acceptodds
Under review as a conference paper at ICLR 2027

ProactGen: Learning Proactive Clarification for Video Generation

Abstract

Recent advances in video generation let users create high-quality videos from natural language descriptions, yet obtaining results that match user intent remains time-consuming: individual generations are computationally expensive, and underspecified prompts force repeated cycles of generation, inspection, and revision. Existing video prompt optimization methods improve generation through automatic rewriting, but cannot supply the ideas and preferences that were not in the initial description. We propose , a video generation framework that learns proactive clarification, eliciting user intent through questions grounded in generated video drafts. Specifically, we train a vision-language model as a questioning policy with reinforcement learning: observing the initial prompt and a video draft, it learns to ask the questions most valuable to downstream generation, rewarded by the alignment between the final generated video and a target reference. We further find that effective questioning does not require drafts of final-output quality, and exploit this property by generating drafts with fewer denoising steps, keeping the added cost of clarification low. Experiments show that improves target alignment over direct generation by at least the gain of the strongest baselines. Moreover, a single round of already exceeds the alignment that the strongest baseline reaches after four refinement rounds, in only 29% of its time.

Primary area:
generative models
Keywords:
Text-to-Video Generation, Proactive Clarification, Reinforcement Learning, User Intent Alignment, Prompt Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_M1jpOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ProactGen, a framework for proactive clarification in text-to-video generation. It uses low-cost video drafts to guide a vision-language model in asking targeted questions about missing user preferences. The questioning policy is trained with GRPO using final video-to-reference alignment as the reward. Experiments on InstanceCap-500 and MiraBench-500 show improved alignment over prompt optimization and proactive clarification baselines, while low-cost drafts reduce the computational overhead.

Critical strengths
  • The paper addresses a meaningful limitation of current video generation systems: users frequently omit important visual and temporal preferences from their initial prompts. The distinction between automatically rewriting existing information and actively acquiring missing user intent is well motivated.
  • The authors evaluate the method on two datasets and compare against several representative baselines, including prompt optimization and proactive clarification approaches. The evaluation includes multiple metrics, ablations on reinforcement learning and draft conditioning, and paired confidence intervals. These experiments provide evidence that the proposed framework improves reference alignment over the evaluated baselines.
Critical weaknesses
  • The contribution of draft-grounded clarification is relatively small and not statistically conclusive. One of the paper's central claims is that observing a generated draft enables the questioning policy to select more effective questions. However, Table 2 shows that the RL-trained policy with draft observation achieves a DreamSim-DTW score of 0.678, compared with 0.675 for the independently trained text-only policy. Moreover, A.6 reports a paired difference of only +0.003. While reinforcement learning clearly improves the questioning policy, the independent benefit of draft observation is not convincingly established. This raises concerns about whether introducing the draft generation stage is truly necessary, especially given its additional computational overhead. Since draft grounding is a key distinguishing contribution of the paper, stronger evidence is needed to justify its effectiveness and cost-efficiency.
  • The questioning policy does not adequately consider user interaction costs. The paper aims to reduce the burden of iterative video generation by proactively clarifying user intent. However, excessive questioning, especially for simple or sufficiently specified prompts, may negatively affect user experience. This raises concerns about its ability to adapt the questioning granularity to task complexity. Since the reward primarily optimizes final video alignment without explicitly penalizing the number of valid questions, the policy may favor unnecessary clarification. The authors should evaluate the trade-off between question budget and alignment, particularly whether the policy can avoid unnecessary questions for simple tasks.
  • Draft-based clarification introduces additional inference overhead compared to the text-only policy (11.13s -> 15.77s in Tab. 4), despite achieving only marginal performance improvements. This raises concerns about its computational efficiency, particularly given the extra costs of draft generation and VLM processing.
  • Could the authors provide more examples demonstrating how ProactGen corrects human actions or camera movements? Such examples would help validate its effectiveness on dynamic aspects of video generation.
Questions for authors

Please see the weaknesses for details. My main concerns are the marginal benefit of draft grounding over text-only RL and the lack of consideration for user interaction costs, which raise questions about the effectiveness and practicality of the proposed framework.

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/4)

We thank the reviewer for the careful and valuable comments. We respond to each concern below:


W1: The contribution of draft-grounded clarification is relatively small and not statistically conclusive. One of the paper's central claims is that observing a generated draft enables the questioning policy to select more effective questions. However, Table 2 shows that the RL-trained policy with draft observation achieves a DreamSim-DTW score of 0.678, compared with 0.675 for the independently trained text-only policy. Moreover, A.6 reports a paired difference of only +0.003. While reinforcement learning clearly improves the questioning policy, the independent benefit of draft observation is not convincingly established. This raises concerns about whether introducing the draft generation stage is truly necessary, especially given its additional computational overhead. Since draft grounding is a key distinguishing contribution of the paper, stronger evidence is needed to justify its effectiveness and cost-efficiency.

A1: We thank the reviewer for raising this central concern. We agree that the single-round DreamSim-DTW comparison on InstanceCap (+0.003 [−0.001, +0.007] in Appendix A.6, Table 3 in the initial submission and Table 4 in the revision) alone cannot establish the independent benefit of observing a draft. We conducted multi-round interaction experiments and ETVA fine-grained intent recovery evaluation to address it, added to Section 4.4 (Figure 3), Appendix A.9.2 (Table 8), and Appendix A.9.4 (Table 11; cost-efficiency is evaluated in A3):

(1.1) Multi-round comparison between draft-based and text-only policies: In the first round, generic questions inferred from the short prompt already yield similar gains (though the 2×2 RL-by-draft interaction is positive at +0.007 [+0.001, +0.012] in Appendix A.6). To test whether observing the draft is needed as interaction continues, we ran both policies for multiple rounds on the same 500 InstanceCap samples:

RoundText-only (RL, text)ProactGenDiff. [95% paired CI]
10.6750.678+0.003 [−0.001, +0.007]
20.6810.683+0.002 [−0.002, +0.007]
30.6830.690+0.007 [+0.002, +0.011]
40.6830.692+0.009 [+0.005, +0.014]
Text-only R5–R120.683–0.686––

As shown, the two policies perform similarly in the first two rounds. After Round 2, the text-only policy plateaus and stays between 0.683 and 0.686 from Round 3 to Round 12. In contrast, ProactGen improves in every round, leads from Round 3 with the paired interval excluding zero, and reaches 0.692 in Round 4, surpassing all 12 rounds of the text-only policy. Based on these results, we hypothesize that once common ambiguities are resolved early on, the text-only policy lacks visibility into the video to locate remaining mismatches, whereas updated drafts offer visual cues for later questions. That said, while our experiment establishes the multi-round empirical gain, it cannot conclusively prove this underlying mechanism, which requires finer-grained question-to-error attribution analysis in future work. Thus, while the draft adds little in the first round, its benefit becomes clear across multiple rounds.

