acceptodds
Under review as a conference paper at ICLR 2027

Editing To Clarify and Harness Reasoning

14 views

Abstract

Multimodal Large Language Models have advanced visual reasoning, yet a purely textual chain of thought remains a bottleneck for questions that require fine-grained focus or view transformations. The "think with images" paradigm narrows this gap, but existing approaches are either constrained by fixed predefined toolkits or produce noisy intermediate images from unified multimodal methods. We pursue a third option: using a dedicated image editing model and decouple it with an understanding model. However, off-the-shelf image editors fail as reasoning assistants with two complementary gaps: a language-side gap, where editors trained as passive instruction-followers cannot map an abstract question to an appropriate visual transformation, and a generation-side gap, where edit correctness degrades as reasoning depth grows. Guided by this analysis, we introduce ETCHR (Editing To Clarify and Harness Reasoning), a question-conditioned, reasoning-aware image editor decoupled from the downstream understanding model and trained with a two-stage recipe targeted at the two gaps: Reasoning Imitation via supervised fine-tuning on edit trajectories, followed by Reasoning Enhancement with VLM-derived rewards for edit correctness and downstream reasoning accuracy. Since the editor is decoupled, ETCHR plugs into different open- and closed-source MLLMs in a training-free manner. Across five task families (fine-grained perception, chart understanding, logic reasoning, jigsaw restoration, and 3D understanding), ETCHR raises average Pass@1 from 55.95 to 60.77 (+4.82) with Qwen3-VL-8B, from 65.08 to 70.55 (+5.47) with Gemini-3.1-Flash-Lite, and from 76.55 to 81.16 (+4.61) with Kimi K2.5.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multimodal Large Language Model, Think with Images

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_2oNPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ETCHR, a framework that decouples image editing from visual understanding to enable MLLMs to "think with images." Addressing the limitations of tool-based methods and unified models, ETCHR employs a dedicated, question-conditioned image editor trained via a two-stage recipe: Reasoning Imitation to map abstract questions to visual transformations, and Reasoning Enhancement using VLM-derived rewards for edit correctness and downstream utility. The system further incorporates an Edit-Verify-Reason inference pipeline where the understanding model validates edits before use. Experiments across five task families demonstrate consistent Pass@1 improvements over baselines when paired with various open- and closed-source MLLMs. However, despite these overall gains, the evaluation lacks a granular, task-stratified analysis of editor-specific failure modes and does not sufficiently validate the reliability of the verification mechanism against human judgment, leaving concerns about error propagation and the true contribution of the editing module unresolved. Should the authors address these concerns, I would consider raising my score.

Critical strengths
  1. Clear presentation. The paper is generally well-structured and clearly presents the proposed framework, training pipeline, and experimental results, making the method relatively easy to follow.

  2. Good decoupling properties. Decoupled design avoids unified-model trade-offs and preserves editing fidelity. Its plug-and-play editor works across diverse MLLMs without fine-tuning, delivering consistent benchmark gains.

Critical weaknesses

1.Insufficient Analysis of the Editor's Contribution and Failure Modes. Although the paper reports overall performance improvements, the experimental analysis does not sufficiently isolate the contribution of the proposed editing module or explain its task-dependent behavior. In particular, the paper provides qualitative examples, but lacks a systematic, task-stratified analysis of editing failures and their downstream effects to establish when and why editing benefits downstream reasoning, and when it introduces detrimental visual changes. More systematic task-level ablations and qualitative analyses are needed to distinguish failures caused by incorrect edits from those caused by the downstream VLM's inability to exploit otherwise useful edited images. Comparisons with controlled editing baselines would provide stronger evidence for the claimed contribution of the editor.

2.Insufficient Validation of the Edit Verification Mechanism. The Edit-Verify-Reason pipeline relies on an VLM to assess the validity of generated edits, but the reliability of this verification step is not sufficiently established. Since VLM-based verifiers may exhibit hallucinations, false acceptance of incorrect edits, or false rejection of valid edits, the verification mechanism may fail to prevent error propagation. The paper would benefit from a dedicated evaluation against human-annotated edit-validity judgments. Further comparisons between verification-enabled inference, verification-disabled inference, and fallback-to-original-image strategies would clarify whether verification reliably improves downstream reasoning. The authors should also discuss the potential for correlated errors between the verifier and the downstream understanding model, and investigate whether structured constraints or independent verification signals can improve reliability.

Questions for authors

Please refer to "Weaknesses" above.

