less bias."

    λ is the stopping threshold controlling how far the edited WTE moves toward the destination hidden state (Algorithm 1, line 4). Setting λ per profession from the known strength of that profession's bias means the reported Fp values are in-sample fits to the target quantity, not out-of-sample predictions. The supplemental automatic method, which does not use this manual input, degrades from Δ=0.121 to Δ=0.442 (Table B.3), confirming that the manually tuned λ largely determines the reported debiasing performance.

full rationale

The paper's core edit-efficacy results are not circular: the WTE is optimized against Stable Diffusion's text-encoder hidden states, while efficacy/generality are scored with the external CLIP ViT-B/32 classifier on generated images, so the metric is not the training objective. There are no load-bearing self-citations. However, two advertised components reduce to their inputs. First, the side-effect-free guarantee is entailed by Algorithm 1's restriction to the target token WTE; Specificity is therefore a tautological metric that should equal Baseline, and the paper's Table 3 values contradict that construction, indicating the guarantee is an input assumption rather than a validated result. Second, the gender-bias section manually sets λ per profession using the known bias strength, so those Fp results are fitted to the target rather than predicted; the automatic variant performs markedly worse. These are partial, load-bearing circularities, so the score is 6 rather than higher; the main efficacy claims retain independent external evaluation.

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

The method relies on domain assumptions about where knowledge is stored in CLIP and how hidden-state distances transfer to image generation. The free parameters are task-specific tuning choices. No new physical or theoretical entities are introduced.

free parameters (3)
  • lambda (optimization strength) = 0.2 or 0.3 for object edits; manually set per profession for gender bias (e.g., lower for CEO, higher for teacher)
    Controls the stopping threshold for the MSE loss, determining how far the WTE moves toward the destination hidden state. Tuned empirically to achieve best editing results; for gender it is manually adjusted based on known bias strength.
  • learning rate = 0.001
    Selected via ablation on 24 samples (Table B.2) as the best trade-off among efficacy, generality, and specificity.
  • alpha_min and alpha in automatic gender method = alpha_min = 2; alpha = max(2, 10 * Delta)
    Introduced for the automatic gender-bias loss to balance counter-stereotypical and stereotypical terms; fitted to the six professions and not independently justified.
assumptions (4)
  • domain assumption The last hidden state of the CLIP text encoder is a sufficient semantic representation for transferring an attribute from a destination prompt to a source token.
    The entire optimization rests on minimizing MSE in this hidden-state space; if this representation is not the right level, the edit may not transfer to generated images. Used in Section 4, Equation 1.
  • domain assumption Implicit priors such as object color are primarily encoded in the word token embedding of the object token.
    Supported only by a linear probe on 200 ChatGPT-generated objects (Section 3). The probe shows correlation, not causation or exclusivity. If priors are distributed across cross-attention or later layers, editing the WTE alone may be insufficient or cause side effects.
  • domain assumption MSE loss between hidden states is an appropriate proxy for edit quality.
    The method optimizes this loss, but it is not derived from image-generation quality. The evaluation on images is external, so this assumption is not directly validated.
  • domain assumption The tokenizer splits target words into known tokens that can be individually edited.
    The method updates the WTE at a specific token index. Multi-word or multi-token objects require joint editing, which is acknowledged as a limitation in Section 5.4 and Figure 7.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Implicit Priors Editing in Stable Diffusion via Targeted Token Adjustment." pith.science (2026). https://pith.science/paper/SQJWL2CW

@misc{pith2026241203400,
  author       = {Pith},
  title        = {Pith review of: Implicit Priors Editing in Stable Diffusion via Targeted Token Adjustment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SQJWL2CW}},
  note         = {Machine review of arXiv:2412.03400}
}
read the original abstract

