acceptodds
Under review as a conference paper at ICLR 2027

Counterfactual Visual On-Policy Distillation

18 views

Abstract

Visual on-policy distillation uses privileged evidence crops to improve a full-image student's fine-detail recognition. However, distilling teacher outputs on individual images does not explicitly distinguish how the target detail and surrounding content contribute to answer preferences. Our idea is to supplement this supervision with controlled detail changes, observing how teacher preferences respond while holding the question and surrounding scene as constant as possible. We propose CF-OPD, counterfactual visual on-policy distillation, which uses semantic replacement to construct verified image pairs with different correct answers. At each student-generated prefix, CF-OPD compares the two crop-conditioned teacher distributions, using their contrast to adjust a student-anchored distribution before mixing it with the current crop teacher. Independent rollouts on both images let each side learn from this paired supervision. We also introduce CF-Bench-X, an evaluation benchmark comprising 595 verified semantic image pairs. Its Flip rate measures the proportion of pairs answered correctly on both images; its Blind rate measures unchanged predictions despite a verified answer change. CF-OPD achieves the highest average accuracy across six visual benchmarks among the compared Qwen3.5-based methods at both scales. At 9B, it improves recognition and paired evaluation over independent visual OPD on the same inputs. Code and dataset will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
on-policy distillation, counterfactual visual supervision, fine-grained multimodal perception

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_3xuKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CF-OPD to improve fine-grained visual recognition by incorporating controlled semantic edits into on-policy distillation. It constructs verified image pairs through controlled semantic edits and uses teacher distribution contrasts at student-generated prefixes to guide training on both original and edited images. Experiments on six visual benchmarks and a paired-edit evaluation set report improvements in recognition accuracy and responses to edited visual details compared with independent OPD baselines.

Critical strengths
  1. A targeted approach to fine-grained visual supervision. The paper addresses the challenge of recognizing answer-determining visual details in on-policy distillation by incorporating teacher distribution contrasts from paired, semantically edited images into the distillation targets. This provides a concrete training approach for improving sensitivity to relevant visual changes, rather than relying solely on independent supervision for individual images.
  2. Empirical improvements on standard and paired-image evaluations. Experiments across six visual benchmarks and two model scales report improvements over independent OPD baselines. The additional paired-image evaluation, using Flip and Blind metrics, directly assesses whether model predictions respond appropriately to changes in answer-relevant visual details. These results provide empirical evidence of the method's practical potential beyond improvements in conventional visual recognition accuracy.
Critical weaknesses
  1. Insufficient isolation of the core contrast mechanism. The central design of CF-OPD is to construct distillation targets using teacher log-probability contrasts between original and edited image crops. However, the reported experiments do not clearly isolate the contribution of this contrast signal by removing it while keeping the paired training data, rollout configuration, and compute budget fixed. Consequently, it remains difficult to determine whether the reported improvements arise specifically from the proposed contrast mechanism or from other factors, such as additional edited data and two-sided training.
  2. Insufficient justification of the counterfactual formulation and paired-data validity. The paper constructs image pairs through controlled semantic edits and applies pixel-level checks, answer verification, and logical audits to filter low-quality examples. While these procedures provide useful quality control, they do not fully rule out unintended semantic changes or editing artifacts that the student could exploit as shortcuts. As a result, the current evidence does not fully establish whether the student learns the intended relationship between answer-relevant visual attributes and predictions, rather than relying on incidental cues introduced by the editing process.
Questions for authors
  1. Can the authors provide a controlled ablation that removes the teacher log-probability contrast term while keeping the paired training data, rollout configuration, and compute budget fixed? If such an experiment has already been conducted, what are the resulting Avg₆, Flip, and Blind scores? This result would directly affect my assessment of whether the proposed contrast mechanism independently contributes to the reported gains.
  2. Can the authors provide additional evidence that the constructed image pairs change the intended answer-relevant attribute while preserving other relevant semantic content? In particular, is there any independent human evaluation or quantitative audit of unintended semantic changes or editing artifacts? Such evidence would materially affect my assessment of whether the observed gains can be attributed to the intended paired supervision.
7 comments
AuthorsOpenReview ↗

Response to Reviewer 3xuK

Thank you for the careful assessment and for recognizing the targeted visual supervision and complementary recognition and paired evaluations. We address each mechanism and data-validity concern below. The revised manuscript with changes highlighted in blue is provided in the Supplementary Material.

