acceptodds
Under review as a conference paper at ICLR 2027

Controlling Image-Space Drift in Inference-Time Noise Optimization

Abstract

Inference-time alignment steers pre-trained diffusion models toward target rewards without updating model weights. Among these methods, optimizing the input noise is especially powerful: by directly modifying the generation process, it can reach rare, high-reward samples. However, this flexibility also makes reward hacking significantly worse, often producing high-scoring images that lose the structure and realism of the base model. Existing approaches try to prevent this by constraining the optimized noise, but we show this is insufficient: the denoising process can amplify small changes in noise into large differences in the generated images. Thus, we propose BIDS (Bounding Image-space Divergence via Score functions), which uses a score-based constraint to limit image-space drift during noise optimization. BIDS uses the Data Processing Inequality to relate image-space divergence to divergence over denoising trajectories, then computes a practical surrogate using only the frozen model's score function. This preserves image structure during reward optimization. Across UNet and DiT architectures, BIDS reduces distributional drift (CMMD) by over 75% and achieves up to an 80.30% human preference win rate. We further extend BIDS to pairwise preferences, enabling alignment to non-differentiable human feedback.

Primary area:
generative models
Keywords:
Diffusion models, inference-time alignment, noise optimization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_X87UOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes BIDS, a score-based regularizer to mitigate reward hacking in inference-time noise optimization for diffusion models. BIDS uses a score-energy surrogate to limit divergence between optimized and base denoising trajectories, requires no model fine-tuning, and is extended to non-differentiable rewards. Experiments across multiple diffusion models and reward functions show improved prompt alignment, reduced distributional drift, and higher preference win rates compared to existing baselines.

Critical strengths
  1. The paper addresses reward hacking in inference-time noise optimization, where higher rewards can lead to unrealistic images or loss of prompt alignment.
  2. BIDS introduces a score-based regularizer without updating the diffusion model or training an additional network. It can be integrated into existing noise optimization methods.
  3. The experiments cover multiple diffusion backbones and reward functions. BIDS achieves better preference win rates and lower distributional drift than several baselines. The additional ablations and early-stopping comparisons are also useful.
Critical weaknesses
  1. DNO already studies reward hacking and designed PRNO to keep optimized noise close to Gaussian statistics. DNO also supports non-differentiable rewards through gradient estimation. Therefore, the main contribution of BIDS seems to be limited.

  2. The paper argues that noise-space regularization is insufficient because small noise perturbations can produce large image changes. However, this only demonstrates the limitation of noise-space Euclidean distance, not noise-distribution regularization. For a fixed diffusion sampler, the image distribution is the pushforward of the noise distribution. The paper does not clearly explain why trajectory-space regularization is necessary instead of improving noise-distribution regularization.

  3. The paper defines KL divergence between optimized and base trajectory distributions, but the practical algorithm optimizes individual noise samples. In experiments, DDIM is not deterministic, but there are also single-step models. For deterministic sampling, it is unclear how this single-trajectory score-energy penalty represents the divergence between the full distributions induced by noise optimization.

Questions for authors
  1. Given the paper's counterexample in Section 3 only shows the limitation of noise-space L2 distance, rather than noise-distribution regularization. Could the authors clarify why trajectory-space regularization is necessary instead of directly regularizing the noise distribution? Is it because trajectory regularization may allow for more effective noise variations, yielding higher rewards for a given level of divergence in the image distribution?

  2. In deterministic DDIM, the same initial noise and prompt always produce the same trajectory and image. How are the optimized and base trajectory distributions defined? How does a penalty computed from a single pair of trajectories approximate the KL divergence between these distributions?

  3. Why should the difference between squared score norms approximate the trajectory KL? Can two very different trajectories have the same score energy?

7 comments
AuthorsOpenReview ↗

Response (1/2)

We thank the reviewer for recognizing the practical contribution and the early-stopping comparisons.

Weakness 1: "DNO already studies reward hacking and designed PRNO"; "DNO also supports non-differentiable rewards through gradient estimation."

Response: We agree with the reviewer that DNO addresses reward hacking through PRNO’s Gaussian noise regularization. As detailed in Appendix A.2, our DNO implementation already includes PRNO regularization. Despite this, Figure 3 shows that structural collapse persists under DNO, whereas BIDS preserves image structure.

