acceptodds
Under review as a conference paper at ICLR 2027

CA-OPD: Confidence-Aware On-Policy Distillation for Structured Visual Prediction

Abstract

Autoregressive vision language models unify heterogeneous perception tasks but are highly susceptible to compounding errors. On-policy distillation (OPD) bridges the training-inference mismatch by training students on their own rollouts. However, unreliable student predictions, especially early in training, can derail the trajectory and degrade the quality of teacher supervision. While recent interleaved distillation methods allow the teacher to verify and replace student tokens, they primarily rely on rigid ranking metrics rather than exact teacher confidence, and they overlook how intervention decisions can inform token-level supervision. To address this, we introduce Confidence-Aware On-Policy Distillation (CA-OPD), a framework that couples reliable rollout construction with adaptive supervision. CA-OPD utilizes teacher confidence to selectively correct unreliable student transitions, gradually transferring rollout control to the student via a strict-to-relaxed schedule. Crucially, CA-OPD aligns knowledge transfer with these intervention decisions: corrected positions receive direct cross-entropy supervision from the teacher's prediction, while retained positions benefit from the teacher's full predictive distribution. Evaluated in a multi-teacher setting for GUI grounding and optical character recognition, CA-OPD substantially improves the Qwen3.5-0.8B baseline across all six target benchmarks, including gains of points on ScreenSpot-Pro and points on OCRBench-v2 English. Controlled studies further show that the gains depend on intervention placement, progressive rollout control, and intervention-aligned supervision, rather than intervention frequency alone.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
on-policy distillation, knowledge distillation, vision-language models, structured prediction, visual grounding, optical character recognition

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_ETAEOpenReview ↗
Rating
2
Confidence
4
Informativeness
1(−1)
Presentation
1(−1)
Soundness
1
Read review
Paper summary

The authors propose CA-OPD, an interleaved on-policy distillation variant tailored for small VLMs that generate structured outputs such as GUI coordinates and OCR text. During student rollouts, any token assigned a probability below a threshold α\alpha by the teacher is overwritten with the teacher's argmax prediction, and generation continues from the corrected prefix. This same accept-or-replace gate also governs the per-token training objective, applying cross-entropy loss at replaced positions and top-kk forward KL divergence at retained ones. Compared to Speculative Knowledge Distillation, the proposed method introduces three main modifications: an absolute-probability acceptance criterion instead of a rank-based one, a gating-dependent loss assignment, and an annealed threshold schedule.The authors evaluate CA-OPD by distilling two domain-specific Qwen3.5-9B teachers into an SFT-initialized Qwen3.5-0.8B student, alongside an additional Qwen3-VL-8B to 2B transfer setting. In the main experiments, CA-OPD outperforms OPD and SKD, but its gains over Offline KD are modest. The largest gain is on ScreenSpot-Pro, while OCR improvements are smaller. Results against Offline KD are mixed in the additional Qwen3-VL setting. Ablations support selective prefix correction. Threshold annealing improves ScreenSpot-Pro, but adds little on some OCR metrics.

Critical strengths
  1. The paper shows why the method works and I think is quiet soundness. In Table 2, random replacement and gating without prefix writeback both stay at the OPD level. The gain appears only when confidence-based placement is combined with writeback. Figure 2(b) shows that most rejected tokens are still in the teacher's top-25, so SKD would accept them.

  2. The comparisons are fair and well reported. All methods share the same initialization, data, and step budget. The stratified analysis defines its buckets before training and uses held-out rollouts to avoid selection bias. Because of this, it is also easy to see where the gains are small, especially compared with Offline KD.

Critical weaknesses
  1. Improvements over Offline KD are modest, especially on OCR tasks. In the Qwen3-VL experiments, CA-OPD also trails Offline KD on several benchmarks.

  2. Building upon SKD, CA-OPD introduces a probability-based gating mechanism, a gate-dependent loss function, and a threshold annealing strategy. Ablation studies demonstrate that this approach enables selective prefix correction, though the additional gains provided by the annealing strategy vary by task. Compared to existing interleaved distillation methods, its overall contribution is relatively limited.

