acceptodds
Under review as a conference paper at ICLR 2027

DuplexBridge: Coordinating Understanding and Generation for Unified Image Editing

Abstract

Complex image editing requires faithful task interpretation and accurate execution, but a failed edit does not reveal which responsibility failed. We present DuplexBridge, which uses a unified model's own understanding capability to enhance prompts for its own generation capability, without an external understanding module. A source-grounded bridge specifies targets, final-state changes, and preservation constraints. Understanding is supervised on fidelity to the original request without observing the generated image; generation is supervised on execution of the condition it actually receives. Reciprocal trust checks selectively admit global task feedback to each role while retaining independently valid local supervision. We introduce CE-Bench, a 1,200-task benchmark measuring joint satisfaction of task-derived change and preservation requirements. Across three model configurations, the full system improves complete compliance over direct editors by 21.75–25.75 percentage points, with gains also observed on KRIS-Bench, RISEBench, ImgEdit, and GEdit-Bench. From a common supervised Wan-image initialization, gated joint training reaches 77.08% complete compliance, compared with 70.92% without gates and 75.00% with both gates retained but generation-local supervision referenced to the original task. These results support separating local responsibilities and selectively sharing task feedback when jointly adapting understanding and generation.

Primary area:
generative models
Keywords:
RL, Image editing, Unified model, Prompt enhancement

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_C4KFOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DuplexBridge, a framework that coordinates the understanding and generation capabilities of unified multimodal models for complex image editing. It introduces a structured bridge that explicitly specifies editing targets, required changes, and preservation constraints, enabling separate supervision of task understanding and editing execution. The framework further employs bidirectional trust gates to selectively route global task rewards to the understanding and generation components while retaining their respective local supervision. Experiments across three unified models demonstrate substantial improvements in complete editing compliance and performance on multiple image editing benchmarks.

Critical strengths
  1. The paper identifies a meaningful challenge in complex image editing: distinguishing failures in task understanding from failures in editing execution. The motivation is clearly presented, and the separation of these two responsibilities provides an intuitive perspective on understanding–generation coordination.

  2. The paper evaluates DuplexBridge across three unified multimodal models and several image editing benchmarks, demonstrating substantial improvements in overall editing performance. The inclusion of component ablations and human evaluations further strengthens the empirical evaluation.

Critical weaknesses
  1. Execution Trust relies on the average execution score, which may allow critical editing requirements to be overlooked. For example, if 8 out of 9 requirements are satisfied and one fails, the gate may still pass if the threshold is below the resulting average score, even when the failed requirement corresponds to the most critical editing operation.

  2. DuplexBridge's main contributions lie in its training supervision and feedback routing design, rather than in a fundamentally new model architecture or optimization algorithm. The novelty of the proposed method relative to prior works such as DARS[1] and DDA-Thinker[2] should be more clearly articulated.

  3. DuplexBridge aims to ensure that complex image editing tasks simultaneously achieve accurate edits and content preservation. However, Appendix M shows that its performance gains become smaller when a task involves a large number of requirements, and some editing capabilities even deteriorate. This suggests that, although the gating mechanism may improve the overall complete compliance rate on certain complex tasks, it does not consistently improve every aspect of editing performance and may even compromise some editing capabilities.

[1] DARS: Dual-Level Credit Assignment RL with Structured Reasoning for Instruction-Based Image Editing

[2] DDA-Thinker: Decoupled Dual-Atomic Reinforcement Learning for Reasoning-Driven Image Editing

Questions for authors

When both gates pass but the final edited image still fails to satisfy the task, the global reward may provide negative feedback to both U and G. However, the current gating mechanism cannot further distinguish the responsibility for such failures. Please see the Weaknesses section for the remaining concerns.

5 comments
AuthorsOpenReview ↗

Response to Reviewer C4KF (Part 1/2)

We thank the reviewer for recognizing our motivation and results across unified models. We address the concerns using new critical-requirement diagnoses, stricter-gate ablations, and three-seed experiments summarized in our General Response.

W1 & Q1. Execution Trust Reliability and Bidirectional Credit Assignment

We agree that mean Execution Trust does not guarantee that every critical requirement is met, and that passing both Partner Trust gates does not establish unique causal responsibility for a final failure. DuplexBridge instead separates role-local supervision and selectively admits global task feedback; the mean score is neither task success nor a causal responsibility label.

(1) Distinct supervision references. U-local evaluates prompt fidelity of the predicted condition C^\hat C against the original task (X,u)(X,u) without viewing the edited image. G-local evaluates execution fidelity and local quality against the condition actually received by G, (X,C^)(X,\hat C), rather than the original instruction. The global reward assesses task fulfillment, preservation, and visual quality under (X,u)(X,u), using fixed reference edits. Prompt Trust controls G-global eligibility; Execution Trust controls U-global eligibility. Neither gate disables independently valid local supervision. Thus, mean Execution Trust screens cross-role feedback for reliable execution evidence; it does not determine complete task success.

(2) Empirical evidence on mean gating and critical failures. In the original Appendix H, Table 25, applying selectors to the same 6,000 frozen rollouts reduces the unexecuted-prompt error rate among U-global-eligible samples from 27.41% to 4.60%, while retaining 64.80% U-global coverage. This directly measures error prevalence in selected feedback, not recall over all critical failures.

To probe the reviewer's "8 of 9 requirements pass" example, we additionally annotated 200 challenging training rollouts that passed the Mean Gate despite at least one failed execution-check group (General Response, Tables C3–C4). Five human annotators achieved 93.5% unanimous agreement on critical-requirement sets. The automatic rubric selector achieved 61.7% requirement-level micro-F1 and 34.0% exact-set match. Only 8/200 (4.0%) of these selected, mean-gate-accepted cases exhibited a human-confirmed critical failure. A failed requirement group therefore need not imply a failed critical edit. Co-occurring execution errors could lower mean scores, but these accepted-only cases cannot establish the overall rejection rate of critical failures.

An additional Rubric Critical Hard Gate rejected 68/200 of these cases; only 3 had human-confirmed critical failures. Critical-failure rates among rejected and retained cases were 4.4% and 3.8%, respectively. The added gate thus excluded many candidates without confirmed critical failures and showed limited selectivity.