Implicit assumptions and priors are often necessary in text-to-image generation tasks, especially when textual prompts lack sufficient context. However, these assumptions can sometimes reflect outdated concepts, inaccuracies, or societal bias embedded in the training data. We present Embedding-only Editing (Embedit), a method designed to efficiently adjust implict assumptions and priors in the model without affecting its interpretation of unrelated objects or overall performance. Given a "source" prompt (e.g., "rose") that elicits an implicit assumption (e.g., rose is red) and a "destination" prompt that specifies the desired attribute (e.g., "blue rose"), Embedit fine-tunes only the word token embedding (WTE) of the target object ("rose") to optimize the last hidden state of text encoder in Stable Diffusion, a SOTA text-to-image model. This targeted adjustment prevents unintended effects on other objects in the model's knowledge base, as the WTEs for unrelated objects and the model weights remain unchanged. Consequently, when a prompt does not contain the edited object, all representations, and the model outputs are identical to those of the original, unedited model. Our method is highly efficient, modifying only 768 parameters for Stable Diffusion 1.4 and 2048 for XL in a single edit, matching the WTE dimension of each respective model. This minimal scope, combined with rapid execution, makes Embedit highly practical for real-world applications. Additionally, changes are easily reversible by restoring the original WTE layers. Our experimental results demonstrate that Embedit consistently outperforms previous methods across various models, tasks, and editing scenarios (both single and sequential multiple edits), achieving at least a 6.01% improvement (from 87.17% to 93.18%).

Figures

Figures reproduced from arXiv: 2412.03400 by the authors.

Figure 1
Figure 1. Illustration of the proposed EMBEDIT modifying the word-token embedding (WTE) of the target word “bear” to shift its interpretation from “brown bear” to “polar bear”. EMBEDIT optimizes the WTE of “bear” by minimizing the distance between the last hidden states of the text encoder in Stable Diffusion for both the original implicit prompt and the explicit prompt. Since EMBEDIT keeps the model weights completely unchan… view at source ↗
Figure 2
Figure 2. EMBEDIT edits implicit assumptions in text-to-image models. Two examples are shown here. In each row, columns 1 and 2 display positive prompts related to the target word, while column 3 shows the generated image for an unrelated object prompt. 2 [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Illustration of single editing and sequential editing [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Illustration of Efficacy, Generality, and Specificity. Images are generated by E [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: A comparison of Efficacy, Generality, Specificity, [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: A comparison of edit performance between E [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Illustration of failure cases where an object consists [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Example of mitigating gender bias on lawyers. Baseline EMBEDIT [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Example of mitigating gender bias on teachers. 7. Conclusions We present EMBEDIT, a simple yet effective approach for modifying implicit assumptions in text-to-image diffusion models by editing word token embeddings (WTEs). Our probing experiments provide intuitive 8 …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

59 extracted references · 44 canonical work pages

  1. [1]

    Per- sistent anti-muslim bias in large language models

    Abubakar Abid, Maheen Farooqi, and James Zou. Per- sistent anti-muslim bias in large language models. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, pages 298–306, 2021. 3

    work page 2021

  2. [2]

    ReFACT: Updating text-to-image models by editing the text encoder

    Dana Arad, Hadas Orgad, and Yonatan Belinkov. ReFACT: Updating text-to-image models by editing the text encoder. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technolo- gies (Volume 1: Long Papers) , pages 2537–2558, Mex- ico City, Mexico, 2024. Association for Computat...

    work page 2024

  3. [3]

    Localizing and edit- ing knowledge in text-to-image generative models

    Samyadeep Basu, Nanxuan Zhao, Vlad I Morariu, So- heil Feizi, and Varun Manjunatha. Localizing and edit- ing knowledge in text-to-image generative models. In The Twelfth International Conference on Learning Rep- resentations, 2024. 3, 7

    work page 2024

  4. [4]

    On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM conference on fairness, ac- countability, and transparency, pages 610–623, 2021

    Emily M Bender, Timnit Gebru, Angelina McMillan- Major, and Shmargaret Shmitchell. On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM conference on fairness, ac- countability, and transparency, pages 610–623, 2021. 7

    work page 2021

  5. [5]

    Easily accessible text-to-image generation am- plifies demographic stereotypes at large scale

    Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tat- sunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation am- plifies demographic stereotypes at large scale. In Pro- ceedings of the 2023 ACM Conference on Fairness, Ac- countability, and Transparency, pages 1493...

    work page 2023

  6. [6]

    Su Lin Blodgett, Solon Barocas, Hal Daum’e, and Hanna M. Wallach. Language (technology) is power: A critical survey of “bias” in nlp. ArXiv, abs/2005.14050,

    arXiv 2005

  7. [7]

    Man is to computer pro- grammer as woman is to homemaker? debiasing word embeddings

    Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, and Adam Kalai. Man is to computer pro- grammer as woman is to homemaker? debiasing word embeddings. In Proceedings of the 30th International Conference on Neural Information Processing Systems , page 4356–4364, Red Hook, NY , USA, 2016. Curran As- sociates Inc. 3

    work page 2016

  8. [8]

    Man is to com- puter programmer as woman is to homemaker? debias- ing word embeddings

    Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. Man is to com- puter programmer as woman is to homemaker? debias- ing word embeddings. Advances in neural information processing systems, 29, 2016. 2, S4

    work page 2016

Show all 59 references
  1. [9]

    What you can cram into a single $&!#* vector: Probing sentence em- beddings for linguistic properties

    Alexis Conneau, Germ ´an Kruszewski, Guillaume Lam- ple, Lo ¨ıc Barrault, and Marco Baroni. What you can cram into a single $&!#* vector: Probing sentence em- beddings for linguistic properties. In Annual Meeting of the Association for Computational Linguistics, 2018. 3

    2018

  2. [10]

    Theories of “gender” in nlp bias research

    Hannah Devinney, Jenny Bj ¨orklund, and Henrik Bj¨orklund. Theories of “gender” in nlp bias research. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, 2022. 7

    2022

  3. [11]

    Probing explicit and implicit gender bias through llm conditional text generation

    Xiangjue Dong, Yibo Wang, Philip S Yu, and James Caverlee. Probing explicit and implicit gender bias through llm conditional text generation. arXiv preprint arXiv:2311.00306, 2023. 3

    2023 arXiv

  4. [12]

    How to probe sentence embeddings in low-resource languages: On structural design choices for probing task evaluation

    Steffen Eger, Johannes Daxenberger, and Iryna Gurevych. How to probe sentence embeddings in low-resource languages: On structural design choices for probing task evaluation. In Proceedings of the 24th Conference on Computational Natural Language Learning, pages 108–118, Online...

    2020

  5. [13]

    Probing for semantic evidence of composition by means of simple classification tasks

    Allyson Ettinger, Ahmed Elgohary, and Philip Resnik. Probing for semantic evidence of composition by means of simple classification tasks. In Proceedings of the 1st Workshop on Evaluating Vector-Space Representations for NLP, pages 134–139, Berlin, Germany, 2016. Asso- ciation...

    2016

  6. [14]

    Fraser, Svetlana Kiritchenko, and Isar Ne- jadgholi

    Kathleen C. Fraser, Svetlana Kiritchenko, and Isar Ne- jadgholi. A friendly face: Do text-to-image systems rely on stereotypes when the input is under-specified? ArXiv, abs/2302.07159, 2023. 7

    2023 arXiv

  7. [15]

    Gallegos, Ryan A

    Isabel O. Gallegos, Ryan A. Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K. Ahmed. Bias and fairness in large language models: A survey. Computa- tional Linguistics, 50(3):1097–1179, 2024. 3

    2024

  8. [16]

    Unified con- cept editing in diffusion models

    Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzy ´nska, and David Bau. Unified con- cept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Com- puter Vision, pages 5111–5120, 2024. 2, 3

    2024

  9. [17]

    Demographic-aware language model fine-tuning as a bias mitigation technique

    Aparna Garimella, Rada Mihalcea, and Akhash Amar- nath. Demographic-aware language model fine-tuning as a bias mitigation technique. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Associa- tion for Computational Linguistics and the 12th Interna- tional...

  10. [18]

    Debias- ing pre-trained language models via efficient fine-tuning

    Michael Gira, Ruisu Zhang, and Kangwook Lee. Debias- ing pre-trained language models via efficient fine-tuning. In LTEDI, 2022. 3

    2022

  11. [19]

    What’s in a name? auditing large language models for race and gender bias

    Amit Haim, Alejandro Salinas, and Julian Nyarko. What’s in a name? auditing large language models for race and gender bias. ArXiv, abs/2402.14875, 2024. 2

    2024 arXiv

  12. [20]

    Aging with grace: Lifelong model editing with discrete key-value adaptors

    Thomas Hartvigsen, Swami Sankaranarayanan, Hamid Palangi, Yoon Kim, and Marzyeh Ghassemi. Aging with grace: Lifelong model editing with discrete key-value adaptors. In Advances in Neural Information Processing Systems, 2023. 3

    2023

  13. [21]

    Clipscore: A reference-free evaluation metric for image captioning

    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, 2021. 5

    2021 arXiv

  14. [22]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5

    2017

  15. [23]

    John Hewitt and Christopher D. Manning. A structural probe for finding syntax in word representations. In Pro- ceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguis- tics: Human Language Technologies, Volume 1 (Long and Shor...

