acceptodds
Under review as a conference paper at ICLR 2027

Don’t Overthink: Adaptive Thinking that Does not Fireback for Image Editing using GRPO

Abstract

Reasoning-augmented image editing operates under a prevailing assumption: more deliberation yields better edits. We show this assumption is fundamentally flawed. Through a diagnostic study of BAGEL, a state-of-the-art unified multimodal model, we identify three root pathologies of SFT-based reasoning: complexity-agnostic outputs that apply identical deliberation depth to trivial and complex edits alike; adherence dependency, where correct reasoning fails to translate into generation gains without sufficient adherence capacity in the generation expert; and verbosity without substance, where tripling the reasoning budget yields no quality improvement (GPT-Score ≈5.2), while structural directives at the same budget deliver a substantial gain to 6.1. These findings converge on a single conclusion: neither adaptive reasoning length nor structured templates alone resolves overthinking, both are jointly necessary. We introduce PRISM (Planning with Reasoning via Instructed Structured Modes), a compound adaptive-and-structured reasoning framework that decomposes pre-generation deliberation into three sequential stages, complexity steering, tailored lens generation, and ordered lens execution, distilled into unambiguous per-sample actions that drive a flow-matching generation expert. The understanding and generation experts are trained jointly via GRPO and Flow- GRPO with variance-guided early-window selection, without any external verifier or multi-turn inference. To power the framework, we introduce REVISE-50K, a dataset of 50,000 samples spanning 50 editing skills with graded complexity labels and disentangled hardness and verbosity axes. PRISM establishes a new state of the art on three reasoning-intensive benchmarks, +13.0% on KRIS, +3.4% on RISE, +17.8% on PICA, while preserving performance on direct editing tasks and surpassing multi-round external verifier systems at 20× lower inference cost.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Diffusion, Image Editing, GRPO, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ih2sOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies why the benefits of reasoning vary across BAGEL image-editing tasks. PRISM predicts task complexity, selects one to three instance-specific reasoning lenses, executes them sequentially, and consolidates their outputs into three actions before a single image-generation step. The understanding and generation experts are trained with GRPO and early-window Flow-GRPO, respectively. The authors also construct the 50,000-example REVISE-50K dataset and report improvements over the listed baselines on KRIS, RISE, and PICABench-Superficial, alongside better direct-editing performance (Table 2). The contribution primarily combines complexity routing, structured planning, generation training, and dataset design, rather than introducing a new foundational optimization algorithm.

Critical strengths
  1. The improvements span multiple settings rather than only difficult edits. In Table 2, PRISM reaches 73.2, 15.27, and 64.32 on KRIS, RISE, and PICA-Superficial, respectively. These correspond to gains of approximately 13.0, 3.4, and 17.8 score units over BAGEL-Think, rather than relative percentages. GEdit-R and ImgEdit also improve from 5.06 and 3.0 to 7.00 and 4.00. Although PRISM does not lead every direct-editing baseline, these results provide encouraging evidence of the combined method's practical potential.
  2. The authors make a substantive effort to isolate sources of improvement and validate supervision. Tables 4 and 9 distinguish the effects of generation RL, the reasoning framework, and training data. In Table 5, always-Heavy structured reasoning achieves 73.4 on KRIS, while the adaptive version recovers GEdit from 6.37 to 7.00, illustrating a useful quality trade-off. The complexity study has three annotators rate the same 1,000 examples, reports teacher–human-consensus quadratic weighted kappa (QWK) of 0.608, and candidly notes inter-annotator agreement of only 0.458 (Section 5.3). The window-variance measurements also provide empirical motivation for the early window; these analyses should not be treated as absent.
Critical weaknesses
  1. The mechanistic interpretation exceeds what the controls establish. Although Table 5 controls GRPO, data, and budget, it lacks an adaptive-but-unstructured variant and corresponding fixed-Light/Moderate results. Table 1 compares fixed word budgets and evaluates textual reasoning rather than final edits, so it cannot rule out simpler budget policies. There is also a concrete inconsistency in the CFG diagnostic: without CFG, Figure 3 calls for comparing KRIS scores of 48.5 and 42.9, whereas Section 3.2 describes the effect as 55.0 to 42.87, crossing CFG settings. The negative effect remains, but attributing the full 12.1-point drop to reasoning alone is not justified. The safer conclusion is that the combined design works in this BAGEL configuration, not that adaptivity and structure are universally “jointly necessary.”
  2. Validation of the evaluation underlying the quality claims remains incomplete. This work reasonably adds preservation as an evaluation dimension, but also switches to a GPT-5.2 judge and replaces the original GEdit protocol with a three-aspect geometric mean. It states that the authors conducted blinded preference comparisons on 200 GEdit and 200 KRIS examples, as well as a 300-example comparison of the old and new protocols, yet preference rates, agreement with human judgments, and uncertainty estimates are not reported. The human study in Section 5.3 validates complexity labels and cannot substitute for output-quality validation. Tables 2–5 also lack confidence intervals for the main scores or repeated-training results, leaving small differences and ranking stability under the new metric uncertain. This is an evidence gap, not proof that the judge is necessarily biased.
Questions for authors
  1. Could you add adaptive free-form CoT and fixed-Light/Moderate controls to Table 5, reporting final editing quality and actual token costs by complexity level? Could you also recompute the effect sizes in Figure 3 and Section 3.2 while holding CFG fixed?
  2. What exactly does the “20-fold” cost reduction measure? Please provide a comparison with EditThinker under matched hardware, resolution, and quality settings, including complexity prediction, lens generation and execution, and action generation in the end-to-end cost.
7 comments
AuthorsOpenReview ↗

Response 1, Part 1

We apologize for posting our response late in the discussion period. We used the time to run all the experiments and comparisons that the reviewers requested, so that each concern is answered with new evidence. We would be grateful for the reviewer's response and feedback at their earliest convenience, which would let us make the best use of the remaining discussion period and address any further concerns in a timely manner.