We also trained two stricter alternatives (General Response, Table C5): Mean + Rubric Critical Hard Gate requires both the mean threshold and all selected critical checks; All-pass Gate requires every execution group to pass. Their CE-Bench CCRs were 74.33% and 68.92%, respectively, versus 77.08% for Mean Gate and 70.92% for Ungated Joint RL. Thus, stricter eligibility does not necessarily improve joint learning. Mean Execution Trust offers a more favorable observed trade-off between removing unreliable global feedback and retaining useful learning signals.

(3) What if both gates pass but the final task fails? Passing both gates means that global losses are eligible, not that the edit succeeds or that U and G receive an identical failure attribution. As specified in Section 4.4 and Appendices E.4–E.5: (i) U-local and G-local remain distinct, task- and received-condition-based objectives; (ii) U-local, U-global, G-local, and G-global normalize group-relative advantages on separate eligible sets. The global RM provides continuous relative preference scores, not binary success labels, so a low absolute score need not yield the same negative advantage for both roles; and (iii) updates alternate, freezing G during U updates and U during G updates. Thus, the roles follow distinct sampling groups, losses, and gradients.

This reduces unreliable cross-role updates without claiming counterfactual causal attribution. Imperfect automatic checks can leave residual ambiguity; the goal is an empirically effective conditional-feedback objective, consistent with Appendix E.4. We will clarify this scope and the separate optimization mechanisms in the revision.

AuthorsOpenReview ↗

Response to Reviewer C4KF (Part 2/2)

W2. Distinction from DARS and DDA-Thinker

We appreciate this request. Our novelty lies in role-local supervision references and global-feedback eligibility when jointly adapting a unified model's native U/G roles, not a new backbone or GRPO optimizer (General Response, Table C1).

Research setting. DDA-Thinker [2] and DARS [1] use separate Thinker–Editor and Planner–Renderer components, respectively. DuplexBridge uses the unified model's own understanding and generation roles through native prompt enhancement (PE), without an external planner.

Optimization direction. DDA-Thinker freezes its editor and optimizes the thinker with Cognitive/Visual Atomic Rewards. DARS and DuplexBridge both optimize planning/understanding and generation, but differ in how feedback is assigned.

Credit assignment and global-feedback use. DARS estimates relative planner/renderer update weights from between-plan and within-plan reward variability, alongside planner prefix gating. DuplexBridge instead evaluates original-task U-local and received-condition G-local against different references, then uses reciprocal Partner Trust to determine eligibility for additional global losses. It reuses local checks without an additional variance-attribution estimator, retaining valid local losses when global feedback is withheld. Accordingly, DARS primarily weights module updates using reward variability, whereas DuplexBridge decides when task-level feedback is sufficiently grounded to update each native role. We do not claim that structured planning, modular rewards, or gating are themselves new.

Controlled evidence (General Response, Table C2). Across three seeds on Wan-Image-7B, Full exceeds DARS-style Routing by 3.39 pp (95% CI: [2.06, 4.72] pp), Original-task G-local by 2.19 pp, Ungated Joint RL by 6.64 pp, and Group-matched Random Gate by 4.97 pp. In the matched Wan-Image setup, DARS-style Routing requires 1,227 GPU-hours to convergence, versus 416 GPU-hours for DuplexBridge (−66.1%). This is a mechanism transplant, not a reproduction of DARS's native complete system; we do not attribute the entire cost gap to variance estimation alone. We will make these differences and their controlled evidence more explicit in the revision.

W3. High Requirement Load and Editing Evaluation

Requirement load is not equivalent to task complexity, and component-wise all-pass metrics differ from Complete Compliance Rate (CCR).

In Appendix M, requirement load is defined as K=Kchange+KpreserveK=K_{\mathrm{change}}+K_{\mathrm{preserve}}, the total number of change and preservation requirement groups that must jointly hold. Larger KK means more obligations, not necessarily more editing operations, nor does it capture target grounding, visual reasoning, operation dependencies, or scene relations. Category composition also differs across KK strata. Thus KK is not a direct measure of overall difficulty.

Change All-pass and Preservation All-pass diagnose which requirement type is satisfied; CCR requires both to pass for the same task. Completing every requested change while damaging protected source content is not a successful edit. Hence the two component metrics help diagnose errors, while CCR directly measures the paper's primary objective: complete instruction compliance.

The original Appendix M, Tables 31–32, reports both component and joint outcomes. Across all 1,200 tasks, Full versus Ungated improves Change All-pass from 80.42% to 83.50%, Preservation All-pass from 76.75% to 81.25%, and CCR from 70.92% to 77.08% (+6.17 pp). For the 180 tasks with K=9K=9, Change All-pass declines from 73.33% to 60.56%, whereas Preservation All-pass and CCR increase from 49.44% to 52.78% (+3.33 pp CCR point estimate). While individual diagnostic metrics may fluctuate under extreme loads, the central goal is joint satisfaction of changes and preservation, not monotonic gains in each marginal measure. We will clarify these definitions and transparently report this load-specific behavior in the revision.

We thank the reviewer for these constructive concerns. The critical-failure analysis, stricter-gate ablations, and three-seed controls further support the proposed supervision and feedback design. Human CCR evaluation on all three backbones also favors Full over Ungated (General Response, Table C6).

References

[1] Cao et al. DARS: Dual-Level Credit Assignment RL with Structured Reasoning for Instruction-Based Image Editing. 2026.

[2] Yang et al. DDA-Thinker: Decoupled Dual-Atomic Reinforcement Learning for Reasoning-Driven Image Editing. 2026.

AuthorsOpenReview ↗

Revision Update for Reviewer C4KF

We thank the reviewer for highlighting the limitations of mean-based execution trust, the need to distinguish DuplexBridge from prior credit-assignment methods, and the challenges of high-requirement-load editing. We have revised the manuscript to address these concerns through additional analysis and controlled experiments.

First, Section 6.3 (pp. 8–9) and Appendix S.2 (Tables 40–42) now provide a human analysis of 200 deliberately selected mean-gate-accepted training rollouts and compare alternative execution gates. Five annotators achieved 93.5% unanimous critical-requirement agreement, while the automatic selector obtained 61.7% micro-F1 and 34.0% exact-set agreement. The new main-text Table 4 shows that Mean Gate achieves 77.08% CCR, compared with 74.33% for Rubric Critical Hard Gate and 68.92% for All-pass Gate. These experiments demonstrate the empirical tradeoff between filtering unreliable feedback and retaining learning signals.

