acceptodds
Under review as a conference paper at ICLR 2027

Learn in Space: On-Policy Distillation via Spatial Divergence for Vision-Centric Reasoning

Abstract

Spatial awareness matters in vision-centric reasoning. Despite the effectiveness of its fine-grained supervision, on-policy distillation (OPD) remains spatially agnostic during optimization. Fundamentally, the same visual evidence can admit many spatially equivalent region descriptions, consequently leading to a vast number of distinct token serializations. Token-level KL treats these serializations as distinct outcomes in vocabulary space, which provides biased learning signals for spatially valid grounding behavior, ultimately limiting performance. To bridge these gaps, we introduce MixOPD, which uses Spatial Divergence to measure the geometric discrepancy between each on-policy student grounding proposal and corresponding teacher-supported evidence distribution estimated via Teacher-Forced Constrained Search, thereby spatially calibrating the token-level KL objective. MixOPD integrates seamlessly into typical OPD frameworks and delivers theoretically grounded performance gains. Extensive experiments demonstrate that MixOPD consistently improves both standard and entropy-aware OPD. It achieves highly competitive performance against representative post-training baselines and recent alternatives, while presenting favorable inference latency and stable training dynamics.

Primary area:
foundation or frontier models, including LLMs
Keywords:
On-Policy Distillation, Vision-Centric Reasoning, Large-Vision Language Model

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_KSZPOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper introduces MixOPD, an on-policy distillation framework for vision-centric reasoning that interleaves explicit grounding with the reasoning process. The main idea is that token-level distillation treats different coordinate sequences as different outputs, even when they point to the same visual evidence. As a result, geometrically valid grounding proposals can still be penalized. To address this, MixOPD estimates a teacher-supported box distribution through an efficient teacher-forced constrained search. It then weights each proposal's coordinate-token loss by its expected 1-IoU against this distribution. The paper also provides a theoretical analysis showing that this weighting acts as a selective local gradient scaling under stated assumptions. Experiments show that MixOPD improves both grounding and visual reasoning on top of standard and entropy-aware OPD, without extra teacher calls or search at inference time.

Critical strengths
  1. Strong motivation: The paper clearly identifies the mismatch between token-level distillation and spatial grounding, where different coordinate sequences can describe the same visual evidence yet receive unnecessary penalties.

  2. Simple and practical method: Spatial calibration integrates naturally with both OPD and EOPD, with no additional search or teacher calls at inference time.

  3. Sufficient abaltion studies: Comparisons with relevant distillation methods and representative post-training baselines adequately demonstrate improvements in both grounding and visual reasoning.

Critical weaknesses
  1. Presentation is a major weakness. The method section introduces too many formal definitions before clearly explaining the algorithm, making a simple method hard to follow. It should better explain why smaller spatial errors lead to lower loss weights. The search procedure is also too brief, with important details left to the appendix, and Figure 3 does little to clarify the intuition. A clear overview and a concrete example would substantially improve readability.

  2. Limited validation across model families: The method is evaluated only on Qwen3-VL students at the 4B and 8B scales. While this demonstrates consistency across model sizes, it does not establish generalization across model families. Experiments with another vision-language model family would strengthen the claim.

Questions for authors
  1. What is the distribution of bounding-box sizes, aspect ratios, and numbers per image in the training data? How does performance vary across these factors?
  2. Does MixOPD remain effective on video or multi-image reasoning tasks, where evidence may span multiple frames or images?
  3. Evaluation is limited within the Qwen3-VL family. Its effectiveness across other model families remains unverified.
4 comments
AuthorsOpenReview ↗

Response to Reviewer#KSZP (Part1/2)

Thank you for your helpful comments. We provide a point-by-point response below.


W1-a:Explain why smaller spatial errors lead to lower loss weights

A: Standard OPD encourages token-level agreement, which need not reflect spatial agreement. In Fig. 2, the student proposal (247, 305, 376, 437) and the teacher-preferred box (252, 298, 382, 441) localize the same button decoration, yet several coordinate positions incur substantial token-level penalties. MixOPD therefore calibrates these local signals using the expected spatial error under the teacher-supported box distribution. Smaller errors reduce unnecessary pressure to match exact coordinate tokens when the proposal already agrees spatially with the teacher, while larger errors retain more of the original distillation signal. We have made this intuition explicit in §3.2.

