acceptodds
Under review as a conference paper at ICLR 2027

G²-OPD: Grounded and Graspable On-Policy Distillation for Dynamic Spatial Reasoning

Abstract

Compact vision-language models are attractive for embodied applications, yet transferring dynamic spatial reasoning from stronger teachers remains challenging. Fixed-trace distillation suffers from a mismatch between teacher-provided reasoning prefixes and the student's own inference trajectories, while outcome-based reinforcement learning provides sparse corrective signals when successful rollouts are rare. We therefore study on-policy distillation (OPD), which supervises student-generated trajectories with dense teacher distributions. However, standard OPD treats reasoning positions uniformly, ignoring whether a prediction is grounded in spatiotemporal evidence and whether the corresponding correction is learnable by the student. To this end, we introduce G-OPD, a Grounded and Graspable OPD framework that adaptively allocates teacher supervision along the student's trajectory. Grounding uses counterfactual video probes to measure sensitivity to visual content, temporal order, and motion. Graspability captures corrections that are both needed and learnable by combining teacher–student disagreement with teacher support for the student's leading candidates. We then combine these two signals through a regularized supervision-allocation rule that redistributes full-vocabulary reverse-KL supervision while preventing excessive concentration. These signals reweight full-vocabulary reverse-KL supervision without task-specific process rewards or additional inference-time modules. Ultimately, G-OPD achieves 51.38% accuracy on VLM4D and 34.49% accuracy on DSR-Bench, improving over baseline by 11.02 and 12.12 percentage points, respectively.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Vision-Language Models, Dynamic Spatial Reasoning, On-Policy Distillation, Knowledge Distillation, Spatiotemporal Reasoning, Embodied AI

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_kkf4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work presents G2G^2-OPD, a grounded‑and‑graspable on‑policy distillation framework to transfer dynamic spatial‑reasoning capabilities from large vision‑language teacher models into compact student models. To address the uniform‑weight limitation of standard on‑policy distillation, G2G^2-OPD adaptively re‑allocates reverse‑KL supervision over the student’s self‑generated trajectories using two complementary signals: a grounding score derived from counterfactual video probes that identifies predictions relying on spatial, temporal‑order and motion evidence, and a graspability score combining teacher‑candidate support and local teacher‑student disagreement to locate learnable corrections. Without introducing extra inference‑time modules or task‑specific rewards, the 2B‑parameter student trained with G2G^2-OPD achieves substantial gains on VLM4D and DSR‑Bench benchmarks compared to baseline distillation and reinforcement‑learning approaches.

Critical strengths
  1. This work brings an important issue from the DPO field into the multimodal cognition field, namely that all tokens share the same training weights. This is indeed a problem that methods of this kind need to address.

  2. To distinguish the weights of different tokens, the paper suggests two signals that are quite interpretable.

  3. A lot of experimental results show that the 2B qwen model using this method has achieved significant performance gains.

Critical weaknesses
  1. The training costs increase noticeably. In fact, the DPO method is often accepted as a more efficient version of the GRPO method. But the approach proposed in the paper significantly raises the training costs of DPO (the teacher network has to run 4 forward passes for each sample), which weakens the original advantage of DPO methods.

  2. The proposed method introduces a large number of hyperparameter designs (evidence weights (alphaspa,alphaord,alphadyn,(gmaxalpha_{spa}, alpha_{ord}, alpha_{dyn}, (g_{max}), graspability's (gamma), candidate set size K, and quantile normalization thresholds), which to some extent limits the further widespread application of this method. The author should further analyze how sensitive the method is to the hyperparameters.

  3. Counterfactual probes are heuristically constructed, so it's hard to say if they'd still work in more general scenarios.

Questions for authors
  1. In fact, within the broader field of multimodal reasoning, there are quite a few ways to optimize DPO using path information. So what advantages does our approach have compared to others?

  2. As far as I know (the author's main experiment also proves this), Qwen3 itself is a really excellent multimodal baseline model. This makes me curious whether the method would still work on other multimodal baseline models (LLava, Qwen3.5, Qwen2.5-VL).

  3. Actually, I have a relative understanding of work in the MLLM4D field. I know that the outputs of MLLM4D include both text and physical-side data. I want to know which tokens are the most critical in your setup. Does this phenomenon also occur with other multimodal models that could act as teachers?

6 comments
AuthorsOpenReview ↗

Response to Reviewer kkf4 (1/3)

Thank you for recognizing the importance of dynamic spatial reasoning, the interpretability of our weighting signals, and the gains achieved by G²-OPD. We clarify the computational trade-off, quantify hyperparameter sensitivity, and provide additional evidence for probe selectivity, broader video understanding, and cross-model generalization. Counterfactual probes shape the training-time allocation; deployment retains the original student inference architecture and the original full-vocabulary reverse-KL teacher target.

W1. Computational overhead and matched teacher-compute comparison

G²-OPD evaluates four frozen-teacher conditions (the original video and three counterfactual views) on the same student-generated response and prefixes. The student generates only once; the additional computation is teacher-side scoring. The student still performs one backward pass per microbatch, with 42 gradient-accumulation microbatches per rank per optimizer update.

MetricUniform OPDG²-OPDRelative overhead
Teacher forward conditions / rollout144.00×
Teacher FLOPs (4 fixed samples)729.861 TFLOPs2919.443 TFLOPs4.00×
Total training + rollout GPU-hours220.73325.30+47.37%
Peak allocated GPU memory (microbenchmark)23.62 GiB24.61 GiB+4.19%
Student backward passes / microbatch111.00×
Student backward passes / update / rank42421.00×

Compared with Uniform OPD, G²-OPD improves by 5.18 points on VLM4D and 5.85 points on DSR-Bench. Teacher-specific FLOPs increase 4×, while total training and rollout GPU-hours increase 47.37% (220.73 → 325.30), because student generation and optimization are shared. Peak memory increases 4.19%: the frozen teacher is reused sequentially under no-grad execution, and intermediate full-vocabulary logits are staged on CPU, avoiding simultaneous GPU residency of four teacher outputs. Teacher FLOPs and peak memory are measured on four fixed paired samples.

To isolate the benefit of allocation from the extra teacher-forward budget, we compare with 4×-rollout Uniform OPD:

MethodTeacher conditions / rolloutRelative rollout budgetVLM4D Acc.DSR Acc.
Uniform OPD (original)11×46.20%28.64%
Uniform OPD (4× rollout)14×48.68%30.52%
G²-OPD (original)41×51.38%34.49%

Under the matched teacher-forward budget, G²-OPD retains gains of 2.70 points on VLM4D and 3.97 points on DSR-Bench, quantifying the benefit of allocating supervision across the student-generated trajectory.

W2. Hyperparameter sensitivity and concentration control

We use one shared configuration across VLM4D and DSR-Bench: αspa=αord=αdyn=0.25\alpha_{\rm spa}=\alpha_{\rm ord}=\alpha_{\rm dyn}=0.25, γ=0.25\gamma=0.25, K=16K=16, and gmax⁡=2g_{\max}=2. The three α\alpha's control spatial, order, and motion evidence; γ\gamma controls the extra graspability emphasis; KK defines the local candidate region. Grounding scores are clipped at their response-wise 95th percentile before positive-score normalization. Teacher support and disagreement are separately scaled by their 5th and 95th percentiles and clipped to [0,1], reducing outlier influence and putting the two signals on comparable scales.