Second, Section 2 and Appendix O.2 (Table 38) clarify that, unlike DARS's variance-based module weighting, DuplexBridge combines original-task U-local supervision, received-condition G-local supervision, and reciprocal global-feedback eligibility. Section 6.2 (Table 3) adds three-seed comparisons supporting this distinction.

Third, we explicitly report the high-load results in Section 6.3 and Table 4. At K=9K=9, the CCR difference interval includes zero, while Change All-pass declines from 73.33% to 60.56%. We do not claim uniform improvement across all editing requirements.

Regarding the question of ambiguous failures when both gates pass, Sections 4.3–4.4 and Appendix E.4 clarify that Partner Trust selects eligible global feedback; it does not uniquely identify counterfactual responsibility for every failure. The revised Limitations (Appendix R) further discusses improving task-critical requirement identification.

We hope these revisions make both the contribution and the empirical boundaries of the method clearer.

Reviewer_vhd1OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents a method for image editing in two stages using unified understanding and generation models: first, build a structured representation from the prompt instruction and the input image using an understanding component, and then generate the target image from the structured representation using a generation component. The authors show that this improves prompt adherence compared to baseline direct editing methods on several benchmarks.

Critical strengths

The two-step approach seems to be more interpretable than the baseline method that directly outputs the target image. The authors have also shown significant improvements compared to baseline methods in compliance rate. Overall, I find the method interesting and the empirical results strong.

Critical weaknesses

I find this paper very difficult to follow – many notations were introduced without explanation, and a lot of the technical details are either missing or buried far away from where they are needed. For example, the training procedure is never clearly explained. There are “structured supervised initialization”, SFT, and RL, but their purposes and exact setup are never mentioned and are only introduced with a single line in the experimental setup. In Section 3.2, the “source resolver” and “compiler” are introduced. The source resolver is only described as “binds targets in X” without mentioning what binding does, how it is done, or what the purpose of this component is. The compiler is only described as what it cannot do, and there is no explanation of what it even is. Similarly, “partner trust” first appears in the first paragraph of Section 4, with a vague description. Granted, I did not read the appendix in detail because it is more than 30 pages and very hard to parse. Overall, the writing style aligns well with an LLM.

Another weakness is that although the proposed method starts with a unified model, its treatment of adding disjoint role-specific parameters might hurt this unification. There is no discussion of the costs (time and memory) of added parameters, or whether the fine-tuned model still retains its original capabilities in the main text, as far as I can tell.

Questions for authors

I suggest that the authors clean up the notation and clarify all the critical concepts in the paper.

8 comments
AuthorsOpenReview ↗

Response to Reviewer vhd1 (Part 1/2)

We thank the reviewer for recognizing DuplexBridge's interpretability and editing gains. We clarify the training stages and components below and will move these explanations into the main text.

W1. Training Procedure and Core Components

1. Native PE, structured SFT, and RL. "Structured supervised initialization" is the SFT warm-start B0B_0, not a separate stage.

  • Native prompt enhancement (PE): The unified model's understanding role U interprets the source image and instruction, then passes a public CE-v1 bridge to its own generation role G. This is the native inference path, not an external planner or training loss.
  • Structured SFT (B0B_0): An offline teacher supplies responses containing six private interpretation fields and a public CE-v1 JSON contract. The complete validated response supervises U, including private fields and JSON; G retains its initial parameters. This defines B0B_0.
  • Alternating U/G RL: Every RL variant starts from B0B_0. During a U-update, G is frozen: U samples candidate bridges and downstream images, learning from prompt fidelity and eligible global feedback. During a G-update, U is frozen: G learns from execution/quality and eligible global feedback on images from a fixed bridge.

U-only, G-only, local-only, ungated, and Full are separately trained controls, not consecutive stages of a single run. On Wan-Image-7B, CE-Bench CCR is 53.83% (direct editing), 56.67% (native PE), 59.42% (structured SFT, B0B_0), and 77.08% (Full joint RL; original Appendix H, Table 23). We will move these stage definitions and update pseudocode into the main text.

2. Source Resolver and field-to-check Compiler (Section 3.2). The Source Resolver identifies the source-image objects or regions named in the bridge; the Compiler converts the received bridge into checks that can be evaluated on the edited image.

  • The Source Resolver reads the source image XX and public contract C^\hat C. It links each target description to a source region or entity and records its type, members, and resolution status. Ambiguous targets may be rejected or left unresolved. It does not inspect the edited image.
  • The Compiler uses fixed templates for scope.targets, ops.from/to, scope.edit/lock, ops.keep/avoid, and preserve/avoid. These produce source-state prerequisites, required output changes, and preservation checks. Template selection is deterministic; source grounding and visual judgments are semantic. Each group passes only when all applicable subchecks pass, and duplicate obligations count once.

Clock example (Figure 2). The digital clock shows 14:35. U adds 90 minutes and specifies 16:05 (4:05 on the analog clock). The Resolver links the analog and digital clocks to different source-image objects. The Compiler then checks that the same analog clock changes from 10:20 to 4:05, while the digital clock and the rest of the protected scene stay unchanged. This yields one change group and separate preservation groups.

G and the Compiler read the same C^\hat C; neither the original instruction nor U's private reasoning can supply missing requirements to the Compiler. G-local therefore measures how well G followed the received condition. We will move this example and Appendix A.3's definitions to Section 3.2.

3. Partner Trust (Section 4.3). Partner Trust is a rule for admitting the existing global task reward to a role's loss; it is not another reward model. Equation (8) defines:

e_U^{\mathrm{glob}} = b_F b_E \mathbf{1}[R_{\mathrm{exec}} \geq \tau_e].$$ Here $e_G^{\mathrm{glob}}$ and $e_U^{\mathrm{glob}}$ are the G-global and U-global eligibility masks. $c_{\mathrm{audit}}$ indicates an available, confident prompt-fidelity check against a verified reference; $v_C$ indicates that the bridge passes the format, semantic-fidelity, clarity, and source-binding checks. $b_F$ means that the global judgment is usable, and $b_E$ that execution checking is reliable and resolvable. $R_{\mathrm{exec}}$ is the mean execution-group score, $\tau_e$ is its threshold, and $\mathbf{1}[\cdot]$ is the pass/fail indicator. **Prompt Trust gates G-global; Execution Trust gates U-global.** Execution-check scores also supervise G-local, but for U-global they determine *eligibility only*: the global reward value comes from a separate frozen pairwise RM. Closing a global gate leaves valid local learning intact. Each of the four losses normalizes advantages over its own eligible samples, without inserting excluded samples as zero rewards. We will bring these definitions into Section 4.
AuthorsOpenReview ↗

Response to Reviewer vhd1 (Part 2/2)

W2. Role-Specific Parameters and Native Capabilities

1. Role-specific updates do not split the unified model into separate models. In Section 3.1, θ\theta and ϕ\phi denote the existing parameters updated for the understanding role U and generation role G; ψ\psi denotes pretrained parameters kept fixed, potentially including a shared backbone. These symbols specify which parameters are updated, not new external understanding/generation networks. During a U-update, only U's active parameter set changes and G stays fixed; the reverse holds during a G-update. Both roles remain part of the same unified model and retain its native U-to-G inference interface.