W1-b: The search procedure is also too brief, with important details left to the appendix, and Figure 3 does little to clarify the intuition.

A: To clarify the search procedure, we have added a step-by-step overview with a concrete grounding example in Fig. 3 (please see page 4 in the revised PDF) and reorganized the teacher-forced constrained search paragraph in §3.2 (L.243-L.248) to provide a detailed description of the algorithm. The new figure and the supplied description illustrate how cached teacher predictions are converted into a teacher-supported box distribution, connecting the token-level mismatch illustrated in Fig. 2 to the spatial calibration shown in Fig. 4.

W2&Q3: Validation limited to the Qwen3-VL model family.

A: Following your suggestion, we added experiments on the InternVL3.5 family. We initialize the student from InternVL3.5-8B and fix InternVL3.5-14B as the teacher model, following the training recipe in Appendix B.3. We compare MixOPD with representative post-training baselines: GRPO, DAPO, OPD, and EOPD. Complete results are provided in Table 2 (page 8) of the revised PDF, with selected results summarized below.

MethodmIoUTreeBVStarZoomHR-4KHR-8KMME-CNMME-ENMMStarCV-B
InternVL3.5-8B0.3442.4772.6041.1066.5061.3361.0363.2269.4581.65
GRPO0.4545.9375.0545.4964.5459.5557.2457.6263.5181.53
DAPO0.4846.1873.6544.0666.3862.8363.6362.1467.8482.80
OPD0.2846.1775.2249.3568.7164.7161.3464.7670.2082.75
MixOPD (OPD)0.4448.8981.5053.3373.1769.8363.6065.8870.8084.12
EOPD0.4148.3978.5352.6369.4266.5863.4166.1269.9183.93
MixOPD (EOPD)0.4550.3782.9056.2975.6370.5865.6966.7870.6984.66

We highlight two observations:

  • Spatial calibration improves grounding under both objectives. On InternVL3.5-8B, MixOPD raises mIoU from 0.28 to 0.44 under standard OPD and from 0.41 to 0.45 under EOPD.

  • Reasoning gains extend beyond Qwen3-VL. MixOPD also improves accuracy on all 9 VQA benchmarks over the corresponding OPD/EOPD baselines, supporting the effectiveness of spatial calibration across both evaluated model families.

AuthorsOpenReview ↗

Response to Reviewer#KSZP (Part2/2)

Thank you for your insightful questions. We provide a point-by-point response below.


Q1: What is the distribution of bounding-box sizes, aspect ratios, and numbers per image in the training data? How does performance vary across these factors?

A: Following your suggestion, we have added statistics on bounding-box sizes, aspect ratios, and counts per training sample, covering about 280k boxes. The full distributions are presented in Figure 16 (page 24 of the revised PDF).

Key statistics are summarized below:

FactorSummary
Box size (COCO area classes)20% small, 31% medium, and 49% large
Aspect ratio (width/height)The distribution peaks near 1:1, with most boxes having aspect ratios between 1:4 and 4:1 and tails extending to very tall (≤1:16) and very wide (≥16:1) boxes
Boxes per sample53.3% contain 1 box and 33.1% contain 2 boxes; 93.9% contain at most 4 boxes

Furthermore, we analyze MixOPD-4B/8B performance on TreeBench across evidence-box size, aspect ratio, and count. The tables report task-category-adjusted accuracy (%).

MixOPD-4B

Box size<4%4–7%7–12%12–25%≥25%
Accuracy51.352.342.655.062.5
Aspect ratio≤1:31:3–1:1.5~1:11.5:1–3:1≥3:1
Accuracy49.850.352.461.350.7
Boxes per image123≥4
Accuracy57.053.247.438.7

MixOPD-8B

Box size<4%4–7%7–12%12–25%≥25%
Accuracy55.258.453.352.349.2
Aspect ratio≤1:31:3–1:1.5~1:11.5:1–3:1≥3:1
Accuracy51.254.752.061.054.0
Boxes per image123≥4
Accuracy62.551.643.341.9

We have three observations:

  1. Box size. Neither model exhibits a consistent monotonic relationship between accuracy and box size. Despite the greater perceptual challenges posed by smaller evidence regions, accuracy shows no consistent decline as box size decreases, suggesting that MixOPD may enhance the base models’ ability to perceive and understand small visual evidence.