Table, figure and appendix numbers in this response refer to the revised paper; where the review cites a number from the submission, we give both.

We thank the reviewer for the detailed and balanced review. We are glad that the reviewer found the improvements broad, spanning reasoning-intensive and direct editing (S1), and our effort to isolate the sources of improvement and to validate the supervision substantive (S2). We have run each of the controls the reviewer requested. We answer W1 together with Q1, then W2 and Q2; all changes are included in the revised paper.


W1 / Q1. Do the controls support adaptivity and structure as jointly necessary?

(a) Adaptive free-form and fixed-level controls (new). We add the three requested controls, so that Table 7 (Table 5 in the submission) becomes a full grid over adaptivity and structure, scored on the final edits. The adaptive free-form variant selects a Light, Moderate or Heavy budget per sample, as PRISM does, but reasons in free form. All rows share the GRPO recipe.

Table R1. Table 7 of the revised paper.

ReasoningAdaptiveStructuredGEditKRIS
None (Flow-GRPO only)––6.9460.1
Free-form, Light budget (new)✗✗6.5061.5
Free-form, Heavy budget✗✗6.2562.8
Free-form, adaptive budget (new)✓✗6.5062.4
Structured, fixed Light (new)✗✓7.0061.4
Structured, fixed Moderate (new)✗✓6.8766.3
Structured, fixed Heavy✗✓6.3773.4
PRISM✓✓7.0073.2

Adaptivity without structure improves neither benchmark. The adaptive free-form variant reaches 6.50 on GEdit and 62.4 on KRIS, within the range of the fixed budgets. Free-form reasoning lowers GEdit below no reasoning at every budget (6.25 to 6.50, against 6.94), whereas structured reasoning at the Light level does not (7.00).

Structure without adaptivity is strong on one benchmark at a time. Fixed Light keeps GEdit at 7.00 but reaches only 61.4 on KRIS; fixed Heavy reaches 73.4 on KRIS but lowers GEdit to 6.37. Only PRISM matches the best fixed level on both. With an adaptive budget, adding structure raises KRIS by 10.8 points; with structure, adding adaptivity raises GEdit by 0.63.

(b) Simpler budget policies, on final edits. Table 1 indeed scores the reasoning text. Table R1 closes this gap: it scores the edited images, and its free-form rows are the simpler budget policies, three fixed budgets and a per-sample adaptive budget. None of them exceeds 62.8 on KRIS or reaches the no-reasoning score on GEdit.

Token cost by complexity level. PRISM's reasoning grows with the predicted complexity: one lens for Light and up to three for Heavy, that is, four to six sequential text calls. Table R2 (Table 15 in Appendix D.11) reports the generated tokens per edit.

Table R2. Generated reasoning tokens per edit, by predicted complexity.

LightModerateHeavyAverage over levels
Free-form, adaptive budget30529057
PRISM22607753

Both adaptive variants scale their reasoning with complexity, by more than three times from Light to Heavy for PRISM (22 to 77 tokens), and PRISM spends no more tokens than free-form reasoning (53 against 57 on average). The two adaptive rows of Table R1 therefore differ in structure, not in budget, and structure adds 10.8 points on KRIS and 0.5 on GEdit. The budget also follows the benchmark: reasoning-intensive benchmarks contain the largest share of Heavy edits (35% on KRIS and 56% on RISE, Figure 12), so PRISM spends its reasoning where it pays off. Adaptivity matters for quality as well as cost: on direct edits, the Heavy reasoning that PRISM avoids lowers GEdit from 7.00 to 6.37 (Table R1).

AuthorsOpenReview ↗

Response 1, Part 2


(c) Effect sizes at fixed CFG. Table R3 reports the effect of reasoning at fixed CFG, computed from the four cells of Figure 3.

Table R3. KRIS for BAGEL, from Figure 3, with the effect of reasoning at fixed CFG.

Without CFGWith CFG
Without reasoning48.555.0
With reasoning42.960.0
Effect of reasoning−5.6+5.0

At fixed CFG, reasoning lowers KRIS by 5.6 points without CFG and raises it by 5.0 with CFG. These within-setting effects support Observation 2 directly: the sign of the effect of reasoning depends on generation adherence. Equivalently, CFG adds 6.5 points without reasoning but 17.1 with it. On GEdit, reasoning lowers the score at either CFG setting. The revision reports these within-setting effects in Section 3.2 and in the caption of Figure 3.

(d) Scope of the claim. We agree, and we state the scope of each claim precisely. Our evidence has two levels of generality. The diagnosis, that reasoning helps reasoning-intensive edits and hurts direct ones, holds on three editors (Table R4; Appendix D.10).

Table R4. Effect of reasoning on three editors. BAGEL uses its own thinking mode; for OmniGen2 and Qwen-Image-Edit, reasoning is generated by Qwen2.5-VL and prepended to the instruction.

BackboneGEdit: no reasoning → reasoningKRIS: no reasoning → reasoning
BAGEL5.95 → 5.71 (−0.24)56.21 → 60.18 (+3.97)
OmniGen25.58 → 5.12 (−0.46)50.52 → 57.38 (+6.86)
Qwen-Image-Edit7.23 → 7.00 (−0.23)64.43 → 68.89 (+4.46)

The remedy, and the finding that both axes are needed, are established for BAGEL trained with our recipe (Table R1). The revision states the second finding for this configuration and replaces "jointly necessary" with "both needed in our experiments".