No additional architectural modules or newly added model parameters are used in our BAGEL, Qwen-Image-Edit-2511, and Wan-Image-7B configurations. We will make the distinction between existing role-specific parameter subsets, shared frozen parameters, and training/checkpoint storage explicit in Section 3.1. We also test the edited checkpoint's original capabilities directly below.

2. Inference time and memory. We measured Wan-Image-7B under matched inference paths:

Table R1. Parameter and inference overhead on Wan-Image-7B.

MetricOriginal (CoT)DuplexBridge Full
Newly added model parameters--0
Direct editing latency (s)48.248.3
Native PE (U-to-G) latency (s)57.456.3
Peak inference GPU memory (GB)6565

For the same inference path, DuplexBridge has comparable latency and peak memory to the original model. The native PE path itself takes 56.3 s, compared with 48.3 s for direct editing on the Full checkpoint: the 8.0-s difference is the cost of performing the additional U-to-G step, not of Partner Trust. Partner Trust is applied during training, not inference. We will report the timing setup and distinguish inference costs from training and judge/RM costs.

3. Retention of understanding and generation capabilities. We evaluated the image-editing-trained Full checkpoint directly, without additional understanding or text-to-image training. This is different from the task-specific DuplexBridge-T2I retraining in Appendix L.

Table R2. Native capability retention on Wan-Image-7B.

CapabilityBenchmarkOriginalFullFull - Original
UnderstandingMMMU58.058.8+0.8
UnderstandingMathVista73.573.1-0.4
UnderstandingMMBench82.782.8+0.1
GenerationGenEval2 AM79.479.3-0.1
GenerationGenEval2 GM32.233.2+1.0
GenerationDPG87.788.6+0.9

Across these six benchmarks, the largest observed decrease is only 0.4 points, while the editing checkpoint improves CE-Bench CCR from 53.83% to 77.08%. Thus, the reported editing improvement comes with largely retained performance on the measured understanding and generation tasks.

Planned Revisions

We will rewrite Sections 3-5 so that readers can follow the method without reconstructing it from the appendix: (i) show the native PE, B0B_0, and alternating-RL stages with their objectives and frozen roles; (ii) explain the Resolver and Compiler through the clock example and explicit inputs/outputs; (iii) define Partner Trust alongside the four loss components and their masks; and (iv) report parameter isolation, inference costs, and capability retention in the main text. We will shorten abstract descriptions, introduce symbols where they first appear, and retain the detailed derivations in the appendix. This implements the plan outlined in General Response C5.

AuthorsOpenReview ↗

Revision Update for Reviewer vhd1

We thank the reviewer again for the detailed suggestions concerning technical clarity, implementation, and preservation of the unified model's native capabilities. We have revised the main manuscript to address these concerns directly, rather than relying on explanations scattered throughout the appendix.

Specifically, Sections 3.1–3.2 (pp. 3–4) now clarify the existing role-specific parameter sets and explicitly describe the inputs, outputs, and operations of the Source Resolver and Checklist Compiler, illustrated by the clock-editing example in Figure 2. Sections 4.3–4.4 (pp. 5–6) define Partner Trust and its eligibility masks, distinguish local rewards from selectively admitted global feedback, and explain the alternating U/G updates. The revised training description makes clear that structured supervised initialization is the SFT warm-start B0B_0, not a separate additional stage. Section 6.2 further reports the progression from native PE through B0B_0 to joint RL.

To address the concern about unification and model overhead, Section 6.6 (p. 10, Table 6) now reports training resources and original-capability retention directly in the main text. Appendix S.4–S.5 (Tables 44–46) provides inference latency, peak GPU memory, and six understanding/generation benchmark results evaluated on the editing-trained checkpoint. The reported configurations introduce no additional model parameters, and the largest observed decrease across these six benchmarks is 0.4 points.

We have also simplified notation and moved essential operational definitions closer to their first appearance. We hope these revisions directly resolve the principal presentation and implementation concerns raised in the review.

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

This paper studies complex instruction-based image editing with unified multimodal models, where the model's own understanding role rewrites the instruction for its own generation role. The understanding role emits a structured JSON bridge specifying targets, operations, and preservation constraints, and the generation role edits from this bridge alone. Joint RL gives each role a local reward: prompt fidelity for understanding, and execution of the received bridge for generation. Reciprocal trust gates decide when a shared global reward from a distilled pairwise reward model is also applied. The paper introduces CE-Bench, a 1,200-task benchmark requiring all change and preservation requirements to pass, and reports large gains on three models, plus public benchmarks, ablations, and human evaluation.

Critical strengths
  1. Important, well-posed problem. When understanding and generation are trained jointly, a failed edit does not reveal which role is responsible. The paper isolates this credit-assignment problem clearly and makes it concrete.

  2. Principled and native design. Each supervision signal is defined by what it observes, so the two roles are judged on their own responsibilities. The unified model enhances prompts for itself without an external planner.

  3. Solid empirical study. Results hold across several recent unified models and public benchmarks, with controlled ablations from a shared initialization, careful human evaluation, and a benchmark that stresses complete compliance with change and preservation requirements.

Critical weaknesses
  1. Evaluation aligned with training. The headline benchmark is self-built and scored with the same requirement checks and aggregation rule as the training reward, so the gains may partly reflect fitting the evaluation protocol. Gains on general editing benchmarks are much smaller.

  2. Incremental and weakly supported gains. Joint planner-renderer training with checklists and gating already exists. The improvement over the closest alternative designs is small, ablations are single runs without variance, and the semantic gate is confounded with extra human-verified supervision.