  2. Aspect ratio. Both models achieve their highest accuracy on moderately wide boxes (1.5:1–3:1), followed by a decline for extremely wide boxes (≥3:1). Notably, the median box aspect ratio in our training data is close to 1:1. We therefore hypothesize that this preference is inherited from the base models rather than introduced by MixOPD.

  3. Box count. Accuracy decreases monotonically as the number of target boxes increases, from 57.0% to 38.7% for 4B and from 62.5% to 41.9% for 8B. This is consistent with the greater difficulty of grounding and reasoning over multiple evidence regions.


Q2: Does MixOPD remain effective on video or multi-image reasoning tasks, where evidence may span multiple frames or images?

A: Thank you for raising this question. We added an evaluation on MuirBench [1] to assess multi-image generalization. Appendix D.1, Multi-image Understanding (page 36) of the revised PDF provides comprehensive quantitative results (Table 9) and qualitative examples (Figure 17). Selected results are summarized below:

[1] MuirBench: A Comprehensive Benchmark for Robust Multi-image Understanding. ICLR, 2025.

MethodBase Model#ParametersAccuracy (%)
TeacherQwen3-VL‡32B72.21
-Qwen3-VL‡235B72.64
BaseQwen3-VL‡8B65.31
OPDQwen3-VL8B67.86
MixOPDQwen3-VL8B74.05
EOPDQwen3-VL8B69.07
MixOPDQwen3-VL8B74.22

‡ Evaluated w/o special system prompts. OPD, EOPD, and MixOPD share the same system prompt (Appendix B.5).

We highlight three observations:

  1. Spatial calibration improves multi-image performance after single-image post-training. Both OPD and EOPD improve over the base models despite using only single-image data for post-training. MixOPD provides further gains over the corresponding OPD/EOPD baselines by 6.19/5.15 pp.

  2. MixOPD achieves competitive multi-image performance with smaller models. Under the EOPD formulation, the 8B variant reaches 74.22% with vision-centric reasoning, exceeding the direct-answering teacher by 2.01 percentage points.

  3. Visually grounded reasoning transfers to multi-image inputs. Figure 17 (p.36) illustrates how the reasoning pattern learned through single-image post-training supports cross-image counting and attribute-based comparison through localized visual evidence. Together with the quantitative results, these examples support the transfer of this reasoning pattern to multi-image scenarios, accompanied by consistent gains in answer accuracy.

These results directly support MixOPD's effectiveness in multi-image scenarios.


Sincerely, we hope our responses are helpful to address your concerns.

The author clearly address my concerns about the loss weight intuition, generalization across model families and multi-image evaluation. They also redesigned the Figure 3 and expanded the search algorithm description, which makes the paper easy to follow. So I raise my score the weak accept.

Dear reviewer#KSZP,

We sincerely appreciate that our feedback addresses your concerns. Your insightful comments are very important to further strengthen this work.

Best regards,

Submission 3146 Authors.

Reviewer_tBLZOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies OPD for vision-centric reasoning with explicit bounding-box grounding. It argues that standard token-level OPD is misaligned with spatial grounding because geometrically similar boxes can have different coordinate token serializations and thus receive inconsistent KL-based supervision. To address this, the paper proposes MixOPD, which computes a Spatial Divergence between a student-generated box and a teacher-supported box distribution. The teacher box distribution is approximated efficiently through teacher-forced constrained search. Experiments on several fine-grained visual reasoning and VQA benchmarks show improvements over OPD, EOPD, RL baselines, and several recent alternatives.