W2. Validation of the evaluation

We add a human evaluation of output quality and of the GEdit-R protocol.

Output quality, judged by humans. Annotators compared edited images side by side, without knowing which model produced each, on 200 GEdit and 200 KRIS samples.

Table R5. Blinded pairwise human preference (n = 400 per comparison; 95% confidence interval in brackets).

ComparisonPRISM preferredBaseline preferredTie
PRISM vs BAGEL-Think84% [80.4, 87.6]8%8%
PRISM vs Qwen-Image-Edit68% [63.4, 72.6]27%5%

Both preferences are significant (p < 0.001, two-sided sign test excluding ties).

Agreement of GEdit-R with humans. On 300 GEdit samples, GEdit-R agrees more closely with the human ranking than the original protocol (Spearman ρ = 0.81 vs 0.64).

The modified protocol affects one benchmark (GEdit). KRIS, RISE, PICA and ImgEdit are scored with their released protocols and judges. PRISM's largest gains over BAGEL-Think, +13.0 on KRIS, +3.4 on RISE and +17.8 on PICA, are therefore measured under unmodified protocols, as is its ImgEdit gain (3.0 to 4.00). The Table 1 diagnostic also holds under Gemini as a second judge (length budgets 6.4 to 6.7, structured variant 7.9).

Uncertainty. We repeated the GPT-based evaluation of every benchmark four times. The judge variability is 0.15 on GEdit, 0.1 on ImgEdit, 1.0 on KRIS, 0.4 on RISE and 1.5 on PICA. Our conclusions rest on differences well above these levels: PRISM's gains over BAGEL-Think are 1.29 on GEdit, 1.0 on ImgEdit, 13.0 on KRIS, 3.4 on RISE and 17.8 on PICA, between 8 and 13 times the judge variability of each benchmark. Within the controls, structure adds 10.8 points on KRIS and adaptivity 0.63 on GEdit (Table R1). Differences within the judge variability, such as 7.00 vs 7.08 on GEdit (Table 6), we report as ties and draw no conclusion from. Appendix D.9 reports the judge variability (Table 13).

Revision: Section 5.3 reports the human study (Table 5) and the GEdit-R validation; Appendix D.8 gives the study protocols, and Appendix D.9 the judge variability (Table 13).

AuthorsOpenReview ↗

Response 1, Part 3


Q2. What the 20× measured, and a matched comparison with EditThinker

Measurement. We timed every method end to end on one A100 GPU at batch size 1, from the input image and instruction to the final edited image, with BAGEL as the editor throughout. For PRISM, the timing includes all four stages: complexity prediction, lens generation, lens execution and action generation. For EditThinker, it includes every verifier call and regeneration.

Table R6. Latency per edit.

MethodSettingImages per editLatency (s)Relative to PRISM
BAGEL512 px, 25 steps111.50.61×
BAGEL-Think512 px, 25 steps116.60.88×
PRISM512 px, 25 steps118.81.00×
BAGEL + EditThinker512 px, 25 steps3.143.42.3×
BAGEL-Think + EditThinker512 px, 25 steps3.563.83.4×
BAGEL-Think + EditThinker1024 px, 50 steps~41839.7×
BAGEL-Think + EditThinker, all 8 rounds1024 px, 50 steps835518.9×

At matched settings, EditThinker takes 2.3 to 3.4 times as long as PRISM. All of PRISM's reasoning stages together add 7.3 s to an 11.5 s edit. EditThinker instead regenerates the image in every round, three to four times per edit on average. The 20× figure corresponds to EditThinker's worst case, all eight rounds at its default setting (1024 px, 50 steps), relative to PRISM at our setting (last row, 18.9×). The revision reports the full range, with the ratio at matched settings as the headline.

Quality at that cost.

Table R7. EditThinker on top of BAGEL and of PRISM.

ModelGEditKRIS
BAGEL5.9556.2
BAGEL + EditThinker6.2262.8
PRISM7.0073.2
PRISM + EditThinker6.7175.6

PRISM alone outperforms BAGEL with EditThinker on both benchmarks (+0.78 on GEdit, +10.4 on KRIS) at less than half its latency. On top of PRISM, the verifier adds 2.4 points on KRIS but lowers GEdit by 0.29, which reproduces, one level up, the overthinking pattern of Section 3.2.

Revision: Section 5.2 adds the comparison with EditThinker (Table 3), and Appendix D.12 gives the full latency measurements and the protocol (Table 18).


We thank the reviewer again for the precise requests. They led to the full grid over adaptivity and structure, scored on final edits, to effect sizes at fixed CFG, to the human evaluation of output quality, and to a matched cost comparison that includes every stage of PRISM. We have also scoped the necessity claim as the reviewer suggested. We hope that these results address the reviewer's concerns, and we would be glad to discuss any remaining questions during the discussion period.

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

This paper examines the assumption that more extensive explicit reasoning before image generation leads to better instruction-based image editing in unified multimodal models.

Through diagnostic experiments on BAGEL, the authors show that enabling chain-of-thought reasoning improves performance on reasoning-intensive benchmarks such as KRIS, but can degrade performance on direct editing tasks such as GEdit. They characterize this degradation as overthinking and identify three limitations of reasoning learned through supervised fine-tuning: complexity-agnostic reasoning, adherence dependency, and verbosity without substance.

Based on these findings, the authors propose PRISM, a framework that combines adaptive reasoning with structured planning through complexity steering, tailored lens generation, and ordered lens execution. They also introduce REVISE-50K, a dataset of 50000 editing samples for reinforcement learning with adaptive reasoning.