AuthorsOpenReview ↗

Response to Reviewer kkf4 (2/3)

Hyperparameter sensitivity:

ParameterSettingVLM4DDSR-Bench
γ\gamma0.1050.6133.59
γ\gamma0.25 (default)51.3834.49
γ\gamma0.5050.8333.87
KK850.5533.45
KK16 (default)51.3834.49
KK3251.0534.08
α\alpha (all channels)0.12550.1733.24
α\alpha (all channels)0.25 (default)51.3834.49
α\alpha (all channels)0.5050.4433.38

Across the tested ranges, maximum variation is 1.21 points on VLM4D and 1.25 on DSR-Bench. The response-local weight rankings are also stable under changes in KK and γ\gamma: Spearman correlations range from 0.9841–0.9944 at initialization and 0.9889–0.9955 at the trained checkpoint; Jaccard similarities range from 0.9059–0.9778.

Theoretical concentration is controlled as well. Corollary 1 shows that, with gmax⁡=2g_{\max}=2 and γ=0.25\gamma=0.25,

NeffT≥4049≈0.816.\frac{N_{\rm eff}}{T}\geq\frac{40}{49}\approx0.816.

Since gt∈[1,2]g_t\in[1,2] and ht∈[1,1.25]h_t\in[1,1.25], the raw priority gthtg_th_t lies in [1,2.5][1,2.5], which yields the bound. The effective number of supervised positions therefore remains at least 81.6% of the response length, characterizing a controlled reallocation of dense supervision.

W3. Counterfactual probe semantics and generality

The three probes isolate distinct evidence requirements of dynamic spatial reasoning:

  • Spatial masking (V): masks the same spatial patches across frames, weakening object-identity and reference-point cues.
  • Temporal-order perturbation (O): reverses or locally reorders observations, testing temporal direction and event order.
  • Local-motion suppression (M): repeats a representative frame within four-frame windows, preserving coarse scene progression while suppressing fine displacement, speed, and orientation changes.

The semantic-by-channel results in Figure 3(a) show content-dependent responses: forward has spatial/order evidence (+0.44/+0.26), moving local-motion evidence (+0.38), closer spatial evidence (+0.94), and before order evidence (+0.46). Relational judgments use multiple channels: approach has motion/order evidence (+1.10/+1.24), and turn has (+0.63/+0.88). These channel-specific responses reveal semantic differentiation across token categories.

Figure 4 quantifies each channel's contribution to response-local allocation: removing the spatial, motion, or order channel decreases the G-only KL allocation uplift by 2.762, 2.084, and 1.445 percentage points, respectively. Figure 6 provides a word-level check: grounding enrichment over matched format tokens is +0.15 for moving, +0.17 for static, and +0.27 for both forward and turn, with positive 95% bootstrap intervals.

We further compare spatial masking against mild photometric perturbations:

GroupSpatial VVPhotometric CCV−CV-C95% CI
All 900.13670.0314+0.1053[0.0695, 0.1459]
New 320.15830.0344+0.1240[0.0665, 0.1930]

Spatial masking produces a larger evidence response than mild photometric changes across both sample groups, with positive confidence intervals.

To test transfer beyond specialized spatial benchmarks, we additionally train G²-OPD using general video data and evaluate the resulting Qwen3-VL-2B on Video-MME:

BenchmarkBase Qwen3-VL-2BG²-OPD-trained Qwen3-VL-2BImprovement
Video-MME61.9%66.7%+4.8 pp

This result provides additional evidence that the training strategy can benefit broader video understanding, beyond the specialized dynamic-spatial tasks. Together, semantic-by-channel responses, word-level enrichment, channel ablation, photometric controls, and Video-MME support the probes as structured measures of spatiotemporal evidence dependence.

AuthorsOpenReview ↗

Response to Reviewer kkf4 (3/3)

Q1. Methodological distinction and simpler weighting baselines

G²-OPD uses the student's own reasoning trajectory to distinguish disagreements about the video from disagreements about wording. Consider two positions with the same teacher–student KL divergence. One concerns whether an object is approaching or departing. The other concerns how to phrase the explanation. KL weighting gives the two positions the same priority based on disagreement, and entropy weighting gives them the same priority if the student is equally uncertain. Random weighting can assign different weights, but the difference carries no information about which judgment depends on the video.

The frame-order probe gives G²-OPD a way to distinguish the two positions even when their KL divergence is the same. If disrupting frame order affects the teacher's score at the direction judgment more than at the wording choice, that difference can raise the relative grounding weight of the direction judgment. Candidate support then helps determine the strength of the correction. Two equally uncertain positions can receive different graspability weights if the teacher's probability mass overlaps more with the student's candidates at one position. The joint weight uses both distinctions, retaining full-vocabulary reverse-KL supervision at every position and keeping the total weight per response fixed.

wt=Tgtht∑t′gt′ht′.w_t=T\frac{g_th_t}{\sum_{t'}g_{t'}h_{t'}}.

Comparison with simpler weighting baselines

MethodVLM4DDSR-Bench
Uniform OPD46.2028.64
Random weighting46.7029.20
Entropy weighting47.1929.69
KL weighting47.9630.31
Grounded-OPD49.1731.92
Graspable-OPD48.4631.29
G²-OPD51.3834.49

Under matched response-wise normalization and weight-concentration constraints, G²-OPD improves over KL weighting by 3.42 and 4.18 percentage points on VLM4D and DSR-Bench, respectively. It also improves over Grounded-OPD by 2.21 and 2.57 points and over Graspable-OPD by 2.92 and 3.20 points. The KL comparison measures the gain over weighting by disagreement alone, and the single-component comparisons show the benefit of using both signals on the same trajectory.

Q2 & Q3. Generalization across teacher–student pairs and token categories

We evaluate an additional pair, Qwen2.5-VL-32B → Qwen2.5-VL-3B, with verified vocabulary alignment:

ModelRoleVLM4D OverallDSR-Bench Micro
Qwen2.5-VL-32BTeacher reference53.03%30.31%
Qwen2.5-VL-3BBase student42.29%19.79%
Qwen2.5-VL-3BG²-OPD-trained student47.58%24.88%
Improvement over base+5.29 pp+5.09 pp

G²-OPD consistently improves both benchmarks with this additional teacher–student pair, extending the evidence beyond the original MLLM-4D/Qwen3-VL setting to Qwen2.5-VL.

The objective is defined over aligned next-token distributions in the shared vocabulary. Physical information serialized in the response—direction, distance, depth, and numerical descriptions—is supervised within this same full-vocabulary objective.

The original semantic-by-channel analysis identifies differentiated evidence for direction/trajectory (forward: V +0.44, O +0.26), motion state (moving: M +0.38), distance/depth (closer: V +0.94), speed (faster: M +0.12, O +0.54), temporal order (before: O +0.46), relative motion (approach: M +1.10, O +1.24), and rotation (turn: M +0.63, O +0.88). At the complete-word level, moving, static, forward, and turn have positive enrichment over matched format tokens (+0.15, +0.17, +0.27, and +0.27, respectively).

We also check token-level transfer on the Qwen2.5-VL-32B → Qwen2.5-VL-3B pair. Direction/trajectory tokens again show positive enrichment under spatial evidence (+0.484, 95% CI [0.233, 0.772]) and temporal-order evidence (+0.301, 95% CI [0.094, 0.514]). Motion-state tokens have a positive local-motion estimate (+0.208, 95% CI [−0.003, 0.448]), although this interval slightly overlaps zero. These statistics indicate which token categories receive greater supervision priority.

Revision: The revised manuscript adds cost and teacher-forward-budget comparisons (Section 5.8, Tables 7 and 11), weighting controls (Section 5.3, Table 5), parameter sensitivity (Appendix D.5, Tables 13–14), and cross-model accuracy and token analyses (Appendix D.6, Tables 15–16).