BIDS instead regularizes denoising trajectories to limit image-space drift. Our key contribution is making this regularization tractable, using changes in the frozen model's score energy as a surrogate for trajectory KL (Eq. 5).

For non-differentiable rewards, DNO Hybrid-2 estimates reward gradients from perturbed samples, which may lead to reward hacking (as shown in Figure 6). BIDS uses the reward only to rank candidates, then differentiates a score-energy preference loss (Algorithm 2). This different update achieves a 63.56% preference win rate over Hybrid-2 with black-box Aesthetic Score across five seeds (Section 5.3).

These results demonstrate that BIDS makes image-distribution-aware regularization practical and effective in mitigating reward hacking, including for non-differentiable rewards.

Weakness 2 / Question 1: "Could the authors clarify why trajectory-space regularization is necessary instead of directly regularizing the noise distribution? Is it because trajectory regularization may allow for more effective noise variations, yielding higher rewards for a given level of divergence in the image distribution?"

Response: Yes. The motivation behind trajectory-space regularization is to allow more effective noise variations that improve reward while preserving the generated content. PRNO’s regularizer penalizes deviations from Gaussian noise statistics but does not capture how noise updates propagate through the sampler and affect the generated images. BIDS addresses this limitation by using score-energy changes along the denoising trajectory as a tractable surrogate for trajectory KL, thereby accounting for the sampler's response to noise perturbations.

Figure 13 provides empirical support for this motivation: BIDS achieves higher Darkness Reward at comparable CLIPScore, indicating better reward optimization while maintaining prompt alignment, even when DNO's regularization strength and stopping step are tuned. Appendix A.4 further shows analytically that, in the linear case, the noise-space penalty corresponding to image-space KL depends on the sampler.

Weakness 3 / Question 2: "The paper defines KL divergence between optimized and base trajectory distributions, but the practical algorithm optimizes individual noise samples. In experiments, DDIM is not deterministic, but there are also single-step models. For deterministic sampling, it is unclear how this single-trajectory score-energy penalty represents the divergence between the full distributions induced by noise optimization."

Response: A deterministic sampler still induces a distribution over trajectories because the initial noise is sampled from a Gaussian distribution. For deterministic DDIM, let Gθ(z,c)G_\theta(z,c) denote the trajectory generated from initial noise zz. The base trajectory distribution is induced by z∼N(0,I)z\sim\mathcal N(0,I). The optimized trajectory distribution is similarly induced by applying the noise-optimization procedure to each initial noise sample before generating its trajectory.

Like DNO, BIDS optimizes individual noise samples at inference time. The optimized distribution is defined implicitly by the mapping from each initial noise sample to its optimized noise, obtained by applying the same optimization procedure. Thus, the population-level formulation describes the distribution induced by repeating the algorithm over different initial noise draws, whereas the practical algorithm operates on individual samples.

Equation (5) defines a population-level score-energy surrogate for trajectory KL. Algorithm 1 regularizes individual noise updates with the paired penalty ∣S0−S∣|S_0-S|, computed from the base and optimized trajectories generated from the same initial noise.

By Jensen's inequality,

∣Ezinit[S0−S]∣≤Ezinit[∣S0−S∣].\left|\mathbb E_{z_{\mathrm{init}}}[S_0-S]\right|\leq\mathbb E_{z_{\mathrm{init}}}[|S_0-S|].

Thus, averaging Algorithm 1's paired penalties over initial noise draws upper-bounds the population score-energy penalty in Eq. 5.

For single-step models, the trajectory contains only one denoising step. BIDS therefore compares the squared norms of the model's predictions before and after noise optimization (Appendix A.2).

AuthorsOpenReview ↗

Response (2/2)

Question 3: "Why should the difference between squared score norms approximate the trajectory KL? Can two very different trajectories have the same score energy?"

Response: We motivate the score-energy surrogate for trajectory KL using the relationship between expected squared score norms and KL divergence for Gaussian distributions with fixed variance.

Consider p=N(0,1)p=\mathcal N(0,1) and q=N(a,1)q=\mathcal N(a,1). Here, pp is the base distribution, and qq changes its mean by aa without changing its variance. The base score is sp(x)=−xs_p(x)=-x, so

Eq[sp(x)2]−Ep[sp(x)2]=(1+a2)−1=a2.\mathbb E_q[s_p(x)^2]-\mathbb E_p[s_p(x)^2]=(1+a^2)-1=a^2.