Critical strengths
  1. The paper clearly illustrates the mismatch between token-level OPD supervision and box-level geometric grounding. The visualizations are helpful and motivate why OPD can be useful for spatial reasoning.
  2. The paper is generally clear, well organized.
Critical weaknesses
  1. The proposed method uses an IoU-derived span-level coefficient to reweight coordinate-token OPD/EOPD losses. Similar ideas have appeared in IoU-aware loss reweighting and privileged grounding distillation. The main distinction claimed by this paper appears to be its use of on-policy, token-dense supervision; however, token-level grounding supervision may also have been explored in prior work, e.g., IoU-PD [1]. The paper would benefit from a concise but explicit comparison and discussion of the differences from these closely related approaches.
  2. Since the paper makes broad claims about improving grounding quality, I would like to see evaluation on standard referring-expression grounding benchmarks, including RefCOCO, RefCOCO+, and RefCOCOg. In particular, reporting the grounding performance of OPD, EOPD, and MixOPD throughout training, using metrics such as [email protected] and mIoU, would help clarify whether the proposed calibration improves general grounding ability.

[1] IoU-PD: IoU-Aware Privileged Distillation for Visual Grounding with Multimodal Large Language Models.

Questions for authors
  1. Figure 1 is currently misleading regarding the core mechanism. It implies that the method directly uses student-GT IoU as the spatial training signal. However, the text indicates that MixOPD actually computes its spatial coefficient using the expected IoU between the student proposal and a teacher-supported box distribution. Could you revise Figure 1 to accurately reflect this actual computational process to prevent reader confusion?
  2. The qualitative examples suggest that MixOPD may be particularly beneficial for small-object grounding. While not strictly required for acceptance, an evaluation on a dataset like RefDrone [2] would be highly informative. Achieving SOTA on it is not necessary, but such results would provide strong additional evidence for the generality and specific strengths of your proposed method.

I would be willing to increase the score if the authors can adequately address these concerns in the revised manuscript.

[2] RefDrone: A Challenging Benchmark for Referring Expression Comprehension in Drone Scenes

4 comments
AuthorsOpenReview ↗

Response to Reviewer#tBLZ (Part1/2)

Thank you for your helpful comments. We provide a point-by-point response below.


W1: Unclear distinction from prior work, particularly IoU-PD.

A: Although IoU-PD and MixOPD share the goal of incorporating spatial awareness into token-level optimization for VLMs, we clarify fundamental differences between them in target tasks, method design, and applicability.

IoU-PD is designed for visual grounding, where the GT box is well defined. MixOPD targets vision-centric reasoning, where multiple valid reasoning trajectories can lead to the same correct answer. Consequently, specifying GT boxes for every valid intermediate visual thought is generally impractical.

These differences motivate distinct method designs. IoU-PD constructs geometric supervision directly from GT box annotations. In contrast, MixOPD derives its spatial supervision proxy from teacher outputs, using GT boxes as optional teacher-side information to improve the quality of the teacher-supported box distribution.

Thus, IoU-PD provides a more direct optimization signal for grounding, but its reliance on GT boxes limits its direct applicability to settings where these annotations are unavailable. MixOPD retains its gains over vanilla OPD/EOPD even without GT box annotations (see Table 10, page 37 of the revised PDF). Following your suggestion, we have added this discussion to Sec 2.1 Related Work (lines 121–125) of the revised PDF to further clarify the distinctions between our method and related approaches.


W2: Insufficient evaluation and analysis of grounding performance in LVLMs through training under OPD, EOPD, and MixOPD.

A: Following your suggestion, we evaluated OPD, EOPD, and their MixOPD counterparts using Qwen3-VL-8B. We include RefCOCO+ from the standard RefCOCO family and broaden the evaluation with HC-RefLoCo, GroundingME, and Ref-Adv-s. We report [email protected] and mIoU every 100 training steps from 100 to 500. The results are summarized below, with training curves and further discussion in Figure 8 (page 10, lines 519–528) of the revised PDF.

Standard OPD formulation

step[email protected]mIoU
RefCOCO+HC-RefLoCoGroundingMERef-Adv-sRefCOCO+HC-RefLoCoGroundingMERef-Adv-s
Base-85.0994.1831.0047.2079.2883.4035.8643.11
100OPD85.7095.9533.9947.4379.9286.8238.5643.43
MixOPD87.3996.6934.6950.2581.0886.8439.9745.43
200OPD83.6497.2132.9949.8078.9788.0537.2644.63
MixOPD86.4296.2738.2851.4481.3386.9843.2946.69
300OPD86.3095.9532.2049.8881.4187.3236.9144.91
MixOPD86.5497.4239.6751.5980.9288.2544.4546.46
400OPD82.5597.0032.7948.6978.6588.1937.4244.10
MixOPD86.6697.2136.8851.8181.3488.3542.4446.53
500OPD84.2495.0131.2047.6579.5286.2135.8343.40
MixOPD86.3097.3237.6851.2981.4488.1342.8146.25