    2019

  16. [24]

    Joniak and Akiko Aizawa

    Przemyslaw K. Joniak and Akiko Aizawa. Gender biases and where to find them: Exploring gender bias in pre- trained transformer-based language models using move- ment pruning. ArXiv, abs/2207.02463, 2022. 3

    2022 arXiv

  17. [25]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll ´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th Euro- pean Conference, Zurich, Switzerland, September 6-12, 2014, Proceed...

    2014

  18. [26]

    Mitigating political bias in language models through reinforced calibration

    Ruibo Liu, Chenyan Jia, Jason Wei, Guangxuan Xu, Lili Wang, and Soroush V osoughi. Mitigating political bias in language models through reinforced calibration. ArXiv, abs/2104.14795, 2021. 3

    2021 arXiv

  19. [27]

    Mace: Mass concept erasure in diffusion models

    Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6430–6440, 2024. 3

    2024

  20. [28]

    Multimodal bias: Assessing gender bias in computer vi- sion models with nlp techniques

    Abhishek Mandal, Suzanne Little, and Susan Leavy. Multimodal bias: Assessing gender bias in computer vi- sion models with nlp techniques. Proceedings of the 25th International Conference on Multimodal Interac- tion, 2023. 7

    2023

  21. [29]

    Nila Masrourisaadat, Nazanin Sedaghatkish, Fatemeh Sarshartehrani, and Edward A. Fox. Analyzing quality, bias, and performance in text-to-image generative mod- els. ArXiv, abs/2407.00138, 2024. 7

    2024 arXiv

  22. [30]

    On measuring so- cial biases in sentence encoders

    Chandler May, Alex Wang, Shikha Bordia, Samuel R Bowman, and Rachel Rudinger. On measuring so- cial biases in sentence encoders. arXiv preprint arXiv:1903.10561, 2019. 3, 7

    1903 arXiv

  23. [31]

    Locating and editing factual associations in gpt

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt. Advances in Neural Information Processing Sys- tems, 35:17359–17372, 2022. 3

    2022

  24. [32]

    Mass-editing mem- ory in a transformer

    Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. Mass-editing mem- ory in a transformer. arXiv preprint arXiv:2210.07229,

    arXiv

  25. [33]

    Fast model editing at scale

    Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. Fast model editing at scale. In International Conference on Learning Rep- resentations, 2022. 3

    2022

  26. [34]

    Editing implicit assumptions in text-to-image diffusion models

    Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision , pages 7053– 7061, 2023. 2, 3, 4, 5, 6, 8

    2023

  27. [35]

    Bias in word embeddings

    Orestis Papakyriakopoulos, Simon Hegelich, Juan Car- los Medina Serrano, and Fabienne Marco. Bias in word embeddings. In Proceedings of the 2020 conference on fairness, accountability, and transparency , pages 446– 457, 2020. 3

    2020

  28. [36]

    Direct and Indirect Effects, page 373–392

    Judea Pearl. Direct and Indirect Effects, page 373–392. Association for Computing Machinery, New York, NY , USA, 1 edition, 2022. 3, 7

    2022

  29. [37]

    Perturbation augmentation for fairer NLP

    Rebecca Qian, Candace Ross, Jude Fernandes, Eric Michael Smith, Douwe Kiela, and Adina Williams. Perturbation augmentation for fairer NLP. In Proceed- ings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages 9496–9521, Abu Dhabi, United Arab Emir...

    2022

  30. [38]

    Learning transferable visual models from natural lan- guage supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sas- try, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural lan- guage supervision. In International conference on ma- chine learni...

    2021

  31. [39]

    Hierarchical text-conditional image generation with clip latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 5

    2022 arXiv

  32. [40]

    On the systematicity of probing contextualized word representa- tions: The case of hypernymy in BERT

    Abhilasha Ravichander, Eduard Hovy, Kaheer Suleman, Adam Trischler, and Jackie Chi Kit Cheung. On the systematicity of probing contextualized word representa- tions: The case of hypernymy in BERT. In Proceedings of the Ninth Joint Conference on Lexical and Computa- tional Sema...

    2020

  33. [41]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of 10 the IEEE/CVF conference on computer vision and pat- tern recognition, pages 10684–10695, 2022. 2, 3, 5, S1

    2022

  34. [42]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Sali- mans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural informati...