(1.2) ETVA fine-grained intent recovery evaluation: To test whether observing the draft already helps capture specific missing details in a single round under a finer-grained metric, we adopt ETVA [1] to decompose full target descriptions into 24,745 atomic verification questions missing from the initial prompt on the same 500 MiraBench samples as in Table 1 (which officially provide structured dense captions):

Method (MiraBench-500)Overall ↑EntityAttributeRelation / ActionCamera
Direct34.5%46.4%32.8%28.6%31.5%
Text-only (RL, text)49.0%62.1%47.7%40.8%39.5%
ProactGen (1 round)50.6%65.2%48.9%42.1%41.7%

As shown, single-round ProactGen recovers 50.6% of missing intent overall, achieving a statistically significant gain over the trained text-only policy (+1.6%, 95% paired CI [+0.3%, +2.9%] excluding zero), and consistently leads the text-only policy across every component (+3.1 pp on entities, +2.2 pp on camera, +1.3 pp on relations/actions, and +1.1 pp on attributes), showing that inspecting the draft helps target missing visual elements that are hard to guess from the text prompt alone. We evaluate the computational and human-time cost-efficiency of introducing the draft in A3 below.

Reference:

[1] Guan, K., et al. (2025). ETVA: Evaluation of Text-to-Video Alignment via Fine-grained Question Generation and Answering. ICCV 2025.

AuthorsOpenReview ↗

Rebuttal by Authors (2/4)

W2: The questioning policy does not adequately consider user interaction costs. The paper aims to reduce the burden of iterative video generation by proactively clarifying user intent. However, excessive questioning, especially for simple or sufficiently specified prompts, may negatively affect user experience. This raises concerns about its ability to adapt the questioning granularity to task complexity. Since the reward primarily optimizes final video alignment without explicitly penalizing the number of valid questions, the policy may favor unnecessary clarification. The authors should evaluate the trade-off between question budget and alignment, particularly whether the policy can avoid unnecessary questions for simple tasks.

A2: Thank you for this valuable suggestion. Following your recommendation, we evaluated the question-budget trade-off via prefix truncation (added in Appendix A.9.2, Table 9), examined real-user interaction burden in a human-in-the-loop study (Section 4.5, Table 3), and added a dedicated Limitations section (Section 5) in the main text to discuss adaptive question budgeting for future work:

(2.1) Question-budget truncation curve: To examine how alignment varies with the number of questions, we kept the first k∈{0,1,…,5}k \in \lbrace 0, 1, \dots, 5 \rbrace questions generated by ProactGen in their original order across all 500 InstanceCap samples and ran consolidation and final generation for each kk (added in Appendix A.9.2, Table 9):

Questions kept (kk)Mean valid answer wordsDreamSim-DTW ↑3D-VAE ↓DnS ↑
0 (Direct)00.658144.03−0.689
15.00.666143.18−0.681
29.90.666143.21−0.678
314.70.669142.92−0.668
419.70.674142.63−0.667
5 (full budget)24.50.678142.34−0.659

As shown, alignment improves steadily across all three metrics from k=0k=0 to k=5k=5 rather than saturating after the first question. However, because the policy is trained to submit up to 5 questions as a single batch without an ordering objective, the learned policy does not order its questions by decreasing importance (for example, questions 4 and 5 still bring clear gains), so this truncation curve does not show a clear trade-off point for post-hoc early stopping.

(2.2) Practical interaction burden and adaptive question counts: While our setup caps answers at 10 words and penalizes invalid questions, the reward does not penalize the number of valid questions, encouraging the policy to ask all 5 questions per round. In our new human-in-the-loop study with 4 participants across 100 targets (Section 4.5, Table 3), where participants were requested to answer in natural language, flag any unanswerable or invalid questions, and rate per-question difficulty and overall mental effort (1–5 ↓), answering a 5-question round remained lightweight in practice: 98.6% of questions were answerable with low per-question difficulty (1.74/5) and substantially lower mental effort than manually revising the prompt from the draft (2.10/5 vs. 4.00/5), while improving final video alignment by +0.021 over Direct. Even so, on prompts that are already simple or sufficiently specified, asking all 5 questions unconditionally can still introduce unnecessary interaction; adapting the number of questions to prompt complexity is an insightful direction that lies beyond the scope of the current paper. Teaching a policy to adjust its question count adaptively requires introducing an explicit question-count penalty into the RL objective and carefully balancing it with the video alignment reward so that the policy avoids unnecessary questions on simple prompts without under-asking on underspecified ones. We have added a Limitations section (Section 5) in the revised main text to highlight this direction and motivate future work.

AuthorsOpenReview ↗

Rebuttal by Authors (3/4)

W3: Draft-based clarification introduces additional inference overhead compared to the text-only policy (11.13s -> 15.77s in Tab. 4), despite achieving only marginal performance improvements. This raises concerns about its computational efficiency, particularly given the extra costs of draft generation and VLM processing.