Critical strengths
  1. The paper provides thorough diagnostic experiments showing that longer reasoning does not necessarily improve editing quality in models such as BAGEL. In particular, the 2 × 2 ablation of CFG and reasoning demonstrates that the benefits of reasoning depend on the generation model’s ability to follow its plan.

  2. PRISM has a coherent design. The authors define nine failure modes and guide the model to select corresponding reasoning perspectives, which helps constrain redundant deliberation. The experimental results also show substantial performance gains.

Critical weaknesses
  1. The reported baseline scores appear inconsistent. Figure 3 reports a GEdit score of 6.18 for BAGEL with classifier-free guidance and without reasoning, whereas Table 2 reports 5.53. The KRIS scores also differ. What accounts for these discrepancies?

  2. The comparison in Table 2 does not isolate the contribution of PRISM from that of additional training. BAGEL and BAGEL-Think are evaluated using their original weights without task-specific fine-tuning, while PRISM is trained on REVISE-50K, which contains 50000 high-quality reasoning-based editing samples. Could the authors add a baseline trained on the same 50000 samples using prompts without reasoning? This would help establish the gains attributable to PRISM itself. If time or computational constraints prevent this experiment, please explain the rationale for the current comparison in more detail.

  3. Section 1 claims that PRISM surpasses a system using multiple rounds of external verification, EditThinker, at 20× lower inference cost. Could the authors provide a quantitative cost comparison and clarify how inference cost is measured?

  4. In Table 4, adding Flow-GRPO to BAGEL improves the GEdit score by 1.59. Adding the full PRISM framework and switching to REVISE-50K yields a score of 7.00, only 0.06 higher than Flow-GRPO without reasoning. Is this marginal gain sufficient to demonstrate the benefit of PRISM’s reasoning on direct editing tasks? Or is its main contribution on these tasks preventing the degradation caused by redundant reasoning?

  5. The paper mentions a blinded pairwise preference study comparing PRISM against BAGEL-Think and Qwen-Image-Edit, but lacks qualitative visual comparisons. Please provide representative side-by-side examples to illustrate the differences in editing quality.

Questions for authors

Please address the questions and missing comparisons raised in the Critical Weaknesses section.

8 comments
AuthorsOpenReview ↗

Response 1, Part 1

We apologize for posting our response late in the discussion period. We used the time to run all the experiments and comparisons that the reviewers requested, so that each concern is answered with new evidence. We would be grateful for the reviewer's response and feedback at their earliest convenience, which would let us make the best use of the remaining discussion period and address any further concerns in a timely manner.

Table, figure and appendix numbers in this response refer to the revised paper; where the review cites a number from the submission, we give both.

We thank the reviewer for the careful reading and the precise questions. We are glad that the reviewer found the diagnostic experiments thorough, in particular the 2 × 2 ablation of CFG and reasoning (S1), and the design of PRISM coherent, with substantial gains (S2). We answer each point below, with the new experiments the reviewer requested, and all changes are included in the revised paper.


W1. Inconsistent baseline scores between Figure 3 and Table 2

We thank the reviewer for the careful check, and we have updated Table 2. Its GEdit entries for BAGEL and BAGEL-Think corresponded to BAGEL without CFG. With CFG, BAGEL's default setting, they are 5.95 and 5.71 under GEdit-R. PRISM (7.00) remains well above both, and the conclusion of Observation 1 is unchanged: reasoning lowers BAGEL's score on direct edits (5.95 → 5.71).

The remaining gap reflects two evaluation protocols. Figure 3 (6.18 for BAGEL with CFG) uses the original GEdit protocol, which scores instruction adherence and perceptual quality. Table 2 uses GEdit-R (Appendix D.5), which adds identity preservation as a third aspect and uses GPT-5.2 as the judge. GEdit-R penalizes edits that alter content outside the edited region, which the two-aspect score cannot detect, and is therefore the stricter of the two. We adopted it for all comparisons in Section 5 because it agrees more closely with human rankings (Spearman ρ = 0.81 vs 0.64 on 300 GEdit samples). In the revision, Observation 1 and Table 2 both use GEdit-R, and the caption of Figure 3 states its protocol.

The KRIS scores agree within judge variability. KRIS is scored with its released protocol in both places. BAGEL-Think scores 60.0 in Figure 3 and 60.18 in Table 2; BAGEL scores 55.0 and 56.21. The difference of about one point between the two evaluation runs matches the judge variability on KRIS, which is 1.0 over four repeated evaluations.

The overthinking effect is not specific to BAGEL. Because these baseline numbers underpin our central observation, we also tested whether the observation depends on BAGEL. We repeated it on two further editors, OmniGen2 and Qwen-Image-Edit, under the protocols of Table 2. For both, the reasoning is generated by Qwen2.5-VL, the multimodal model on which both editors build, and prepended to the editing instruction.

Table R1. Effect of reasoning on three editors (GEdit-R and KRIS, as in Table 2; Appendix D.10).

BackboneGEdit: no reasoning → reasoningKRIS: no reasoning → reasoning
BAGEL5.95 → 5.71 (−0.24)56.21 → 60.18 (+3.97)
OmniGen25.58 → 5.12 (−0.46)50.52 → 57.38 (+6.86)
Qwen-Image-Edit7.23 → 7.00 (−0.23)64.43 → 68.89 (+4.46)

All three editors show the same pattern: reasoning helps where the edit must be inferred (KRIS, +4.0 to +6.9) and hurts where the instruction already states it (GEdit, −0.23 to −0.46). On direct edits, the model spends computation on deliberation the edit does not need, and the result degrades.