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

This paper studies on-policy distillation for dynamic spatial reasoning in compact vision-language models. A Qwen3-VL-2B student learns from a frozen MLLM-4D-8B teacher using reasoning prefixes generated by the student itself. The proposed method, G²-OPD, reweights the full-vocabulary reverse KL loss using two signals. Grounding measures how the teacher’s predictions change under spatial masking, temporal reordering, and local motion suppression. Graspability combines teacher–student disagreement with the teacher’s support for the student’s top-K candidates. The resulting weights are normalized within each response, while the teacher targets and the student’s inference architecture remain unchanged.

The method achieves 51.38% on VLM4D and 34.49% on DSR-Bench, improving over uniform OPD by 5.18 and 5.85 percentage points, respectively. The paper includes a complete ablation of grounding and graspability, along with an analysis of the resulting supervision weights.

Critical strengths
  1. The gains over uniform OPD are meaningful and supported by useful component ablations. Table 4 reports all four combinations of grounding and graspability. Each component improves over the uniform OPD baseline, and the combined method performs best on both benchmarks. The gains of 5.18 percentage points on VLM4D and 5.85 percentage points on DSR-Bench provide more direct evidence for the proposed weighting method than the larger gains over the base student model alone. This is the paper’s strongest experimental evidence.

  2. The method improves a compact student model without adding inference-time modules. The counterfactual teacher evaluations are used only during training. The method also preserves the original teacher distribution and retains supervision at every response position. These choices make it a practical modification to an existing OPD pipeline, with the added complexity concentrated in how training supervision is allocated.

Critical weaknesses
  1. The training comparisons need to be documented more clearly. Appendix D identifies the teacher, student, training dataset, and evaluation protocol, and Table 6 specifies the weighting parameters. However, the provided PDF does not give the complete optimization schedule, effective batch size, video input settings, rollout settings, or training budget for each comparison. Some details are deferred to the run logs. The benchmark tables also do not report variability across runs. One particularly important example is the comparison between TA-OPD and the proposed Graspable-OPD variant. Their VLM4D scores are nearly identical, but their DSR-Bench scores differ substantially: 23.48 versus 31.29 (Tables 1–2 and Table 4). Since the graspability mechanism explicitly builds on TA-OPD’s use of disagreement and support, this discrepancy needs an explanation. At present, it is difficult to determine how much of the difference comes from the weighting rule and how much might come from differences in the training setup.

  2. The mechanistic analyses provide more direct evidence about the behavior of the weights than about the claimed interpretation of the signals. Figure 4 removes individual grounding channels when recomputing allocation statistics on fixed completions. It does not measure downstream accuracy after training without those channels. Likewise, showing that graspability weights are high when both support and disagreement are high largely follows from the definition of the weighting rule. The proposition concerning candidate support provides useful motivation, but it does not establish how easily the student network can learn a particular correction. The complete two-factor ablation supports the usefulness of the reweighting, so this concern does not invalidate the empirical results. However, the distinct value of the temporal-order and local-motion probes over simpler visual weighting has not been fully isolated. A matched spatial-only baseline or a VA-OPD-style baseline would clarify the incremental contribution. Without such evidence, I would interpret “graspability” as a useful proxy signal and narrow the mechanistic claims accordingly.

Questions for authors
  1. Clarification of training controls: Could you provide, or point to, the complete configurations for Tables 3 and 4, including optimization, rollout generation, video sampling, and update budgets? Which settings are shared across methods? Are the reported benchmark results from single runs? What is the measured training overhead of the additional teacher evaluations?

  2. Clarification of the closest baseline: Were the TA-OPD results obtained using the same teacher, student, data, and training schedule as Graspable-OPD? Given their nearly identical VLM4D scores, how do you explain the gap between 23.48 and 31.29 on DSR-Bench?

  3. Evidence for the temporal extensions: Do you have downstream accuracy results comparing spatial-only grounding with the full spatial/temporal-order/local-motion scheme under the same training setup, or a matched comparison with VA-OPD? Figure 4 reports changes in allocation statistics and therefore does not directly establish the contribution of the additional temporal channels to accuracy.

5 comments
AuthorsOpenReview ↗

Response to Reviewer qvER (1/3)

Thank you for emphasizing the value of the complete two-factor ablation and the unchanged student inference architecture. We clarify the shared training controls and run variability and add downstream evidence for the temporal grounding channels.

W1 & Q1. Training controls, variability, and computational cost

The main controlled experiments use MLLM-4D-8B as teacher, Qwen3-VL-2B as student, and MLLM4D-R1-30k for training. The shared method and scoring configuration is:

ComponentConfiguration
Distillation objective / coefficientFull-vocabulary reverse KL; GKD coefficient 1
Teacher targetOriginal teacher distribution; target strength 1.0
Grounding coefficients / capαspa=αord=αdyn=0.25;gmax⁡=2\alpha_{\rm spa}=\alpha_{\rm ord}=\alpha_{\rm dyn}=0.25;\\ g_{\max}=2
GraspabilityK=16,γ=0.25K=16,\\ \gamma=0.25
Quantile normalizationGrounding: positive-score Q0.95Q_{0.95}; support/disagreement: Q0.05/Q0.95Q_{0.05}/Q_{0.95}
Spatial probe50% spatial patches; one deterministic mask shared across frames
Temporal-order probe50% reversal, 50% local shuffle
Temporal windowsFour frames for local shuffle and local-motion suppression
Weight normalization / reductionMean one per response; global active response-token mean
Gradient path / teacher executionStudent distribution in KL; four frozen-teacher conditions run sequentially

The optimization and rollout settings are also shared across controlled OPD comparisons:

ParameterConfiguration
OptimizerFused AdamW; weight decay 0.1; β=(0.9,0.95)\beta=(0.9,0.95); ϵ=10−8\epsilon=10^{-8}
Learning rate / schedule5×10−65\times10^{-6}, cosine decay, 0.1% warmup, gradient clipping 0.1
Batch size1 per GPU per microbatch; nominal global batch size 336 per update
Accumulation / distributed training42 microsteps, 8 training ranks, ZeRO-2
Video input32 frames per video
Rollout samplingTemperature 1.0, top-pp 1.0, top-kk disabled, max. 4,096 new tokens
Training / rollout budget357 updates; generation batch 96; 12 microsteps per generation batch; 8 generations per prompt

For Uniform, Grounded, Graspable, and G²-OPD, teacher, student, data, evaluation splits, answer scoring, base OPD objective, and optimization schedule are fixed; only the weighting signal or grounding channels change. Offline distillation and GRPO in Table 3 are different training paradigms. The original benchmark results in Tables 3 and 4 are from single runs. We report mean ± standard deviation across three seeds, with paired gains computed from per-seed differences:

BenchmarkUniform OPD (%)G²-OPD (%)Paired gain (pp)
VLM4D46.05 ± 0.3051.16 ± 0.44+5.10 ± 0.41
DSR-Bench28.52 ± 0.3134.36 ± 0.46+5.83 ± 0.46
AuthorsOpenReview ↗

Response to Reviewer qvER (2/3)

Computational overhead. Each student response is generated once. The four frozen-teacher conditions score the same response prefixes from the single student-generated rollout. The teacher conditions run sequentially under no-grad execution, and intermediate full-vocabulary logits are staged on CPU. The student performs one backward pass per microbatch, with 42 gradient-accumulation microbatches per rank per update.

MetricUniform OPDG²-OPDRelative overhead
Teacher forward conditions / rollout144.00×
Teacher FLOPs (4 fixed samples)729.861 TFLOPs2919.443 TFLOPs4.00×
Total training + rollout GPU-hours220.73325.30+47.37%
Peak allocated GPU memory (microbenchmark)23.62 GiB24.61 GiB+4.19%
Student backward passes / microbatch111.00×
Student backward passes / update / rank42421.00×

Teacher FLOPs increase 4×, but total training and rollout GPU-hours increase 47.37%; peak memory increases 4.19% in the fixed-sample microbenchmark. Student generation and optimization are shared across the teacher conditions. Teacher FLOPs and peak memory are measured on four fixed paired samples.