A3: Thank you for raising this practical concern. We agree that the text-only policy is faster in a single server-side round (11.13 s vs. 15.77 s in Table 4 of the initial submission, now Table 6 in Appendix A.8); to evaluate efficiency across budgets, we have updated the time–alignment plot in Figure 3 (Section 4.4) to include 1–12 text-only rounds and incorporated measured human answering time in Appendix A.8 (Table 7):

(3.1) Stage breakdown and matched-compute comparison: The +4.64 s single-round server difference comes from the 8-step draft (4.25 s) and VLM visual processing (~0.4 s). To compare both policies visually across compute budgets, we added the 1–12 round text-only curve to Figure 3, with key points summarized below:

MethodR1R2R3R4R11 (Text best)R12
Text-only (System time / DS-DTW)11.1 s / 0.67515.3 s / 0.68120.0 s / 0.68325.0 s / 0.68362.0 s / 0.68667.4 s / 0.685
ProactGen (System time / DS-DTW)15.8 s / 0.67824.7 s / 0.68333.6 s / 0.69042.5 s / 0.692––

As shown in Figure 3 and the table above, under server budgets below 20 s, 1–2 text-only rounds are faster, and the two curves cross around 25 s (0.683). Beyond 25 s, Text-only plateaus at 0.686 even after 62.0–67.4 s (11–12 rounds), whereas ProactGen reaches 0.690 in 33.6 s (R3) and 0.692 in 42.5 s (R4), achieving higher alignment in less total compute time.

(3.2) Combined latency with measured human answering time: To evaluate total wall-clock cost with a real user in the loop, we measured an average human answering time of 108.8 s per 5-question round in our human study (Section 4.5, Table 3) and replaced the simulated-user time in Appendix A.8 (Table 6), as reported in Table 7:

Method & Rounds (Wan2.2 framework)Server timeHuman timeCombined totalDS-DTW ↑
Text-only (1 round)8.6 s108.8 s117.4 s0.675
ProactGen (1 round)13.2 s108.8 s122.0 s0.678
Text-only (2 rounds)10.9 s217.5 s228.4 s0.681
ProactGen (2 rounds)19.5 s217.5 s237.0 s0.683
Text-only (3 rounds)13.6 s326.3 s339.9 s0.683
ProactGen (3 rounds)25.9 s326.3 s352.2 s0.690
Text-only (4 rounds)16.7 s435.0 s451.7 s0.683
ProactGen (4 rounds)32.2 s435.0 s467.2 s0.692

As shown, once user answering time is included, the share of time spent generating the draft becomes very small (+4.6 s is only 3.8% of single-round time). Because ProactGen reaches higher alignment in fewer interaction rounds, the human time saved by avoiding even one extra round far outweighs the cost of draft generation (for instance, 3 rounds of ProactGen save ~100 s of total time over 4 rounds of Text-only while achieving higher alignment).

AuthorsOpenReview ↗

Rebuttal by Authors (4/4)

W4: Could the authors provide more examples demonstrating how ProactGen corrects human actions or camera movements? Such examples would help validate its effectiveness on dynamic aspects of video generation.

A4: Thank you for this helpful suggestion. We have added multi-frame qualitative examples illustrating how ProactGen corrects human actions and camera movements in Appendix A.9.6 (Figure 5), provided the corresponding full-motion videos at an anonymous page (https://anonymous.4open.science/r/proactgen-case-study-page-1012/index.html), and reported dataset-wide quantitative results on action and camera recovery in Appendix A.9.4 (Table 11):

(4.1) Qualitative examples on human actions and camera movements: We include four examples from our MiniMax-H3 [1] evaluations (Sections 4.5–4.6) as multi-frame sequences in Appendix A.9.6 (Figure 5), comprising two human-action examples (Samples 377 and 148) and two camera-movement examples (Samples 078 and 059). To allow direct inspection of temporal motion in video form, we also provide side-by-side full videos (target reference, low-step draft, Direct baseline, and ProactGen final video) together with the complete initial prompts, question-answer pairs, and refined prompts at https://anonymous.4open.science/r/proactgen-case-study-page-1012/index.html.

(4.2) Quantitative validation on human actions and camera movements: To verify systematic gains on these dynamic aspects beyond individual examples, we evaluate on the same 500 MiraBench samples as in Table 1 using the independent ETVA [2] atomic verification protocol, measuring recovery rates specifically on missing Relation / Action items (n=497n = 497 samples) and Camera items (n=391n = 391 samples):

Method (MiraBench-500)ETVA Relation / Action ↑ETVA Camera ↑
Direct28.6%31.5%
Text-only (RL, text)40.8%39.5%
ProactGen (1 round)42.1%41.7%

As shown, across the full benchmark, ProactGen raises missing relation/action recovery from 28.6% to 42.1% (+13.5 pp over Direct) and camera recovery from 31.5% to 41.7% (+10.2 pp over Direct), outperforming both Direct and the trained text-only policy on both aspects.

References:

[1] MiniMax. (2026). MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities.

[2] Guan, K., et al. (2025). ETVA: Evaluation of Text-to-Video Alignment via Fine-grained Question Generation and Answering. ICCV 2025.


Thank you again for your constructive comments; please let us know if you have any further questions.

Reviewer_4KSjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ProactGen, a framework for proactive clarification in text-to-video generation. Given an initial prompt, the system generates a low-cost video draft, asks clarification questions conditioned on the prompt and sampled draft frames, and incorporates the user’s answers into a refined prompt for final generation. The questioning policy is trained with GRPO using similarity between the final generated video and a hidden reference video as the reward. Experiments with simulated users show improved target alignment over direct generation, automatic prompt optimization, and an adapted proactive clarification baseline. The paper also investigates reinforcement learning, draft observation, and reduced-budget drafts through ablations.

Critical strengths
  1. The paper addresses a practical limitation of prompt rewriting: rewriting cannot reliably recover preferences absent from the initial request. Combining clarification with observations of the generator’s actual output is a well-motivated approach.

  2. Optimizing questions through their effect on the final video is a compelling design. Keeping the generator, simulated user, and prompt refinement model frozen helps attribute improvements to the questioning policy.

  3. Using reduced-budget drafts is a useful engineering choice. The comparison with full-budget drafts suggests that clarification can tolerate lower visual quality, and the reported time–alignment tradeoff is favorable relative to the evaluated baselines.

Critical weaknesses
  1. The benefit for real users remains unverified. All interactions use a simulated user with access to reference frames and a detailed target caption. This setting is reproducible, but differs substantially from users who may have incomplete, evolving, or difficult-to-articulate preferences. Reported latency includes model-generated answers rather than human response time, so it does not establish an end-to-end user efficiency advantage.