AuthorsOpenReview ↗

Response to Reviewer 3xuK — W1: Isolating the teacher contrast from data and two-sided training

↑: higher is better; ↓: lower is better.

Thank you for emphasizing the controlled test of the contrast mechanism. We added a strict contrast-removal experiment that retains the paired data, rollout procedure, and optimization budget, and repeated it with three seeds.

1. Removing only the contrast. Full CF-OPD and α=0 start from the same initialization and use the same mixed training data. Both independently generate responses on the original and edited members of each pair. The current and counterpart crops are scored by the same EMA teacher at the current student prefix. In the control, we still execute both teacher-view computations but multiply the log-probability contrast by zero. λ=0.15, candidate support, normalization, tail handling, and all other target settings are unchanged. This preserves the teacher computation used to construct the target in both configurations.

2. Matched training components. Additional edited examples and two-sided training are present in both configurations. Matching these components enables a direct test of the contrast contribution. The independent-OPD baseline provides a complementary same-input comparison; the strict α=0 comparison is the one used to isolate the contrast while retaining the mixture and two teacher views.

3. Full results across six recognition benchmarks and CF-Bench-X. Each method uses seeds 42, 43, and 44. Entries are mean±sample SD; scores are percentages.

MetricCF-OPDα=0Independent OPD
V* ↑93.72±0.9190.23±0.6090.58±0.91
Zoom ↑62.72±0.9563.59±0.7163.55±0.43
HR-4K ↑86.96±0.3885.42±0.1484.67±0.40
HR-8K ↑83.67±1.0482.71±0.5282.96±0.85
MME-EN ↑72.96±0.0871.45±0.2471.63±0.35
MME-CN ↑70.26±0.4969.05±0.2069.66±0.30
Avg₆ ↑78.3798±0.120677.0729±0.096377.1738±0.3444
Flip ↑75.85±0.3971.76±1.1872.38±0.51
Blind ↓17.42±0.6819.78±1.2419.44±0.79

Statistics and differences are calculated before rounding.

Against α=0, CF-OPD improves mean Avg₆ by 1.31 points, Flip by 4.09 points, and Blind by 2.35 points in the favorable direction. All three seeds improve these aggregate metrics. Mean accuracy is higher on five standard benchmarks, supporting the gain in the six-benchmark average. The independent-OPD comparison gives +1.21 Avg₆, +3.47 Flip, and −2.02 Blind points. These results demonstrate improvements in both conventional recognition and correct responses to edited details.

4. Mechanistic interpretation. The counterpart crop supplies an alternative visual observation at the same question and student prefix. The contrast changes token preferences according to how the teacher responds to that observation. The controlled results support a contribution from this target component beyond the benefit of merely exposing the model to more images or training on both sides. At inference, the model still uses only the full image; no counterfactual image or crop teacher is required.

Locations. §3.3–3.4, §4.4 Table 3, and Appendix A. The revised manuscript with changes highlighted in blue is in the Supplementary Material.

AuthorsOpenReview ↗

Response to Reviewer 3xuK — W2: Counterfactual validity, semantic preservation, and shortcuts

↑: higher is better; ↓: lower is better.

Thank you for highlighting the validity of counterfactual supervision. We added human evaluation of intended answer changes, editing quality, and visual artifacts, together with concrete examples and controlled training results.

1. What “counterfactual” specifies here. A pair is designed to change an answer-relevant visual attribute while keeping the question fixed between its two members. The intended answers differ. The corresponding evidence crops are scored by the same teacher at the same student-generated prefix; on the edited side, the crop roles are exchanged. This construction supplies an explicit alternative visual attribute and its corresponding answer for paired supervision.

2. Why the automatic checks need empirical validation. Pixel/locality checks constrain where changes occur, blind answer reading checks the resulting image–answer relation, and the logic audit checks execution and consistency. A localized edit can still change the wrong object, introduce a cue, or alter another relevant attribute. We therefore manually evaluated all 5,190 accepted training and test pairs, costing US$467.

CriterionTraining (4,595)Test (595)
Intended answer change ↑4,227 (91.99%)559 (93.95%)
Editing problem ↓207 (4.50%)12 (2.02%)
Visible artifact ↓103 (2.24%)9 (1.51%)
Artifact may cue the answer ↓81 (1.76%)6 (1.01%)

The intended answer change is observed in 91.99% of training pairs and 93.95% of test pairs. The measurements separately identify unsuccessful/problematic edits, visible artifacts, and artifacts that may cue the answer. In the test set, visible artifacts occur in 1.51% of pairs and potentially answer-cueing artifacts in 1.01%.