We additionally compare against 4×-rollout Uniform OPD to match the total teacher-forward budget:

MethodTeacher conditions / rolloutRelative rollout budgetVLM4D Acc.DSR Acc.
Uniform OPD (original)11×46.20%28.64%
Uniform OPD (4× rollout)14×48.68%30.52%
G²-OPD (original)41×51.38%34.49%

G²-OPD retains gains of 2.70 points on VLM4D and 3.97 points on DSR-Bench under the matched teacher-forward budget. This matches teacher-forward conditions; the GPU-hour table separately reports end-to-end training and rollout cost.

Q2. TA-OPD Reproduction and DSR-Bench Mechanism Analysis

We reproduced TA-OPD following Wang et al. (2026) and report our own runs in Tables 1–2. We used the same teacher, student, training data, optimization schedule, and evaluation protocol as Graspable-OPD. TA-OPD selects positions by teachability, while Graspable-OPD uses teacher support and disagreement to continuously reweight distillation at every response position. Both use one intact-video teacher condition. Grounding adds three counterfactual conditions in full G²-OPD.

Graspable-OPD scores 48.46% versus 48.51% for TA-OPD on VLM4D. On DSR-Bench, it raises Micro from 23.48% to 31.29% (+7.81 pp) and Macro from 23.23% to 31.56% (+8.33 pp). Compared with Uniform OPD, it gains 2.26 pp on VLM4D Overall and 2.65 pp on DSR-Bench Micro.

DSR-Bench question structure. Across 1,435 questions, many combine coordinate constraints and time intervals, and all specify a reference perspective. These interdependent judgments motivate retaining supervision across the reasoning trajectory.

CharacteristicNumber of questionsPercentage
Explicit object-coordinate constraints1,22485.3%
Explicit temporal intervals1,15380.3%
Specified reference perspective1,435100%
Total questions1,435100%

Teacher support and disagreement. Within high-disagreement positions, disagreement-only weighting slightly favors low support (1.124 vs. 1.106). The product favors high support (1.110 vs. 0.990).

Weighting strategyHigh disagreement + high supportHigh disagreement + low support
Uniform1.0001.000
Support-only1.0120.868
Disagreement-only1.1061.124
Product (ours)1.1100.990

High/low is defined at 0.5 after response-local scaling. The high- and low-support columns include 45 and 58 responses, respectively.

AuthorsOpenReview ↗

Response to Reviewer qvER (3/3)

KL supervision. On fixed DSR-Bench prefixes, high-support, high-disagreement positions gain 1.609 pp of KL share (95% CI [1.28, 2.09]).

RegionResponses containing the regionMean normalized weightChange in KL share (pp)
High disagreement + high support131.101+1.609
High disagreement + low support261.000-0.632
Low disagreement + high support260.991-0.058
Low disagreement + low support100.990-0.296

The weighting comparison shows how support and disagreement affect priority. The DSR-Bench analysis reports the resulting KL-share changes across the four support–disagreement groups.

Graspability coefficient. Increasing γ shifts more KL supervision to high-support, high-disagreement positions at 100% coverage.

γSupervision coverageKL share of high-support, high-disagreement positions
0100%12.00%
0.1100%12.33%
0.25 (ours)100%12.80%
0.5100%13.52%
1.0100%14.77%

Averages include all 26 fixed DSR-Bench responses, with zero shares assigned when the high-support, high-disagreement group is absent.

W2 & Q3. Downstream contribution of temporal grounding channels and post-training alignment

Figure 4 in the paper isolates each channel's effect on response-local allocation by recomputing weights on fixed completions. We complement this mechanism diagnostic with trained variants using different grounding configurations while keeping graspability fixed:

Grounding configurationGraspabilityVLM4D Acc.DSR Acc.
Spatial only (V)Off47.85%30.10%
Spatial only (V)On49.72%32.40%
Spatial + Order (V+O)On50.50%33.31%
Spatial + Motion (V+M)On50.72%33.52%
Spatial + Order + Motion (V+O+M)On51.38%34.49%

With graspability fixed, adding temporal order to spatial grounding improves accuracy by 0.78/0.91 points, and adding local motion improves it by 1.00/1.12 points on VLM4D/DSR-Bench. Combining both temporal channels gives the best results on both benchmarks. These trained variants directly quantify the downstream accuracy of each grounding configuration.

We also compare the cold-start and step-357 student checkpoints on the same 58 videos, questions, response prefixes, tokens, and teacher views:

ScopeResponsesInitial KLFinal KLKL reduction95% CI
Overall581.06190.50000.5619[0.4872, 0.6368]
VLM4D321.31060.51150.7991[0.7463, 0.8528]
DSR-Bench260.75580.48580.2700[0.2319, 0.3089]

Mean relative KL reduction is 50.07%; 75.05% of token positions show reduced KL, and token-mixture KL decreases from 1.0773 to 0.4395. The trained checkpoint also reduces fixed-context KL on 32 additional videos spanning real first-person, real third-person, synthetic direction, and synthetic false-positive-control scenarios.

Proposition 1 characterizes candidate support as distributional compatibility. We complement this interpretation with paired checkpoint replay, which measures post-training distribution alignment, and the grounding-channel ablation, which quantifies downstream accuracy. Together, these analyses connect the allocation signals to optimization outcomes and benchmark performance.

Revision: The revised manuscript adds complete training settings (Appendix C.4, Table 10), three-seed results (Appendix D.4, Table 12), and temporal-channel training ablations (Section 5.7, Table 6). Appendix E reports probe controls and KL analyses (Tables 17–21).

Reviewer_tXiTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces G2-OPD for transferring dynamic spatial reasoning to a compact vision-language model. It reweights on-policy distillation using counterfactual video sensitivity and teacher-supported student disagreement, while retaining full-vocabulary reverse-KL supervision. Results on two benchmarks and component ablations support its practical value. Overall, I lean toward acceptance, although the interpretation of the weighting signals and the computational fairness of the comparisons need stronger validation.

Critical strengths

The method addresses a clear limitation of uniform OPD through supervision allocation tailored to spatial content, temporal order, and local motion, without adding inference-time modules.

With the same 2B student, G2-OPD improves over Uniform OPD by 5.18 points on VLM4D and 5.85 points on DSR-Bench. Both components help individually, and their combination performs best.

Critical weaknesses

W1. Limited validation of grounding and learnability Counterfactual sensitivity does not necessarily establish correct visual grounding, since perturbations can introduce distribution shifts. Likewise, candidate support and disagreement do not directly establish learnability. The propositions justify distributional quantities and allocation, but do not connect the scores to subsequent learning gains. Fixed-completion analyses partly confirm behavior built into the formulas; measuring actual improvement at high-scoring positions would strengthen the mechanism claims.

W2. Unclear incremental value under matched training cost Graspability builds on TA-OPD, and the distinct value of the combined allocation needs sharper isolation. The reported comparisons do not establish whether gains persist under matched teacher-compute budgets despite the additional counterfactual probes. Random weighting and simple disagreement-based weighting would also help test the specificity of the design. Evaluation with only one student architecture further limits evidence for generality.

Questions for authors

Please refer to the Weaknesses section above.

3 comments
AuthorsOpenReview ↗

Response to Reviewer tXiT (1/3)

Thank you for recognizing the gains over Uniform OPD and the motivation for trajectory-level supervision allocation. We add evidence for the interpretation and post-training effects of the two signals, and isolate the benefit of allocation from extra teacher computation and generic weighting.