  2. A central interaction baseline is missing. The paper does not compare against showing the draft to a user and requesting free-form corrections. This comparison would help establish whether proactive question selection provides value beyond simply exposing the draft and allowing feedback. A simulated user with a controlled response budget may already be informative.

  3. Evidence for the independent contribution of draft grounding is limited. The trained draft-conditioned policy achieves 0.678 DreamSim-DTW versus 0.675 for the trained text-only policy, and the paired confidence interval includes zero. The training-by-draft interaction supports the usefulness of learning to exploit drafts, but the stronger claim that draft observation itself reliably improves performance is less firmly established. The text-only base policy is also a strong baseline at 0.672. This might relates to the perception of the draft by the questioning VLM.

  4. Evaluation is closely related to the training objective. The primary DreamSim-DTW metric uses the same DreamSim representation as the training reward. Independent metrics provide useful supporting evidence, but reference similarity does not fully capture perceived quality, temporal coherence, or user satisfaction. Human evaluation and more targeted video intent similarity evaluation would strengthen the assessment.

  5. Generalization are insufficiently characterized. The questioning policy is optimized against a single frozen generator and may learn generator-specific defaults, failure modes, and responsiveness to prompt refinements. Its effectiveness under a substantially different generator therefore remains unverified. Moreover, the reported latency is configuration-dependent: slower generators would increase absolute draft and final-generation costs, while the relative efficiency advantage would depend on whether inexpensive drafts remain informative. Evaluation also uses a single video generator; transfer to other generators, longer videos, or different generation settings remains unclear.

Questions for authors

Please see weaknesses

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/5)

We thank the reviewer for the thorough and insightful review. We address each point below:


W1: The benefit for real users remains unverified. All interactions use a simulated user with access to reference frames and a detailed target caption. This setting is reproducible, but differs substantially from users who may have incomplete, evolving, or difficult-to-articulate preferences. Reported latency includes model-generated answers rather than human response time, so it does not establish an end-to-end user efficiency advantage.

A1: Thank you for raising this point. We agree that real users often have incomplete, evolving, or inherently hard-to-verbalize preferences, which we now explicitly discuss in a new Limitations section (Section 5) in the revised main text. Because such open-ended states lack a shared ground truth for controlled cross-method comparison (and non-verbalizable nuances exceed text interaction itself), they are difficult to benchmark directly. Within a controllable setup, we tried our best to move closer to real usage by using a reference video as the hidden intent and replacing the simulated user with real participants (Section 4.5 Table 3, Appendix A.8 Table 7, and Appendix A.9.5). While fixed-target, this study still narrows the gap from simulation by testing real-user answerability, subjective effort, and end-to-end latency with human response times:

(1.1) Human-in-the-loop study: To verify whether real users can easily answer ProactGen's questions and improve generation under a shared target video, we replaced the simulated user in our pipeline with 4 human participants across 100 randomly sampled InstanceCap targets in a counterbalanced design (2 participants per target under ProactGen, totaling 200 sessions and 1,000 questions). We also transferred the policy zero-shot to MiniMax-H3 [1] to evaluate on a higher-fidelity video generator (comparison with draft-based free-form corrections is reported in A2). Each participant was requested to answer in natural language, flag any unanswerable or invalid questions, and rate per-question difficulty and overall mental effort (1–5 ↓):

Setting (100 targets, 200 sessions)AnswerableDifficulty (1–5 ↓)Effort (1–5 ↓)Response timeDirectProactGenGain vs. Direct
Human-in-the-loop (MiniMax-H3)98.6% (986/1000)1.742.10108.8 s0.6830.704+0.021

As shown, participants answered 98.6% (986/1,000) of the questions with low difficulty (1.74/5) and low mental effort (2.10/5) in 108.8 s per round, improving final alignment by +0.021 over Direct, confirming that the learned questions are easy to answer and effective for real users.

(1.2) End-to-end latency with measured human response time: To evaluate end-to-end user efficiency within our main efficiency framework (Section 4.4), we replace the simulated-user time in Appendix A.8 (Table 4 in the initial submission, now Table 6) with the 108.8 s human response time per 5-question round measured in (1.1) above. For 5-turn Proactive T2I, which likewise asks 5 questions in total, we substitute the same 108.8 s human answering time for comparison:

Method & Rounds (InstanceCap-500, Wan2.2)Server timeHuman timeCombined totalDS-DTW ↑
Direct7.1 s0 s7.1 s0.658
RAPO++-SSPO (1 / 4 rounds, non-interactive)21.2 / 55.1 s0 s21.2 / 55.1 s0.665 / 0.669
Proactive T2I (Ag3+GT, 5 turns = 5 questions)91.9 s~108.8 s~200.7 s0.662
Text-only (1 round)8.6 s108.8 s117.4 s0.675
ProactGen (1 round)13.2 s108.8 s122.0 s0.678
Text-only (2 rounds)10.9 s217.5 s228.4 s0.681
ProactGen (2 rounds)19.5 s217.5 s237.0 s0.683
Text-only (3 rounds)13.6 s326.3 s339.9 s0.683
ProactGen (3 rounds)25.9 s326.3 s352.2 s0.690
Text-only (4 rounds)16.7 s435.0 s451.7 s0.683
ProactGen (4 rounds)32.2 s435.0 s467.2 s0.692

As shown, while interactive clarification takes longer than non-interactive SSPO, a single ProactGen round exceeds the ceiling of user-free rewriting (0.678 vs. 0.669). Among interactive methods, once user answering time is included, the share of time spent generating the draft becomes very small (+4.6 s is only 3.8% of single-round time). Because ProactGen reaches higher alignment in fewer interaction rounds, the human time saved by avoiding even one extra round far outweighs the cost of draft generation (for instance, 3 rounds of ProactGen save ~100 s of end-to-end user time over 4 rounds of Text-only and exceed all 12 text-only rounds; see A3).

Reference:

[1] MiniMax. (2026). MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities.

AuthorsOpenReview ↗

Rebuttal by Authors (2/5)

W2: A central interaction baseline is missing. The paper does not compare against showing the draft to a user and requesting free-form corrections. This comparison would help establish whether proactive question selection provides value beyond simply exposing the draft and allowing feedback. A simulated user with a controlled response budget may already be informative.