EOPD formulation

step[email protected]mIoU
RefCOCO+HC-RefLoCoGroundingMERef-Adv-sRefCOCO+HC-RefLoCoGroundingMERef-Adv-s
Base-85.0994.1831.0047.2079.2883.4035.8643.11
100EOPD88.7295.6432.1051.2182.2785.1737.3545.44
MixOPD85.4594.8033.1946.7679.8984.3337.9441.93
200EOPD88.0096.3735.1951.1581.8885.9040.0946.20
MixOPD86.5496.4834.2949.4380.5086.4739.8744.67
300EOPD87.1595.8531.3050.5481.9886.2535.9845.65
MixOPD89.4596.5834.7952.1183.2687.1839.6347.05
400EOPD86.6695.5433.0951.5181.4786.4537.1946.36
MixOPD88.2496.9038.0851.7482.8987.1043.0046.49
500EOPD86.0695.9531.7050.3281.0786.7435.6245.09
MixOPD87.6397.0038.6852.4882.5387.7543.4147.28

We highlight two key observations: 1. Under standard OPD, MixOPD outperforms OPD in 37 of 40 checkpoint–metric comparisons and improves both metrics over the base model on all four benchmarks at every evaluated checkpoint.

  1. Under EOPD, MixOPD generally trails the baseline at steps 100–200, but surpasses it on both metrics across all four benchmarks at steps 300–500, supporting complementary gains in later-stage grounding performance.
AuthorsOpenReview ↗

Response to Reviewer#tBLZ (Part2/2)

Thank you for your insightful questions. We provide a point-by-point response below.


Q1: Revise Figure 1 to clarify MixOPD’s computational process.

A: According to your suggestion, we have redrawn Figure 1 and revised its caption to show that spatial calibration depends on the teacher-supported box distribution, rather than the GT evidence box, providing clearer intuition for the readers. Please see Figure 1, page 1 of the revised PDF.


Q2: Additional evaluation on RefDrone would help assess MixOPD’s effectiveness in small-object grounding.

A: Following your suggestion, we conducted a cross-domain evaluation of MixOPD and its corresponding OPD/EOPD baselines on RefDrone. The results are summarized below; complete results and further discussion are provided in Appendix D.1 (Cross-Domain Evaluation), Table 8, page 35 of the revised PDF.

MethodBase ModelParamsF1_instAcc_instF1_imgAcc_imgN-acc
-Qwen3-VL†30B-A3B56.8839.9851.0935.58-
TeacherQwen3-VL32B57.4740.5451.4935.8740.38
-Qwen3-VL†235B58.7941.9352.1636.89-
StudentQwen3-VL4B52.9336.1846.8431.6132.69
OPDQwen3-VL4B52.4435.8046.9931.8235.26
MixOPDQwen3-VL4B54.3737.5348.8233.4236.54
EOPDQwen3-VL4B54.4637.6147.9632.6635.90
MixOPDQwen3-VL4B55.8338.9249.2933.8637.82
StudentQwen3-VL8B50.8034.4045.8631.7662.82
OPDQwen3-VL8B51.1134.6445.4931.2957.69
MixOPDQwen3-VL8B53.5036.8348.1933.5457.69
EOPDQwen3-VL8B52.0135.4947.3132.7556.41
MixOPDQwen3-VL8B55.1138.3348.9734.1255.13

† denotes results reported in the RefDrone paper.

We highlight three key observations:

  1. Benefiting from improved learning dynamics for intermediate grounding, MixOPD improves all four instance- and image-level grounding metrics over the corresponding OPD/EOPD baselines at both parameter scales.
  2. In contrast to the VQA results, the calibrated students still trail the 32B teacher on these grounding metrics. We attribute this gap in part to the VQA-only training data: localization improves through supervision of intermediate visual thoughts, providing a less direct learning signal than dedicated grounding tasks.
  3. The 32B teacher has substantially lower no-target accuracy (N-acc) than the original 8B student (40.38% vs. 62.82%). The 8B MixOPD variants also exhibit lower rejection accuracy than the original student. This may reflect the transfer of the teacher's tendency to localize plausible but non-matching candidates.

