acceptodds
Under review as a conference paper at ICLR 2027

Endogenous Reprompting: Self-Evolving Cognitive Alignment for Unified Multimodal Models

Abstract

Unified Multimodal Models (UMMs) exhibit strong understanding, yet this capability often fails to effectively guide generation. We identify this as a Cognitive Gap: the model lacks the understanding of how to enhance its own generation process. To bridge this gap, we propose Endogenous Reprompting, a mechanism that transforms the model’s understanding from a passive encoding process into an explicit generative reasoning step by generating self-aligned descriptors during generation. To achieve this, we introduce SEER (Self-Evolving Evaluator and Reprompter), a training framework that establishes a two-stage endogenous loop using only 300 samples from a compact proxy task, Visual Instruction Elaboration. First, Reinforcement Learning with Verifiable Rewards (RLVR) activates the model’s latent evaluation ability via curriculum learning, producing a reliable endogenous reward signal. Second, Reinforcement Learning with Model-rewarded Thinking (RLMT) leverages this signal to optimize the generative reasoning policy. Experiments show that SEER consistently outperforms state-of-the-art baselines in evaluation accuracy, reprompting efficiency, and generation quality, without sacrificing general multimodal capabilities.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Unified Multimodal Models, Chain-of-Thought Reasoning, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_WjZ9OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SEER, a framework designed to bridge the cognitive gap between understanding and generation in Unified Multimodal Models (UMMs). The proposed approach introduces Endogenous Reprompting, which enables a unified model to translate user instructions into generation-oriented prompts aligned with its own generative capabilities. SEER consists of two stages: (1) Reinforcement Learning with Verifiable Rewards (RLVR), which trains an internal evaluator through curriculum learning and pairwise visual comparisons, and (2) Reinforcement Learning with Model-rewarded Thinking (RLMT), which optimizes the reprompting policy using feedback from the internal evaluator.

Critical strengths

The paper addresses an important challenge in unified multimodal models: effectively transferring visual understanding capabilities to image generation.

Critical weaknesses
  1. A central claim of the paper is that using the internal comprehension capability of UMMs to generate the reprompt can better match the image generator's generative priors than an external LLM. The author uses Table 2 to demonstrate that SEER outperforms external reprompters such as GPT-5.2, Gemini3, and PromptEnhancer. However, this experiment only shows that SEER's reprompting effect is better, and cannot fully prove that the advantage comes from the consistency of the model's internal representations because Seer has also introduced other training mechanisms. A suggestion is training the respective Reprompter for each of the two models, then exchange generators for testing. Only when the model significantly outperforms the cross-combination during matching can it better support Model-Specific Alignment.

  2. The internal evaluator carries the risk of reward hacking and self-reinforcement bias. Reprompter learns to generate visual features preferred by the Evaluator, but these features do not necessarily align with human preferences. Additionally, the Evaluator and Reprompter share correlated errors, causing the training process to continuously amplify these errors.

Questions for authors
  1. Could the authors provide more evidence to support the claim that using the internal comprehension capability of UMMs to generate the reprompt can better match the image generator's generative priors than an external LLM?
  2. Could the authors provide additional evidence of generalization on larger, independent benchmarks with more diverse and compositional visual instructions?
3 comments

Thank you for recognizing the importance of transferring visual understanding to image generation and for suggesting the cross-model reprompter exchange. We hope our response will resolve your concerns.

W1 & Q1 — Cross-model reprompter exchange.

Following your suggestion, we trained reprompters for Harmon-1.5B and Show-o2-1.5B with the same 300 examples and SEER procedure, using feedback from their respective UMMs, then swapped their reprompts across fixed generators.

ReprompterHarmon generator (CLIP / ImageReward)Show-o2 generator (CLIP / ImageReward)
Harmon-trained28.63 / 1.13 (matched)27.55 / 0.62 (cross)
Show-o2-trained27.84 / 0.94 (cross)28.48 / 0.89 (matched)

Both generators perform better with their matched reprompters on CLIP and ImageReward. This supports model-specific alignment.

W2 — Reward hacking and correlated errors.

We checked evaluator decisions against independent human judgments using eight image pairs sampled every 10 RLMT steps, obtaining 90.4% agreement. Blind human evaluation also favors SEER over the Harmon base (0.85, Table 3). These results provide independent evidence that the learned improvements align with human preferences rather than only the internal evaluator's judgments.