AuthorsOpenReview ↗

Response 1, Part 2

We answer W2 and W4 together, since both concern the attribution in Table 6 (Table 4 in the submission).


W2. Isolating the contribution of PRISM from the additional training data

We ran the requested baseline. We trained BAGEL on the same REVISE-50K samples without reasoning, in two ways: with supervised fine-tuning, as in BAGEL's own training, and with Flow-GRPO, using exactly the recipe of PRISM's generation expert (same reward, learning rate and schedule).

Table R2. Same training data, with and without reasoning.

ModelTraining on REVISE-50KReasoningGEditKRIS
BAGEL–✗5.9556.2
BAGELSFT✗5.1260.7
BAGELFlow-GRPO✗6.8163.1
PRISMGRPO (Reasoning) + Flow-GRPO (Image Editing)✓7.0073.2

At matched data, reasoning accounts for most of the gain on KRIS. The data alone raise KRIS by 4.5 points with SFT and by 6.9 with Flow-GRPO; PRISM adds a further 10.1 points on the same samples. On GEdit, SFT on REVISE-50K lowers the score (5.12), Flow-GRPO raises it to 6.81, and PRISM reaches 7.00.

Data and reasoning are complementary. Together with Table 6, the new baseline completes a 2 × 2 design over training data and reasoning.

Table R3. GEdit / KRIS for each combination of data and reasoning. All four models train the generation expert with the same Flow-GRPO recipe.

Trained on OmniEditTrained on REVISE-50K
No reasoning6.94 / 60.16.81 / 63.1
PRISM reasoning7.08 / 66.87.00 / 73.2

Reasoning raises KRIS on either dataset (+6.7 on OmniEdit, +10.1 on REVISE-50K), and REVISE-50K raises KRIS with or without reasoning (+3.0 and +6.4). The dataset adds twice as much when the model reasons, because its graded complexity labels and reasoning-intensive instructions are what the reasoning stages learn from. The gain of PRISM over BAGEL therefore does not come from additional data alone. Table 2 compares released systems, each trained on its own data; Tables R2 and R3 give the controlled comparison.

Revision: Table 6 adds the two baselines trained on REVISE-50K without reasoning, and Section 5.3 discusses them.


W4. The GEdit gain of PRISM over Flow-GRPO

On direct edits, the aim of PRISM is not to gain from reasoning but to avoid the loss that reasoning otherwise causes; its gains are meant to come from edits that require reasoning. This is why Table 6 reports one benchmark of each kind.

Without adaptivity, reasoning lowers GEdit. Against the same model without reasoning (6.94), free-form CoT lowers GEdit to 6.25 and structured reasoning at a fixed Heavy level to 6.37 (Table 7). BAGEL's own reasoning lowers it from 5.95 to 5.71 (Table 2), and the same holds for OmniGen2 and Qwen-Image-Edit (Table R1). PRISM is the only reasoning variant that does not lose on GEdit (7.00), while it adds 10.1 points on KRIS at matched data (Table R2).

At matched data, the difference on GEdit is small but not negative. The 0.06 compares models trained on different data (Flow-GRPO on OmniEdit against PRISM on REVISE-50K). At matched data, PRISM's reasoning changes GEdit by +0.14 on OmniEdit and +0.19 on REVISE-50K (Table R3). We do not present these as reasoning gains on direct edits, since they are close to the judge variability of 0.15 on GEdit (over four repeated evaluations), but they confirm that adaptive reasoning does not hurt there. PRISM also reasons least where reasoning helps least: it assigns a single lens to edits it predicts as Light, which generate 22 reasoning tokens on average, against 77 for Heavy edits.

AuthorsOpenReview ↗

Response 1, Part 3


W3. Quantitative cost comparison with EditThinker

We measured inference cost as wall-clock latency per edit. We timed every method on one A100 GPU at batch size 1, from the input image and instruction to the final edited image, with BAGEL as the editor throughout. The timing covers all four of PRISM's reasoning stages and all of EditThinker's verifier calls and regenerations.

Table R4. Latency per edit.

MethodSettingImages per editLatency (s)Relative to PRISM
BAGEL512 px, 25 steps111.50.61×
BAGEL-Think512 px, 25 steps116.60.88×
PRISM512 px, 25 steps118.81.00×
BAGEL + EditThinker512 px, 25 steps3.143.42.3×
BAGEL-Think + EditThinker512 px, 25 steps3.563.83.4×
BAGEL-Think + EditThinker1024 px, 50 steps~41839.7×
BAGEL-Think + EditThinker, all 8 rounds1024 px, 50 steps835518.9×

At matched settings, EditThinker takes 2.3 to 3.4 times as long as PRISM. PRISM's reasoning decodes text, which adds 7.3 s to an 11.5 s edit. EditThinker instead regenerates the image in every round, three to four times per edit on average. The 20× figure corresponds to EditThinker's worst case, all eight rounds at its default setting (1024 px, 50 steps), relative to PRISM at our setting (last row, 18.9×). The revision reports the full range, with the ratio at matched settings as the headline.

PRISM is also the stronger editor.

Table R5. EditThinker on top of BAGEL and of PRISM.

ModelGEditKRIS
BAGEL5.9556.2
BAGEL + EditThinker6.2262.8
PRISM7.0073.2
PRISM + EditThinker6.7175.6

PRISM alone outperforms BAGEL with EditThinker on both benchmarks (+0.78 on GEdit, +10.4 on KRIS) at less than half its latency. On top of PRISM, the verifier adds 2.4 points on KRIS but lowers GEdit by 0.29, which reproduces, one level up, the overthinking pattern of Section 3.2.

Revision: Section 5.2 adds the comparison with EditThinker (Table 3), and Appendix D.12 gives the full latency measurements and the protocol (Table 18).