The Gaussian KL formula gives

=\frac12\left(\mathbb E_q[s_p(x)^2]-\mathbb E_p[s_p(x)^2]\right).$$ Thus, in this fixed-variance Gaussian case, the difference in expected score energies is proportional to marginal KL. This motivates Eq. 5, which compares score energies before and after optimization across denoising steps to construct a computable surrogate for trajectory KL. **Different trajectories can indeed have the same score energy.** Squared score norms discard directional information. For example, under a two-dimensional standard Gaussian, the score is $s(x)=-x$. The points $(1,0)$ and $(0,1)$ both have squared score norm $1$. Although these points differ, they have the same base-model density. Furthermore, summing score energies across timesteps allows different per-step contributions to produce the same total. Therefore, equal score energy does not imply identical trajectories or guarantee small trajectory KL. Eq. 5 uses score energy as a practical surrogate rather than an exact characterization of trajectory divergence.
Reviewer_X87UOpenReview ↗

Thank you

Thank you for the detailed clarification and additional empirical comparisons. I have no concerns about contribution or novelty now. However, my remaining question is the justification for using score-energy differences as a surrogate. As the authors acknowledge, different trajectories can have identical score energies. It is also unclear to me why score-energy differences should be preferred over other regularizers, like ∑twt∣xtopt−xtbase∣22\sum_t w_t |x_t^{\mathrm{opt}}-x_t^{\mathrm{base}}|_2^2? Is the choice more empirically motivated than theoretically justified as an approximation to trajectory KL? Thank you again!

Reviewer_VU76OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies inference-time optimization of diffusion models, where the noise inputs are optimized while model weights remain fixed. The first contribution is a new, simple regularizer based on score energy along the denoising trajectory. It penalizes changes in this energy from the initial trajectory to reduce reward hacking. The second contribution is a simple pairwise method for optimizing non-differentiable rewards. It compares two generated images and uses a preference loss to update both noise sequences. The same score-energy statistic provides a differentiable surrogate reward for these updates. The authors test both methods across several diffusion models and rewards. They measure distributional drift and prompt alignment, and conduct human and automated preference evaluations. The results suggest less reward hacking and a better balance between reward and image fidelity.

Critical strengths
  1. The proposed methods are simple and appear easy to implement within an existing noise-optimization pipeline. Using the frozen model's score energy for both regularization and preference optimization is an appealing design. This simplicity makes the approach practically attractive.

  2. The reported empirical results are promising across several diffusion models and rewards. The image comparisons and preference evaluations suggest that the regularizer reduces visible reward hacking while retaining reward gains. This provides evidence of practical value beyond the method's conceptual motivation.

Critical weaknesses
  1. I do not see how Eq.5 follows from Eq.4. The Gaussian argument on page 5 does not seem sufficient. For example, let p=N(0,1)p=\mathcal{N}(0,1) and q=N(a,1−a2)q=\mathcal{N}(a,1-a^2), with 0<a<10<a<1. The base score is sp(x)=−xs_p(x)=-x. Its expected squared norm is 11 under both distributions, but DKL(q∥p)=−12log⁡(1−a2)D_{\mathrm{KL}}(q\|p)=-\tfrac{1}{2}\log(1-a^2) can be arbitrarily large. This is a simple example rather than a full diffusion construction, but it shows why matching expected score energy alone cannot control KL. Eq.~4 is a reasonable starting point; what is missing is a justification for replacing its KL term with the proposed penalty. I would like the authors to state assumptions under which this replacement is valid. The empirical correlation in Appendix A.7 does not establish that connection.

  2. I have a similar concern about the pairwise method: why should its update improve the actual reward? Take a scalar example with S(z)=z2S(z)=z^2 and reward r(z)=(z−1.4)2r(z)=(z-1.4)^2 to maximize. Start at zw=2z_w=2 and zl=1z_l=1, with reference energies 44 and 11. The rewards are 0.360.36 and 0.160.16, so the winner is correctly identified. With β=1\beta=1, one update gives zw′=2−2αz_w'=2-2\alpha and zl′=1+αz_l'=1+\alpha. The reward changes are −2.4α+4α2-2.4\alpha+4\alpha^2 and −0.8α+α2-0.8\alpha+\alpha^2. Both are negative for every 0<α<0.60<\alpha<0.6, even with an arbitrarily small learning rate, while the surrogate loss improves. This does not mean the method cannot work in practice. But my counterexample above shows that the update can worsen both candidates for general rewards and score functions. The authors should study what assumptions are needed for the method to work.