Q2 — Independent WISE evaluation.

We evaluate generation with and without SEER reprompting on WISE (Niu et al., ICML 2026), with 1,000 prompts covering 25 subdomains. Both settings use the same trained model and generation settings.

Domain / metric ↑Direct generationWith SEER repromptingChange
WiScore (weighted)0.55890.6453+0.0864
Cultural commonsense0.57060.6975+0.1269
Temporal reasoning0.54220.6162+0.0740
Spatial reasoning0.60640.7008+0.0944
Biology0.62900.6300+0.0010
Physics0.54000.5960+0.0560
Chemistry0.42550.4760+0.0505
CLIP23.21924.084+0.865
ImageReward−0.398−0.158+0.240

The largest improvements are in cultural commonsense and spatial/temporal reasoning. Biology is nearly unchanged. WiScore was evaluated using a Qwen-based judge.

Thanks for your rebuttal. I maintain my score.

Thank you for your time and thoughtful review. We sincerely appreciate your positive assessment and consideration of our work.

Reviewer_TVuLOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the gap between understanding and generation in unified multimodal models (UMMs), a problem also discussed in prior work. It explores improving generation through prompt rewriting guided by the model’s own understanding.

The proposed framework uses two stages of reinforcement learning:

  1. Evaluator training: Strengthen the model’s ability to judge which image better satisfies a user’s instruction.
  2. Prompt-rewriting optimization: Fix the trained evaluator and use its feedback as a reward to optimize the prompt-rewriting policy.

The visual generation components remain frozen throughout training.

Using 300 training samples, SEER achieves 92% evaluator accuracy and an 85% human-evaluation win ratio against the base model on the authors’ visual instruction benchmark. It also outperforms the evaluated external prompt rewriters while largely preserving performance on general multimodal benchmarks.

Critical strengths
  1. The paper addresses a meaningful question for UMM development: how understanding can more effectively support generation. The introduction uses concrete examples to explain the difficulty of translating abstract user requirements into visual outputs. The proposed task, covering six instruction categories, makes this broader motivation experimentally accessible.

  2. The two-stage framework is intuitive and reasonably motivated. Training the evaluator before using its feedback avoids assuming that the original model already provides reliable judgments. Evaluating rewritten prompts through their generated images also aligns the optimization objective with instruction following rather than linguistic quality alone.

  3. The experiments provide direct evidence of improved outputs on the studied task. Comparisons using the same generator help assess different rewriters, while blind human evaluation provides validation beyond the internal reward. The reported human-evaluation win ratios of 85% against the base model and 78% when applying the method to Show-o2 support its practical potential.

  4. The paper is generally well organized, with diagrams and examples that make the training workflow understandable. The appendices provide additional data-construction details and prompt templates. Evaluations of general understanding and generation capabilities also address whether targeted training compromises other functions, an important consideration for UMMs.

Critical weaknesses
  1. The experiments do not fully establish the proposed explanation of model-specific alignment. SEER is adapted using feedback from the target generator, but the paper does not report an external rewriter receiving comparable adaptation. The observed advantage therefore does not clearly distinguish benefits of unified representations from those of generator-specific training.

  2. The contribution of individual design choices remains unclear. The evaluator comparisons do not isolate the effects of the alignment warm-up or reinforcement learning relative to supervised training. Since the updated understanding parameters also provide generation conditioning, the overall gains cannot be attributed exclusively to changes in the rewritten prompt text.

  3. The evidence is strongest for the custom visual instruction task, while broader generation benefits remain less established. GenEval and DPG-Bench are evaluated with rewriting disabled, appropriately testing capability preservation rather than rewriting effectiveness. The numbers of subjects, image pairs, and generation trials in the main evaluation also require clarification to assess stability and coverage.

  4. Evaluator reliability as the rewriting policy improves remains insufficiently examined. Training negatives mainly omit the requested modification, whereas improved outputs may require finer distinctions. The first-round evaluator achieves 59% accuracy on a separate, harder test set. Second-round training helps, but the paper does not show whether internal rewards remain aligned with human judgments throughout policy optimization.

Questions for authors
  1. What evidence supports attributing the gains specifically to shared understanding–generation representations? Could the authors clarify which conclusions follow from the current comparisons and which remain hypotheses, particularly relative to an independently trained rewriter receiving comparable generator feedback?