W1. Grounding validity, graspability, and downstream learning effects

Grounding tracks task-relevant visual evidence

Grounding measures how much the teacher's support for a sampled token changes when a particular source of spatiotemporal evidence is disrupted. The semantic-by-channel analysis in Figure 3(a) shows that the response is content-dependent: forward responds to spatial and order evidence (+0.44 and +0.26), moving to local-motion evidence (+0.38), and closer primarily to spatial evidence (+0.94). Relational judgments such as approach depend jointly on motion and order (+1.10 and +1.24), while turn responds to both (+0.63 and +0.88). Figure 6 also shows positive grounding enrichment over matched format tokens for moving (+0.15), static (+0.17), forward (+0.27), and turn (+0.27).

We further test whether high GG identifies claims that require visual evidence. On 16 controlled depth-motion videos involving approach/departure judgments, we annotate this requirement and evaluate the ranking induced by GG.

MetricResult
AUROC0.834
Precision@Top-20%100%
Spearman correlation0.579
Evidence-demand rate in bottom 20%21.9%

High-GG positions concentrate on claims requiring visual evidence, while the bottom 20% contains substantially fewer such claims.

Counterfactual semantic consistency. In a separate temporal-reversal test, we reverse each video while keeping answer-option order unchanged. The answer should change for 48 motion question–video pairs and stay the same for 16 static left–right relation pairs and 64 color-control pairs. Depth prompts fix the camera and the ball's physical size to reduce scale ambiguity. The conditional option probe compares A/B/C/D logits after <answer>, normalized over the four options. We report free-generation results separately below.

GroupQuestion-video pairsOriginal acc.Reversed acc.Both correctChange rateMean JS (nats)
Motion48100.0%100.0%100.0%100.0%0.655731
Static relation16100.0%100.0%100.0%0.0%0.000053
Color control64100.0%100.0%100.0%0.0%0.000001

All motion predictions reverse, while static-relation and color predictions remain unchanged. A predetermined free-generation cross-check (indices=[0,3,4,7], with 12 motion, 4 static, and 16 color cases) reproduces the same pattern:

GroupCasesOriginal acc.Reversed acc.Both correctChange rate
Motion12100.0%100.0%100.0%100.0%
Static relation4100.0%100.0%100.0%0.0%
Color control16100.0%100.0%100.0%0.0%

We also compare spatial masking with mild photometric perturbations to distinguish structural sensitivity from generic appearance sensitivity:

GroupSpatial VVPhotometric CCV−CV-C95% CI
All 900.13670.0314+0.1053[0.0695, 0.1459]
New 320.15830.0344+0.1240[0.0665, 0.1930]

The positive paired differences, together with the evidence-demand annotations and temporal-reversal controls, show that grounding responds to task-relevant visual structure and controlled semantic changes.

Graspability and actual post-training alignment

Proposition 1 gives candidate support a distribution-level interpretation. If AtA_t is the student's candidate set and st=∑a∈Atqt(a)s_t=\sum_{a\in A_t}q_t(a) is teacher mass on that set, then

min⁡supp⁡(ν)⊆AtDKL(ν∥qt)=−log⁡st.\min_{\operatorname{supp}(\nu)\subseteq A_t}D_{\rm KL}(\nu\Vert q_t)=-\log s_t.

Larger support therefore means a closer teacher match is available within the student's current candidate region, characterizing local distributional compatibility. We assess the resulting learning effect through paired checkpoint replay and downstream training ablations.

To measure actual post-training alignment, we compare the cold-start and step-357 checkpoints on identical videos, questions, response prefixes, tokens, and teacher views:

ScopeResponsesInitial KLFinal KLKL reduction95% CI
Overall581.06190.50000.5619[0.4872, 0.6368]
VLM4D321.31060.51150.7991[0.7463, 0.8528]
DSR-Bench260.75580.48580.2700[0.2319, 0.3089]

Across the 58 fixed responses, mean relative KL reduction is 50.07% (95% CI [45.95%, 53.95%]); 75.05% of token positions show reduced KL (95% CI [69.86%, 79.79%]), and token-mixture KL decreases from 1.0773 to 0.4395.

AuthorsOpenReview ↗

Response to Reviewer tXiT (2/3)

Within these 58 responses, token positions with higher initial graspability scores (HH) showed larger relative KL reductions after training. The mean relative reduction was 56.6% in the highest HH quintile and 25.1% in the lowest, a gap of 31.5 percentage points. After stratifying positions by initial KL quartile within each response, the mean difference between high- and low-HH positions within strata was +4.03 pp (95% CI [0.18, 7.81], bootstrapped by response). Higher initial HH was thus associated with greater subsequent improvement in alignment with the teacher distribution.

On 32 additional videos excluded by video ID and path, mean KL reduction is positive in all four scenario groups: real first-person (0.8177), real third-person (0.8891), synthetic direction (0.9002), and synthetic false-positive controls (0.7494).

We additionally train grounding-channel variants while keeping graspability fixed:

Grounding configurationGraspabilityVLM4DDSR-Bench
Spatial only (V)Off47.8530.10
Spatial only (V)On49.7232.40
Spatial + Order (V+O)On50.5033.31
Spatial + Motion (V+M)On50.7233.52
Spatial + Order + Motion (V+O+M)On51.3834.49

With graspability fixed, adding order improves accuracy by 0.78/0.91 points and adding local motion by 1.00/1.12 points on VLM4D/DSR-Bench; the full three-channel variant performs best. This downstream training evidence complements the fixed-completion allocation diagnostics. Candidate support characterizes local compatibility, checkpoint replay measures post-training distribution alignment, and the trained ablations quantify benchmark-level effects.

W2. Incremental value under matched training cost and simpler baselines

Direct comparison with reproduced TA-OPD

We reproduced TA-OPD following Wang et al. (2026) with the same teacher, student, training data, optimization schedule, and evaluation protocols as our variants. TA-OPD selects positions by teachability, while Graspable-OPD uses teacher support and disagreement to continuously reweight distillation at every response position. Graspable-OPD achieves comparable performance on VLM4D (48.46% vs. 48.51%), while improving DSR-Bench Micro from 23.48% to 31.29% (+7.81 pp) and Macro from 23.23% to 31.56% (+8.33 pp). Compared with Uniform OPD, Graspable-OPD also improves VLM4D Overall by 2.26 points and DSR-Bench Micro by 2.65 points, with corresponding Macro improvements of 4.30 and 2.38 points.

Why combine teacher support and disagreement?

Among high-disagreement positions, disagreement-only weighting slightly favors low support (1.124 vs. 1.106). The product favors high support (1.110 vs. 0.990).

Weighting strategyHigh disagreement + high supportHigh disagreement + low support
Uniform1.0001.000
Support-only1.0120.868
Disagreement-only1.1061.124
Product (ours)1.1100.990

High/low is defined at 0.5 after response-local scaling. The high- and low-support columns include 45 and 58 responses, respectively.

On fixed DSR-Bench prefixes, high-support, high-disagreement positions gain 1.609 percentage points of KL share (95% CI [1.28, 2.09]).

RegionResponses containing the regionMean normalized weightChange in KL share (pp)
High disagreement + high support131.101+1.609
High disagreement + low support261.000-0.632
Low disagreement + high support260.991-0.058
Low disagreement + low support100.990-0.296

The CI uses the 13 responses containing this group. KL share measures scalar loss contributions. The weighting comparison shows how support and disagreement affect priority. The DSR-Bench analysis reports the resulting KL-share changes across the four support–disagreement groups.

Soft weighting versus hard selection

A separate top-10% hard-selection diagnostic retains 12.9% of DSR-Bench positions on average and removes 70.3% of the original direct KL loss.

Nominal hard retention ratioActual average coverageOriginal KL contribution removed
5%11.1%73.7%
10%12.9%70.3%
20%25.6%52.2%
50%54.7%40.2%
100%100.0%0.0%

Coverage includes rounding up and retaining at least one position per response. Removed KL is the sum of direct loss terms at excluded positions.