Questions for authors

I am curious why such a simple preference optimization method in section 4.1 works in your experiments. My counterexample shows that the update can worsen both candidates, but there may be a class of problems where it reliably improves the reward. Diffusion noise optimization might belong to that class. Under what assumptions on the scores and rewards would each update improve the actual reward?

3 comments

We thank the reviewer for recognizing the method's simplicity and empirical value.

Weakness 1: "I do not see how Eq.5 follows from Eq.4"; "state assumptions under which this replacement is valid."

Response. We clarify the Gaussian assumptions motivating the score-energy surrogate in Eq. 5. This surrogate is motivated by the trajectory KL objective in Eq. 4.

Consider p=N(0,1)p=\mathcal N(0,1) and q=N(a,1)q=\mathcal N(a,1), where optimization changes the mean while preserving the variance. The base score is sp(x)=−xs_p(x)=-x. Therefore,

=(1+a^2)-1=a^2.$$ The Gaussian KL formula gives $$D_{\mathrm{KL}}(q\|p) =\frac{a^2}{2} =\frac12\left(\mathbb E_q[s_p(x)^2]-\mathbb E_p[s_p(x)^2]\right).$$ Thus, **for Gaussian distributions with fixed variance, the expected score-energy difference is proportional to the marginal KL divergence**. As the reviewer's counterexample illustrates, this relationship does not generally hold when the variance also changes. We therefore use the score-energy difference as a practical surrogate motivated by this restricted Gaussian relationship, rather than an exact expression for trajectory KL. We have clarified this assumption in Proposed Approach (Section 4) of the revised manuscript. > **Weakness 2 and Question:** "why should its update improve the actual reward?"; "Under what assumptions on the scores and rewards would each update improve the actual reward?" **Response:** We clarify sufficient conditions under which the update in Algorithm 2 improves the actual reward of the preferred image. **Smooth rewards.** For a fixed prompt, let $r(z)$ denote the actual reward of the image generated from noise $z$, and let $S(z)$ denote its trajectory score energy from Eq. 8. Algorithm 2 uses the actual reward to identify the preferred candidate $z_w$ and updates it in the direction $$z_w'=z_w-\eta\nabla S(z_w),$$ where $\eta>0$ incorporates the learning rate and the positive scaling factors in the preference loss. To allow score energy to approximate reward imperfectly near the preferred candidate, write $$r(z)=-\lambda S(z)+e(z),\qquad \lambda>0,$$ where $e(z)$ represents the discrepancy between the actual reward and the scaled negative score energy $-\lambda S(z)$. For locally smooth $r$ and $S$, a sufficient condition at $z_w$ is $$\|\nabla e(z_w)\|<\lambda\|\nabla S(z_w)\|.$$ This condition permits an imperfect surrogate: the error gradient must be smaller than the scaled score-energy gradient. It does not require a global proportional relationship between reward and negative score energy. Since $\nabla r=-\lambda\nabla S+\nabla e$, the Cauchy–Schwarz inequality gives $$-\nabla r(z_w)^\top\nabla S(z_w)\geq\lambda\|\nabla S(z_w)\|^2-\|\nabla e(z_w)\|\,\|\nabla S(z_w)\|>0.$$ Thus, the score-energy descent direction is an ascent direction for the actual reward. For a sufficiently small learning rate, the preferred candidate's reward increases, and consequently the best reward in the pair also increases. In the reviewer's counterexample, $r'(2)=1.2$ and $S'(2)=4$, so the update decreases the actual reward. **Non-differentiable rewards.** For a reward such as JPEG compression, maximizing $r$ means minimizing the encoded file size. Improvement can instead be characterized directly through the finite update. If $S(z_w')<S(z_w)$ and $$|e(z_w')-e(z_w)|<\lambda[S(z_w)-S(z_w')],$$ then $$r(z_w')-r(z_w)=\lambda[S(z_w)-S(z_w')]+e(z_w')-e(z_w)>0.$$ This condition requires the change in approximation error to be smaller than the reward improvement associated with the reduction in score energy, without requiring reward gradients.