  2. Can the authors use existing checkpoints or targeted comparisons to separate improvements from prompt rewriting and changes in generation conditioning? How important is the initial alignment warm-up compared with directly training the evaluator on visual instruction judgments?

  3. How does evaluator–human agreement change as the policy produces increasingly competitive images? Does the harder-set evaluation suggest a practical criterion for updating the evaluator? When constructing second-round training pairs, how is the preference for newly generated images over the original positives verified?

  4. What are the expected benefits and limitations of enabling rewriting on standard generation benchmarks or broader instruction sets? Could the authors also specify the main evaluation’s numbers of subjects, image pairs, and generation trials, and provide an appropriate measure of uncertainty?

4 comments

Thank you for noting that the two-stage framework is “intuitive and reasonably motivated” and that blind human evaluation provides evidence beyond the internal reward. We hope our response will resolve your concerns.

W1 & Q1 — Model-specific alignment.

Following WjZ9's suggested cross-model exchange (full results in WjZ9, W1 & Q1), we swapped the SEER reprompts of Harmon-1.5B and Show-o2-1.5B across fixed generators. Both start from Qwen2.5-1.5B-Instruct but are aligned with their respective generators. With reprompters trained on the same 300 examples, matched pairings perform better: Harmon 28.63 / 1.13 vs. 27.84 / 0.94 and Show-o2 28.48 / 0.89 vs. 27.55 / 0.62 (CLIP / ImageReward). This supports generator-specific matching.

W2 & Q2 — Reprompting versus conditioning, and the warm-up.

Reprompting. Table 4 shows modest changes in direct generation after SEER training, with reprompting disabled (GenEval 0.72 → 0.74; DPG-Bench 80.5 → 81.1). To measure the additional effect of reprompting, we evaluate the trained SEER model on WISE with and without reprompting. Enabling it improves weighted WiScore by 0.0864, CLIP by 0.865, and ImageReward by 0.240. This separates the benefit of reprompting from changes in generation conditioning introduced by training.

Warm-up. Removing the initial basic-alignment warm-up and directly training the evaluator on instruction comparisons gives 96.7% training accuracy but 82% test accuracy. The full curriculum reaches 92% test accuracy (Table 1). The warm-up therefore improves generalization on the held-out evaluation set.

W3 & Q4 — Evaluation coverage and generalization.

The test set contains 40 instructions with five new base prompts each, split into 40 / 60 / 40 / 60 inputs across ID-Simple, ID-Hard, OOD-Simple, and OOD-Hard. Each method generates one image per input using a matched seed, with no exclusions; all six annotators judge every pair. Appendix A.2 reports Fleiss' κ = 0.72 and significance tests. We also evaluate the full pipeline on 1,000 WISE prompts (WiScore +0.0864, CLIP +0.865, ImageReward +0.240). Per-cell sizes and tie rates are detailed in 8M7P (W1 & Q1).

W4 & Q3 — Evaluator reliability during RLMT and Round 2.

We sampled eight image pairs every 10 RLMT steps and compared evaluator decisions with independent human judgments, obtaining 90.4% agreement. On the harder SEER-Eval2 set of closely matched images, the Round-1 evaluator achieves only 59% accuracy. Training it further with harder comparisons raises accuracy to 76% (Table 13). For Round 2, we use Round-1 generations as candidate positives and check their preference over the original positives through VLM-assisted and human verification before RLVR.

Dear Reviewer,

Thank you again for your thoughtful and constructive feedback on our submission. We sincerely appreciate your time and effort throughout the review process.

As the November 18 discussion deadline approaches, we would like to kindly follow up on our previous responses and additional experiments. We hope they have adequately addressed your concerns.

If you have had a chance to review our responses, we would greatly appreciate knowing whether any concerns remain unresolved. We would be more than happy to provide further clarification or additional information before the discussion period ends.

Thank you once again for your time and consideration.

Best regards,

The Authors of Paper 3068

Reviewer_TVuLOpenReview ↗

Response to author

Thanks for the rebuttal! The WISE experiment answered my main question about whether prompt rewriting helps, so I’m raising my score from 2 to 3.

Thank you for your time and thoughtful review! We sincerely appreciate your positive assessment and consideration of our work!