W5. Qualitative side-by-side comparisons

We add side-by-side examples in Appendix D.13 (Figure 16), comparing PRISM with BAGEL, BAGEL-Think and Qwen-Image-Edit; the last two are the baselines of the preference study. The nine examples range from direct edits (removal, addition) to edits that require cultural knowledge, an inferred target state, spatial or logical reasoning, and counting. Asked to make a crossing passable, PRISM switches the traffic light to green, whereas BAGEL and BAGEL-Think leave it red, and Qwen-Image-Edit regenerates the whole scene. PRISM is also the only model that completes the visual pattern and the only one that reduces four rabbits to three. Figure 1 shows the full reasoning chain of PRISM for two further edits.

We add a blinded human preference study, reported below and in Section 5.3 (Table 5). Annotators compared the edited images side by side, without knowing which model produced each, on 200 GEdit and 200 KRIS samples.

Table R6. Blinded pairwise human preference (n = 400 per comparison; 95% confidence interval in brackets).

ComparisonPRISM preferredBaseline preferredTie
PRISM vs BAGEL-Think84% [80.4, 87.6]8%8%
PRISM vs Qwen-Image-Edit68% [63.4, 72.6]27%5%

Revision: Appendix D.13 adds the side-by-side comparison (Figure 16), referenced in Section 5.3; Section 5.3 (Table 5) and Appendix D.8 report the human preference study.


We thank the reviewer again for the precise questions. They led to the matched-data baseline, which shows that most of PRISM's gain on reasoning-intensive edits comes from its reasoning rather than from the additional data, to a measured cost comparison with EditThinker, and to consistent baseline entries across Figure 3 and Table 2. We hope that these results address the reviewer's concerns, and we would be glad to discuss any remaining questions during the discussion period.

Reviewer_6sVoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PRISM, a framework that adapts the amount and structure of reasoning to each editing task. It first predicts the task complexity, then generates task-specific reasoning lenses, executes them sequentially, and converts the results into concrete editing actions. The framework is built on BAGEL, with GRPO used for the understanding expert and Flow-GRPO for the generation expert. The authors also introduce REVISE-50K, a dataset covering 50 editing skills with different reasoning complexity levels. Experiments on five benchmarks show improvements over BAGEL and several existing editing methods, especially on reasoning-intensive tasks.

Critical strengths
  • Interesting motivation. The observation that more reasoning can actually hurt image editing is interesting and worth studying.

  • Strong experimental results. PRISM shows clear improvements on reasoning-heavy benchmarks, particularly KRIS and PICA. The results are quite encouraging.

  • Reasonably thorough ablations. The authors make a good effort to separate the effects of structured reasoning, Flow-GRPO, and the new dataset. REVISE-50K is also a useful contribution.

Critical weaknesses
  • The novelty is somewhat limited. The main idea feels like a combination of existing techniques: adaptive reasoning, structured CoT, and GRPO/Flow-GRPO. The lens mechanism is essentially a way of organizing reasoning through prompts, rather than a fundamentally new reasoning approach.
  • A central argument is that adaptivity and structure are both necessary. However, Table 5 is missing the most important baseline: adaptive reasoning without structure.
  • Could the model simply generate structured editing actions directly, without going through lens proposal and execution? How much improvement comes specifically from the lens mechanism? What happens if the model directly generates tailored actions without the intermediate lenses?
  • Several key results depend on VLM judges. Interestingly, the paper mentions human preference studies in the Ethics Statement but doesn't report their actual results.
  • The paper claims 20× lower inference cost, but I couldn't find a clear latency or computational cost comparison. Also, PRISM still involves several sequential reasoning steps, so calling it a "single forward pass" seems misleading.
  • In Appendix D.7, Table 10 says all action formats use three actions, but Prompt 12 specifies four fixed roles. Which setting was actually used?
Questions for authors

Please see the above weaknesses.

6 comments
AuthorsOpenReview ↗

Response 1, Part 1

We apologize for posting our response late in the discussion period. We used the time to run all the experiments and comparisons that the reviewers requested, so that each concern is answered with new evidence. We would be grateful for the reviewer's response and feedback at their earliest convenience, which would let us make the best use of the remaining discussion period and address any further concerns in a timely manner.

Table, figure and appendix numbers in this response refer to the revised paper; where the review cites a number from the submission, we give both.

We thank the reviewer for the careful reading and the constructive comments. We are glad that the reviewer found the observation that reasoning can hurt editing worth studying (S1), the gains on reasoning-intensive benchmarks encouraging (S2), and the ablations and REVISE-50K useful (S3). We address each point below, and all changes are included in the revised paper.


W1. Novelty

Our central contribution is the discovery of overthinking in image editing: reasoning can hurt the edit. Adaptive reasoning adapts how much a model reasons, and structured CoT fixes how it reasons, but both lines were developed for understanding: AdaptThink and Tree of Thoughts for language models, and FAST, TON and LLaVA-CoT for vision-language question answering. In generation, ShortCoTI shortens the reasoning of text-to-image models for efficiency. None of them studies image editing, where the reasoning does not produce an answer but must be executed by a generator. Both lines assume that deliberation does no harm once its amount or format is chosen. We show that in image editing this assumption fails: enabling reasoning lowers the score on direct edits while raising it on reasoning-intensive ones.

The phenomenon is general, not specific to BAGEL. To verify this, we repeated the diagnosis on two further editors, OmniGen2 and Qwen-Image-Edit. On all three editors, reasoning lowers GEdit and raises KRIS (Table R1; Appendix D.10).

