acceptodds
Withdrawn from ICLR 2027

PRISM: Scope-Aware Multi-Agent Prompt Induction for Promptless Question Grounded Visual Synthesis

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Conventional Text-to-Image (T2I) synthesis and optimization assume that a visual prompt is already available. In reading comprehension (RC) illustration synthesis, however, the input consists of a passage and its questions, with no visual prompt specified. The system must first determine what evidence each illustration should depict and then induce prompts for synthesis. We formulate this prompt induction challenge as Question Grounded Visual Synthesis (QGVS). Its dual-scope constraints require question-scope evidence selection and answer secrecy for each illustration, alongside passage-scope visual consistency across the set. We introduce PRISM, a multi-agent system for RC illustration synthesis. Its scope-aware workflow jointly plans the full illustration set: question-scoped agents select evidence and construct renderer prompts, while passage-wide style planning and cross-question entity routing coordinate visual consistency. The direct mode uses frozen pretrained models. Across four RC dataset sources, direct PRISM improves visual quality over the strongest comparison (GeoMean 66.0 versus 60.6), with especially large consistency gains while preserving distinct question content. This advantage persists across the multiple tested VLM planners and T2I renderer sizes. Further experiments show that the synthesized illustrations yield larger question-answering gains for simulated developing and beginning readers. Finally, GRPO post training yields a PRISM variant with a different performance profile.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
illustration generation, prompt optimization, reading comprehension

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_kF8rOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces PRISM, a multi-agent framework for generating question-specific illustrations for reading comprehension. It coordinates evidence selection, shared visual style, and cross-question entity consistency while restricting direct answer disclosure. Experiments across four reading comprehension datasets show improved overall illustration quality compared with existing methods. Downstream evaluations suggest that the illustrations improve VLM question-answering accuracy when textual evidence is limited.

Critical strengths
  1. The paper clearly defines the task of generating question specific illustrations while maintaining consistency across a passage.
  2. Shared style planning and entity routing provide explicit mechanisms for coordinating visual appearance across different question specific scenes.
  3. The direct framework operates with frozen pretrained models, without requiring task-specific parameter updates.
Critical weaknesses
  1. Evidence of downstream usefulness remains limited. Figure 3 shows: with the full passage available, PRISM decreases accuracy on three of four student VLMs and provides no consistent improvement over text alone.
  2. Improvements under reduced text settings may primarily reflect information supplementation. The generator accesses the full passage and correct answers, whereas students receive incomplete text. Without a control supplying equivalent evidence as additional text, the benefit of visual presentation remains unclear.
  3. The improvement over a simpler monolithic approach is modest relative to the additional computational cost. Monolithic prompt induction achieves a GeoMean of 65.5 versus PRISM’s 66.0, with substantially lower runtime.
Questions for authors

Could the authors clarify the practical value of the generated illustrations beyond improvements in visual quality? Given the inconsistent gains with full passages, it remains unclear whether the benefits under reduced-text conditions arise from visual presentation or simply from supplying missing evidence. A comparison with equivalent textual support would help establish whether these benefits justify the additional computational cost.

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

This paper introduces the QGVS task: given a reading comprehension passage and its questions, generate one illustration per question without a pre-existing image prompt. Each illustration must depict evidence relevant to its question without revealing the answer, while the full set must maintain consistent style and entity identity. The authors propose PRISM, which uses a single VLM in multiple roles to plan style, per-question content, and shared entities, then generates style and entity reference images as conditioning for a T2I renderer, followed by an answer-leakage check. On the evaluated benchmarks, PRISM shows clear improvements in cross-illustration consistency; simulated weaker readers also show modest gains in question-answering accuracy when provided with the generated illustrations.