A2: Thank you for suggesting this important baseline. Following your recommendation, we evaluated a Draft + Free-form baseline (showing the draft to the user and requesting free-form corrections) under both a controlled-budget simulated-user setting and a human-in-the-loop setting. Both comparisons have been added in Appendix A.9.3 (Table 10) and Section 4.5 (Table 3):

(2.1) Simulated-user setting with controlled response budgets: To test whether learned question selection adds value beyond exposing the draft for free-form feedback, we provided the simulated user with the same reference input (p0p_0 and 8 reference video frames) along with the same 8-step Wan2.2 draft across all 500 InstanceCap samples. We requested single-round free-form corrections under a budget of 10/25/50 words (matching ProactGen's maximum budget of 5 questions at up to 10 words each), followed by the same consolidation and generation pipeline:

Method (InstanceCap-500, Wan2.2)Word budgetMean wordsDreamSim-DTW ↑3D-VAE ↓DnS ↑
Direct––0.658144.03−0.689
Draft + Free-form10 words9.10.662143.79−0.684
Draft + Free-form25 words20.00.664143.79−0.681
Draft + Free-form50 words36.50.672143.58−0.668
ProactGen (1 round, 5 questions)50 words (5 questions)24.50.678142.34−0.659

As shown, free-form draft correction improves steadily with larger word budgets (reaching 0.672 under a 50-word budget), confirming that it is a strong baseline. Even so, ProactGen outperforms free-form corrections at every budget across all three alignment metrics while using fewer words on average than the 50-word baseline (24.5 vs. 36.5 words), as targeted questions direct feedback toward the missing details that matter most to generation.

(2.2) Human-in-the-loop comparison against free-form revision: To further compare interaction effort and alignment with real users, the 4 participants in A1 were requested to complete counterbalanced ProactGen and unrestricted free-form draft revision tasks across the same 100 targets on MiniMax-H3 and rate overall mental effort (1–5 ↓) after each session:

Human interaction mode (n = 100)User response time ↓Full session time ↓Mental effort (1–5 ↓)Final DS-DTW ↑
Direct (non-interactive)0 s––0.683
Draft + Free-form revision119.9 s182.1 s4.000.697
ProactGen (answering 5 questions)108.8 s172.4 s2.100.704
Diff. (ProactGen − Free-form)−11.1 s−9.7 s−1.90+0.007

As shown, writing free-form corrections requires users to inspect the draft, decide which aspects to prioritize for revision, and compose edits from scratch, resulting in high mental effort (4.00/5). Answering ProactGen's 5 targeted questions achieves higher alignment (0.704 vs. 0.697) while saving 11.1 s per round and reducing mental effort from 4.00 to 2.10, confirming clear practical value beyond simply exposing the draft.

AuthorsOpenReview ↗

Rebuttal by Authors (3/5)

W3: Evidence for the independent contribution of draft grounding is limited. The trained draft-conditioned policy achieves 0.678 DreamSim-DTW versus 0.675 for the trained text-only policy, and the paired confidence interval includes zero. The training-by-draft interaction supports the usefulness of learning to exploit drafts, but the stronger claim that draft observation itself reliably improves performance is less firmly established. The text-only base policy is also a strong baseline at 0.672. This might relates to the perception of the draft by the questioning VLM.

A3: We thank the reviewer for raising this point. We agree that the single-round DreamSim-DTW comparison (+0.003, with the paired confidence interval crossing zero) does not by itself establish the independent benefit of observing a draft. We conducted multi-round interaction experiments and fine-grained intent recovery evaluation to examine the contribution of the draft (Section 4.4 Figure 3, Appendix A.9.2 Table 8, and Appendix A.9.4 Table 11):

(3.1) Multi-round comparison between draft-based and text-only policies: To test whether observing the draft helps as interaction continues, we ran both trained policies for multiple rounds on the same 500 InstanceCap samples:

RoundText-only (RL, text)ProactGenDiff. [95% paired CI]
10.6750.678+0.003 [−0.001, +0.007]
20.6810.683+0.002 [−0.002, +0.007]
30.6830.690+0.007 [+0.002, +0.011]
40.6830.692+0.009 [+0.005, +0.014]
Text-only R5–R120.683–0.686––

As shown, the two policies perform similarly in the first two rounds. After Round 2, the text-only policy plateaus and stays between 0.683 and 0.686 from Round 3 to Round 12. In contrast, ProactGen improves in every round, leads from Round 3 with the paired interval excluding zero, and reaches 0.692 in Round 4, surpassing all 12 rounds of the text-only policy. Based on these results, we hypothesize that once common ambiguities are resolved early on, the text-only policy lacks visibility into the video to locate remaining mismatches, whereas updated drafts offer visual cues for later questions. That said, while our experiment establishes the multi-round empirical gain, it cannot conclusively prove this underlying mechanism, which requires finer-grained question-to-error attribution analysis in future work. Thus, while the draft adds little in the first round, its benefit becomes clear across multiple rounds.

(3.2) ETVA fine-grained intent recovery evaluation: On the more sensitive ETVA atomic intent recovery metric across the same 500 MiraBench samples as in Table 1 (detailed in A4 below), draft observation also yields a statistically significant advantage in a single round: single-round ProactGen recovers 50.6% of missing user intent versus 49.0% for the trained text-only policy (paired diff. +1.6%, 95% CI [+0.3%, +2.9%] excluding zero), showing that observing the draft already helps capture more specific missing details in Round 1.

Reviewer_F1y8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ProactGen, a video-generation system that asks users clarifying questions before producing the final video. A vision-language model observes an initial prompt and a cheaply generated draft, then asks about missing details. The questioner is trained with reinforcement learning using the final video’s similarity to a hidden reference as its reward. Experiments with simulated users show improvements over direct generation and several prompt-refinement methods. The paper also studies whether drafts generated with fewer denoising steps can reduce the additional processing time.

Critical strengths
  1. The paper addresses a useful problem. Automatic prompt rewriting cannot reliably recover preferences that the user never supplied. Asking targeted questions is a sensible approach, and training the questioner through the resulting video gives it a direct signal about which information helps generation.

  2. The experiments include useful controls. The trained and untrained policies, each evaluated with and without drafts, help separate the components. Shared generation noise and paired confidence intervals also make the comparisons more informative. In particular, the results provide clear evidence that training improves the draft-conditioned policy.