  3. Design does not fully match the motivation. The execution gate averages group scores, so edits that violate a protected region can still pass, which is exactly the motivating failure. Gains are weakest on the most complex edits, and humans see no improvement in preservation or quality.

Questions for authors
  1. Training Cost. What are the training compute, the number of judge and reward-model calls, and the number of verified reference bridges used, compared with ungated joint RL?

  2. Judge overfitting. Can you evaluate CE-Bench with an independent judge that is unrelated to the training checks, or extend human evaluation to all three models?

5 comments
AuthorsOpenReview ↗

Response to Reviewer NuPQ (Part 1/3)

We thank the reviewer for recognizing DuplexBridge's motivation, native unified-model design, and results across backbones. We address evaluation independence, novelty and robustness, supervision fairness, computational cost, and reliability on complex edits with additional three-seed ablations, convergence-cost accounting, human evaluation across three backbones, and critical-failure diagnostics.

W1 & Q2: CE-Bench Independence, Motivation, and Cross-Backbone Human Verification

First, CE-Bench uses an evaluator from a different model family and independently designed rubrics, rather than reusing the training reward model or its scoring rules. Training-time local semantic checks use Qwen3.5-27B, while global task feedback is provided by a frozen pairwise RM built on Qwen3.5-27B. CE-Bench instead uses Gemini 3.8 Flash to judge independently specified original-task requirements, requiring all change and preservation groups to pass.

Their supervision targets and scoring roles are distinct:

  • U-local / G-local: Evaluate fidelity to the original task and execution of the condition actually received by G, respectively.
  • Execution Trust: The mean execution-check score determines U-global eligibility. Execution checks do not replace the global reward, although their change/invariant scores contribute to G-local supervision.
  • Global reward: A frozen pairwise RM supplies continuous preference feedback under the original instruction against fixed baseline edits.
  • CE-Bench: Independently specifies task-derived requirements without reading the predicted bridge and measures complete compliance using all-pass CCR.

Thus, local execution supervision, partner-gated global feedback, and final CE-Bench evaluation are not identical reward signals. Both training and evaluation employ fine-grained requirement verification, but differ in model family, rubric source, and supervision target.

Second, we have extended human evaluation to all three unified-model backbones, directly addressing the requested cross-model verification (General Response, Table C6). Human CCR improves from Ungated to Full on BAGEL: 68.37% to 75.00% (+6.63 pp); Qwen-2511: 73.98% to 80.10% (+6.12 pp); and Wan-Image-7B: 69.39% to 76.14% (+6.75 pp). Human–automatic agreement for Full is 94.39%, 95.41%, and 93.40%, respectively; the absolute human–automatic CCR gap is 0.51 pp for each Full configuration and at most 1.53 pp across all nine variants.

These cross-backbone CCR differences are descriptive because each variant is evaluated on its own human-decidable subset. For a matched comparison, the original Section 6.5 and Appendix N evaluate 194 commonly decidable Wan-Image tasks: human CCR improves from 69.07% to 75.77% (+6.70 pp, 95% CI [1.03, 12.37]); automatic CCR improves by +6.19 pp (95% CI [1.55, 10.82]); and the automatic-minus-human gain bias is -0.52 pp (95% CI [-3.61, 2.58]). These results provide cross-backbone human support for the reported gains and reduce concerns about direct training-judge reuse.

Third, CE-Bench addresses our specific research objective. DuplexBridge seeks to leverage the native understanding capability of unified multimodal models to improve edits requiring tightly coupled task interpretation, reasoning, target grounding, and execution. General-purpose editing benchmarks often emphasize more direct visual modifications or do not jointly stress these requirements. CE-Bench therefore covers source-grounded editing, text and structured content, reasoning-grounded editing, compositional editing, and reference-conditioned multi-image editing, requiring complete satisfaction of both change and preservation groups.

CE-Bench is our primary evaluation of understanding-intensive complete compliance; public benchmarks provide complementary generalization evidence. On Wan-Image, KRIS increases from 76.27 to 80.13, RISE from 26.11 to 29.17, ImgEdit from 4.34 to 4.36, and GEdit EN from 7.58 to 7.74. Different gain magnitudes are consistent with differing evaluation emphases, while the public benchmarks indicate broader transfer beyond CE-Bench. We will clarify this benchmark motivation, evaluator/rubric independence, and the new human results in the revision.

AuthorsOpenReview ↗

Response to Reviewer NuPQ (Part 2/3)

W2: Distinctions from Joint RL, Checklist/Gating Methods, and Robustness

We agree that joint Planner–Renderer RL, fine-grained checklists, and gating have precedents. Our contribution is not any one of these components in isolation, but a responsibility-aware learning rule for a unified model's own U/G co-adaptation: what each role should be supervised for, and when the shared global outcome is eligible to update it.

(1) Research setting and credit assignment. DDA-Thinker [2] optimizes a thinker using Cognitive/Visual Atomic Rewards while freezing the separate editor. DARS [1] jointly optimizes a separate planner and renderer through post-render planner diagnostics, prefix gating, and update weights based on between-plan/within-plan reward variability. Although DARS and DuplexBridge both address joint-learning credit assignment, DuplexBridge studies native understanding and generation roles within one unified model and uses different supervision references and feedback-eligibility decisions.

Specifically, original-task U-local judges whether U faithfully represents the original request without inspecting the rendered image. Received-condition G-local judges whether G executed the structured condition it actually received, rather than assigning to G requirements that U omitted. Reciprocal Partner Trust admits the extra global loss to G through Prompt Trust and to U through Execution Trust, while retaining independently valid local losses when a global gate closes.

Accordingly, our mechanism is not merely a different weight on the same terminal reward: it separates the information boundaries of local responsibility and selectively admits cross-role global feedback. Unlike DARS's variance-based module weighting over multiple plans and generated outputs, partner-trust eligibility reuses local checks without requiring an additional between-/within-plan variance-attribution estimate. In a common Wan-Image setting, the DARS-style Routing transplant requires 1,227 GPU-hours to convergence versus 416 GPU-hours for DuplexBridge (-66.1%). This is neither a comparison to the complete original DARS system nor an isolation of variance-estimation overhead alone. See General Response C1 and original Appendix O.2.