3. Inspectable semantic-preservation failures. Appendix H shows the original image, editing instruction, edited image, and observed error for each example.

ExampleRequested editObserved failure
V* #145Move the shopping cart to the dog’s rightAn extra cart appears on the left; the intended spatial answer does not change
HR-4K #522Move the bird onto the car roofAn inset introduces a different car, changing the object referent
HR-8K #458Turn the foreground child toward the cameraAnother child is cut off and a bicycle disappears

These cases, confirmed in the human evaluation, directly show that an edit can fail to change the intended answer or can modify non-target semantic content. Successful examples in Appendix F cover color, text, material, pattern, identity, and state changes. The examples make the observed editing outcomes inspectable, and the full-pool counts quantify their prevalence.

4. Evidence about what the student learns. Under fixed paired inputs, two-sided rollouts, and teacher-mixture settings, adding the contrast increases mean Flip from 71.76 to 75.85, reduces Blind from 19.78 to 17.42, and raises Avg₆ from 77.0729 to 78.3798. The improvement in six-benchmark average recognition accompanies the paired gain. Both configurations use the same edited images, so the comparison directly demonstrates the additional contribution of the contrast. The human quality measurements and controlled objective experiment jointly support the effectiveness of the paired supervision.

For the studied multimodal fine-tuning use, we consider the measured residual error proportions acceptable when pairs supplement original training examples. For paired evaluation, we also consider the measured error rates acceptable in this setting.

Locations. §3.2–3.4, §4.4 Table 3, Appendix G Table 6, and Appendices F/H. The blue-marked revision is in the Supplementary Material.

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

This paper proposes CF-OPD, a method that improves visual recognition by comparing teacher predictions on original and edited images. The authors report better results on six visual benchmarks and introduce CF-Bench-X, a new benchmark containing 595 image pairs.

Critical strengths

Clear motivation: This paper presents a clear and reasonable idea for improving visual recognition by comparing original and edited images. The method is simple to understand and does not add extra inference costs.

The authors report promising results on six visual benchmarks using both 4B and 9B models. Their method achieves the highest average accuracy among the compared Qwen3.5-based methods.

The authors conduct useful ablation studies on different training settings, loss functions, and data configurations. These experiments help explain which settings improve performance.

Critical weaknesses

The authors do not fully demonstrate that the counterfactual contrast is responsible for the improvements. Their method changes both the contrast signal and the teacher–student mixture, but they do not test a version without contrast while keeping the mixture unchanged.

The authors do not clearly explain several important experimental settings. In particular, the difference between "OPD" and "Plain OPD" in Table 3 is unclear, even though their results differ significantly, and several training details are missing.

The authors show that performance strongly depends on the training data mixture. When training only on paired images, CF-OPD performs slightly worse than standard OPD on average accuracy, suggesting that the improvement may depend heavily on including original training data.

Questions for authors

What distinguishes “OPD” from “Plain OPD” in Table 3, particularly regarding their targets, losses, vocabulary-tail treatment, and optimization settings?

What is the independently measured error rate of the accepted training pairs and CF-Bench-X pairs?

8 comments
AuthorsOpenReview ↗

Response to Reviewer QCTw

Thank you for the constructive review and for recognizing the motivation, unchanged inference procedure, results at both model scales, and ablation studies. We respond to every weakness and question separately. The revised manuscript with changes highlighted in blue is provided in the Supplementary Material.

AuthorsOpenReview ↗

Response to Reviewer QCTw — W1: Contrast versus the teacher–student mixture

↑: higher is better; ↓: lower is better.

Thank you for identifying that the original comparison changes more than the contrast term. We added the requested version without contrast while keeping the teacher–student mixture unchanged, and repeated it with seeds 42, 43, and 44.

Experimental settings. We fix the teacher-mixture ratio at λ=0.15 for CF-OPD and α=0. These two configurations share the same initialization, training data, candidate support, normalization, tail handling, two-sided rollouts, optimization settings, and computation of both teacher views. We also include independent OPD with the same training inputs.

All six benchmarks and paired metrics. The table reports mean±sample SD over the same three seeds. Scores are percentages; higher is better except for Blind.