Reviewer_8M7POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SEER. It first employs RLVR to train a unified multimodal model’s comprehension capabilities, effectively turning it into an internal pairwise evaluator. Subsequently, using the relative rewards provided by this evaluator, it utilizes RLMT to optimize a "reprompting" policy—applied prior to the generation phase—thereby enhancing the model’s ability to execute complex visual instructions while keeping the visual generator frozen. The core contribution of this work is refining the cognitive translation process that bridges comprehension and generation.

Critical strengths

(1) This paper focuses on a problem in UMMs: strong comprehension capabilities alongside weak generation abilities.

(2) Unlike methods that directly use internal comprehension modules to optimize generation parameters, SEER employs internal evaluation signals to optimize the reprompting policy for intermediate layers.

Critical weaknesses

(1) The scope of the primary evaluation is too limited. Appendix A.1.3 merely states that the test set consists of 40 instructions, split evenly between in-distribution (ID) and out-of-distribution (OOD) samples, with a ratio of approximately 2:3 between simple and hard tasks. However, Tables 2 and 3 further subdivide these into various cells—categorized by ID/OOD status, task difficulty (simple/hard), and different baselines—meaning some conclusions may be based on as few as a dozen or so instructions. The paper fails to specify several key details: the number of prompts per instruction, the number of images generated per sample, the number of random seeds used, whether all six annotators evaluated every single pair, the proportion of tied pairs, and the actual number of valid samples per cell.

(2) The current reward mechanism applies only to the final reprompt and its resulting image. It does not directly constrain the <think> process to make a causal contribution to the final answer. It is entirely possible for the model to learn a suboptimal reasoning trace while relying primarily on a direct mapping within the answer component. The paper lacks several critical ablation studies: an answer-only version (omitting the CoT and outputting only the <answer>); a version using a fixed-template CoT; a negative control where the correspondence between reasoning and the answer is randomized; a supervised fine-tuning (SFT) baseline using the same 300 samples for the reprompt; and an RLMT variant that replaces the internal evaluator with an external one. Furthermore, the existing data-scale ablation reports only the evaluator's accuracy—omitting the impact of using 50, 150, or 300 samples on the final human win ratio—which makes it impossible to substantiate the stronger claim that "300 samples yield high-quality generative reasoning."

(3) In Table 4, reprompting is explicitly disabled for both GenEval and DPG-Bench, serving only to verify that Stages 1 and 2 do not compromise the model's direct generation capability. This merely indicates the absence of significant adverse side effects; it does not demonstrate that SEER provides any benefit for general text-to-image (T2I) generation. Given that reprompting is the core contribution of the paper, it is problematic that this mechanism is disabled when evaluating on standard benchmarks.

(4) UniCorn also employs internal model-based evaluation or self-generated supervision to enhance generation. The paper distinguishes its approach by stating, "They optimize the generation process, whereas we optimize the reprompting." While this distinction is valid, it lacks a direct comparison based on an equivalent data budget and the same underlying backbone model. In particular, it would be valuable to compare the performance of "using an internal evaluator to directly optimize generation" versus "using an internal evaluator to optimize reprompting," tested on the same benchmark (such as Harmon or Show-o2).

(5) If efficiency is to be considered a contribution, the current metrics are insufficient. The average of 22.94 tokens shown in Table 2 accounts only for the final reprompt. It fails to include the additional <think> tokens generated during SEER’s reasoning process, the total number of output tokens, end-to-end latency, or the overhead associated with image generation.

Questions for authors

(1) The appendix only specifies that the test set contains 40 instructions. Please clarify how many base prompts correspond to each instruction, the number of images generated by each method, the number of random seeds used, and the effective sample size for each cell in Table 2 and Table 3.

(2) The core mechanism of the paper is reprompting, yet standard generation benchmarks only validate direct generation. Please provide the full results for the SEER pipeline on these benchmarks to demonstrate that the method also yields net gains for general generation tasks.

14 comments

Thank you for highlighting SEER’s use of internal evaluation signals to optimize reprompting rather than directly updating the generator. We hope our response will resolve your concerns.

W1 & Q1 — Test size and human evaluation.

The test set contains 40 instructions, each paired with five distinct base prompts (200 inputs in total; 100 ID and 100 OOD). For each input, each method generates one image with the same seed across methods. Thus, each comparison in Tables 2–3 contains 200 image pairs, with no exclusions. All six annotators rated every pair, with wins/ties/losses scored as 1/0.5/0 and averaged over inputs and annotators. The table shows the tie rates pooled across all 10 baseline comparisons in Tables 2–3. The overall tie rate varies from 10% to 36% across individual comparisons. Inter-annotator agreement is Fleiss' κ = 0.72 (Appendix A.2).