  3. Experiments cover only 0.8B and 2B students with short training schedules. Generalization to larger students and longer training remains unclear. Since the gate uses absolute teacher probabilities, its robustness to differences in teacher calibration also needs further evaluation.

Questions for authors
  1. Robustness against stronger baselines: Given that plain Offline KD closely matches CA-OPD and even outperforms it on several Qwen3-VL benchmarks, can the authors demonstrate consistent gains over stronger offline objectives (e.g., reverse KL distillation [1] or LLaVA-KD [2]) and mixed-data online methods without confidence gating (e.g., GKD with λ<1\lambda < 1 [3] or DistiLLM [4])?

  2. Necessity of the annealing schedule: Scheduled sampling [5] already achieves a gradual transition from teacher to student prefixes and is cited in the paper without empirical comparison. Does annealing provide a consistent advantage over a tuned fixed threshold across tasks and runs?

  3. SKD comparison: Could the authors include the stronger (K=1K=1) SKD configuration from Table 10 in the comprehensive benchmark comparisons? How much of an advantage does CA-OPD still maintain when compared against this baseline?

References

[1] Y. Gu, L. Dong, F. Wei, M. Huang. MiniLLM: Knowledge Distillation of Large Language Models. ICLR 2024.

[2] Y. Cai et al. LLaVA-KD: A Framework of Distilling Multimodal Large Language Models. ICCV 2025.

[3] R. Agarwal et al. On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes. ICLR 2024.

[4] J. Ko, S. Kim, T. Chen, S.-Y. Yun. DistiLLM: Towards Streamlined Distillation for Large Language Models. ICML 2024.

[5] S. Bengio, O. Vinyals, N. Jaitly, N. Shazeer. Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks. NeurIPS 2015.

10 comments
AuthorsOpenReview ↗

Response to Weakness 1

Dear Reviewer:

Thank you for your careful reading and constructive feedback.

Gains over Offline KD. Under the same initialization, training data, and optimizer budget, CA-OPD achieves higher mean scores than Offline KD on all six GUI/OCR benchmarks in the main setting (Table 1). The improvements are +0.91 on OCRBench-v2 English, +0.31 on OCRBench-v2 Chinese, and +0.33 on CC-OCR, alongside +0.45 on ScreenSpot-v2 and +1.69 on ScreenSpot-Pro.

The aggregate OCR scores also encompass different capabilities. In Appendix D.1, CA-OPD achieves higher mean scores in all eight OCRBench-v2 capability categories, with gains ranging from +0.09 to +2.29 points. The largest improvements occur in Spotting (+2.29) and Referring (+1.39). On ScreenSpot-Pro, improvements span all six application categories, ranging from +1.03 to +2.17 points. The tables below reproduce these results, averaged over three training runs.

OCRBench-v2 by capability category (Table 7).

CategoryOffline KDCA-OPDDifference
Recognition69.0269.42+0.40
Referring46.9348.32+1.39
Spotting22.3424.63+2.29
Extraction75.4876.16+0.68
Parsing49.2149.30+0.09
Calculation36.3336.59+0.26
Understanding54.4655.29+0.83
Reasoning37.5538.00+0.45

Referring, Spotting, and Calculation are English-only.

ScreenSpot-Pro by application category (Table 6).

CategoryOffline KDCA-OPDDifference
CAD40.0141.63+1.62
Creative38.3839.41+1.03
Dev38.9441.11+2.17
OS43.7545.75+2.00
Office64.2665.60+1.34
Scientific44.9147.03+2.12

These results show that the main-setting gains extend across multiple capabilities and applications. On OmniDocBench, CA-OPD preserves performance close to the SFT initialization: 70.43 versus 70.49, compared with 70.40 for Offline KD.

In the additional Qwen3-VL setting (Table 16), CA-OPD scores below Offline KD on ScreenSpot-Pro, CC-OCR, and OCRBench-v2 Chinese by 0.28, 0.29, and 0.46 points, respectively. Meanwhile, it exceeds both OPD and SKD on five of the six GUI/OCR benchmarks, including gains of +1.96 and +1.78 on ScreenSpot-Pro. Thus, the evidence supports improvements over Offline KD across the evaluated categories in the main setting, while the additional setting demonstrates improvements over on-policy baselines with task-dependent performance relative to Offline KD.