MetricCF-OPDα=0Independent OPD
V* ↑93.72±0.9190.23±0.6090.58±0.91
Zoom ↑62.72±0.9563.59±0.7163.55±0.43
HR-4K ↑86.96±0.3885.42±0.1484.67±0.40
HR-8K ↑83.67±1.0482.71±0.5282.96±0.85
MME-EN ↑72.96±0.0871.45±0.2471.63±0.35
MME-CN ↑70.26±0.4969.05±0.2069.66±0.30
Avg₆ ↑78.3798±0.120677.0729±0.096377.1738±0.3444
Flip ↑75.85±0.3971.76±1.1872.38±0.51
Blind ↓17.42±0.6819.78±1.2419.44±0.79

Statistics and differences are calculated before rounding.

Against α=0, CF-OPD improves mean Avg₆ by 1.31 points and Flip by 4.09 points, and reduces Blind by 2.35 points. Against independent OPD, the corresponding differences are +1.21, +3.47, and −2.02 points. Avg₆, Flip, and Blind improve in each seed against both controls. CF-OPD achieves the best mean accuracy among the three methods on five of the six standard benchmarks, as well as the best Avg₆, Flip, and Blind.

These results show that counterfactual contrast further improves average recognition accuracy and paired performance with the teacher-mixture ratio held fixed.

Locations. §3.3 Eq. (4), §4.4 Table 3, and Appendix A. The revised manuscript with changes highlighted in blue is in the Supplementary Material.

AuthorsOpenReview ↗

Response to Reviewer QCTw — W2: Clarifying OPD, Plain OPD, and their training recipes

↑: higher is better; ↓: lower is better.

Thank you for pointing out that the names were insufficiently explained. We have clarified the data-introduction procedures and distinguish them from the choice of distillation target.

What the names mean. Both recipes originate from Qwen3.5-9B. Plain OPD trains on original Vision-OPD examples and counterfactual-pair examples together from the start. OPD target first learns from original-image data, then adds counterfactual pairs while retaining the original examples. Both recipes use the edited images as training inputs and apply independent visual OPD to each example.

AspectPlain OPDOPD target
Originating baseQwen3.5-9BQwen3.5-9B
Data introductionOriginal examples and pairs from the startOriginal examples first, then original examples plus pairs
Uses edited imagesYesYes
Uses teacher contrast across the pairNoNo
Target and lossCurrent-crop EMA teacher; JSDSame independent-OPD formulation
Support and tailStudent top-100 plus residual tailSame

Complete results. Scores are percentages. CF-OPD is included to show the separate effect of using the paired target in the mixed-data setting.

RecipeV* ↑Zoom ↑HR-4K ↑HR-8K ↑MME-EN ↑MME-CN ↑Avg₆ ↑Flip ↑Blind ↓
Plain OPD91.1063.3179.5075.6271.2670.0275.1370.620.2
OPD target91.1062.9686.2585.2571.1369.4377.6971.121.0
CF-OPD93.7263.0887.7583.5072.2170.4678.4577.615.5

OPD target’s Avg₆ is 2.56 points above Plain OPD. The two recipes both include counterfactual examples and differ in when those examples enter training. Original-image OPD before adding pairs obtains 77.62 Avg₆; the independent OPD target after adding pairs obtains 77.69. The complete CF target obtains 78.45 with the same mixed inputs.

Training details. Both independent-OPD recipes use an EMA crop teacher with update rate 0.05, JSD with equal mixture weights, the student-selected top-100 candidates and a residual vocabulary-tail bucket. Their optimizer uses AdamW with learning rate 2×10⁻⁶, β=(0.9,0.999), weight decay 0.01, and gradient clipping at 1. The student generates responses from full images; the teacher scores evidence crops at the student’s prefixes.

Locations. The recipe distinction is described in §4.1 and Appendix A; the historical configurations appear in Appendix A, Table 4, and the new controlled comparison in §4.4, Table 3. The blue-marked revision is in the Supplementary Material.

Reviewer_beS8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Counterfactual Visual On-Policy Distillation (CF-OPD) for improving fine-detail recognition in vision-language models. It constructs original–edited image pairs with different verified answers to the same question. At student-generated prefixes, a crop-conditioned EMA teacher scores both images’ evidence crops. Their clipped log-probability contrast adjusts a student-anchored distribution, which is then mixed with the current crop teacher’s distribution for distillation. Training uses independent rollouts on both pair members; inference requires only the full-image student.