AuthorsOpenReview ↗

Response to Reviewer tXiT (3/3)

Increasing γ shifts more KL supervision to high-priority positions at 100% coverage.

γSupervision coverageKL share of high-support, high-disagreement positions
0100%12.00%
0.1100%12.33%
0.25 (ours)100%12.80%
0.5100%13.52%
1.0100%14.77%

Averages include all 26 fixed DSR-Bench responses, with zero shares assigned when the high-support, high-disagreement group is absent.

KL at excluded positions. On fixed prefixes, we compare initial and final checkpoints at positions excluded by the top-10% gate. On DSR-Bench, KL falls by 32.2% on average (95% CI [22.1%, 41.4%]).

DatasetMean relative KL reduction95% CIResponses showing improvement
DSR-Bench32.2%[22.1%, 41.4%]20/26
VLM4D58.5%[56.9%, 60.1%]32/32
All46.7%[41.0%, 51.9%]52/58

These checkpoint comparisons measure the learning effect of full G²-OPD training.

DSR-Bench question structure. Coordinate constraints, time intervals, and reference perspectives often enter the same judgment, motivating supervision across the reasoning trajectory.

CharacteristicNumber of questionsPercentage
Explicit object-coordinate constraints1,22485.3%
Explicit temporal intervals1,15380.3%
Specified reference perspective1,435100%
Total questions1,435100%

Teacher-compute accounting

Reproduced TA-OPD and H-only Graspable-OPD each use one intact-video teacher condition.

Full G²-OPD adds three counterfactual views of the same student rollout. The student performs one backward pass per microbatch, with 42 accumulation microbatches per rank per update.

MetricUniform OPDG²-OPDRelative overhead
Teacher forward conditions / rollout144.00×
Teacher FLOPs (4 fixed samples)729.861 TFLOPs2919.443 TFLOPs4.00×
Total training + rollout GPU-hours220.73325.30+47.37%
Peak allocated GPU memory (microbenchmark)23.62 GiB24.61 GiB+4.19%
Student backward passes / microbatch111.00×
Student backward passes / update / rank42421.00×

G²-OPD gains 5.18/5.85 points over Uniform OPD on VLM4D/DSR-Bench. Teacher FLOPs increase 4×. End-to-end GPU-hours rise 47.37% because student generation and optimization are shared. Sequential no-grad teacher forwards and CPU staging of full-vocabulary logits limit peak memory growth to 4.19%. FLOPs and peak memory use four fixed paired samples.

The 4×-rollout Uniform OPD baseline matches the number of teacher-forward conditions:

MethodTeacher conditions / rolloutRelative rollout budgetVLM4D Acc.DSR Acc.
Uniform OPD (original)11×46.20%28.64%
Uniform OPD (4× rollout)14×48.68%30.52%
G²-OPD (original)41×51.38%34.49%

At matched teacher-forward count, G²-OPD gains 2.70 points on VLM4D and 3.97 on DSR-Bench. End-to-end GPU-hours are reported separately.

We also compare against generic weighting and controlled OPD variants:

MethodVLM4DDSR-Bench
Uniform OPD46.2028.64
Random weighting46.7029.20
Entropy weighting47.1929.69
KL weighting47.9630.31
Graspable-OPD48.4631.29
Grounded-OPD49.1731.92
G²-OPD51.3834.49

Adding grounding to Graspable-OPD improves VLM4D and DSR-Bench by 2.92 and 3.20 percentage points, respectively. G²-OPD also exceeds random, entropy, and KL weighting by 4.68/5.29, 4.19/4.80, and 3.42/4.18 points, respectively. These reweighting baselines use matched response-wise normalization and weight-concentration constraints.

Finally, we evaluate Qwen2.5-VL-32B as the teacher and Qwen2.5-VL-3B as the student, with verified vocabulary alignment:

ModelRoleVLM4D OverallDSR-Bench Micro
Qwen2.5-VL-32BTeacher reference53.03%30.31%
Qwen2.5-VL-3BBase student42.29%19.79%
Qwen2.5-VL-3BG²-OPD-trained student47.58%24.88%
Improvement over base+5.29 pp+5.09 pp

G²-OPD improves the base student by 5.29 points on VLM4D Overall and 5.09 points on DSR-Bench Micro.

Revision: The revised manuscript adds grounding-channel ablations (Section 5.7, Table 6), cost analysis (Section 5.8, Tables 7 and 11), and weighting controls (Section 5.3, Table 5). Appendix E reports probe-validity and checkpoint-alignment analyses (Tables 17–21).

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

This paper proposes G²-OPD, an on-policy distillation method for improving dynamic spatial reasoning in compact vision-language models. It combines two signals to weight teacher supervision: grounding, estimated through counterfactual video perturbations, and graspability, estimated from teacher–student disagreement and candidate support. Experiments on VLM4D and DSR-Bench show improvements over standard distillation methods. However, the methodological novelty is limited, and the proposed grounding and graspability signals are not sufficiently validated beyond their effects on benchmark accuracy. The lack of computational cost analysis and broader generalization experiments further weakens the contribution.

Critical strengths

The paper addresses a relevant problem in transferring spatial reasoning capabilities from large models to smaller ones.

The method is clearly presented, with detailed mathematical formulations and supporting analysis.

The experiments show consistent improvements over uniform OPD, and the ablation studies demonstrate the contributions of both weighting components.

Critical weaknesses

Limited methodological novelty. The proposed method mainly combines existing ideas from counterfactual visual supervision and teacher–student disagreement-based weighting.

Insufficient validation of the grounding signal. The method measures grounding through changes in teacher confidence under video perturbations. However, sensitivity to perturbations does not necessarily indicate correct spatial reasoning. The paper lacks direct evidence that higher grounding scores correspond to more accurate or relevant visual evidence.

Unclear computational trade-offs. G²-OPD requires four teacher evaluations per student rollout, introducing additional computation compared with uniform OPD. However, training time, memory consumption, and computational costs are not reported. It remains unclear whether the performance gains justify the additional training overhead.

Limited generalization. All controlled distillation experiments use a single teacher–student pair, MLLM-4D-8B and Qwen3-VL-2B, evaluated on two benchmarks. The effectiveness of the weighting strategy across different architectures, model scales, and tasks remains unverified.

Insufficient comparison with simpler weighting strategies. Although the ablations demonstrate the individual contributions of grounding and graspability, the paper does not compare against simpler alternatives, such as uncertainty-based or KL-based token weighting. This makes it difficult to determine whether the proposed signals are necessary for the observed improvements.

Questions for authors

Can the authors provide direct evidence that the grounding scores correlate with correct visual reasoning rather than general sensitivity to input perturbations?

What are the training time and computational overhead compared with uniform OPD? How do the methods compare under matched training budgets?

Can the authors evaluate the method with additional teacher–student pairs and compare it against simpler token-weighting strategies?

4 comments
AuthorsOpenReview ↗

Response to Reviewer HGEq (1/3)

Thank you for recognizing the relevance of dynamic spatial reasoning, the clarity of the formulation, and the consistent improvements over Uniform OPD. We clarify the task-specific contribution, add controlled evidence for grounding, and report the computational and generalization results requested. G²-OPD retains the original full-vocabulary teacher target and student inference architecture. Its contribution is a response-position allocation rule that uses counterfactual evidence during training.

W1. Methodological distinction: task-specific joint allocation

G²-OPD is designed for dynamic spatial understanding, where the same frames can support opposite answers when their order changes. A clip and its reversal contain exactly the same frames, yet an approaching object becomes a departing one. VA-OPD is a representative method for counterfactual visual supervision, and TA-OPD is a representative teachability-aware approach to weighting supervision based on teacher–student disagreement.