SubsetID-SimpleID-HardOOD-SimpleOOD-HardOverall
Cases per comparison40604060200
Tie rate, pooled across baseline comparisons25%10%20%13%~16%

W2 — Reasoning ablations and data scale.

Reasoning format. Blind human comparisons on the test set favor CoT over answer-only (0.59), CoT over fixed-template CoT (0.66), and answer-only over fixed-template CoT (0.56). These results support the benefit of input-dependent reasoning.

SFT and external evaluator. We additionally trained a reprompter with SFT on Gemini 3-generated reprompts, using the same 300 examples. Despite learning from the strongest external reprompter in Table 2, SEER achieves a 0.70 human win ratio against this SFT baseline. This suggests that, with limited supervision, imitating even strong reprompts generalizes less effectively than optimizing through image-based feedback. Regarding the evaluator choice, our cross-model experiment in WjZ9 (W1 & Q1) shows that reprompters trained with their respective UMMs' feedback perform better with their matched generators, supporting model-specific alignment.

Training-set size. Table 7 reports evaluator accuracies of 0.60 / 0.67 / 0.92 with 50 / 150 / 300 training examples. Human win ratio on the test set also favors the 300-example reprompting policy over the 50-example policy (0.77) and 150-example policy (0.73).

W3 & Q2 — Full reprompting results on WISE.

GenEval and DPG-Bench mainly evaluate explicitly specified visual attributes and relations, while SEER focuses on translating implicit understanding into effective generation guidance. We therefore evaluate the full pipeline on WISE, covering 1,000 knowledge-intensive prompts. Compared with direct generation using the same trained model, enabling reprompting improves weighted WiScore 0.5589 → 0.6453 (+0.0864), CLIP 23.219 → 24.084 (+0.865), and ImageReward −0.398 → −0.158 (+0.240). The six-domain results are given in WjZ9 (Q2).

W4 — UniCorn and generator-side optimization.

UniCorn uses a Proposer–Solver–Judge process to construct self-generated training supervision. SEER instead uses the internal evaluator's image preferences as RLMT rewards for reprompting. Because the methods differ, Direct-Gen tests the optimization target rather than reproducing UniCorn. On the same Harmon backbone, SEER optimizes understanding/reasoning parameters with the visual generator frozen, while Direct-Gen updates the MAR generator with the LLM and VAE frozen. Both use matched image-generation and evaluator-call budgets.

Validation checkpointSEER CLIPSEER ImageRewardDirect-Gen CLIPDirect-Gen ImageReward
128.921.2127.331.05
229.021.2128.481.13
329.261.2228.981.20
429.171.2429.121.27

SEER reaches stronger performance earlier, while Direct-Gen improves more gradually and eventually surpasses SEER in ImageReward (1.27 vs. 1.24). Unlike SEER, Direct-Gen can also improve the generator itself, potentially achieving a higher performance ceiling. We freeze the visual generator to isolate how the UMM's understanding is translated into effective generation guidance, avoiding confounding effects from direct generator optimization.

W5 — What the efficiency numbers measure.

Our main efficiency claim is training-data efficiency with 300 examples. Table 2 reports the average word count of the final reprompt: 22.94 for SEER, 46.23 for GPT-5.2, and 33.06 for Gemini 3. The internal reasoning-token counts of GPT-5.2 and Gemini 3 were not accessible in our evaluation, so we report comparable final-reprompt lengths rather than total reasoning-token usage.

Thanks for the authors' response. I still have some concerns.

(1) Test size is still too small. Authors should provide large-size evaluation (e.g., larger than 5,000) to demonstrate the effectiveness of the proposed method.

(2) The main table only compares zero-shot, phase 1 Only, and SEER. It lacks the comparison of other UMM methods. It is very important to demonstrate the effectiveness of proposed methods compared to previous works.

(3) This paper only conduct comparison on the private Visual Instruction Test set and lacks comparison with other methods on large-scale benchmarks or other well-known benchmarks. I concern whether experiments in these paper can correctly illustrate the effectiveness of the proposed method.

Thank you for your timely reply. We hope these additional analyses will address your remaining concerns.