The paper also introduces an automated editing and verification pipeline and CF-Bench-X, containing 595 evaluation pairs. Experiments cover six visual benchmarks and Qwen3.5 models at 4B and 9B. At 9B, CF-OPD achieves 78.45 average accuracy versus 77.69 for independent OPD on the same training inputs, while improving paired correctness from 71.1% to 77.6% and reducing unchanged predictions from 21.0% to 15.5%.

Critical strengths
  1. A concrete mechanism for using semantic changes during distillation. Comparing the same teacher at the same student prefix provides a practical token-level signal from an answer-changing intervention. Training on both pair members makes productive use of the edited images while retaining ordinary full-image inference. This is a meaningful extension of crop-based visual OPD.

  2. Evaluation includes a useful same-input comparison and complementary metrics. Table 2 compares CF-OPD with independent OPD trained on the same images, addressing the important alternative explanation that improvements come solely from additional edited data. Reporting correctness on both pair members, alongside unchanged predictions and standard accuracy, also avoids treating any answer change as successful grounding.

Critical weaknesses
  1. The experiments do not fully isolate the contribution of the counterfactual contrast. Equation 4 changes both the contrast signal and the target’s student–teacher mixture relative to ordinary OPD. The sweeps do not include an α = 0 control with the same λ, candidate support, and tail treatment. Consequently, the same-input comparison establishes the benefit of the overall target construction but does not determine how much comes from the paired contrast rather than the altered distillation target. This matters because the average-accuracy improvement over that baseline is 0.76 points, with no reported variation across training seeds. A matched comparison with evidence-removal target reconstruction, such as the cited VAD method, would further clarify the benefit specific to semantic replacement.

  2. The reliability and independence of CF-Bench-X need stronger validation. Training and evaluation pairs use the same construction and verification procedure, and no independent human audit or measured verifier error rate is reported. Withholding intended labels from the reader is useful, but does not establish label correctness, answer uniqueness, or freedom from editing artifacts. The paired gains could therefore partly reflect adaptation to the construction pipeline. Moreover, Object and State contain only 23 and 14 pairs, limiting conclusions about these harder changes. Since paired evaluation supplies the clearest improvement, these uncertainties materially affect the assessment.

Questions for authors
  1. Controlled ablation: What happens with α = 0 while retaining the default λ = 0.15 and identical support, tail handling, training inputs, and optimization budget? Does the counterfactual contrast still provide a reproducible improvement over this control?

  2. Benchmark clarification and validation: Was any independently sampled subset manually audited for answer correctness, uniqueness, and unintended changes? Please report the audit size and error rates, or provide equivalent independent validation.

  3. Experimental reliability: How were hyperparameters and checkpoints selected, and was CF-Bench-X excluded from that selection? Can you provide repeated-run uncertainty for the principal CF-OPD versus same-input OPD comparison?

6 comments
AuthorsOpenReview ↗

Response to Reviewer beS8

Thank you for the detailed assessment and for recognizing the shared-prefix semantic contrast, two-sided training, and complementary evaluation metrics. We address each weakness and question separately below. The revised manuscript with changes highlighted in blue is provided in the Supplementary Material.

AuthorsOpenReview ↗

Response to Reviewer beS8 — W1: Isolating the contrast, seed variation, and evidence removal

↑: higher is better; ↓: lower is better.

Thank you for identifying the distinction between the contrast and the changed distillation target. We added a strict contrast-removal control and repeated all three methods with seeds 42, 43, and 44.

1. Separating contrast from the teacher mixture. In the new α=0 control, λ remains 0.15. The student-selected top-100 support, target normalization, tail handling, initialization, training data, two-sided rollouts, and optimization settings remain the same as full CF-OPD. Both teacher views are still evaluated; only the contrast coefficient is set to zero. This control retains the student–teacher mixture and directly tests the contrast contribution in Eq. (4).

2. Complete benchmark results and variation. The following reports mean±sample SD across the three seeds. Scores are percentages; SD and differences are percentage points. Avg₆ is the equal-weight six-benchmark average; Flip is better when higher and Blind when lower.

MetricCF-OPDα=0Independent OPD
V* ↑93.72±0.9190.23±0.6090.58±0.91
Zoom ↑62.72±0.9563.59±0.7163.55±0.43
HR-4K ↑86.96±0.3885.42±0.1484.67±0.40
HR-8K ↑83.67±1.0482.71±0.5282.96±0.85
MME-EN ↑72.96±0.0871.45±0.2471.63±0.35
MME-CN ↑70.26±0.4969.05±0.2069.66±0.30
Avg₆ ↑78.3798±0.120677.0729±0.096377.1738±0.3444
Flip ↑75.85±0.3971.76±1.1872.38±0.51
Blind ↓17.42±0.6819.78±1.2419.44±0.79

Statistics and differences are calculated before rounding.