Thanks for addressing my questions! I'll maintain my current acceptance score.

Dear Reviewer,

We want to thank you for your thoughtful review and constructive feedback. Your comments have helped us improve the clarity and technical presentation of our paper, and we will incorporate these improvements into the revised manuscript.

Best regards,
Authors

Reviewer_U57XOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose BIDS, which they claim is an inference-time optimization objective designed to mitigate structural collapse in noise-space optimization. Motivated by the observation that small perturbations in noise can lead to substantial changes in generated images, they shift the regularization target to the image space. By leveraging the Data Processing Inequality, the authors upper-bound image-space divergence using trajectory divergence and construct a score-based surrogate objective to preserve image structure during reward optimization. They further claim to demonstrate that BIDS can be extended to discrete preference optimization, enabling alignment with non-differentiable rewards without offline fine-tuning or numerical gradient estimation.

Critical strengths

The paper is clearly written and easy to follow.

Critical weaknesses
  1. The novelty of the paper is limited, as numerous existing works have explored noise optimization. The authors need to clearly distinguish their method from prior work.

  2. The main experiments are primarily conducted on UNet-based architectures, and it remains unclear whether the proposed method is effective for flow-matching models. Given that flow-matching models are currently mainstream in the field, I believe it is necessary to evaluate the proposed method on such models. Specifically, the authors should report results on SD3.5, SD3, and FLUX.

  3. The paper contains errors in mathematical notation, as the symbols are not formatted according to their respective mathematical meanings. This reduces the readability of the paper.

Questions for authors

The authors should explain why they did not conduct experiments on flow-matching architectures. To the best of my knowledge, since last year, the research community has increasingly focused on flow-matching models rather than UNet-based diffusion models.

6 comments

We thank the reviewer for the clear assessment of the presentation. We address the contribution and experimental scope below.

Weakness 1: "The novelty of the paper is limited, as numerous existing works have explored noise optimization. The authors need to clearly distinguish their method from prior work."

Response: We agree that prior work has established noise optimization as an effective inference-time approach. However, as the reviewer also noted, BIDS shifts regularization to image space which mitigates reward hacking prevalent in prior approaches, while preserving prompt fidelity. Our key contribution is making this regularization tractable through a score-based surrogate.

The following table distinguishes BIDS from the noise-optimization baselines evaluated in our paper:

MethodUser-specified reward optimizationBlack-box rewardsRequires auxiliary informationImage-distribution drift objective
ReNO [1]✓✗✗✗
DNO [2]✓✓✗✗
InitNO [3]✗✗✗✗
DyMO [4]✓✗✓✗
BIDS✓✓✗✓

ReNO, DNO, and InitNO regularize in the noise space. Noise-space regularization can still lead to reward hacking, as Figure 3 shows qualitatively for DNO, while BIDS reduces this reward hacking by regularizing the output image space. These distinctions establish that the novelty of BIDS lies in its regularization formulation, its tractable surrogate, and the empirical results demonstrating its effectiveness in mitigating reward hacking.

References

[1] Eyring et al., ReNO: Enhancing One-Step Text-to-Image Models through Reward-based Noise Optimization, NeurIPS 2024.

[2] Tang et al., Inference-Time Alignment of Diffusion Models with Direct Noise Optimization, ICML 2025.

[3] Guo et al., InitNO: Boosting Text-to-Image Diffusion Models via Initial Noise Optimization, CVPR 2024.

[4] Xie and Gong, DyMO: Training-Free Diffusion Model Alignment with Dynamic Multi-Objective Scheduling, CVPR 2025.

Weakness 2 / Question: "it remains unclear whether the proposed method is effective for flow-matching models."; "The authors should explain why they did not conduct experiments on flow-matching architectures."

Response: We have extended our evaluation to SANA-0.6B [1], an efficient flow-matching DiT model whose published results are competitive with SD3-medium and FLUX-dev (Table 7 of the SANA paper). Using the score estimate derived in Appendix A.13, we compare BIDS with DNO on Darkness optimization across 45 white-animal prompts and 50 HPDv2 prompts.

White-animal prompts (45).