Table R1. Effect of reasoning on three editors. BAGEL uses its own thinking mode; for OmniGen2 and Qwen-Image-Edit, reasoning is generated by Qwen2.5-VL and prepended to the instruction.

BackboneGEdit: no reasoning → reasoningKRIS: no reasoning → reasoning
BAGEL5.95 → 5.71 (−0.24)56.21 → 60.18 (+3.97)
OmniGen25.58 → 5.12 (−0.46)50.52 → 57.38 (+6.86)
Qwen-Image-Edit7.23 → 7.00 (−0.23)64.43 → 68.89 (+4.46)

The finding determines the design. If overthinking were a matter of length, adapting the length would cure it. Table 1 shows that it does not: tripling the budget leaves reasoning quality flat (5.12 to 5.26), whereas a structural directive at the same length raises it to 6.10. The ordering is unchanged under Gemini as an independent judge (6.4 to 6.7 against 7.9). PRISM therefore adapts an axis that length-adaptive methods leave fixed, what the model reasons about, in addition to how much. Our answer to W2 shows that both are needed.

Lenses are learned, not prompted. If the lenses only organized reasoning through prompts, prompting alone would deliver the gain. Three results show that it does not:

  • The format pays off only through training. Tailored actions score below fixed roles when only prompted (6.48 vs 7.48, Table 12) and overtake them only after GRPO (7.99 vs 7.69).
  • The model goes beyond the seed vocabulary. 86% of the lenses PRISM generates on the evaluation benchmarks are not among the nine seed lenses (Section 3.3), e.g. Reflection and Visual Coherence in Figure 1. The vocabulary only initializes the proposals; the stage rewards for lens relevance (R2) and grounded execution (R3) shape them.
  • GRPO alone does not explain the gain. Free-form CoT trained with the same GRPO recipe, budget and data reaches 62.8 on KRIS, against 73.4 with structure (Table 7).

What we claim. We do not claim GRPO, Flow-GRPO or chain-of-thought as contributions. We claim (i) the diagnosis that reasoning can harm editing, and that its benefit depends on structure and generation adherence rather than on length; (ii) a formulation in which both the reasoning budget and the reasoning perspectives are predicted per sample and trained jointly with the generation expert inside one model, whereas prior reasoning editors either fix the reasoning (BAGEL-Think) or delegate it to an external verifier (ReasonEdit, EditThinker); and (iii) REVISE-50K, whose graded complexity labels make this training possible. Table 6 shows that each contributes a separable share of the gain.

AuthorsOpenReview ↗

Response 1, Part 2


W2. Adaptive reasoning without structure

We agree that this is the most informative control for our central claim, and we add it, together with the fixed-budget variants on both sides, so that Table 7 (Table 5 in the submission) becomes a full grid over adaptivity and structure. In the adaptive free-form variant, the model selects a Light, Moderate or Heavy reasoning budget per sample, as PRISM does, but reasons in free form; the free-form CoT of Table 5 corresponds to the fixed Heavy budget. On the structured side, the complexity is fixed to each level.

Table R2. Table 7 of the revised paper. All rows share the GRPO recipe.

ReasoningAdaptiveStructuredGEditKRIS
None (Flow-GRPO only)––6.9460.1
Free-form, Light budget (new)✗✗6.5061.5
Free-form, Heavy budget✗✗6.2562.8
Free-form, adaptive budget (new)✓✗6.5062.4
Structured, fixed Light (new)✗✓7.0061.4
Structured, fixed Moderate (new)✗✓6.8766.3
Structured, fixed Heavy✗✓6.3773.4
PRISM✓✓7.0073.2

Adaptivity without structure improves neither benchmark. With an adaptive budget, free-form CoT reaches 6.50 on GEdit and 62.4 on KRIS. Routing short reasoning to direct edits lifts GEdit only to the level of the shortest fixed budget (6.50), and KRIS stays within the range of the fixed budgets (61.5 to 62.8): adaptivity changes how much the model reasons, not how well, as Table 1 predicts. Moreover, free-form reasoning lowers GEdit below no reasoning at every budget we tested (6.25 to 6.50, against 6.94), whereas structured reasoning at the Light level does not (7.00). On direct edits, what hurts is unstructured reasoning, not only its length.

Structure without adaptivity is strong on one benchmark at a time. Fixed Light keeps GEdit at 7.00 but underthinks hard edits (KRIS 61.4); fixed Heavy reaches 73.4 on KRIS but overthinks direct edits (GEdit 6.37). PRISM matches the best fixed level on each benchmark at once. The same holds within one benchmark: on KRIS samples labelled Light, fixed Heavy scores 62 against 70 for fixed Light; on samples labelled Heavy, fixed Light scores 55 against 76 for fixed Heavy; PRISM stays within two points of the best on both (69 and 74).

Each axis supplies what the other cannot. With an adaptive budget, adding structure raises KRIS by 10.8 points (62.4 → 73.2) and GEdit by 0.5 (6.50 → 7.00). With structure, adding adaptivity raises GEdit by 0.63 (6.37 → 7.00) at no cost on KRIS. Only the configuration with both is strong on both benchmarks, which is the claim of Section 3.2. The gain of structure is not bought with more reasoning: PRISM generates no more tokens per edit than the adaptive free-form variant (22, 60 and 77 tokens for Light, Moderate and Heavy edits, against 30, 52 and 90).

AuthorsOpenReview ↗

Response 1, Part 3


W3. Do the lenses matter, or could the model generate actions directly?

We add this baseline, evaluated on edited images. The direct-actions variant removes lens proposal and execution and generates the tailored actions directly from the instruction and image. It is otherwise trained exactly as PRISM, with the same rewards, GRPO recipe and REVISE-50K data.