VA-OPD measures how teacher scores change with access to fine-grained visual detail. That comparison alone leaves open whether the teacher uses frame order to distinguish approaching from departing. Our temporal-order probe varies frame order while keeping the frames themselves fixed. We also probe spatial content and local motion separately, so that evidence for object relations, motion direction, and speed can contribute separately to the grounding weight. In the reversal test reported in W2, the teacher changes its motion predictions while preserving its static-relation and color judgments.

The three evidence scores enter the grounding multiplier separately:

gt=min⁡gmax⁡,,1+αspaV^t+αordO^t+αdynM^t.g_t=\min\\{g_{\max},\\,1+\alpha_{\rm spa}\hat V_t+ \alpha_{\rm ord}\hat O_t+\alpha_{\rm dyn}\hat M_t\\}.

An evidence-sensitive motion judgment can be important even when the teacher places little probability mass on the student's current candidates. TA-OPD uses local teachability to select positions for distillation. A motion judgment that falls below its selection cutoff receives no direct distillation loss at that position, even if it matters for the answer. G²-OPD uses candidate support and local disagreement to adjust the strength of supervision. A position with limited candidate support still receives full-vocabulary reverse-KL supervision, and its video evidence can still affect its weight through gtg_t. When teacher support and local disagreement are both high, the position receives a larger graspability boost.

Let AtA_t be the student's top-KK candidates and st=∑a∈Atqt(a)s_t=\sum_{a\in A_t}q_t(a) the teacher mass on them. Proposition 1 characterizes the closest match to the teacher available within those candidates. The minimum is attained by the teacher distribution conditioned on AtA_t. We combine candidate support with local disagreement dt=DKL(qtUt∥ptUt)d_t=D_{\rm KL}(q_t^{U_t}\Vert p_t^{U_t}), where UtU_t is the union of the teacher's and student's top-KK sets and both distributions are renormalized on UtU_t.

min⁡supp⁡(ν)⊆AtDKL(ν∥qt)=−log⁡st,ht=1+γd^ts^t.\min_{\operatorname{supp}(\nu)\subseteq A_t}D_{\rm KL}(\nu\Vert q_t) =-\log s_t, \qquad h_t=1+\gamma\hat d_t\hat s_t.

The joint allocation lets video evidence and local compatibility affect the same training decision. With ut=1/Tu_t=1/T, Proposition 2 gives the allocation below. It redistributes a fixed total weight across the response, with weights detached during the student update.

\rho^{\ast}=\arg\max_{\rho\in\Delta_T} \left\\{\sum_t\rho_t(\log g_t+\log h_t)-D_{\rm KL}(\rho\Vert u)\right\\},

ρt∗=gtht∑t′gt′ht′,wt=Tρt∗,∑twt=T.\rho_t^{\ast}=\frac{g_th_t}{\sum_{t'}g_{t'}h_{t'}}, \qquad w_t=T\rho_t^{\ast}, \qquad \sum_t w_t=T.

Downstream grounding-channel ablation

Grounding configurationGraspabilityVLM4DDSR-Bench
Spatial only (V)Off47.8530.10
Spatial only (V)On49.7232.40
Spatial + Order (V+O)On50.5033.31
Spatial + Motion (V+M)On50.7233.52
Spatial + Order + Motion (V+O+M)On51.3834.49

Adding graspability to spatial grounding improves VLM4D and DSR-Bench by 1.87 and 2.30 percentage points, respectively. With graspability fixed, adding temporal order improves the two benchmarks by 0.78 and 0.91 points, and adding local motion improves them by 1.00 and 1.12 points. Both temporal probes therefore add downstream value beyond spatial grounding in this setting, and the full three-channel configuration performs best on both benchmarks.

AuthorsOpenReview ↗

Response to Reviewer HGEq (2/3)

W2 & Q1. Grounding validity and semantic selectivity

Grounding quantifies the teacher's dependence on task-relevant visual evidence. We evaluate its semantic selectivity by testing whether high scores identify claims requiring visual evidence and whether teacher predictions respond consistently to controlled semantic changes.

Association with visual-evidence requirements. On 16 controlled depth-motion videos involving approach/departure judgments, we annotate whether each claim requires visual evidence and evaluate the ranking induced by GG.

MetricResult
AUROC0.834
Precision@Top-20%100%
Spearman correlation0.579
Evidence-demand rate in bottom 20%21.9%

High-GG positions are concentrated on claims requiring visual evidence, while the bottom 20% contains substantially fewer such claims.

Semantic consistency under temporal reversal. In a separate test, we reverse each video while keeping answer-option order fixed. The answer should change for 48 motion question–video pairs and stay the same for 16 static left–right relation pairs and 64 color-control pairs. For depth prompts, the question fixes the camera and ball size. The conditional option probe compares the single-token logits for A/B/C/D after the assistant prefix <answer>, normalized over these four options. We report free-generation behavior separately below.

GroupPairsOriginal acc.Reversed acc.Both correctChange rateMean JS (nats)
Motion48100.0%100.0%100.0%100.0%0.655731
Static relation16100.0%100.0%100.0%0.0%0.000053
Color control64100.0%100.0%100.0%0.0%0.000001

The teacher reverses all motion predictions while preserving the invariant controls. A predetermined free-generation subset (indices=[0,3,4,7]; 12 motion, 4 static, 16 color cases), preselected before score and outcome inspection, reproduces the same pattern: change rates are 100%, 0%, and 0%, respectively.

Structural vs. photometric sensitivity. We compare spatial masking with mild photometric perturbations:

GroupSpatial VVPhotometric CCV−CV-C95% CI
All 900.13670.0314+0.1053[0.0695, 0.1459]
New 320.15830.0344+0.1240[0.0665, 0.1930]

Spatial masking produces a consistently larger evidence response than mild photometric changes. Together with the evidence-demand annotations and temporal-reversal controls, these results show that GG tracks task-relevant visual evidence across complementary tests.

AuthorsOpenReview ↗

Response to Reviewer HGEq (3/3)

W3 & Q2. Computational overhead and matched teacher-compute comparison

G²-OPD evaluates four frozen-teacher conditions—the original video and three counterfactual views—on one student-generated rollout. Student generation and the backward workload are shared across all four conditions.

MetricUniform OPDG²-OPDRelative overhead
Teacher forward conditions / rollout144.00×
Teacher FLOPs (4 fixed samples)729.861 TFLOPs2919.443 TFLOPs4.00×
Total training + rollout GPU-hours220.73325.30+47.37%
Peak allocated GPU memory (microbenchmark)23.62 GiB24.61 GiB+4.19%
Student backward passes / microbatch111.00×
Student backward passes / update / rank42421.00×

Relative to Uniform OPD, G²-OPD improves accuracy by 5.18 points on VLM4D and 5.85 points on DSR-Bench. Teacher FLOPs increase 4×, but end-to-end training and rollout cost increases 47.37%, because student generation and optimization are shared. Peak memory rises 4.19%: the frozen teacher is reused sequentially under no-grad execution, and intermediate full-vocabulary logits are staged on CPU between conditions, keeping the four teacher outputs from occupying GPU memory simultaneously. Teacher FLOPs and peak memory are measured on four fixed paired samples.

To isolate allocation from teacher-forward budget, we compare with 4×-rollout Uniform OPD:

MethodTeacher conditions / rolloutRelative rollout budgetVLM4D Acc.DSR Acc.
Uniform OPD (original)11×46.20%28.64%
Uniform OPD (4× rollout)14×48.68%30.52%
G²-OPD (original)41×51.38%34.49%

Under the matched teacher-forward budget, G²-OPD retains gains of 2.70 points on VLM4D and 3.97 points on DSR-Bench. This controls teacher-forward conditions; the GPU-hour table separately reports end-to-end cost.