DNO γDNO Darkness Reward ↑BIDS Darkness Reward ↑BIDS Win Rate ↑
0.10.7430.76167.8%
10.5430.76172.2%

The two comparisons use the same 45 prompts and BIDS images. BIDS remains preferred when weaker DNO regularization increases DNO's achieved reward.

HPDv2 prompts (50).

MetricDNO (γ=1)BIDS (Ours)
Mean Darkness Reward ↑0.7270.901
Win Rate ↑7.0%93.0%

Both methods use the same SANA sampler, identical initial noise (seed 0), and 50 optimization iterations over all 20 noise inputs. BIDS uses β=0.08 throughout. Each pair is evaluated in both image orders with GPT-6.1-Sol. Section 5.2 now reports the white-animal results in Table 3. Appendix A.14 gives the settings, HPDv2 results in Table 10, and base/DNO/BIDS qualitative comparisons from the white-animal prompts.

Key Takeaway: On SANA-0.6B, BIDS achieves higher mean Darkness Reward and better image preferences than DNO on both prompt sets, including the weaker DNO regularization setting.

[1] Xie et al., SANA: Efficient High-Resolution Text-to-Image Synthesis with Linear Diffusion Transformers, ICLR 2025. arXiv:2410.10629.

Weakness 3: "symbols are not formatted according to their respective mathematical meanings."

Response: To address this concern and to improve clarity, we have added a notations table at the beginning of the appendix to define the notation and mathematical meanings, with a footnote in Problem Formulation pointing to it.

AuthorsOpenReview ↗

Urgent Request to the Reviewer

Dear Reviewer,

Thank you again for your thoughtful review and constructive feedback.

As the reviewer discussion phase is approaching its end, we wanted to kindly follow up on our rebuttal. We have carefully addressed your concerns by clarifying the novelty of BIDS, adding new experiments on flow-matching, and improving the mathematical notation in the revised manuscript.

We would be very grateful if you could take a moment to review our responses and let us know whether they address your concerns or if any further clarification would be helpful.

We sincerely appreciate your time, efforts, and service to the research community.

Regards,

Authors

Dear Authors,

Thank you for the authors’ response. However, the authors have not directly addressed the second issue raised in the Critical Weaknesses, which I consider a fatal limitation of the paper. Although the authors state in the rebuttal:

“We have extended our evaluation to SANA-0.6B [1], an efficient flow-matching DiT model whose published results are competitive with SD3-medium and FLUX-dev (Table 7 of the SANA paper).”

To the best of my knowledge, the prevailing consensus in this field is that a strong method should be compatible with flow-matching architectures, whereas the UNet architecture belongs to an earlier generation of frameworks. Therefore, evaluating the method on SANA-0.6B alone does not sufficiently address my concern regarding its compatibility with modern flow-matching models.

Regarding the notation issue raised in the third point of my Critical Weaknesses, although the authors redefined the symbols in the appendix, they still did not follow the ICLR formatting requirements by using the appropriate fonts and styles according to the meanings of the symbols.

Furthermore, given that the proposed method does not appear to be compatible with flow-matching architectures, I still believe that its novelty is limited. Thank you again for the authors’ efforts and response. Unfortunately, considering the paper’s overall quality and the issues not directly addressed in the rebuttal, I am inclined to maintain my original rating. Since flow-matching models are now widely recognized as a mainstream generative architecture, I strongly encourage the authors to incorporate a flow-matching framework in future work to improve the method’s compatibility and generality.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Manuscript Revisions

We thank the reviewers for recognizing the paper's clear presentation, practical simplicity, and empirical value. We have revised the manuscript, with changes highlighted in blue.

  • New flow-matching experiments. Section 5.2 evaluates BIDS on SANA-0.6B for Darkness optimization. On 45 white-animal prompts, BIDS achieves 67.8% and 72.2% Win Rates against DNO with two regularization strengths (Table 3). On 50 HPDv2 prompts, BIDS achieves a 93.0% Win Rate (Table 10). These Win Rates use GPT-6.1-Sol. Appendix A.14 gives the settings and white-animal comparisons in Figure 15.
  • Clarified surrogate motivation. Proposed Approach (Section 4) now states the fixed-covariance Gaussian assumption relating expected score energy to marginal KL.
  • Added notation reference. Table 4 defines the main symbols, with a footnote in Problem Formulation pointing to it.

Our responses below address each reviewer's questions in detail.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.