Relative to α=0, CF-OPD improves Avg₆ by 1.31 points and Flip by 4.09 points, and reduces Blind by 2.35 points. All three seeds improve these three aggregate metrics. Relative to same-input independent OPD, the corresponding mean changes are +1.21, +3.47, and −2.02 points. Mean accuracy improves on five of six benchmarks against either control, alongside higher Avg₆ and consistent paired-evaluation gains. The original +0.76 Avg₆ comparison establishes the overall target’s benefit; the new α=0 experiment specifically tests the contrast while retaining the mixture. The repeated-run results are reported separately from the original model.

3. Comparison with evidence removal. We reproduced VAD at 9B and evaluated it on all six recognition benchmarks and CF-Bench-X. The table compares that reproduction with the submitted CF-OPD-9B model; Δ=CF-OPD−VAD.

MethodV* ↑Zoom ↑HR-4K ↑HR-8K ↑MME-EN ↑MME-CN ↑Avg₆ ↑Flip ↑Blind ↓
VAD-9B90.0561.6687.0085.7574.7970.5978.3168.724.0
CF-OPD-9B93.7263.0887.7583.5072.2170.4678.4577.615.5
Δ (CF−VAD)+3.67+1.42+0.75−2.25−2.58−0.13+0.14+8.9−8.5

CF-OPD improves Avg₆ by 0.14 points, from 78.31 to 78.45. It outperforms VAD on V*, ZoomBench, and HRBench-4K by 3.67, 1.42, and 0.75 points, respectively, and remains competitive on HRBench-8K and MME-RealWorld EN/CN. It also improves Flip by 8.9 points and reduces Blind by 8.5 points. Evidence removal tests a prediction after suppressing evidence; semantic replacement supplies an alternative visible attribute with a different answer. This comparison supports the usefulness of semantic replacement for paired responses. The strict α=0 experiment provides the controlled attribution to our contrast term; the VAD comparison provides complementary method-level evidence.

Locations. The target and loss are in §3.3, Eq. (4); the new controlled runs are in §4.4, Table 3; the input settings are in Appendix A. The complete VAD comparison is provided above. The revised manuscript with changes highlighted in blue is in the Supplementary Material.

AuthorsOpenReview ↗

Response to Reviewer beS8 — W2: Benchmark reliability, pipeline dependence, and small categories

↑: higher is better; ↓: lower is better.

Thank you for emphasizing that benchmark reliability is particularly important when paired evaluation provides the clearest gains. We added human evaluation of the complete accepted training and test sets, and report the category and benchmark-source breakdowns below.

1. Validation beyond automatic acceptance. We manually evaluated all 5,190 pairs: 4,595 training pairs and all 595 CF-Bench-X pairs, at a cost of US$467. The human evaluation directly checks intended answer changes and editing quality in the accepted training and test data.

CriterionTraining (4,595)Test (595)
Intended answer change ↑4,227 (91.99%)559 (93.95%)
Editing problem ↓207 (4.50%)12 (2.02%)
Visible artifact ↓103 (2.24%)9 (1.51%)
Artifact may cue the answer ↓81 (1.76%)6 (1.01%)

The intended answer change is achieved in 91.99% of training pairs and 93.95% of test pairs. The test set has 2.02% editing problems, 1.51% visible artifacts, and 1.01% potentially answer-cueing artifacts. These measurements provide direct human evidence of the quality of the accepted pairs.

2. Independence and possible pipeline adaptation. CF-Bench-X images, questions, and edits are excluded from training and from hyperparameter and checkpoint selection. We assess performance on six unedited public recognition benchmarks together with paired evaluation, and compare CF-OPD with controls using the same training inputs. The new contrast control improves mean Avg₆ from 77.0729 to 78.3798. This supports benefits beyond responding to the editing pipeline, while the human evaluation quantifies residual issues in the paired benchmark.

3. Every edit category. The following uses the submitted CF-OPD-9B weights (run sw_replayb40_a100, Avg₆=78.45, overall Flip=77.65%, Blind=15.46%) and the same-input independent OPD baseline. These are the weights used for the manuscript’s category figure.