AuthorsOpenReview ↗

Response to Weakness 2

Dear Reviewer:

Thank you for your careful reading and constructive feedback.

Contribution beyond interleaved generation. CA-OPD connects two decisions that arise together in structured visual prediction: which token should enter the prefix, and how the student should learn at that position. Once a coordinate digit or character is committed, subsequent predictions depend on it. Our design uses the same teacher-support decision to govern both prefix construction and token-level supervision.

SKD uses teacher rank to accept or replace proposals and applies a common divergence objective across positions (SKD, Alg. 1). CA-OPD instead uses absolute teacher probability: rejected proposals receive teacher-argmax writeback and CE toward the committed target, while retained proposals receive distributional KL. Thus, verification also determines the form of supervision.

The ablation in Table 2 directly tests this additional role. Keeping the probability gate and argmax writeback unchanged, replacing rejected-position CE with KL reduces ScreenSpot-Pro from 45.92 to 45.26 and OCRBench-v2 English from 52.47 to 51.75:

Rejected-position supervisionSS-ProOCRv2-EN
CE (CA-OPD)45.9252.47
KL45.2651.75

This supports the value of gate-dependent supervision beyond trajectory modification alone.

Our fixed-prefix diagnostic further examines the local consequences of probability-versus-rank disagreement. On SFT-init rollouts from 128 randomly sampled training inputs per domain, we select the first actual-change position per sequence in each disagreement group. From the same preceding prefix, we either retain the student proposal or write back the teacher's argmax, then continue greedily without further intervention. At α≈0.25:

Gate disagreementDomainPairsMean replacement benefit [95% CI]
CA rejects, K=25 acceptsGrounding125+68.70 [27.60, 115.04]
CA rejects, K=25 acceptsOCR80+0.0590 [0.0068, 0.1202]
K=1 rejects, CA acceptsGrounding42+6.44 [−26.90, 40.63]
K=1 rejects, CA acceptsOCR44−0.0043 [−0.0332, 0.0303]

Table note. Benefit is distance reduction in 0–1000 coordinates (grounding) or increased 1−NED against the training reference (OCR), intervals are descriptive 95% paired-bootstrap CIs. Both arms share a 1024-token cap.

At proposals admitted by K=25 but rejected by CA, replacement improves mean downstream outcomes. At supported alternatives retained by CA but rejected by K=1, additional replacement shows no clear average benefit. These local results illustrate why the verification criterion matters, while the CE-to-KL ablation tests the additional use of verification to assign supervision.

Annealing varies this joint decision over training, changing both which proposals are retained and where CE versus KL is applied. We present the fixed-threshold and schedule comparisons in our response to Question 2.

AuthorsOpenReview ↗

Response to Weakness 3

Dear Reviewer:

Thank you for your careful reading and constructive feedback.

Scope and teacher calibration. Our experiments target distillation into compact students, covering Qwen3.5-0.8B and Qwen3-VL-2B with their respective teachers. The additional Qwen3-VL setting extends the evaluation beyond the main model configuration: CA-OPD outperforms both OPD and SKD on five of the six GUI/OCR benchmarks. Our empirical conclusions concern these student sizes and the evaluated training budgets, larger students and substantially longer CA-OPD training remain outside the present evaluation.

For teacher confidence, we assess whether the support signal leads to useful interventions through downstream outcomes. In our fixed-prefix diagnostic at α≈0.25, teacher-argmax replacement improves grounding click success by 13.60 percentage points [6.40, 20.80] and OCR 1−NED by 0.0744 [0.0210, 0.1345], with descriptive 95% paired-bootstrap intervals. Both branches use the same preceding prefix and continuation budget.

These results provide direct evidence of intervention utility under the evaluated teachers. We interpret the probability floor as a teacher-support threshold rather than a calibrated probability of task correctness. Absolute-probability gating can nevertheless be affected by changes in teacher calibration, Appendix I discusses the potential need to retune the floor under systematic teacher–student mismatch.