(1) Evaluation scale.

For UMM text-to-image generation, evaluation on benchmarks with around 1,000 prompts is common practice and provides a meaningful assessment of generation performance. Representative benchmarks include:

Benchmark# Prompts
GenEval553
WISE (ICML 2026)1,000
DSG-1K1,000
DPG-Bench1,065
GenAI-Bench1,600
PartiPrompts1,600
T2I-CompBench2,400

Our evaluation includes GenEval, DPG-Bench, and WISE, rather than relying only on our own Visual Instruction Test set. GenEval and DPG-Bench assess direct-generation capability preservation, while WISE evaluates the full reprompting pipeline. We agree that evaluation at a larger scale would provide further evidence, but standardized benchmarks also offer complementary task coverage.

(2) Comparisons with existing UMMs.

Table 1 reports the Stage-1 evaluator ablation, not the main generation comparison. Our generation comparisons appear in Tables 2–3. Specifically, Table 3 compares SEER against Harmon (Base), BAGEL, BAGEL-Think, BLIP3-o, and Show-o2, with human win rates of 0.85 / 0.75 / 0.69 / 0.81 / 0.81, respectively. Table 2 also compares five external reprompting methods using the same image generator.

(3) Public benchmark comparisons on WISE.

WISE (ICML 2026) is a widely adopted public benchmark with 1,000 prompts across 25 subdomains. It has received over 200 indexed citations and has been widely used in image-generation research. Our earlier WISE results used a Qwen2.5-VL-7B evaluator for computational efficiency. For a more consistent comparison with existing UMMs, we additionally use TorchUMM's WISE evaluation protocol, with Qwen2.5-VL-72B-Instruct as the judge. The comparison results are shown below:

ModelWISE WiScore ↑
Show-o2 (1.5B)0.3349
Show-o2 (7B)0.3595
BAGEL (14B)0.3989
SEER P21 (direct generation)0.3935
SEER P21 (+ reprompting)0.5107

Under the 72B evaluation protocol, reprompting improves WiScore from 0.3935 to 0.5107 (+0.1172), above the listed UMM baselines. The improvement is consistent with our earlier 7B evaluation (0.5589 → 0.6453, +0.0864).

Reviewer_2XbPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SEER is a two-stage reinforcement learning method. It first trains an image evaluator on labeled image pairs, then freezes the evaluator and uses its judgments of generated images to train a prompt rewriter. The aim is to turn abstract instructions into concrete visual descriptions. Training updates the model's understanding and reasoning parameters, which also help interpret text for image generation. Parameters specific to visual generation remain frozen. The authors report improvements on their own task.

Critical strengths

S1. Human evaluation supports better instruction following.

Blind human comparisons favor SEER on the authors' test set. The comparisons include the original model with chain-of-thought prompting, supporting a benefit from the additional training.

S2. Improvements are reported on two models.

The authors report gains on both Harmon and Show-o2 using 300 filtered examples as additional training data. This provides initial evidence that the approach can be applied to more than one model.

Critical weaknesses

W1. The advantage over directly fine-tuning the base model is unclear.

The paper does not compare SEER with direct fine-tuning of the same base model on the same filtered images and their associated instructions. Prior work such as PromptEnhancer also trains prompt rewriters using feedback on generated images. Table 2 uses the same image generator for all rewriters, but equivalent training on this task is not reported for the external rewriters. It is therefore unclear whether the gains come from SEER's design or from additional training on the target task.

W2. The public benchmark experiments do not test the full method.

The main evidence comes from the authors' own test set. The GenEval and DPG-Bench experiments disable prompt rewriting and mainly check whether the model retains its existing generation abilities. These results do not show how much the core rewriting step helps on an independent public benchmark.

W3. The data split is difficult to verify.

Of the 40 test instructions, 20 were seen during training and are paired with new subjects; the other 20 are described as unseen concepts. The first group tests whether known instructions can be applied to new subjects, rather than whether the model can handle new instructions. The current anonymous repository does not include the actual training and test data, so sample overlap and the split between seen and unseen concepts cannot be checked directly. The available evidence does not establish data leakage.

Questions for authors

Q1. How does SEER compare with direct fine-tuning? Please compare against the same base model fine-tuned on the same filtered images and instructions. Does SEER still outperform an external prompt rewriter trained on the same inputs with the same evaluator feedback? Please use similar computational budgets for these comparisons.