SubsetPairsOPD Flip ↑CF Flip ↑OPD Blind ↓CF Blind ↓
Text26677.8281.5819.5516.17
Color/pattern17379.1986.1313.878.67
Surface6445.3159.3832.8117.19
Shape/material5565.4574.5518.1812.73
Object/spatial2330.4347.8352.1739.13
State/pose1450.0042.8642.8650.00
All59571.0977.6521.0115.46

We have narrowed the descriptions of Object/spatial (23 pairs) and State/pose (14 pairs) to match their sample sizes, and added qualitative examples showing the intended changes, evidence regions, and observed outcomes.

4. Every benchmark-source subset of CF-Bench-X. These are paired-evaluation subsets, not the standard recognition scores.

SubsetPairsOPD Flip ↑CF Flip ↑OPD Blind ↓CF Blind ↓
V*3582.8691.4314.298.57
Zoom8862.5070.4522.7318.18
HR-4K12378.0584.5516.2610.57
HR-8K11172.9780.1817.1212.61
MME-EN7062.8671.4331.4321.43
MME-CN16870.2474.4023.2118.45
All59571.0977.6521.0115.46

The submitted CF-OPD model improves Flip and reduces Blind in all six source subsets, demonstrating consistent paired-evaluation gains across benchmark sources. We consider the measured error rates acceptable for the multimodal fine-tuning and paired-evaluation settings studied here.

Locations. Data separation and benchmark scope: §3.5; category analysis: §4.3; human evaluation: Appendix G, Table 6; qualitative examples: Appendices F/H, including Figures 8–10. The blue-marked revision is in the Supplementary Material.

Comments on the paper

5 comments
AuthorsOpenReview ↗

Summary of Revisions for the Area Chairs and Reviewers

Dear Area Chairs and Reviewers,

We sincerely thank you for your time, careful assessment, and constructive suggestions. Your comments guided additional experiments and clarifications. We respectfully summarize these below, grouping shared concerns. W and Q denote Weakness and Question numbers. The revised manuscript with changes highlighted in blue is provided in the Supplementary Material; our individual replies include the detailed results.

1. Controlled ablations and repeated runs Addresses: beS8 (W1, Q1, Q3); QCTw (W1); 3xuK (W1, Q1).

We added nine Qwen3.5-9B runs: CF-OPD, strict α=0, and same-input independent OPD, each with seeds 42, 43, and 44. The α=0 control removes only the teacher contrast while retaining λ=0.15, candidate support, tail handling, initialization, paired inputs, two-sided rollouts, and optimization settings.

Results are mean±sample SD across three seeds (%); higher Flip and lower Blind are better.

MethodAvg₆FlipBlind
CF-OPD78.3798±0.120675.85±0.3917.42±0.68
α=077.0729±0.096371.76±1.1819.78±1.24
Independent OPD77.1738±0.344472.38±0.5119.44±0.79

CF-OPD improves Avg₆, Flip, and Blind over both controls in every seed. Relative to same-input OPD, the within-seed differences are +1.21±0.43, +3.47±0.19, and −2.02±1.47 percentage points, respectively. We report all six benchmarks and their variability, supporting reproducible gains under the tested setting. α=1 and λ=0.15 were fixed heuristically before these runs; CF-Bench-X was excluded from hyperparameter and checkpoint selection, and terminal weights were prespecified. See §4.4, Table 3; Appendix A; and beS8 response, Tables R1/R4.

2. Comparison with evidence removal Addresses: beS8 (W1).

We added a supplementary comparison using our VAD-9B reproduction, with all six benchmark scores and a Δ row in the beS8 response, Table R2. Submitted CF-OPD and VAD obtain 78.45 and 78.31 Avg₆. CF-OPD achieves 77.6 versus 68.7 Flip and 15.5 versus 24.0 Blind: +8.9 and −8.5 percentage points. Each method leads on three benchmarks, while CF-OPD performs better on paired evaluation.

3. Human evaluation and editing failures Addresses: beS8 (W2, Q2); QCTw (Q2); 3xuK (W2, Q2).

We evaluated all 5,190 image pairs: 4,595 training and 595 test pairs, costing US$467. The assessment covers intended answer changes, editing problems, visible artifacts, and possible answer cues.