    2022

  35. [43]

    LINSPECTOR: Multilingual probing tasks for word representations

    G ¨ozde G¨ul S ¸ahin, Clara Vania, Ilia Kuznetsov, and Iryna Gurevych. LINSPECTOR: Multilingual probing tasks for word representations. Computational Linguistics, 46 (2):335–385, 2020. 3

    2020

  36. [44]

    Jisu Shin, Hoyun Song, Huije Lee, Soyeong Jeong, and Jong C. Park. Ask llms directly, ”what shapes your bias?”: Measuring social bias in large language models. ArXiv, abs/2406.04064, 2024. 2

    2024 arXiv

  37. [45]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Mah- eswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning , pages 2256–2265. PMLR, 2015. 3

    2015

  38. [46]

    The biased artist: Exploiting cultural biases via homoglyphs in text-guided image generation models

    Lukas Struppek, Dominik Hintersdorf, and Kristian Ker- sting. The biased artist: Exploiting cultural biases via homoglyphs in text-guided image generation models. ArXiv, abs/2209.08891, 2022. 7

    2022 arXiv

  39. [47]

    What are the biases in my word embedding? In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pages 305–311, 2019

    Nathaniel Swinger, Maria De-Arteaga, Neil Thomas Heffernan IV , Mark DM Leiserson, and Adam Tauman Kalai. What are the biases in my word embedding? In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pages 305–311, 2019. 2, 3

    2019

  40. [48]

    Massive editing for large language models via meta learning

    Chenmien Tan, Ge Zhang, and Jie Fu. Massive editing for large language models via meta learning. In Inter- national Conference on Learning Representations, 2024. 3

    2024

  41. [49]

    Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N

    Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Neural Information Processing Systems, 2017. 3

    2017

  42. [50]

    Nationality bias in text generation

    Pranav Narayanan Venkit, Sanjana Gautam, Ruchi Pan- chanadikar, Ting-Hao ’Kenneth’ Huang, and Shomir Wilson. Nationality bias in text generation. In Con- ference of the European Chapter of the Association for Computational Linguistics, 2023. 3

    2023

  43. [51]

    Do NLP models know numbers? probing numeracy in embeddings

    Eric Wallace, Yizhong Wang, Sujian Li, Sameer Singh, and Matt Gardner. Do NLP models know numbers? probing numeracy in embeddings. In Proceedings of the 2019 Conference on Empirical Methods in Natu- ral Language Processing and the 9th International Joint Conference on Natural ...

    2019

  44. [52]

    White men lead, black women help? benchmarking language agency social bi- ases in llms, 2024

    Yixin Wan and Kai-Wei Chang. White men lead, black women help? benchmarking language agency social bi- ases in llms, 2024. 2

    2024

  45. [53]

    ”kelly is a warm person, joseph is a role model”: Gender biases in llm-generated reference letters

    Yixin Wan, George Pu, Jiao Sun, Aparna Garimella, Kai- Wei Chang, and Nanyun Peng. ”kelly is a warm person, joseph is a role model”: Gender biases in llm-generated reference letters. ArXiv, abs/2310.09219, 2023. 2

    2023 arXiv

  46. [54]

    Editing massive concepts in text-to- image diffusion models

    Tianwei Xiong, Yue Wu, Enze Xie, Yue Wu, Zhenguo Li, and Xihui Liu. Editing massive concepts in text-to- image diffusion models. ArXiv, abs/2403.13807, 2024. 7

    2024

  47. [55]

    Sohee Yang, Elena Gribovskaya, Nora Kassner, Mor Geva, and Sebastian Riedel. Do large language models latently perform multi-hop reasoning? In Proceedings of the 62nd Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers) , pages 10210–10229,...

    2024

  48. [56]

    Diverse diffusion: Enhancing image diversity in text-to-image generation

    Mariia Zameshina, Olivier Teytaud, and Laurent Naj- man. Diverse diffusion: Enhancing image diversity in text-to-image generation. ArXiv, abs/2310.12583, 2023. 7

    2023 arXiv

  49. [57]

    Deep learning on a healthy data diet: Finding important examples for fairness

    Abdelrahman Zayed, Prasanna Parthasarathi, Gonc ¸alo Mordido, Hamid Palangi, Samira Shabanian, and Sarath Chandar. Deep learning on a healthy data diet: Finding important examples for fairness. In AAAI Conference on Artificial Intelligence, 2022. 3

    2022

  50. [58]

    dog” but the test prompt is “puppy

    Jieyu Zhao, Tianlu Wang, Mark Yatskar, Ryan Cotterell, Vicente Ordonez, and Kai-Wei Chang. Gender bias in contextualized word embeddings. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lan- guage Tech...

    2019

  51. [2022]

    Association for Computational Linguistics. 3

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.