acceptodds
Withdrawn from ICLR 2027

When Steering Works: Predicting Debiasing Benefit from Denoiser Responses

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Semantic steering can reduce stereotypes in text-to-image generation, but it can also change content that the prompt asks for. Can we predict which prompts will benefit before decoding their images? We introduce Directional Gain (), which compares two denoiser responses at the same latent state: the change caused by steering and the change caused by replacing a stereotype prompt with its counterfactual. We measure agreement between these directions relative to random edits. Across SD3.5, FLUX, and Qwen-Image, with 1,295 test prompts per model, predicts improvement better than response size. At the top 20%, selection by gives larger stereotype reductions and usually smaller prompt-matching losses. Its advantage in combined-score improvement over response size is 4.7–20.7 points per selected prompt under Qwen3-VL; GPT-5.6 also favors in all three models. Averaged over six single-probe choices, one probe retains 89–93% of the mean improvement obtained by six-probe selection. It uses fewer denoiser evaluations than generating a clean–steered image pair and requires no image decoding or judging. With the steering settings fixed, the ranking also selects prompts that benefit on new generation seeds.

Primary area:
interpretability and explainable AI
Keywords:
text-to-image models, social bias, bias mitigation, activation steering, representation editing, interpretability, diffusion models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_LTKNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes a new method for predicting the effect of applying semantic steering to mitigate stereotypes in text-to-image models. This prediction can be made before final image generation, allowing users to apply semantic steering efficiently only to prompts for which the intervention is effective. The paper introduces a metric called Directional Gain (G). G measures the alignment between two directions: the change in the denoiser output induced by steering and the change induced by replacing a stereotypical prompt with its counterfactual counterpart. This allows the effectiveness of steering to be estimated without decoding or evaluating the generated images. Experiments on SD3.5, FLUX, and Qwen-Image show that G can predict the trade-off score between prompt fidelity and stereotype reduction better than the magnitude of the change in the model output.

Critical strengths
  1. Design of an efficient metric for estimating steering effectiveness. The proposed Directional Gain does not require final image decoding, so the effect of the semantic intervention can be efficiently estimated without waiting for the slow image generation process in diffusion models.
  2. Evaluation across multiple models and settings. The authors evaluate the proposed method on three different text-to-image models (SD3.5, FLUX, and Qwen-Image), using 1,295 test prompts per model. They also consider two independent image judges, different generation seeds, and varying numbers of denoiser probes. These experiments provide useful evidence of the consistency and computational efficiency of the proposed metric.
Critical weaknesses
  1. Insufficient motivation, practicality, and generality of the problem setting. This work focuses on the different effects of semantic steering in diffusion models and aims to predict the effect of an intervention without evaluating generated images. However, I am deeply concerned that the authors do not sufficiently explain when this prediction is necessary or how it can contribute to reducing bias. In particular, the proposed method recommends applying steering only to prompts with high prediction scores and generating images from the original prompts for those with low scores. Although avoiding ineffective interventions can be meaningful in terms of computational efficiency, I still wonder whether it is desirable to generate biased images from unmodified prompts for unselected cases, given the original goal of reducing stereotypes. For example, SAFREE (Yoon et al., 2025) and Concept Removal Guidance (Choi et al., 2026) can adaptively control the range and strength of interventions based on information extracted from semantic representations and denoiser predictions. Based on these prior works, it is unclear whether merely predicting if an intervention is necessary offers advantages over methods that adaptively adjust interventions.

    Furthermore, I am concerned about the practicality and generality of the proposed method. In this work, stereotype–counterfactual pairs are needed for both semantic steering and the computation of Directional Gain. It is not clearly stated how such pairs should be prepared for real-world text-to-image generation, making the method dependent on a strong assumption.

    Moreover, the authors use a steering method they designed, in which the same prompt pairs are used both to construct the intervention direction and to compute the reference direction for G. Consequently, it remains unclear whether the proposed metric generalizes to other existing steering methods or whether its effectiveness is specific to the intervention setup used in this paper.