Critical weaknesses
  1. The experiments do not clearly establish the added value of observing a draft. The trained text-only questioner performs almost as well as ProactGen, and the paired confidence intervals do not establish a clear benefit from adding the draft on any of the three metrics (Tables 2–3). The text-only system is also faster because it skips draft generation. This leaves the central question unresolved: does the system need the draft, or does most of the improvement come from asking the user for more information? The text-only baseline should appear in the main efficiency comparison, including multiple rounds at matched time budgets. The external clarification comparison also needs clarification: the original Proactive text-to-image paper below reports its strongest results from Ag3, which does not use the belief graph to generate questions, while this submission describes a belief-graph adaptation without clearly identifying the variant.

    • Hahn et al. Proactive Agents for Multi-Turn Text-to-Image Generation Under Uncertainty.
  2. The evaluation does not show whether the improvement matters to real users. The simulated user has access to both a reference video and its detailed caption and is instructed to provide informative answers. This is a useful controlled setting, but it does not establish that people can answer the questions easily or prefer the resulting videos. That gap matters because the primary evaluation metric uses the same DreamSim model as the training reward, although with different frame sampling and aggregation. The additional metrics provide some independent support, but they do not establish the practical value of the small advantage over trained text-only questioning. Similarly, the timing results measure automated processing with a simulated respondent, so they do not show how much time a person would save.

Questions for authors

see weakness above

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/4)

We thank the reviewer for the thoughtful and constructive feedback. We respond to each point below:


W1-1: The experiments do not clearly establish the added value of observing a draft. The trained text-only questioner performs almost as well as ProactGen, and the paired confidence intervals do not establish a clear benefit from adding the draft on any of the three metrics (Tables 2–3). The text-only system is also faster because it skips draft generation. This leaves the central question unresolved: does the system need the draft, or does most of the improvement come from asking the user for more information? The text-only baseline should appear in the main efficiency comparison, including multiple rounds at matched time budgets.

A1: We thank the reviewer for raising this central question. We agree that the single-round comparison alone cannot establish the value of the draft, and have conducted multi-round alignment and efficiency experiments to address it. We have added these results to Section 4.4 (Figure 3) and Appendix A.9.2 (Table 8).

(1.1) Multi-round comparison between draft-based and text-only policies: To test whether the draft is needed as interaction continues, we ran both policies across multiple rounds on the same 500 samples (DreamSim-DTW, higher is better):

RoundText-onlyProactGenDiff. [95% paired CI]
10.6750.678+0.003 [−0.001, +0.007]
20.6810.683+0.002 [−0.002, +0.007]
30.6830.690+0.007 [+0.002, +0.011]
40.6830.692+0.009 [+0.005, +0.014]
Text-only R5–R120.683–0.686––

As shown, the two policies perform similarly in the first two rounds. After Round 2, the text-only policy plateaus and stays between 0.683 and 0.686 from Round 3 to Round 12. In contrast, ProactGen improves in every round, leads from Round 3 with the paired interval excluding zero, and reaches 0.692 in Round 4, surpassing all 12 rounds of the text-only policy. Based on these results, we hypothesize that once common ambiguities are resolved early on, the text-only policy lacks visibility into the video to locate remaining mismatches, whereas updated drafts offer visual cues for later questions. That said, while our experiment establishes the multi-round empirical gain, it cannot conclusively prove this underlying mechanism, which requires finer-grained question-to-error attribution analysis in future work. Thus, in the first round the two policies gain similarly and the draft contributes little; the value of the draft appears as interaction continues, driving sustained alignment gains across later rounds.

(1.2) Text-only policy in the efficiency comparison: To compare both policies at matched time budgets, we have added the text-only policy (1–12 rounds) to the time–alignment plot in Figure 3, with key rounds summarized below:

MethodR1R2R3R4R11 (Text best)R12
Text-only (System time / DS-DTW)11.1 s / 0.67515.3 s / 0.68120.0 s / 0.68325.0 s / 0.68362.0 s / 0.68667.4 s / 0.685
ProactGen (System time / DS-DTW)15.8 s / 0.67824.7 s / 0.68333.6 s / 0.69042.5 s / 0.692––

As shown, at similar system time budgets, two text-only rounds (15.3 s, 0.681) score slightly above one ProactGen round (15.8 s, 0.678), and the two methods match around 25 s (0.683). Beyond 25 s, three and four ProactGen rounds (0.690 in 33.6 s; 0.692 in 42.5 s) exceed the best text-only result at any round (0.686 in 62.0 s) in less system time. We note that these figures measure server-side system processing time and exclude user answering time; as pointed out in W2, human answering time requires separate measurement, which we report in A4.

AuthorsOpenReview ↗

Rebuttal by Authors (2/4)

W1-2: The external clarification comparison also needs clarification: the original Proactive text-to-image paper below reports its strongest results from Ag3, which does not use the belief graph to generate questions, while this submission describes a belief-graph adaptation without clearly identifying the variant.

  • Hahn et al. Proactive Agents for Multi-Turn Text-to-Image Generation Under Uncertainty.

A2: We thank the reviewer for pointing this out. Our main text and appendix did not accurately identify the adapted variant or the changes we made to it. The Proactive T2I [1] baseline in our main results is adapted from Ag3, the principle-prompted agent that performs best in the original paper's automatic evaluation, rather than the belief-graph-guided Ag1/Ag2. We have corrected this in Section 4.1, Section 4.2, and Appendix A.7 (Table 5), and added results for all three variants.