Q2. Does prompt rewriting help on an independent public benchmark? Please report results for the full SEER method, SEER with rewriting disabled, and the main baselines on a public benchmark that tests how well generated images follow instructions.

Q3. How were the training and test sets separated? Please provide the training and test examples, explain where the images and labels come from, and describe how you checked for duplicates and defined unseen concepts. Was the test set used to select model checkpoints or adjust training settings?

3 comments

Thank you for recognizing the improved instruction following in our blind human evaluations and the gains on both Harmon and Show-o2. We hope our response will resolve your concerns.

W1 & Q1 — Same-data training and reprompter adaptation.

Direct generator SFT. We fine-tuned the original model using the same 300 filtered image–instruction examples. This control did not improve the automatic metrics: CLIP and ImageReward changed by −0.565 and −0.063, respectively. In contrast, SEER aims to improve generation by translating the UMM's understanding into generation guidance. We therefore freeze the visual generator to study this understanding-to-generation pathway without confounding effects from direct generator optimization.

Cross-model reprompter exchange. Following WjZ9's suggestion (full results in WjZ9, W1 & Q1), we swapped the SEER reprompts of Harmon-1.5B and Show-o2-1.5B across fixed generators. Both models start from Qwen2.5-1.5B-Instruct but are aligned with their respective generators. Their reprompters were trained on the same 300 examples, each using feedback from its own UMM. Matched reprompters score higher on both metrics: Harmon 28.63 / 1.13 vs. 27.84 / 0.94; Show-o2 28.48 / 0.89 vs. 27.55 / 0.62 (CLIP / ImageReward). This supports model-specific alignment. We also compare against a reprompter trained with SFT on Gemini 3-generated prompts using the same 300 examples (see 8M7P, W2).

W2 & Q2 — Reprompting on a public benchmark.

We evaluated SEER on WISE, a public benchmark with 1,000 prompts across 25 subdomains. Using the same trained checkpoint and generation settings, we compared direct generation with generation after SEER reprompting. Weighted WiScore rises from 0.5589 to 0.6453 (+0.0864), CLIP from 23.219 to 24.084 (+0.865), and ImageReward from −0.398 to −0.158 (+0.240). This measures the contribution of reprompting on an independent public benchmark. The domain breakdown is in WjZ9 (Q2).

W3 & Q3 — Data construction, overlap, and checkpoint selection.

We generated 600 candidate examples, applied GPT-5 semantic filtering, and human-checked the remaining candidates; 300 image–instruction examples were retained (Appendix A.1.2). In the instruction-supervision stage of RLVR, a verified positive image is paired with an image generated from the base prompt without the requested modification. We randomize image order and set the Yes/No label accordingly (Appendix A.1.4). RLMT uses the 300 instruction–base-prompt inputs, with no reference reprompts. The test set was built separately and human-labeled; examples appear in Table 6 and Figs. 6–8.

A GPT-6-assisted duplicate check found no repeated base prompts between training and test. ID instructions are paired with new subjects, while OOD instructions cover unseen concepts. Thus, no complete (instruction, base-prompt) pair overlaps, and no images are shared between training and test. Both training stages follow fixed 10-epoch schedules, and we use the final checkpoints rather than selecting or tuning them on the test set.

Dear Reviewer,

Thank you again for your thoughtful and constructive feedback on our submission. We sincerely appreciate your time and effort throughout the review process.

As the November 18 discussion deadline approaches, we would like to kindly follow up on our previous responses and additional experiments. We hope they have adequately addressed your concerns.

If you have had a chance to review our responses, we would greatly appreciate knowing whether any concerns remain unresolved. We would be more than happy to provide further clarification or additional information before the discussion period ends.

Thank you once again for your time and consideration.

Best regards,

The Authors of Paper 3068

Dear Reviewer,

Thank you again for your thoughtful and constructive feedback on our submission. We sincerely appreciate your time and effort throughout the review process.

As the November 18 discussion deadline approaches, we would like to kindly follow up on our previous responses and additional experiments. We hope they have adequately addressed your concerns.

If you have had a chance to review our responses, we would greatly appreciate knowing whether any concerns remain unresolved. We would be more than happy to provide further clarification or additional information before the discussion period ends.

Thank you once again for your time and consideration.

Best regards,

The Authors of Paper 3068

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.