  2. Lack of comparison with baselines. This work justifies the validity of G by comparing it with the magnitude of the denoiser response and the text-only predictor. I do not think comparisons with these baselines are sufficient to establish the necessity and advantages of the proposed metric.

    First, M considers only the magnitude of the change in the denoiser output and does not incorporate any information about the reference direction. Therefore, while comparing G with M is reasonable as an ablation, it is insufficient to demonstrate the superiority of the proposed metric. For example, Concept Removal Guidance (Choi et al., 2026) introduces a metric that estimates the presence of unwanted concepts based on the relationship between the denoiser's noise predictions and the guidance direction. Similarly, SAFREE (Yoon et al., 2025) uses projections onto concept subspaces and cosine similarity in text representations to adaptively adjust interventions. These prior works suggest that stronger alternative metrics leveraging the geometric properties of denoiser predictions and text representations should be considered.

    Second, the text-only predictor relies solely on numerical features, such as the norms and distances of text representations, without sufficiently considering the semantic content of the prompt pairs. Figure 4 shows that high-G examples often involve visually explicit attribute changes, such as gender swaps, whereas low-G examples tend to involve changes in personal traits or contextual information. Such differences might be predictable to some extent simply by analyzing the semantic content of the prompt pairs using an LLM. Furthermore, an LLM-based predictor would not require any denoiser evaluations from the image generation model and could be evaluated in advance or asynchronously, independently of the image generation process. These practical advantages make comparisons with stronger LLM-based semantic baselines particularly important. In particular, if G can identify cases in which an intervention appears promising based on textual information but fails to produce the intended effect in practice, such evidence would help establish the value of leveraging model-specific denoiser responses that cannot be inferred from textual semantics alone.

    Finally, M and G are compared using the same number of probes despite having different computational costs. M requires only the denoiser outputs before and after steering, whereas G additionally requires denoiser evaluations for the stereotype–counterfactual prompt pair. Thus, M has a lower computational cost per probe. Under the same computational budget, M could potentially use more denoising steps or random seeds. Therefore, a comparison under the same computational budget is necessary.

Choi et al. (2026). Concept Removal Guidance: Evidence-Calibrated Negative Guidance for Safe Diffusion Sampling. ICML 2026.

Yoon et al. (2025). SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image and Video Generation. ICLR 2025.

Questions for authors

Please refer to Weaknesses 1 and 2 above for my main questions and concerns.

Reviewer_jJo5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates which prompts benefit from a fixed debiasing intervention before images are decoded. Directional Gain compares the steering-induced denoiser change with a stereotype-to-counterfactual response at the same latent state. Across three generators, it predicts benefit better than response magnitude. Experiments also examine single-probe cost, two automated judges, and reuse across generation seeds. The reported primary-judge correlations are modest, approximately 0.19-0.24, making the selection results particularly important for understanding practical value.

Critical strengths
  1. The paper evaluates prompt-selection decisions, including performance under new seeds, demonstrating the utility of the diagnostic beyond correlation alone.

  2. Two judges, headroom baselines, and within-prompt-type analyses address several plausible alternative explanations. The discussion also acknowledges substantial remaining fidelity losses.

Critical weaknesses
  1. Single-probe numerical stability requires further assessment. Appendix B.5 reports per-probe correlations of 0.77/0.76 for SD3.5/FLUX when the probe-step set is expanded. Six-probe scores retain correlations of 0.91/0.93, and correlations with benefit change by at most 0.03. Changing the accelerator alters cosine values by up to 0.46. These findings show sensitivity to numerical context, while its effect on single-probe selection remains unclear. I would recommend reporting top-20% selection overlap and downstream benefit across probe-step sets, hardware, and precision.

  2. The combined score can indicate a benefit despite substantial degradation of requested content. In Table 3, prompt-matching rates for G-selected Qwen-Image prompts decrease by 24.1 percentage points under the primary judge. I would recommend comparison of selection under an explicit fidelity-loss constraint.