(2.1) What we changed in Ag3: For a fair text-to-video comparison, we keep the released Ag3 pipeline unchanged except for video wording and backbone models, and additionally provide the simulated user with 8 GT reference video frames:

Original Ag3 [1]Our adaptation
How questions are generatedAn LLM writes the next question from the dialogue history and questioning principles; no belief graph is usedUnchanged, with image wording replaced by video wording
Backbone modelsGemini 1.5 Pro for the agent and simulated user; Imagen 3 for generationQwen3.8-27B and Wan2.2, the same models used by ProactGen
What the simulated user seesThe GT caption and a private belief graph built from itUnchanged, plus 8 GT video frames so that it has the same visual access as ProactGen's simulated user
Interaction formatOne question per turn, prompt updated after each answer, answers not length-limitedUnchanged; we report the prompt after turn 5 and turn 15

(2.2) Results of all variants: To compare across variants, we evaluated Ag1, Ag2, and Ag3 on InstanceCap-500 under the original protocol without GT video frames (with Ag3 + GT video corresponding to the baseline in Table 1):

MethodTurnsDreamSim-DTW ↑3D-VAE ↓DnS ↑Pipeline time / sample
Ag15 / 150.656 / 0.652143.50 / 142.87−0.693 / −0.699156.1 / 408.3 s
Ag25 / 150.653 / 0.652143.32 / 143.14−0.697 / −0.694137.9 / 343.1 s
Ag35 / 150.658 / 0.657143.90 / 143.54−0.688 / −0.69094.8 / 222.5 s
Ag3 + GT video5 / 150.662 / 0.668143.55 / 143.37−0.685 / −0.67899.3 / 242.4 s
ProactGen1 (5 questions)0.678142.34−0.65915.8 s

As shown, adding GT video frames strengthens Ag3, and with a single 5-question round, ProactGen outperforms all Proactive T2I variants across all three metrics at much lower pipeline time.

Reference:

[1] Hahn, M., et al. (2025). Proactive Agents for Multi-Turn Text-to-Image Generation Under Uncertainty. ICML 2025.

AuthorsOpenReview ↗

Rebuttal by Authors (3/4)

W2-1: The evaluation does not show whether the improvement matters to real users. The simulated user has access to both a reference video and its detailed caption and is instructed to provide informative answers. This is a useful controlled setting, but it does not establish that people can answer the questions easily or prefer the resulting videos. That gap matters because the primary evaluation metric uses the same DreamSim model as the training reward, although with different frame sampling and aggregation. The additional metrics provide some independent support, but they do not establish the practical value of the small advantage over trained text-only questioning.

A3: Thank you for raising both the real-user question and the metric concern. To address them, we added a human-in-the-loop study, a blinded human preference evaluation, and an independent ETVA intent recovery evaluation in Section 4.5 (Table 3) and Appendix A.9.4–A.9.5 (Table 11) to verify real-user usability, human visual preference, and fine-grained intent fulfillment:

(3.1) Human-in-the-loop study: To verify whether real users can easily answer ProactGen's questions and benefit from the resulting videos, we replaced the simulated user in our pipeline with 4 human participants across 100 randomly sampled InstanceCap targets in a counterbalanced design (2 participants per target under ProactGen, totaling 200 sessions and 1,000 questions). We also transferred the policy zero-shot to MiniMax-H3 [1] to evaluate on a higher-fidelity video generator. Each participant was requested to answer the questions in natural language, flag any unanswerable or invalid questions, and rate per-question difficulty and overall mental effort (1–5 ↓):

Setting (100 targets, 200 sessions)AnswerableDifficulty (1–5 ↓)Effort (1–5 ↓)DirectProactGenGain vs. Direct
Human-in-the-loop (MiniMax-H3)98.6% (986/1000)1.742.100.6830.704+0.021

As shown, participants answered 98.6% (986/1,000) of the questions with low difficulty (1.74/5) and low mental effort (2.10/5), improving final video alignment by +0.021 over Direct, confirming that the learned questions are easy to answer and effective in practice.

(3.2) Blinded human preference evaluation: To test whether automatic metric gains translate into human-perceptible improvements, we conducted a blinded side-by-side preference evaluation across three comparisons (100 pairs each, balanced across 4 evaluators). Evaluators were requested to choose which anonymized candidate video better matched the target video without method identities:

Blinded human comparison (n = 100 each)WinsLossesTiesWin rate (excl. ties)
ProactGen (1 round) vs. Direct6429768.8%
ProactGen (1 round) vs. Proactive T2I (15 turns)5836661.7%
ProactGen (4 rounds) vs. Text-only (4 rounds)5140956.0%

As shown, single-round ProactGen clearly outperformed Direct (68.8% excluding ties) and 15-turn Proactive T2I (61.7%), and achieved more wins than 4-round Text-only (51 vs. 40, with 9 ties; 56.0%), indicating that the alignment gains correspond to perceptible visual improvements.

(3.3) Independent ETVA intent recovery evaluation: To directly measure how much missing user intent is recovered under an independent evaluation metric (alongside 3D-VAE, DnS, and the five MiraBench ViCLIP scores in the main paper), we adopt ETVA [2] to decompose target descriptions into 24,745 atomic verification questions missing from the initial prompt. Because ETVA relies on human-annotated multi-aspect dense captions to build its question inventory, we run this evaluation on the same 500 MiraBench samples as in Table 1, which officially provide structured dense captions:

Method (MiraBench-500)Overall ↑EntityAttr.Rel./ActionCamera
Direct34.5%46.4%32.8%28.6%31.5%
Text-only (RL, text)49.0%62.1%47.7%40.8%39.5%
ProactGen50.6%65.2%48.9%42.1%41.7%

On this independent metric, single-round ProactGen recovers 50.6% of missing intent (+16.1 pp over Direct and +1.6 pp over RL Text-only). Across components, ProactGen achieves the largest gains over Direct on entities (+18.9 pp; +3.1 pp over Text-only) and attributes (+16.1 pp), while also improving relations/actions (+13.5 pp) and camera (+10.2 pp), outperforming both Direct and Text-only on every component.

References:

[1] MiniMax. (2026). MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities.

[2] Guan, K., et al. (2025). ETVA: Evaluation of Text-to-Video Alignment via Fine-grained Question Generation and Answering. ICCV 2025.

AuthorsOpenReview ↗

Rebuttal by Authors (4/4)

W2-2: Similarly, the timing results measure automated processing with a simulated respondent, so they do not show how much time a person would save.