(2) Three-seed mechanism evidence. We repeated controlled ablations from the same B0 initialization and supervision configuration (General Response, Table C2). Full achieves 77.14% mean CCR (SD: 0.10 pp), with mean gains over:

  • Ungated Joint RL: +6.64 pp (95% CI [5.62, 7.66]).
  • Group-matched Random Gate: +4.97 pp (95% CI [4.37, 5.57]).
  • DARS-style Routing: +3.39 pp (95% CI [2.06, 4.72]).
  • Original-task G-local: +2.19 pp (95% CI [1.88, 2.51]).

The Original-task G-local control retains reciprocal gates while changing G's local supervision reference. Ungated and Full retain the same local rewards and global RM but differ in semantic global-feedback eligibility. These controls therefore support separate incremental benefits from received-condition execution supervision and semantic feedback selection, rather than attributing all full-system gains over direct editing to the gates. The reported paired-difference intervals all exclude zero across the three training seeds.

(3) No extra verified-reference supervision. All 5,000 training tasks have a verified reference contract, C∗C^*, and Ungated Joint RL and Full use the same task pool and verified references (100% coverage). Controlled variants also share B0 initialization, local checks, the global RM, and reference access. Full receives no exclusive verified reference annotations; gating changes the candidates and GRPO groups eligible for global loss, not the available supervision budget.

In summary, DuplexBridge combines responsibility-specific local objectives with reciprocal global-loss eligibility for native U/G co-adaptation, rather than simply combining established checklists and gating. The new three-seed, supervision-matched, and convergence-cost results support its incremental effectiveness. We will move the mechanism comparison from Appendix O.2 and key controlled evidence into the main text.

AuthorsOpenReview ↗

Response to Reviewer NuPQ (Part 3/3)

Q1: Training Cost, Judge/RM Calls, and Verified References

We measured cumulative resources for Ungated Joint RL and Full, each trained separately to its own convergence endpoint from the same initialization.

Table R-N1. Wan-Image-7B training cost and supervision budget.

MetricUngated Joint RLDuplexBridge Full
Policy training GPU-hours960416
Qwen3.5-27B judge calls192k78k
Global RM pairwise calls82k36k
Verified C∗C^* / coverage5k / 100%5k / 100%

Full uses 56.7% fewer policy-training GPU-hours, 59.4% fewer judge calls, and 56.1% fewer global-RM calls, with higher CCR. Both settings use the same judges, RM, training tasks, and verified references. These numbers measure cumulative resources to separate convergence endpoints, not matched step counts or total rollout budgets; policy-training GPU-hours do not include all external judge/RM service costs.

W3: Mean Execution Trust, Critical Failures, and Complex Editing

We agree that mean Execution Trust does not guarantee every critical requirement passes. Its purpose is to determine whether the global reward is eligible to update U, not to substitute the mean VQA score for final task utility. Execution checks also contribute to G-local execution supervision; for U-global, they act only through the eligibility gate. The frozen pairwise RM still evaluates task fulfillment, preservation, and visual quality under the original instruction.

Gate reliability. Original Appendix H, Table 25 reports that, on the same frozen pool of 6,000 rollouts, the unexecuted-prompt rate among U-global-eligible samples decreases from 27.41% to 4.60% with Execution Trust. We additionally analyzed 200 difficult training rollouts that passed the Mean Gate despite at least one failed execution group (General Response C3). Only 8/200 (4.0%) contained a human-confirmed critical failure. A rubric-based Critical Hard Gate additionally rejected 68 of these cases, but only 3 of the rejected cases contained confirmed critical failures; the other 65 lacked a confirmed critical failure. Thus, the stricter selector substantially reduces feedback coverage without clear critical-failure discrimination in this subset.

End-to-end RL: Mean Gate achieves 77.08% CCR, versus 74.33% with Mean + Rubric Critical Hard Gate and 68.92% with an All-pass Gate. In these tested variants, requiring stricter execution eligibility does not improve final editing performance; the mean gate provides a more favorable empirical balance between filtering unreliable global feedback and retaining useful supervision.

High requirement load and human preferences. Original Appendix M, Tables 31-32 show a +8.06-pp CCR gain at K=7 (95% CI [4.72, 11.39]). At K=9, CCR increases from 49.44% to 52.78% (+3.33 pp, 95% CI [-0.56, 7.22]); this interval includes zero. At the same load, change all-pass decreases from 73.33% to 60.56% (-12.78 pp, 95% CI [-17.78, -7.78]). We transparently report this specific regression to provide a complete picture of the empirical tradeoffs at extreme requirement loads; notably, the additional hard/all-pass gates do not remedy this, underscoring the inherent difficulty of this regime.

Original Appendix N, Table 36 reports human net preference for Full over Ungated: instruction following +17.77 pp (95% CI [8.63, 26.90]), preservation +0.51 pp ([-8.12, 9.14]), and visual quality -2.03 pp ([-11.68, 7.11]). Because both intervals span zero, these results demonstrate that DuplexBridge substantially improves complex instruction following (+17.77 pp) without causing statistically significant degradation in baseline visual quality or preservation. Thus, the supported conclusion is improved complete task compliance and human preference, even though fine-grained execution at extreme requirement loads remains an open challenge.

We thank the reviewer again. The new cross-backbone human evaluation, three-seed controls, matched supervision accounting, and gate diagnostics directly address reliability, incremental value, and evaluation fairness.

References

[1] Cao et al. DARS: Dual-Level Credit Assignment RL with Structured Reasoning for Instruction-Based Image Editing. 2026.

[2] Yang et al. DDA-Thinker: Decoupled Dual-Atomic Reinforcement Learning for Reasoning-Driven Image Editing. 2026.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General Response to All Reviewers (Part 3/3)

C4. Supervision Fairness, Evaluation Reliability, and Capability Retention

(1) Supervision fairness and compute. Ungated Joint RL and Full use the same 5,000 training tasks and access to verified reference bridges C∗C^* for all tasks (100% coverage). When independently trained to convergence, Full reduces policy-training compute from 960 to 416 GPU-hours (−56.7%), judge calls from 192k to 78k, and global RM pairwise calls from 82k to 36k. Details appear in our response to NuPQ.

(2) Independent evaluation and human validation across three backbones. Training-time local semantic judgments use Qwen3.5-27B, and the global RM is based on this model. CE-Bench instead uses Gemini 3.8 Flash, from a different model family, with evaluation rubrics independently constructed from the original tasks, rather than the model-predicted bridges or training-RM outputs.

To address concerns about evaluation-protocol bias and cross-backbone reliability, we extend human verification to all three unified-model backbones, comparing human CCR, automatic CCR, and per-output agreement.