Sincerely, we hope our responses are helpful to address your concerns.

Thanks for your detailed response. Table 10 and Table 8 together address my concerns regarding the contribution. I also appreciate the authors presentation of the additional results. My main concerns have been addressed. I am willing to increase my score.

Dear reviewer#tBLZ,

We sincerely appreciate that our feedback addresses your concerns. Your insightful comments are very important to further strengthen this work.

Best regards,

Submission 3146 Authors.

Reviewer_42LHOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies a mismatch between token-level on-policy distillation and visual grounding: different coordinate strings can identify essentially the same evidence. MixOPD estimates a teacher-supported distribution of boxes and uses the spatial discrepancy between that distribution and each student proposal to reweight grounding-token supervision. Experiments with Qwen3-VL 4B and 8B under standard OPD and EOPD report improvements in localization and visual reasoning.

Critical strengths
  1. The paper identifies a concrete supervision mismatch: exact coordinate tokens do not reflect whether two predicted boxes refer to the same visual evidence. Its span-level calibration is well motivated and integrates into both standard OPD and EOPD without requiring teacher search at inference.
  2. The matched comparisons show useful gains in both grounding and answer accuracy across two student sizes. For example, under standard OPD at the 8B scale, adding spatial calibration raises grounding mIoU from 0.28 to 0.46 and VStar accuracy from 87.26% to 90.40%.
  3. The empirical analysis goes beyond headline averages. The paper examines teacher privileges, search settings, training dynamics, and three-seed paired comparisons, and it checks how closely the practical search reproduces the scalar signal from more expensive teacher search.
Critical weaknesses
  1. The theoretical guarantee concerns local gradient magnitude on a fixed rollout. It does not establish that the unchanged gradient direction corrects an inaccurate box or that global optimization improves grounding. The contribution should therefore be framed as geometry-selective attenuation rather than a general guarantee of spatially correct credit assignment.
  2. Training relies on a large privileged teacher, refined evidence annotations, and ground-truth answers. The teacher-privilege ablation is informative, but a matched OPD comparison under weaker privileges is needed to show that the incremental benefit of spatial calibration persists when this supervision is less available.
  3. A simpler geometry-aware control is missing. Direct box supervision, an IoU-weighted token objective, or a geometric reward under the same teacher, data, and compute would clarify how much of the gain requires the proposed teacher-box distribution and constrained search.
  4. The practical search is an approximation to candidate-conditioned teacher decoding. Its end-to-end coefficient agreement with beam search is moderate, and the retained candidate sets differ substantially. Hyperparameter ablations alone do not show whether approximation errors on difficult proposals affect downstream training.
  5. The strongest evidence concerns spatially demanding benchmarks and matched OPD variants. Several general-purpose gains are small or do not survive strict per-seed multiplicity correction, and the margin over the strongest alternative is narrower. The broad performance claims should be qualified accordingly.
Questions for authors
  1. What corrective behavior, if any, does the method guarantee when the base OPD gradient points in an unhelpful direction for an incorrectly grounded proposal?
  2. Under matched teacher inputs and compute, how does MixOPD compare with a simple IoU-weighted objective or direct box supervision?
  3. How much of the gain over OPD remains without ground-truth-answer privilege or evidence-annotation refinement? What is the cost of obtaining those training inputs?
  4. What are the search failure rate and the distribution of coefficient errors on difficult or teacher-misgrounded examples? Can these errors be related to downstream performance?
4 comments
AuthorsOpenReview ↗

Response to Reviewer#42LH (Part1/2)

Thank you for your constructive comments. We respond point by point.


W1: The theory guarantees local attenuation, not grounding correction or global improvement.

A: Our guarantee is explicitly local (Theorem 3.8, §3.3; Remark E.35): under its assumptions, calibration attenuates unnecessary coordinate-token pressure on valid proposals while retaining supervision for sufficiently separated errors. Table 2 shows grounding and reasoning gains across both objectives and model families. At step 500, Figure 8 shows gains on both grounding metrics across all four benchmarks under both objectives. Thus, the theory characterizes local calibration, while experiments demonstrate its practical benefits.