Critical strengths
  • The paper addresses a practically meaningful and under-explored scenario: generating illustrations for reading comprehension questions requires deciding what to depict from the passage and questions alone. The authors formalize this as a standalone task and identify its core tension: each illustration must serve its own question without leaking the answer, while the full set must share a coherent visual style and character identity. This framing clearly distinguishes QGVS from single-image prompt optimization and story visualization.
  • The system design is clear and easy to follow. The responsibilities, inputs, and outputs of each module are well specified; together with the appendix, readers can largely reconstruct the full pipeline.
  • The evaluation covers multiple data sources and planner backbones. Human ratings further confirm that the method improves cross-illustration consistency without making different questions' illustrations look repetitive.
Critical weaknesses
  • Limited methodological contribution. PRISM is essentially an engineering combination of existing components: LLM planning, style templates, subject reference images, and self-review, applied to a new task, with all "Agents" being the same model processing each passage as a batch. The ablation results show that a merged-planning variant (Monolithic Prompt Induction) performs nearly on par with the full system (65.5 vs. 66.0 GeoMean) while being considerably faster. Removing the Style Agent drops PRISM below existing baselines. This suggests that the gains stem primarily from the shared reference-image conditioning rather than the scope-aware decomposition. The paper does not clarify whether the baselines could use the same reference-image conditioning.
  • Missing comparison with native sequential-image generation capabilities. Models such as Wan 2.7 and Seedream already natively support sequential-image generation or multi-reference consistency, which substantially overlaps with what PRISM achieves through reference images. The paper argues that other renderers lack reference-image support, but this is based only on a limited exploratory trials with SD 3.5 + IP-Adapter. If single-call LLM planning combined with native sequential generation can achieve comparable results, the contribution would largely remain at the level of task definition.
  • Insufficient evaluation robustness and comparability. Automatic scoring relies on a single Qwen-family judge (same model family as the planner), with no cross-model verification. The human evaluation does not include any method that explicitly maintains cross-image consistency (e.g., sequential-image generation or multi-round editing), so the observed consistency advantage may partly reflect differences in task objectives and generation conditions rather than PRISM's planning mechanism.
Questions for authors
  • Could you add a simple-pipeline baseline: single-call LLM planning with the same style/entity reference images, or a native sequential-image model? This would clarify what portion of PRISM's advantage comes from the scope-aware decomposition versus the reference-image conditioning.
  • In the reading comprehension experiment, illustrations are generated from the full passage and the ground-truth answer, yet the simulated weak readers only see truncated text. The observed accuracy gains may come from the illustrations recovering deleted information rather than from the visual presentation itself. Could you provide an ablation with answer-blind illustration generation, as well as a text-only control that supplies the same selected evidence?
Reviewer_S59VOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies question grounded visual synthesis for reading comprehension. PRISM induces illustration prompts from passages and question-answer pairs while coordinating shared appearance, question-specific evidence, entity references, and answer-secrecy review. On 240 passages and 856 questions, direct PRISM obtains a GeoMean of 66.0 versus 60.6 for the strongest direct baseline, with human ratings supporting substantial consistency gains. Optional GRPO mainly changes the consistency-diversity tradeoff. The task formulation and system organization are useful, but secrecy reliability and the incremental value of the complete decomposition remain insufficiently established.

Critical strengths

The distinction between question-level semantics and passage-level appearance is clear. Entity routing provides a concrete connection between these scopes, and Appendix A documents inspectable interfaces.

Automatic and human evaluations provide complementary evidence. In Table 5, PRISM scores 4.91 on consistency and 4.34 overall, while diversity remains 4.39 versus the highest 4.43. This supports coherent appearance without necessarily repeating question content.

The paper includes structural ablations, source-level results, and multiple model pairings. It appropriately limits renderer generalization and post-training claims; Appendix C.5 explicitly disclaims calibrated human-reader simulation and significant MCQA method differences.

Critical weaknesses

Answer secrecy is central, yet human evaluation finds leakage in 1.7% of direct and 3.4% of GRPO illustrations. Figure 11, Q2, visibly shows “NO POO / NO SHAMPOO” in the GRPO image, revealing the correct definition. This demonstrates an actual audit failure. Equation (7) describes the reviewer model's decision, not verified absence of leakage. Failure analysis and downstream results excluding confirmed leaks are missing; the human leakage rubric is also narrower than the full audit contract.

Table 2 provides limited support for the necessity of the complete decomposition. Monolithic induction achieves 65.5 versus 66.0 at 2.19× versus 3.86× runtime; removing the Content Agent yields 65.8 and higher LAS. The monolithic variant's retained references and contracts need specification, together with evidence that decomposition improves meaningful content distinctions under comparable resources.

For Figure 3, reduced profiles receive approximately 50% or 25% of the text, whereas illustrations are planned from the full passage and private answers. Gains can therefore reflect restored information. The acknowledged simulation caveats are appropriate, but an equivalent textual-evidence condition, effective denominators, and uncertainty are needed to isolate the contribution of visual organization.

Six of thirteen aggregate components come from one Judge, several categories receive renderer prompts, and consistency/diversity share embedding similarities. Human evaluation supports overall visual gains, but the 0.5-point monolithic gap needs a limited prompt-blind or aggregation-sensitivity check before being treated as strong mechanistic evidence.

Questions for authors

How did the Figure 11 disclosure pass review? Please clarify retry/termination rules and whether MCQA gains persist after excluding confirmed leaks.

Does the monolithic variant use identical style and entity references? Can a matched-budget comparison on the existing subset assess question-content distinctions?

Can you provide Figure 3's valid sample counts, baseline accuracies, paired intervals, and an equivalent textual-evidence control, plus one prompt-blind Judge check?

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

Related papers

Discussion (0)

Sign in to comment.