Table C6. Three-backbone human verification of CE-Bench evaluation.

BackboneVariantNHuman CCR (%) ↑Auto CCR (%) ↑Agreement (%) ↑
BAGELDirect19648.4748.9890.31
Ungated19668.3768.8893.37
Full19675.0074.4994.39
Qwen-2511Direct19657.1458.1691.84
Ungated19673.9874.4994.39
Full19680.1079.5995.41
Wan-Image-7BDirect19553.3354.3688.72
Ungated19669.3970.9291.33
Full19776.1476.6593.40

NN counts human-decidable outputs per method. Human CCR, Auto CCR, and agreement are evaluated on the same decidable subset within each row; subsets can differ across methods.

Consistent human gains: Full achieves 75.00% / 80.10% / 76.14% human CCR on BAGEL / Qwen-2511 / Wan-Image-7B, versus 68.37% / 73.98% / 69.39% for Ungated. Small automatic–human gaps: For each Full model, the absolute Auto–Human CCR difference is 0.51 pp; the largest difference among all nine configurations is 1.53 pp. High instance-level agreement: Full's human–automatic agreements are 94.39% / 95.41% / 93.40%.

Together with independent evaluation models and rubric construction, these results support CE-Bench reliability and reduce concerns that gains arise solely from alignment with the training judge.

(3) Native capability retention. The image-editing-trained Wan-Image-7B Full checkpoint remains close to the original model on six non-target understanding and generation benchmarks, with a maximum observed decline of 0.4 points. See our response to vhd1.

C5. Planned Revisions for Clarity and Technical Presentation

We appreciate the reviewers' concerns about unclear training stages, insufficiently introduced components, and dense notation. We will:

  • Training pipeline: Distinguish native PE, structured SFT initialization (B0B_0), and alternating U/G RL, specifying each stage's objectives and supervision.
  • Core components: Define the inputs, outputs, and functions of the Source Resolver, Checklist Compiler, and Partner Trust at first mention in Sections 3.2–4, with the existing worked example.
  • Notation and organization: Remove redundant abstract phrasing, standardize symbols, and move essential algorithms, definitions, and controlled evidence from the appendix into the main text.
  • Implementation and capability: Clarify the role-specific parameter-update boundary, report compute and memory overhead, and present original-capability retention results.

Further details are provided in our response to Reviewer vhd1. These revisions aim to make the learning objectives, training procedure, and empirical evidence clearer and directly verifiable.

References

[1] Cao et al. DARS: Dual-Level Credit Assignment RL with Structured Reasoning for Instruction-Based Image Editing. 2026.

[2] Yang et al. DDA-Thinker: Decoupled Dual-Atomic Reinforcement Learning for Reasoning-Driven Image Editing. 2026.

AuthorsOpenReview ↗

General Response to All Reviewers (Part 2/3)

C3. Reliability of Mean Execution Trust

Reviewers NuPQ and C4KF question whether averaging execution-group scores can overlook an individual critical failure. We therefore examine residual critical errors and compare mean gating with stricter alternatives.

DuplexBridge does not use the mean VQA execution score as final task utility. The execution checks evaluate G's fidelity to the received condition C^\hat C, providing G-local supervision and determining U-global eligibility. A separate global RM evaluates final task fulfillment, preservation, and visual quality against the original instruction.

(1) Fixed-rollout diagnosis (original Appendix H, Table 25, p. 36). Applying the gates to the same pool of 6,000 frozen rollouts reduces the unexecuted-prompt error rate among U-global-eligible samples from 27.41% to 4.60%.

(2) Human identification of critical requirements. We deliberately selected 200 challenging training rollouts that passed the Mean Gate despite failing at least one execution-check group. Five human annotators labeled critical requirements, and we assessed an automatic rubric selector.

Table C3. Human agreement and critical-requirement identification.

Requirement LoadCasesHuman Unanimous Agreement ↑Rubric Micro-F1 ↑Exact Set Match ↑
K = 3/511094.5%71.7%46/110
K = 7/99092.2%52.3%22/90
Overall20093.5%61.7%68/200

Unanimous agreement requires all five annotators to select exactly the same critical-requirement set; F1 is requirement-level micro-F1.

Although human agreement is high, the automatic selector achieves only 34.0% exact-set matching, with lower alignment at higher requirement loads.

(3) Residual critical failures and additional filtering. We further require every automatically selected critical requirement to pass on these same 200 rollouts.

Table C4. Additional filtering by a Rubric Critical Hard Gate.

Human VerificationRubric AcceptedRubric RejectedTotal
Critical Requirement Failed538
No Critical Requirement Failed12765192
Total13268200
Critical Failure Rate3.8%4.4%4.0%

Only 8/200 (4.0%) of these deliberately selected, Mean-Gate-accepted boundary cases contain a human-confirmed critical failure. One possible explanation is that critical failures often co-occur with other execution errors and hence lower the mean score, making rejection more likely. The rubric hard gate additionally rejects 68 cases, of which only 3 exhibit human-confirmed critical failures; the other 65 (95.6%) do not. Thus the additional hard constraint excludes many potentially useful feedback samples without clearly improving critical-failure selectivity in this subset.

(4) End-to-end RL with stricter gates. We compare the additional critical-requirement constraint with an All-pass Gate, which requires every execution group to pass.

Table C5. Execution Trust gating ablations on Wan-Image-7B.

Training VariantCE-Bench CCR (%) ↑Δ vs. Mean (pp)K=9 Change All-pass (%) ↑K=9 Preservation All-pass (%) ↑K=9 CCR (%) ↑
Ungated Joint RL70.92−6.1773.3349.4449.44
DuplexBridge Mean Gate77.08—60.5652.7852.78
Mean + Rubric Critical Hard Gate74.33−2.7553.8950.5650.56
All-pass Gate68.92−8.1751.1146.6745.56

All scores are percentages; differences are pp. These are per-checkpoint endpoints, distinct from the three-seed results in Table C2.

The rubric and all-pass alternatives reduce overall CCR by 2.75 and 8.17 pp, respectively, without improving K=9 Change All-pass or CCR. Excessively strict acceptance may discard informative global-feedback groups, thereby weakening learning. Mean Execution Trust achieves a more favorable empirical balance between filtering unreliable feedback (4.60% error within U-global-eligible samples) and retaining usable supervision.