Reviewer_RBc5OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CA-OPD, a distillation framework. During student generation, a teacher evaluates each proposed token against a probability threshold, replacing rejected proposals with its highest-probability prediction. The framework applies cross-entropy loss to corrected positions and truncated forward-KL loss to accepted ones. As training progresses, the threshold is gradually relaxed to give the student greater control over its generation trajectory. Experiments using task-specific teachers demonstrate improvements over controlled distillation baselines, complemented by analyses of intervention strategies, threshold scheduling, knowledge retention, and computational overhead.

Critical strengths
  • CA-OPD achieves notable improvements over competing methods under matched training data and optimization budgets.
  • The paper provides separate analyses of prefix correction, loss functions, threshold selection, and scheduling strategies. Reporting variability across three runs and training costs further strengthens the transparency and interpretability of the experimental results.
Critical weaknesses
  • Rejected proposals do not always result in prefix modifications. Rejection may occur even when the student's proposed token matches the teacher's argmax, in which case only the supervision objective changes to cross-entropy. Consequently, the reported statistics do not clearly distinguish actual trajectory corrections from changes in the training loss.

  • The intervention placement comparisons are not conducted under fully matched budgets.

Questions for authors
  • What proportion of rejected positions leads to an actual change in the proposed token? Please distinguish these cases from instances where the token remains unchanged but receives CE supervision, with results broken down by domain and training stage.

  • Does random intervention remain less effective when the number and temporal distribution of actual token replacements are matched? Comparing against a rank-based gating strategy with identical CE/KL supervision would further clarify the contribution of the proposed gate.

  • How does CA-OPD compare with Offline KD under the same wall-clock training budget?

6 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 1

Dear Reviewer:

Thank you for your careful reading and constructive feedback.

Rejection versus actual token change. We separately record gate rejection, actual token change, and supervision-only change. For a student proposal yty_t, let yt∗=arg⁡max⁡vpT(v∣ht)y_t^*=\arg\max_v p_T(v\mid h_t). We define

rt=1[pT(yt∣ht)<αt]r_t = \mathbf{1}[p_T(y_t \mid h_t) < \alpha_t], ct=rt1[yt≠yt∗]c_t = r_t \mathbf{1}[y_t \ne y_t^{\ast}], and ut=rt1[yt=yt∗]u_t = r_t \mathbf{1}[y_t = y_t^{\ast}], so that rt=ct+utr_t = c_t + u_t.

so that rt=ct+utr_t=c_t+u_t. Here, ctc_t changes the committed token, whereas utu_t leaves the token unchanged and assigns CE supervision.

Across the instrumented 471-step training run, 49,226 of 50,417 rejections changed the token (97.64%), while 1,191 were CE-only events (2.36%). The domain and training-stage breakdown is:

DomainTraining stepsRejection rateActual-change rateCE-only rateActual changes / rejections
Grounding1–1576.95%6.62%0.33%95.21%
Grounding158–3145.90%5.75%0.15%97.49%
Grounding315–4715.20%5.16%0.04%99.17%
OCR1–1575.39%5.22%0.17%96.89%
OCR158–3145.46%5.39%0.07%98.66%
OCR315–4714.86%4.80%0.06%98.80%

The first three rates use all response tokens within each domain–stage group as the denominator, the final column uses rejected tokens. We aggregate token counts before computing ratios.

Actual token changes account for 95.21–99.17% of rejections across all six domain–stage groups. Thus, gate rejection predominantly results in prefix modification.

AuthorsOpenReview ↗

Response to Weakness 2 and Question 2

Dear Reviewer:

Thank you for your careful reading and constructive feedback.

Random and rank-based intervention controls. We added a disagreement-conditioned random control targeting approximately matched actual-change budgets. At positions where the student proposal differs from the teacher's argmax, this control intervenes with a probability determined by the training step and domain. Selected positions receive teacher-argmax writeback and CE supervision, retained positions receive KL. Selection within this pool does not depend on the proposal's teacher probability. The control retains teacher-based disagreement detection and selects 93.82% of its disagreement positions overall.