CriterionTrainingTest
Intended answer change4227 (91.99%)559 (93.95%)
Editing problem207 (4.50%)12 (2.02%)
Visible artifact103 (2.24%)9 (1.51%)
Artifact may cue answer81 (1.76%)6 (1.01%)

Issue categories may overlap. We consider these residual rates acceptable for the studied use of pairs as supplementary supervision in multimodal fine-tuning. We also added representative failures showing the original image, edit instruction, edited image, and error type. See §3.2; Appendix G, Table 6; Appendix H, Figures 8–10.

4. OPD definitions and the role of original data Addresses: QCTw (W2, W3, Q1).

We clarified that both recipes originate from Qwen3.5-9B. Plain OPD mixes original and counterfactual images from the start; OPD target trains on original images before adding counterfactual pairs while retaining original data. We also organized the existing data-mixture results: CF-OPD obtains 78.45 Avg₆ with mixed data versus 75.09 with paired-only data. With mixed inputs, CF-OPD exceeds OPD by 0.76 Avg₆ and 6.5 Flip points, with Blind 5.5 points lower. With paired-only inputs, Avg₆ is 0.20 lower, but Flip improves by 3.8 and Blind decreases by 4.6. These results support retaining original data for general accuracy and using pairs for targeted visual-change supervision. See Appendix A, Table 4.

5. Scope of small-category conclusions Addresses: beS8 (W2).

We clarified the limited coverage of Object/spatial (23 pairs) and State/pose (14 pairs), narrowed the corresponding claims, and provided qualitative examples to make successes and failures inspectable. See §4.3 and Appendices F/H.

We are grateful for the opportunity to address these concerns and respectfully submit the additional evidence for your consideration. Thank you again for your time and thoughtful feedback.

AuthorsOpenReview ↗

Response on Human-Validated Paired Evaluation

Thank you for your encouraging assessment of the additional controls and for this helpful follow-up. We retained the pair-level human inspection records for intended answer changes and answer-label quality. Based on these records, we report the three filtered evaluations below.

Human validation and uniform evaluation. The answer-label check requires that the reference answers for both the original and edited images are correct, uniquely supported by the corresponding image and question, and unambiguous. We separately check whether the intended answer change occurs. The combined subset satisfies all of these conditions.

We reuse the same previously evaluated checkpoints and saved predictions: CF-OPD, the α=0 control, and independent OPD with seeds 42, 43, and 44, together with the VAD-9B reproduction. For each filtering criterion, exactly the same pair IDs are retained for every method and seed. The reference labels and scoring procedure remain unchanged. Results for the three repeated methods are mean±sample SD; ↑: higher is better; ↓: lower is better.

1. Excluding the 36 pairs without the intended answer change: 559 pairs.

MethodFlip ↑Blind ↓
CF-OPD75.37±0.4117.59±0.68
α=0 control71.20±1.2920.16±1.29
Independent OPD71.85±0.5519.92±0.88
VAD-9B68.5224.15

2. Excluding the 42 pairs with answer-label issues: 553 pairs. Both answer labels in every retained pair pass the correctness, uniqueness, and unambiguity checks described above.

MethodFlip ↑Blind ↓
CF-OPD75.47±0.4217.66±0.58
α=0 control71.25±1.3020.19±1.31
Independent OPD71.91±0.5520.01±0.85
VAD-9B68.5424.05

3. Applying both exclusions: 553 pairs. The 36 pairs without the intended change are all included among the 42 pairs with answer-label issues. Their union therefore contains 42 pairs, leaving the same 553 pairs as in the second evaluation. Every pair in this combined subset has the intended answer change and correct, unique, unambiguous labels on both sides.

On the combined human-validated subset, CF-OPD improves Flip by 4.22, 3.56, and 6.93 percentage points over α=0, independent OPD, and VAD, respectively, while reducing Blind by 2.53, 2.35, and 6.39 points. Differences are calculated before rounding. CF-OPD achieves higher Flip and lower Blind than both repeated controls in each of seeds 42, 43, and 44 under all three filtering conditions. These results show that its paired-evaluation advantage is retained when both the intended answer change and the correctness and unambiguity of both labels are human-validated.

Area_Chair_Hqd3OpenReview ↗

Discussion phase will end on Nov 18

This is a kind reminder that discussion phase will end on Nov. 18.

Dear Reviewer beS8, thank you for your active engagement in the discussion. Could you please check the most recent response from the authors as well?

Dear Reviewer QCTw and 3xuK, could you please check authors' rebuttal and update your thoughts? Your early feedback would make the discussion more smooth and informative.

Thank you.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.