AuthorsOpenReview ↗

General Response to All Reviewers (Part 1/3)

We thank the reviewers for recognizing the motivation and empirical findings of DuplexBridge. To address concerns about novelty, robustness, feedback reliability, and technical clarity, we provide three-seed mechanism ablations, convergence-cost comparisons, human analyses of critical requirements, human evaluation across all three backbones, and original-capability retention results. We address shared concerns here and reviewer-specific questions separately.

C1. Efficient Credit Assignment within a Unified Understanding–Generation Model

Research setting. DDA-Thinker [2] and DARS [1] employ separate Thinker–Editor and Planner–Renderer components, respectively. DuplexBridge instead studies joint adaptation of the native understanding (U) and generation (G) roles of a single unified multimodal model, without an external planner.

Optimization. DDA-Thinker freezes the editor and updates only the thinker. DARS and DuplexBridge both study credit assignment when understanding/planning and generation are jointly optimized.

Credit assignment. DARS estimates module update weights from between-plan and within-plan reward variability using an M×KM\times K nested rollout bank (O(n2)O(n^2) samples when both dimensions scale as nn). DuplexBridge instead differentiates responsibilities through original-task U-local and received-condition G-local supervision, then uses reciprocal Partner Trust to determine eligibility for additional global losses. It does not require a separate cross-module reward-variance attribution estimator.

This distinction is also reflected in training efficiency: in the common Wan-Image setting, DARS-style Routing requires 1,227 GPU-hours to convergence versus 416 GPU-hours for DuplexBridge (−66.1%), while our three-seed CCR improves by 3.39 percentage points (pp). These costs compare the DARS-style mechanism transplanted into a shared Wan-Image configuration, not the original DARS end-to-end system.

Table C1. Research setting and credit-assignment mechanisms.

DimensionDDA-Thinker [2]DARS [1]DuplexBridge
Research settingSeparate Thinker–EditorSeparate Planner–RendererNative U/G in one unified model
OptimizationEditor frozen; thinker updatedJoint planner–renderer updatesJoint U/G adaptation
Credit assignmentCognitive/Visual Atomic RewardsBetween-/within-plan variability-based module weightingDistinct local references + Partner-Trust-gated global losses
Attribution costNo joint-module routingM×KM\times K nested rollout bankReuse local checks; no additional variance estimator

C2. Three-Seed Ablations: Consistent Gains from the Core Mechanisms

We conducted three-seed controlled ablations on Wan-Image-7B, using the same supervised initialization B0B_0, verified-reference access, local checks, and global RM across variants.

Table C2. Three-seed mechanism ablations on Wan-Image-7B.

Training VariantCCR Mean ± SD (%) ↑Δ vs. Full (pp)95% CI of Δ (pp)
Joint RL, Local-only67.83 ± 0.52−9.31[−10.83, −7.78]
Joint RL, Ungated70.50 ± 0.36−6.64[−7.66, −5.62]
Group-matched Random Gate72.17 ± 0.30−4.97[−5.57, −4.37]
DARS-style Routing73.75 ± 0.44−3.39[−4.72, −2.06]
Original-task G-local74.94 ± 0.17−2.19[−2.51, −1.88]
DuplexBridge (Full)77.14 ± 0.100.00—

Semantic selection. Full exceeds Ungated by 6.64 pp and Group-matched Random Gate by 4.97 pp, supporting the contribution of semantic global-feedback eligibility. Credit assignment. Full exceeds DARS-style Routing by 3.39 pp (95% CI for Full minus DARS-style: [2.06, 4.72] pp) with lower convergence compute. Local responsibility. Full exceeds Original-task G-local by 2.19 pp, supporting execution supervision against the condition actually received by G. All reported paired-difference CIs exclude zero, indicating consistent gains across these repeated runs.

AuthorsOpenReview ↗

Revision Summary

We thank all reviewers for their constructive feedback. We have substantially revised the manuscript to improve the clarity of the method, strengthen the controlled evidence, and directly address concerns about evaluation reliability, feedback selection, computational cost, and preservation of native model capabilities.

The principal revisions are summarized below.

1. Training procedure and technical definitions (Sections 3–5). We clarified the relationship between native prompt enhancement, structured SFT initialization (B0B_0), and alternating U/G RL. Sections 3.1–3.2 now explicitly define the Source Resolver and Checklist Compiler, including their inputs, outputs, and the clock-editing example in Figure 2. Sections 4.3–4.4 explain Partner Trust, the four loss components, eligibility masks, and alternating role updates.

2. Novelty and repeated-training evidence (Section 6.2; Table 3; Appendix S.1, Table 39). We clarified the distinction from DARS and DDA-Thinker in Section 2 and Appendix O.2 (Table 38). We also added three-seed experiments covering six training variants. DuplexBridge achieves 77.14 ± 0.10% CCR, exceeding Ungated Joint RL, Group-matched Random Gate, DARS-style Routing, and Original-task G-local by 6.64, 4.97, 3.39, and 2.19 percentage points, respectively.

3. Execution Trust and critical requirements (Section 6.3; Table 4; Appendix S.2, Tables 40–42). We added a human analysis of 200 selected training rollouts, including critical-requirement identification and automatic-rubric reliability. We further evaluated Mean Gate, Rubric Critical Hard Gate, and All-pass Gate, obtaining 77.08%, 74.33%, and 68.92% CCR, respectively. We explicitly report the remaining challenges at high requirement loads.

4. Evaluation independence and human verification (Sections 5 and 6.5; Table 5; Appendix S.3, Table 43). We clarified that CE-Bench uses Gemini 3.8 Flash and independently constructed original-task rubrics, distinct from the Qwen3.5-27B-based training judgments. Human verification now covers all three unified-model backbones, with Full human–automatic agreement between 93.40% and 95.41%.

5. Supervision fairness, computational cost, and capability retention (Section 6.6; Table 6; Appendix S.4–S.5, Tables 44–46). We confirmed identical access to 5,000 verified reference contracts for Ungated and Full. We added policy-training GPU-hours, judge/RM calls, inference latency and memory, and evaluations of six non-target understanding/generation metrics on the editing-trained checkpoint.

In total, the revised manuscript presents four new main-text evidence tables (Tables 3–6) and eight detailed supplementary tables (Tables 39–46). We also revised the Limitations in Appendix R to discuss critical-requirement identification and future extensions.

We hope these changes make the proposed mechanism, its empirical support, and its limitations substantially easier to evaluate directly from the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.