W2&Q3: Gains without GT-answer privilege or evidence refinement, and input preparation costs.

A: We added matched OPD/EOPD comparisons without GT-box privilege, fixing the teacher, data, answer privilege, and training recipe (Table 10, p.37).

No.MethodGT-box privilegemIoUTreeBVStarZoomMME-EN
1OPD—0.2648.9787.4351.2066.84
2MixOPD—0.4150.0489.1854.6367.69
3OPD✓0.2849.9687.2652.6667.31
4MixOPD✓0.4653.2590.4057.8368.27
5EOPD—0.3950.7087.7852.5867.62
6MixOPD—0.4252.0292.1558.5868.34
7EOPD✓0.4251.5289.7054.3667.87
8MixOPD✓0.4853.9992.5059.7669.98

The results show:

  • Without GT-box privilege, MixOPD improves every metric over both matched baselines (Rows 1–2, 5–6), gaining 1.75/3.43 pp over OPD and 4.37/6.00 pp over EOPD on VStar/Zoom.
  • It also matches or exceeds both baselines trained with GT-box privilege on every metric (Rows 2–3, 6–7).

Thus, GT-box privilege strengthens teacher guidance but is not essential to spatial calibration's gains.

Our source datasets provide GT answers, requiring no extra answer annotation. Automated evidence refinement (§B.1, pp.23–24; Figure 14) takes under one hour on eight H800 GPUs for all ≈145K samples, a one-time cost shared across runs.


W3&Q2: Simple IoU weighting, box supervision, or geometric rewards under matched teacher inputs, data, and compute.

A: We added a single-box IoU-weighted control with coefficient 1−IoU(bj,b^jMAP)1-\mathrm{IoU}(b_j,\hat b_j^{\mathrm{MAP}}) on Qwen3-VL 4B. Intermediate proposals may target reference objects or comparison candidates outside annotated answer-supporting regions (§3.2; Figure 21). Direct GT-box supervision can therefore misalign with these steps; we instead use the teacher's MAP box conditioned on the student's reasoning context.

MethodSpatial Supervision ProxymIoUTreeBVStarZoomHR-4KHR-8KMME-rw-CNMME-rw-ENMMStarCV-B
OPD-0.2647.0084.2949.1981.0875.6764.0467.4670.2285.60
MixOPDMAP Box0.4249.4687.9655.2780.6377.9665.0668.4570.8087.17
MixOPDTeacher-box Distribution0.4551.0389.7055.3881.4279.3866.1868.6770.4287.91
MethodSpatial Supervision ProxymIoUTreeBVStarZoomHR-4KHR-8KMME-rw-CNMME-rw-ENMMStarCV-B
EOPD-0.3949.7989.1853.4981.8878.2564.9667.6970.2988.43
MixOPDMAP Box0.4351.2889.5354.7981.9679.3466.5569.0170.8990.05
MixOPDTeacher-box Distribution0.4552.1890.9256.4982.8378.7967.6568.8471.5889.37

Both variants share teacher inputs, data, training recipe, and TFCS candidates; only the spatial coefficient changes. Distributional calibration improves 9/10 OPD and 7/10 EOPD metrics, raising mIoU 0.42→0.45 and 0.43→0.45, respectively. This supports retaining teacher-supported alternatives, especially for fine-grained tasks (as discussed in §A.2).


W4: The downstream impact of search approximation remains unclear, particularly on difficult proposals.

A: TFCS makes online spatial calibration feasible. Table 6 reports 67 s/span for width-128 beam search versus 0.01 s for TFCS. At 256 rollouts/cycle and 4.55 spans/rollout, serial extrapolation yields 21.7 hours/cycle and ≈512 days/145K-sample epoch for beam search alone, making full training prohibitive under this execution model. Beyond hyperparameter ablations, Figure 13 directly evaluates coefficient fidelity under same-support autoregressive rescoring and end-to-end beam search. TFCS avoids extra teacher decoding, with matched training gains documented in Table 2 and Figure 8.

AuthorsOpenReview ↗

Response to Reviewer#42LH (Part2/2)

W5: Qualify broad claims given modest general-purpose gains, limited corrected significance, and smaller margins over strong alternatives.