  3. Both judges apply an automated stereotype rubric, so their agreement does not establish human-perceived improvement. I would recommend human validation of a stratified sample.

  4. The within-group results in Table 22 are informative, but the advantage over magnitude is uncertain for several groups, warranting caution in interpreting the aggregate ranking gains. A text baseline incorporating contrast type would help quantify how much ranking value requires denoiser probes.

Questions for authors
  1. Does single-probe selection retain its advantage across probe-step sets, accelerators, and precision settings? Please include selection overlap, since small ranking changes near the cutoff can alter which prompts are steered.
Reviewer_73kHOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Directional Gain (G), a pre-generation diagnostic that predicts which prompts will benefit from a fixed semantic steering intervention for debiasing text-to-image models. The steering edit adds a stereotype-to-counterfactual text-embedding difference to the neutral prompt’s encoder output, and G decides which prompts should receive this edit. At a shared intermediate latent state, G computes the cosine similarity between the denoiser’s response to the steering edit and its response to swapping the stereotype prompt for its counterfactual, standardized against random edits. Prompts are ranked by G, only the top fraction is steered, and the rest are generated without modification. Across three generators, G predicts steering benefit better than response magnitude and selects prompts with larger stereotype reductions and smaller prompt-fidelity losses, while remaining cheap to compute and reusable across generation seeds.

Critical strengths

A well-motivated and practical problem formulation. Unlike many previous studies that proposed debiasing techniques, this paper asks a complementary question: given a fixed intervention, which prompts should receive it?

Critical weaknesses
  • The approach is validated with only one simple steering rule and lacks comparison with established debiasing methods. The contribution is framed as a general pre-generation diagnostic for “a specified semantic intervention,” but G is evaluated only with the authors’ own edit (adding a stereotype-to-counterfactual embedding difference at the text-encoder output). Since G is computed from the denoiser response to this particular edit, its usefulness for other interventions is not established. The paper should show whether G remains predictive when wrapped around established debiasing methods such as SEGA, ITI-GEN, LightFair, or FairPro [1]. In particular, [1] addresses stereotype mitigation in LLM-based T2I models including Qwen-Image, but it is neither cited nor discussed. A practical question also remains: does selective steering with G outperform simply applying a stronger debiasing method to all prompts? Currently, the baselines are limited to response magnitude, a text-feature predictor, a clean-image predictor, and random selection.
  • Evaluation of image quality and alignment is limited. Steering can reduce stereotypes while degrading visual quality or prompt fidelity. The paper measures prompt fidelity only through a binary VLM prompt-matching check, and even G-selected prompts show substantial matching losses (e.g., 24.1 pp in Qwen-Image at top 20%, Table 3). Image quality is not measured at all. The authors should report standard quality and alignment metrics (e.g., FID or aesthetic scores, CLIPScore, GenEval) for clean, steered-all, and G-selected generations.
  • Reliance on VLM judges without human validation. All outcomes, including stereotype ratings and prompt matching, come from Qwen3-VL and GPT-5.6. Agreement between the two judges is moderate (ρ ≈ 0.68–0.75, Table 12), and the paper provides no evidence that either judge agrees with human perception of stereotypes. The authors acknowledge this limitation, but the main claims depend entirely on these judges.

[1] Park et al. “Aligned but Stereotypical? How System Prompts Shape Demographic Bias in LLM-Based Text-to-Image Models” arXiv:2512.04981

Questions for authors
  • Can G be applied to other debiasing interventions (e.g., SEGA, LightFair, FairPro [1])? If so, does it remain predictive, and how does G-based selective steering compare with applying these methods to all prompts?
  • Could you report image quality and alignment metrics (e.g., FID, CLIPScore, GenEval) for clean, fully steered, and G-selected generations?
  • Is there evidence, either from prior work or a small human study, that Qwen3-VL or GPT-5.6 stereotype ratings agree with human judgments?
  • G requires a stereotype–counterfactual pair for every prompt. How would the method be used in practice when the relevant stereotype is not known in advance?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.