We also trained CA-OPD with rank-1 gating, keeping argmax writeback and the CE/KL supervision rule unchanged. This control intervenes at every teacher–student disagreement and tests the alternative verification criterion.

Across the instrumented 471-step runs, CA-OPD changes 49,226 tokens, compared with 50,513 for the random control, a 2.61% difference in counts. Their actual-change rates are 5.45% and 5.60%. Dividing training into equal early, middle, and late stages, the respective rates are 5.84% / 5.55% / 4.96% for CA-OPD and 5.92% / 5.73% / 5.14% for the random control. Thus, actual modification counts and their coarse distribution over training are approximately matched.

The following table collects the intervention statistics and reported benchmark scores:

MethodActual changesChange rateSS-v2SS-ProOCR-ENOCR-ZHCC-OCROmniDocRefCOCOMMBench
Disagreement-conditioned random50,5135.60%87.8345.3552.1352.8566.3469.7681.8379.45
CA-OPD with rank-1 gating52,6955.83%89.0144.5552.2652.8966.4569.8981.6780.14
CA-OPD49,2265.45%88.4145.9252.4753.2267.3270.4381.9080.16

Change rates use all response tokens in each instrumented run as the denominator.

Across the reported scores, CA-OPD exceeds the random control on all eight benchmarks, including +0.57 on ScreenSpot-Pro and +0.98 on CC-OCR. Relative to rank-1 gating with the same supervision rule, its scores are higher on seven of eight benchmarks, including +1.37 on ScreenSpot-Pro and +0.87 on CC-OCR, rank-1 gating scores higher on ScreenSpot-v2. This pattern shows CA-OPD's competitiveness against both intervention alternatives, with the strongest reported advantages on ScreenSpot-Pro and CC-OCR.

AuthorsOpenReview ↗

Response to Question 3

Dear Reviewer:

Thank you for your careful reading and constructive feedback.

Comparison under an estimated time-matched budget. We target the same wall-clock budget using the per-step runtime estimates reported in the paper to determine the extended offline training budget. Table 18 reports 59.3 s/step for CA-OPD and 35.3 s/step for Offline KD, giving Offline KD approximately 1.68 times the original update budget. We evaluate the extended offline forward-KL run and report its scores alongside the original main-setting results:

MethodSS-v2SS-ProOCR-ENOCR-ZHCC-OCROmniDocRefCOCOMMBench
Offline KD (original budget)87.9644.2351.5652.9166.9970.4081.8679.99
Offline KD (estimated time-matched budget)88.2145.5851.1053.3166.4770.5681.7079.75
CA-OPD88.4145.9252.4753.2267.3270.4381.9080.16

Across these reported scores, CA-OPD remains competitive with the extended-budget Offline KD baseline, with advantages of +0.34 on ScreenSpot-Pro, +1.37 on OCRBench-v2 English, and +0.85 on CC-OCR. Extended-budget Offline KD scores higher on OCRBench-v2 Chinese and OmniDocBench by 0.09 and 0.13 points, respectively.

Reviewer_o4DZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Confidence-Aware On-Policy Distillation (CA-OPD) for structured visual prediction. During student-generated rollouts, the teacher evaluates each proposed token and replaces those receiving insufficient probability support. A progressively relaxed threshold gradually transfers rollout control to the student. Corrected positions receive cross-entropy supervision on the teacher’s prediction, while retained positions use distributional distillation. Experiments on GUI grounding and OCR benchmarks demonstrate improvements over standard OPD and SKD in the main setting. Additional experiments examine intervention placement, threshold scheduling, supervision choices, and another teacher–student pair.

Critical strengths

1)The research problem is meaningful and advancing. The manuscript is well-written and easy to follow. 2) The experimental analysis is relatively thorough, including component ablations, intervention diagnostics, three-run results, training-cost measurements, and an additional teacher–student configuration.