A: Our claim is consistent gains from spatial calibration, strongest on spatially demanding tasks (§4.2). Table 2 also shows positive mean gains in all 24 general-purpose comparisons (four benchmarks × two objectives × three student settings). On Qwen3-VL-4B, MME-CN/EN gains survive panel-wise Holm correction in all three seeds under both objectives (Figure 19). §D.3 distinguishes reproducible from inconclusive gains and acknowledges narrower margins over ZwZ (Figure 18). This supports competitive overall performance with benchmark-dependent gains.


Q1: What correction is guaranteed when OPD's gradient is unhelpful for an incorrect proposal?

A: Calibration rescales rather than redirects gradients: G~j=ωjGj\widetilde G_j=\omega_jG_j, ωj∈[0,1]\omega_j\in[0,1] (Theorem 3.8). EOPD's FKL adds recovery pressure toward retained teacher-supported alternatives under-covered by the student. Under its assumptions, Corollary E.38(ii) guarantees positive calibrated recovery pressure at gated positions on sufficiently separated errors; E.38(iii) extends this to the combined objective if OPD does not cancel FKL. MixOPD calibrates this signal's strength (Remark E.39), with conditional local guarantees rather than unconditional box correction.


Q4: Search failure rates, coefficient-error distributions on difficult or teacher-misgrounded examples, and their relation to downstream performance?

A: In the offline audit (§A.3), TFCS falls back on 1/1,000 well-formed spans (0.1%), retaining the OPD/EOPD objective. Against width-128 beam search, mean/median absolute coefficient errors are 0.144/0.079 (Figure 13). §A.2 (Figures 10–12) separately analyzes annotation-relative calibration errors across teacher/student grounding-error regimes. These diagnostics assess search fidelity and teacher reliability. Matched training results (Table 2; Figure 8) demonstrate overall effectiveness but do not establish a direct link between these diagnostics and learning dynamics.


Sincerely, we hope our responses are helpful to address your concerns.

Thanks for the detailed response and additional experiments. The matched comparisons without GT-box privilege and the MAP-box control substantially address my main concerns about the contribution of spatial calibration. I also appreciate the clarification of the theoretical scope. These additions strengthen the empirical support for the method, and I am raising my recommendation from weak accept to accept.

Dear reviewer#42LH,

We sincerely appreciate that our feedback addresses your concerns. Your insightful comments are very important to further strengthen this work.

Best regards,

Submission 3146 Authors.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We sincerely thank the reviewers and the AC for their time and careful reading. Reviewers found the motivation clear and concrete (KSZP, tBLZ, 42LH), the method simple and plug-and-play with no inference-time cost (KSZP, 42LH), and the experiments and analyses thorough (KSZP, 42LH).

Scope of revision. The proposed method, theory, and all empirical results reported in the original submission remain unchanged. Following the reviewers' suggestions, we add a cross-model-family validation and a grounding performance analysis to the main text, and refine its presentation in a few places. In addition, the appendix adds new results and analyses, offering a more complete understanding of our work. Revised content is highlighted in blue.

The changes to our manuscript are summarized below:


TypeAdd/ChangeLocationAddresses
FigureTableLinesPage#42LH#tBLZ#KSZP
Main Text
PresentationFigure 1 redrawn1--1-Q1-
PresentationOverview figure of teacher-forced constrained search with a concrete example; reorganized method description3-243–2484–5--W1
PresentationDiscussion of IoU-aware weighting methods (e.g., IoU-PD)--121–1253-W1-
ExperimentsCross-model-family Validation (InternVL3.5)-2401-4028--W2, Q3
AnalysisGrounding performance throughout training (RefCoCo+ etc.)8-519–52810-W2-
Appendix
AnalysisTraining-box statistics16--24--Q1
ExperimentsEvaluation on small-object grounding (RefDrone)-8-35-Q2-
ExperimentsEvaluation on multi-image understanding (MuirBench)179-36--Q2
ExperimentsAblation on GT-box privilege-10-37W2, Q3W1-
ExperimentsAblation on spatial supervision formulation-14-38W3, Q2--

All remaining comments are addressed in our individual responses. We would be happy to discuss further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.