A4: Thank you for pointing this out. Our original timing figures in Section 4.4 and Appendix A.8 measured server-side processing time; in our human-in-the-loop study (A3), we measured real human response time and mental effort to directly evaluate how much time a person saves in practice (Section 4.5 Table 3 and Appendix A.8 Table 7):

(4.1) Main efficiency comparison with measured human response time: To evaluate total user wall-clock time within our main efficiency comparison (Section 4.4), we replace the simulated-user time in Appendix A.8 (Table 4 in the initial submission, now Table 6) with the 108.8 s human response time per 5-question round measured in our human study (A3). For 5-turn Proactive T2I, which likewise asks 5 questions in total, we substitute the same 108.8 s human answering time for comparison:

Method & Rounds (InstanceCap-500, Wan2.2)Server timeHuman timeCombined totalDS-DTW ↑
Direct7.1 s0 s7.1 s0.658
RAPO++-SSPO (1 / 4 rounds, non-interactive)21.2 / 55.1 s0 s21.2 / 55.1 s0.665 / 0.669
Proactive T2I (Ag3+GT, 5 turns = 5 questions)91.9 s~108.8 s~200.7 s0.662
Text-only (1 round)8.6 s108.8 s117.4 s0.675
ProactGen (1 round)13.2 s108.8 s122.0 s0.678
Text-only (2 rounds)10.9 s217.5 s228.4 s0.681
ProactGen (2 rounds)19.5 s217.5 s237.0 s0.683
Text-only (3 rounds)13.6 s326.3 s339.9 s0.683
ProactGen (3 rounds)25.9 s326.3 s352.2 s0.690
Text-only (4 rounds)16.7 s435.0 s451.7 s0.683
ProactGen (4 rounds)32.2 s435.0 s467.2 s0.692

As shown, while interactive clarification takes longer than non-interactive SSPO, a single ProactGen round exceeds the ceiling of user-free rewriting (0.678 vs. 0.669). Among interactive methods, once user answering time is included, the share of time spent generating the draft becomes very small (+4.6 s is only 3.8% of single-round time). Because ProactGen reaches higher alignment in fewer interaction rounds, the human time saved by avoiding even one extra round far outweighs the cost of draft generation (for instance, 3 rounds of ProactGen save ~100 s of total time over 4 rounds of Text-only and exceed all 12 text-only rounds).

(4.2) Single-round time and effort savings over manual free-form revision: To further test whether answering targeted questions saves user time compared with manually editing the prompt after viewing a draft, the same 4 participants were requested to complete counterbalanced ProactGen and free-form revision tasks across the 100 targets on MiniMax-H3 [1] and rate overall mental effort (1–5 ↓) after each session:

Human interaction mode (n = 100)User response time ↓Full session time ↓Mental effort (1–5 ↓)Final DS-DTW ↑
Direct (non-interactive)0 s––0.683
Draft + Free-form revision119.9 s182.1 s4.000.697
ProactGen (answering 5 questions)108.8 s172.4 s2.100.704
Diff. (ProactGen − Free-form)−11.1 s−9.7 s−1.90+0.007

As shown, answering 5 targeted questions takes 108.8 s on average, saving 11.1 s per round over writing free-form revisions, while reducing mental effort from 4.00 to 2.10 and achieving higher alignment (0.704 vs. 0.697).

Reference:

[1] MiniMax. (2026). MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities.


Thank you again for your review. We look forward to any further discussion and are happy to address any additional questions.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

Dear Reviewers, ACs, and SACs,

We thank all reviewers for their time, careful reading, and valuable suggestions, and we appreciate their recognition of our problem motivation of proactively acquiring missing user intent beyond prompt rewriting (F1y8, 4KSj, M1jp), our outcome-driven RL design grounded in low-cost few-step drafts (F1y8, 4KSj), and our controlled multi-dataset evaluation (F1y8, M1jp).

To address the reviewers' concerns, we conducted additional experiments and updated the manuscript. The main concerns and key takeaways are summarized below:

Core ConcernReviewersHow Addressed & Key TakeawayReference
Value & efficiency of drafts over text-only RLF1y8, 4KSj, M1jpMulti-round experiments show that text-only RL quickly plateaus after two rounds, whereas observing the draft lets ProactGen keep targeting remaining visual errors and surpass 12 rounds of text-only RL in less time.F1y8 (1/4), 4KSj (3/5), M1jp (1/4, 3/4); Sec. 4.4, Fig. 3, App. A.9.2
Real-user benefit, latency & Free-form baselineF1y8, 4KSj, M1jpIn a 100-target human study, real users answered 98.6% of questions easily; compared with manually editing prompts after viewing the draft (Free-form), ProactGen is faster, cuts mental effort nearly in half (2.10 vs. 4.00), and yields better alignment.F1y8 (3/4, 4/4), 4KSj (1/5, 2/5), M1jp (2/4, 3/4); Sec. 4.5, Tab. 3, App. A.8, A.9.3
Evaluation beyond DreamSimF1y8, 4KSjA 300-pair blinded human preference study and an independent ETVA fine-grained intent recovery evaluation both confirm clear perceptual and semantic gains over baselines and text-only RL.F1y8 (3/4), 4KSj (4/5); Sec. 4.5, Tab. 3, App. A.9.4
Cross-generator & longer-video generalization4KSjWithout retraining, the frozen policy transfers zero-shot to HunyuanVideo-1.5 and longer 5s MiniMax-H3 videos with consistent alignment gains.4KSj (5/5); Sec. 4.6, App. A.9.5
Baseline variants, question budget & dynamic casesF1y8, M1jpAdded all Proactive T2I variants (Ag1–Ag3), analyzed question-budget truncation alongside a new Limitations section on adaptive stopping, and provided multi-frame and full-video examples correcting human actions and camera motion.F1y8 (2/4), M1jp (2/4, 4/4); Sec. 4.1, 5, App. A.7, A.9.2, A.9.6

All revisions are marked in red in the updated manuscript. We address each reviewer's comments point by point below, and thank all Reviewers, ACs, and SACs for their time and consideration.

AuthorsOpenReview ↗

Gentle Follow-up as Discussion Period Ends

Dear Reviewers F1y8, 4KSj, M1jp, and Area Chair,

Thank you again for your time and effort in reviewing our paper.

As the author-reviewer discussion period is in its final hours, we wanted to gently follow up on our point-by-point responses and revised manuscript. We remain actively available and would be glad to clarify any remaining questions or points that could help with your final assessment before the discussion window closes.

Best regards, Authors of Submission 25111

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.