Critical weaknesses
  1. The technical novelty appears incremental. Teacher verification and token replacement are already central to SKD (ICLR 2025). CA-OPD mainly introduces probability-based gating, threshold annealing, and different losses for corrected and retained tokens. The contribution beyond this combination needs clearer justification.
  2. The method strongly depends on teacher confidence being a reliable indicator of correctness. Low probability for a student proposal does not necessarily mean that the teacher’s preferred token is correct. More analysis of teacher errors, confidence calibration, and robustness under domain shift would strengthen the proposed intervention strategy.
  3. The advantages over stronger baselines are sometimes small and inconsistent. For example, CA-OPD improves OmniDocBench over Offline KD by only 0.03 points in the main setting. In the additional Qwen3-VL experiment, it underperforms Offline KD on ScreenSpot-Pro, CC-OCR, and OCRBench-v2 Chinese. These results limit the claim of a broadly reliable advantage.
  4. The accuracy–efficiency trade-off is not fully established. Appendix H reports that CA-OPD requires approximately 1.5× the training time of OPD and 1.7× that of Offline KD. Since comparisons primarily match optimizer steps, evaluation under equal wall-clock or computational budgets is needed to determine whether the additional complexity is worthwhile.
Questions for authors

Please refer to the weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Weakness 1

Dear Reviewer:

Thank you for your careful reading and constructive feedback.

Novelty relative to SKD. Our starting point is the chain of discrete commitments in structured visual prediction. Once a coordinate digit or character enters the prefix, subsequent predictions depend on it. This connects two questions: whether the teacher supports committing a proposal, and how the student should learn when the teacher rejects it.

SKD accepts proposals within the teacher's top-K, resamples rejected tokens from the teacher, and applies the same divergence objective across positions (SKD, Alg. 1). CA-OPD uses absolute teacher probability to decide whether to retain a proposal or commit the teacher's argmax. The same decision assigns CE toward the committed target at rejected positions and distributional KL at retained positions, coupling trajectory construction with token-level supervision.

A. Gate-dependent supervision. Switching rejected-position supervision from CE to KL reduces performance while keeping the gate and argmax writeback.

VariantSS-Pro ↑OCRv2-EN ↑
CA-OPD: argmax writeback, CE at rejected positions45.9252.47
KL at rejected positions45.2651.75
Sampled writeback43.6151.69

B. Probability versus rank gating. On SFT-init rollouts from 128 randomly sampled training inputs per domain, we select the first actual-change position per sequence in each disagreement group. From the identical preceding prefix, we either retain the proposal or commit the teacher's argmax, then continue greedily without further intervention. Both threshold endpoints are tested on this same snapshot.

Gate disagreementα≈DomainPairsMean replacement benefit [95% CI]
CA rejects, K=25 accepts0.50Grounding126+63.44 [21.17, 114.67]
CA rejects, K=25 accepts0.50OCR83+0.0684 [0.0181, 0.1248]
CA rejects, K=25 accepts0.25Grounding125+68.70 [27.60, 115.04]
CA rejects, K=25 accepts0.25OCR80+0.0590 [0.0068, 0.1202]
K=1 rejects, CA accepts0.25Grounding42+6.44 [−26.90, 40.63]
K=1 rejects, CA accepts0.25OCR44−0.0043 [−0.0332, 0.0303]

Table note. Benefit is distance reduction in 0–1000 coordinates (grounding) or increased 1−NED against the training reference (OCR), intervals are descriptive 95% paired-bootstrap CIs. Both arms share a 1024-token cap.

The distinction from SKD concerns both the verification criterion and what verification controls. The fixed-prefix results show why the criterion matters: CA intervenes at low-support proposals admitted by K=25, where replacement improves average outcomes, and retains supported non-top-1 proposals whose replacement shows no clear average benefit. Beyond these rollout decisions, CA uses rejection to assign direct CE supervision, whereas SKD applies its divergence objective uniformly. The CE-to-KL ablation supports the value of this additional use of verification. Our contribution is a teacher-support decision that jointly governs prefix construction and the form of supervision.

AuthorsOpenReview ↗

Response to Weakness 2

Dear Reviewer:

Thank you for your careful reading and constructive feedback.