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

ETCHR trains a question-conditioned image editor to create intermediate visual evidence for a separate multimodal language model. Stage I performs supervised imitation on question-edit trajectories. Stage II applies preference-based RL using edit-correctness and downstream-answer rewards. During inference, the understanding model verifies the edit before either reasoning from it or falling back to the original image. The paper reports gains across five task families and three MLLMs.

Critical strengths
  1. The improvements transfer across three different understanding models. Table 1 reports positive gains for Qwen3-VL, Gemini, and Kimi, including large improvements on Maze, Jigsaw, ViewSpatial, and DL3DV. Since the training rewards use Qwen, transfer to Gemini and Kimi provides useful evidence against purely Qwen-specific optimization.
  2. On the controlled 100-sample subsets, the proposed editor outperforms Nano Banana 2 mainly on Logic, Jigsaw, and 3D tasks. This suggests that task-oriented editing can add value beyond attaching a generic frontier image editor.
Critical weaknesses
  1. The experiments do not establish the claimed general reasoning mechanism. The generation-side diagnosis supplies the shortest path directly in the prompt, measuring the ability to render a known path rather than derive it. Task-level prompts explicitly specify operations such as drawing a maze path, restoring a jigsaw, or changing viewpoint, and evaluation uses the same five transformation families seen during training. No held-out edit family demonstrates the claimed general meta-capability. Moreover, RL adds only approximately 0.5 points over SFT, and reflection adds less than one point while slightly hurting some 3D results.
  2. Evaluation integrity and fairness are insufficiently established. Training and evaluation draw from related parent datasets, including V* and DL3DV, without reporting scene-level deduplication or split IDs. Qwen is used for data filtering, both rewards, and judging, creating shared model bias. Unsupported tasks are assigned numerical values for tool baselines and included in averages, substantially depressing those baselines. No matched-compute textual-deliberation control, shuffled-edit control, latency analysis, confidence intervals, or multiple training seeds are provided.
Questions for authors
  1. Are all evaluation images, scenes, questions, and procedural-generation seeds strictly disjoint from SFT and RL data? Please provide split IDs or hashes, especially for DL3DV, V*, COCO/Spatial-SSRL, and the in-house logic tasks.
  2. Do gains persist on a completely held-out transformation family and under controls using shuffled, uninformative, or plausible-but-incorrect edits?
  3. How much inference latency and compute does ETCHR add, including the 9B 30-step editor, verification call, and final reasoning call? How does this compare with an equal-compute additional MLLM deliberation call?
  4. Are the RL-over-SFT and reflection gains significant under paired bootstrap intervals and multiple training seeds?
  5. Do edit correctness and 3D geometry remain reliable when evaluated by independent human or non-Qwen judges?
Reviewer_ydcZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ETCHR trains an image editor to produce images that help another model answer visual questions. Training combines supervised fine-tuning with reinforcement learning. At inference, the answering model checks the edit, uses both images if the edit passes, and otherwise uses only the original. The paper evaluates three answering models across five task families and reports average improvements of 4.61–5.47 percentage points.

Critical strengths
  1. The editor works with different answering models without retraining them. The improvements on Gemini and Kimi are useful evidence that the benefit extends beyond Qwen3-VL, which provides the training rewards.

  2. Table 1 shows improvements across quite different tasks. The jigsaw result with Kimi stands out, increasing from 44.5 to 70.5. Table 2 also provides a useful comparison with Nano Banana 2 using the same answering models.

Critical weaknesses
  1. Table 1 does not clearly establish an advantage over other reasoning approaches. The systems use different backbones, and the averages include tasks marked as outside the supported scope of the tool baselines. Table 2 is a better controlled comparison, but it is unclear whether Nano Banana 2 receives explicit editing instructions generated by the answering model. Section 2 already shows that such instructions help, so this detail matters.

  2. The verification step is not convincing for 3D tasks. The prompt in Appendix B asks whether the relevant objects are visible, which does not establish that the generated view preserves the scene geometry. Table 3 also shows a small drop in 3D accuracy after verification. The paper acknowledges this drop, but does not measure how often the verifier accepts a geometrically incorrect edit.

Questions for authors
  1. What prompts and verification procedure were used for Nano Banana 2 in Table 2? Did the answering model generate explicit editing instructions, as in the analysis in Section 2?

  2. Do you have an assessment of incorrect 3D edits that pass verification? A few checked examples and the corresponding verifier decisions would help explain what this step can reliably detect.

  3. How were the training and evaluation data separated for DL3DV and the COCO-based tasks? Were source videos, scenes, and images kept separate?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.