W4 & W5 & Q3. Generalization and simpler weighting baselines

We evaluate Qwen2.5-VL-32B as the teacher and Qwen2.5-VL-3B as the student, with verified vocabulary alignment:

ModelRoleVLM4D OverallDSR-Bench Micro
Qwen2.5-VL-32BTeacher reference53.03%30.31%
Qwen2.5-VL-3BBase student42.29%19.79%
Qwen2.5-VL-3BG²-OPD-trained student47.58%24.88%
Improvement over base+5.29 pp+5.09 pp

Finally, we compare against simpler reweighting strategies under the same response-wise normalization and matched weight-concentration constraints:

MethodVLM4DDSR-Bench
Uniform OPD46.2028.64
Random weighting46.7029.20
Entropy weighting47.1929.69
KL weighting47.9630.31
Spatial-only grounding + graspability49.7232.40
Grounded-OPD (V+M+O)49.1731.92
Graspable-OPD48.4631.29
G²-OPD51.3834.49

G²-OPD exceeds random, entropy, and KL weighting by 4.68/5.29, 4.19/4.80, and 3.42/4.18 points on VLM4D/DSR-Bench, respectively; it also exceeds spatial-only grounding plus graspability by 1.66/2.09 points. These results and the additional teacher–student evaluation address the concerns about generic weighting and single-pair generalization.

Revision: The revised manuscript adds weighting comparisons (Section 5.3, Table 5), grounding-channel ablations (Section 5.7, Table 6), and training-cost results (Section 5.8, Tables 7 and 11). Appendix E reports probe controls, and Appendix D.6 reports the second teacher–student pair.

AuthorsOpenReview ↗

Follow-up to Reviewer HGEq

Dear Reviewer HGEq,

Thank you for your thoughtful review. We have revised the manuscript and added experiments addressing your concerns about methodological novelty, grounding validity, computational cost, and generalization.

Our core novelty lies in a joint supervision-allocation framework designed for dynamic spatial reasoning, integrating spatiotemporal evidence relevance with student-aware correction compatibility along on-policy trajectories. Unlike VA-OPD-style visual supervision, G²-OPD explicitly distinguishes spatial content, temporal order, and local motion, capturing dynamic judgments such as approaching versus departing when frame order changes. Unlike TA-OPD's hard teachability-based selection, our continuous allocation preserves full-vocabulary reverse-KL supervision at every response position while redistributing a fixed response-level supervision budget.

Our additional experiments support these contributions:

  • Grounding Validity and Temporal Contributions: Visual-evidence annotations yield an AUROC of 0.834 and Precision@Top-20% of 100%. Frame reversal changes all 48 motion judgments while preserving all 80 static and color controls. Trained ablations confirm the additional benefits of temporal order and local motion (Section 5.7, Table 6; Appendix E).
  • Computational Trade-offs: Teacher FLOPs increase 4×, while total training and rollout GPU-hours increase by 47.37% and peak memory by 4.19%. At matched teacher-forward counts, G²-OPD outperforms 4×-rollout Uniform OPD by 2.70 pp on VLM4D and 3.97 pp on DSR-Bench (Section 5.8, Tables 7 and 11).
  • Simpler Weighting and Cross-Model Generalization: G²-OPD exceeds KL weighting by 3.42/4.18 pp on VLM4D/DSR-Bench under matched weighting controls. With Qwen2.5-VL-32B as teacher and Qwen2.5-VL-3B as student, it improves over the base student by 5.29/5.09 pp (Section 5.3, Table 5; Appendix D.6, Table 15).

These results support our task-specific joint allocation design and provide additional evidence for its grounding validity, computational trade-offs, and cross-model applicability. We hope these revisions and the detailed rebuttal address your concerns. If any concerns remain unresolved, we would greatly appreciate your feedback and would be happy to provide further clarifications or additional experimental evidence as needed.

Thank you again for your consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Note to AC: Rebuttal Summary

Dear Area Chair,

We sincerely appreciate your time and effort in overseeing the review process for our submission, "G²-OPD: Grounded and Graspable On-Policy Distillation for Dynamic Spatial Reasoning." We provide this summary to highlight how our revision addresses the reviewers' feedback and further strengthens the contributions of our work.

We first thank the reviewers for recognizing the importance of dynamic spatial reasoning, the clarity of our formulation, and the consistent improvements over Uniform OPD. Our revised manuscript and rebuttal address the primary concerns as follows:

  • Methodological Novelty & Joint Supervision Allocation: We clarified the task-specific contribution of G²-OPD beyond existing visual supervision and teachability-aware weighting approaches. By separately probing spatial content, temporal order, and local motion, our grounding mechanism identifies distinct sources of spatiotemporal evidence. Graspability further combines teacher support and local teacher–student disagreement to identify corrections that are both needed and compatible with the student's current predictions. Their joint allocation continuously redistributes a fixed response-level supervision budget while retaining full-vocabulary reverse-KL supervision at every response position. Additional comparisons show that G²-OPD outperforms KL-based weighting by 3.42 and 4.18 percentage points on VLM4D and DSR-Bench, respectively, demonstrating the value of task-specific joint allocation over generic reweighting.

  • Grounding Validity & Graspability: To validate the interpretation of our signals, we added controlled evidence-demand annotations (AUROC 0.834; Precision@Top-20% 100%), temporal-reversal experiments (all 48 motion predictions change while all 80 static-relation and color controls remain unchanged), and spatial-versus-photometric perturbation controls. We further connected graspability to actual learning outcomes through checkpoint replay on 58 fixed responses, showing a 50.07% mean relative KL reduction. Positions in the highest graspability quintile exhibit a 56.6% KL reduction, compared with 25.1% in the lowest quintile. Trained grounding-channel ablations additionally confirm the downstream contributions of temporal-order and local-motion evidence, improving accuracy by 0.78/0.91 and 1.00/1.12 points, respectively, over spatial-only grounding with graspability.

  • Experimental Controls & Computational Efficiency: We provided complete training and rollout configurations, reproduced TA-OPD under the same teacher, student, data, optimization schedule, and evaluation protocol, and added three-seed results with paired gains of +5.10±0.41 points on VLM4D and +5.83±0.46 points on DSR-Bench. We also quantified the computational trade-off: although teacher FLOPs increase 4×, end-to-end training and rollout GPU-hours increase by 47.37%, while peak allocated memory increases by only 4.19%. Importantly, under matched teacher-forward counts, G²-OPD still outperforms 4×-rollout Uniform OPD by 2.70 points on VLM4D and 3.97 points on DSR-Bench.

  • Robustness & Generalization: We added systematic hyperparameter sensitivity analyses, showing maximum accuracy variations of only 1.21 and 1.25 points across the tested configurations on VLM4D and DSR-Bench, respectively. We further evaluated an additional teacher–student pair, Qwen2.5-VL-32B → Qwen2.5-VL-3B, where G²-OPD improves the base student by 5.29 and 5.09 points on the two benchmarks. An additional experiment on general video understanding also demonstrates a 4.8-point improvement on Video-MME.

Following the rebuttal, we have received encouraging feedback from the reviewers. Reviewer HGEq acknowledged that most concerns had been addressed and raised the rating to 3. Reviewer qvER maintained a Weak Accept recommendation after noting that the additional clarifications and experiments addressed most concerns. Reviewer kkf4 also raised the rating, acknowledging the authors' responses. Reviewer tXiT initially recommended Weak Accept, and we provided additional analyses addressing the concerns raised in the review.

We believe that the revised manuscript and additional experimental evidence substantially strengthen the methodological justification, empirical validation, and practical assessment of G²-OPD. We hope this summary assists you in your final assessment.

Thank you for your time and consideration.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.