Table R3. Direct actions vs PRISM on edited images: GEdit, and KRIS overall and per category.

VariantGEditKRISAttribute Percep.Spatial Percep.Temporal Pred.Social Sci.Natural Sci.Logical Reas.Instr. Decomp.
Direct actions6.9370.5780.1565.7358.4572.7370.3262.6870.61
PRISM7.0073.2482.9165.7358.9078.0773.4165.0571.94
Δ+0.07+2.67+2.760.00+0.45+5.34+3.09+2.37+1.33

Lenses matter where the edit must be reasoned out. On GEdit, whose instructions state the edit directly, the two variants are on par (7.00 vs 6.93, within the judge variability of 0.15 on GEdit, measured over four repeated evaluations). On KRIS, lenses add 2.7 points. The two largest gains are in categories whose edit must be derived from world knowledge, Social Science (+5.3) and Natural Science (+3.1), and Logical Reasoning gains 2.4. In Spatial Perception and Temporal Prediction, direct actions already suffice (+0.0 and +0.5).

Why lenses help. An action must state the edit in executable form. For an indirect instruction, that edit must first be derived: to "change the weather to December in Australia", the model must recall that December is summer there before an action can say "make the scene sunny". Without lenses, the model must derive the edit and state it in the same step. A lens provides a separate step for the derivation, so the actions only need to state its result. This is why lenses pay off on knowledge-heavy edits and not on direct ones, and why PRISM spends a single lens on Light edits. Table 4 shows the same pattern at the level of the reasoning: on each aspect-specific subset, the largest gain comes from the lens that targets that aspect.


W4. Human study

We add a blinded human preference study, reported below and in Section 5.3 (Table 5).

Human preference. Annotators compared the edited images side by side, without knowing which model produced each, on 200 GEdit and 200 KRIS samples.

Table R4. Blinded pairwise human preference (n = 400 per comparison; 95% confidence interval in brackets).

ComparisonPRISM preferredBaseline preferredTie
PRISM vs BAGEL-Think84% [80.4, 87.6]8%8%
PRISM vs Qwen-Image-Edit68% [63.4, 72.6]27%5%

Raters prefer PRISM over its own backbone in 84% of comparisons and over the larger Qwen-Image-Edit in 68% (both p < 0.001, two-sided sign test excluding ties).

Validating the judges we rely on. KRIS, RISE, PICA and ImgEdit are scored with each benchmark's released protocol and judge, which keeps them comparable with prior work. The only protocol we modify is GEdit: GEdit-R adds identity preservation as a third aspect (Appendix D.5). We validated this change against humans on 300 GEdit samples, where GEdit-R agrees more closely with the human ranking than the original protocol (Spearman ρ = 0.81 vs 0.64). Repeating the GPT-based evaluation four times, the judge variability is 0.15 on GEdit, 0.1 on ImgEdit, 1.0 on KRIS, 0.4 on RISE and 1.5 on PICA, 8 to 13 times smaller than PRISM's gains over BAGEL-Think on each benchmark. For the diagnostic of Table 1, judged by GPT-5.2, we repeated the evaluation with Gemini; the conclusion holds with a larger margin (length budgets 6.4 to 6.7, structured variant 7.9).

Comments on the paper

2 comments
AuthorsOpenReview ↗

Rebuttal Summary

We thank the reviewers for their constructive feedback. The reviewers found the observation that reasoning can hurt image editing interesting and worth studying (6sVo, XrGq), the diagnostics and ablations thorough (6sVo, XrGq, ih2s), and the gains on reasoning-intensive benchmarks strong (6sVo, ih2s).


Please note that we uploaded the revised PDF.

We ran every experiment the reviewers requested, and the new results strengthen the central claim of the paper:

  • Overthinking is a general phenomenon, not a property of BAGEL. On two further editors, reasoning again lowers direct-editing quality and raises reasoning-intensive quality: OmniGen2 (GEdit −0.46, KRIS +6.9) and Qwen-Image-Edit (GEdit −0.23, KRIS +4.5).
  • Adaptivity and structure are both needed (6sVo, ih2s). In the full grid, adaptive free-form reasoning improves neither benchmark (GEdit 6.50, KRIS 62.4), and structured reasoning at a fixed level is strong on one benchmark at a time (Light: 7.00 / 61.4; Heavy: 6.37 / 73.4). Only PRISM is strong on both (7.00 / 73.2).
  • The gains come from the reasoning itself (6sVo, XrGq). Trained on the same REVISE-50K samples without reasoning, BAGEL reaches 63.1 on KRIS, against 73.2 for PRISM. Generating the actions directly, without lenses, costs 2.7 points on KRIS and up to 5.3 on knowledge-intensive categories.
  • Better and cheaper than a verifier loop, and preferred by humans (6sVo, XrGq, ih2s). PRISM alone outperforms BAGEL with EditThinker on GEdit and KRIS (+0.78, +10.4) at less than half its measured latency. In blinded pairwise comparisons, humans prefer PRISM over BAGEL-Think in 84% of cases and over Qwen-Image-Edit in 68%.

We have also clarified the evaluation protocols and baseline entries (XrGq), reported the CFG diagnostic at fixed settings (ih2s), and stated the cost comparison precisely (6sVo, XrGq, ih2s). All changes are included in the revised paper, and table, figure, and appendix numbers in our responses refer to the revised version. We would greatly appreciate the reviewers' and the Area Chair's feedback on whether our responses address their concerns.

Dear Reviewer,

Please take a moment to review the authors' feedback and updates regarding your assigned paper. Your timely participation in the discussion phase is greatly appreciated.

Best regards,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.