Teacher confidence and intervention reliability. Teacher probability measures support for a student proposal. We directly test its value for intervention: at a rejected position, does committing the teacher's argmax improve the downstream outcome relative to retaining the student's proposal?

Using the same 128 randomly sampled training inputs per domain, we select each sequence's first CA rejection that actually changes the token, regardless of the SKD decision. We hold the preceding prefix fixed and greedily continue both branches under the same 1024-token cap. Both threshold endpoints are evaluated on the same SFT-init snapshot.

α≈Outcome measurePairsBetter / worse / unchangedMean benefit [95% CI]
0.50Grounding click success (percentage points)12620 / 3 / 103+13.49 [6.35, 20.63]
0.25Grounding click success (percentage points)12521 / 4 / 100+13.60 [6.40, 20.80]
0.50OCR 1−NED8749 / 25 / 13+0.0822 [0.0325, 0.1388]
0.25OCR 1−NED8446 / 25 / 13+0.0744 [0.0210, 0.1345]

The estimates are positive in both domains at both tested thresholds, with descriptive 95% paired-bootstrap intervals above zero. Grounding success denotes a click inside the reference bounding box, OCR uses the training reference. As a post-hoc sensitivity check, excluding the nine OCR pairs with at least one capped continuation at α≈0.25 still yields +0.0578 [0.0032, 0.1204] over 75 pairs.

These paired results support teacher probability as a useful signal for selecting interventions with positive average downstream utility under the evaluated teachers and data. The per-pair counts also quantify adverse interventions. Calibration and domain-shift considerations are discussed in Appendix I.

AuthorsOpenReview ↗

Response to Weakness 3

Dear Reviewer:

Thank you for your careful reading and constructive feedback.

Magnitude and consistency of gains. In the main setting, CA-OPD achieves higher mean scores than Offline KD on all six benchmarks below under the same initialization, training data, and optimizer budget (Table 1):

BenchmarkCA-OPD − Offline KD
ScreenSpot-v2+0.45
ScreenSpot-Pro+1.69
OCRBench-v2 English+0.91
OCRBench-v2 Chinese+0.31
CC-OCR+0.33
OmniDocBench+0.03

The gains extend across application and capability categories. The following tables reproduce the Offline KD and CA-OPD results from Appendix D.1, with differences added.

ScreenSpot-Pro by application category (Table 6).

CategoryOffline KDCA-OPDDifference
CAD40.0141.63+1.62
Creative38.3839.41+1.03
Dev38.9441.11+2.17
OS43.7545.75+2.00
Office64.2665.60+1.34
Scientific44.9147.03+2.12

OCRBench-v2 by capability category (Table 7).

CategoryOffline KDCA-OPDDifference
Recognition69.0269.42+0.40
Referring46.9348.32+1.39
Spotting22.3424.63+2.29
Extraction75.4876.16+0.68
Parsing49.2149.30+0.09
Calculation36.3336.59+0.26
Understanding54.4655.29+0.83
Reasoning37.5538.00+0.45

Referring, Spotting, and Calculation are English-only. CA-OPD exceeds Offline KD in all six application categories and all eight capability categories, showing that the improvements cover multiple subsets.

Preservation on OmniDocBench. Here, our focus is whether full-page document capability is retained after distillation. CA-OPD scores 70.43 compared with 70.49 for the SFT-initialized student, a change of −0.06 points. Offline KD scores 70.40, while OPD and SKD score 68.57 and 68.36. Thus, CA-OPD keeps document performance close to the initialization and Offline KD while achieving the grounding and OCR gains above.

In the additional Qwen3-VL setting, the differences from Offline KD are −0.28 on ScreenSpot-Pro, −0.29 on CC-OCR, and −0.46 on OCRBench-v2 Chinese. The improvement over on-policy baselines nevertheless carries over: CA-OPD exceeds both OPD and SKD on five of the six reported GUI/OCR benchmarks. For example, its ScreenSpot-Pro gains are +1.96 over OPD and +1.78 over SKD (Table 16). Together, the results support gains across categories and preservation of document capability in the main setting, alongside improvements over on-policy baselines in the additional setting, with task-dependent trade-offs relative to Offline KD.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.