- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
LEGO-OPD constructs a multimodal distillation teacher by combining an LLM language prior with a VLM’s image-induced prediction residual. A prefix-dependent calibration rule controls the strength of the visual update using a source-referenced KL budget. Experiments on Qwen3 models show improvements in both multimodal and text-only reasoning, supported by ablations on calibration and expert selection.
Critical strengths
- The factorized teacher construction is well motivated: it separates the language prior from the VLM’s image-induced prediction change rather than combining complete teacher distributions. The generalized Bayesian formulation makes this construction clear, while acknowledging that the residual is a model-dependent proxy for visual evidence.
- The empirical gains are encouraging. With the same expert pair, LEGO-OPD improves the multimodal average over H-OPD from 43.47 to 46.28 and the text average from 55.57 to 56.68. The comparison against fixed coefficients in Table 2 provides additional evidence for adaptive calibration.
Critical weaknesses
- The practical calibration only approximates the stated KL target. Figure 2 usefully documents deviations for larger updates, but the diagnostic covers , whereas the strongest results use . The calibration error and coefficient-clipping frequency at this setting remain unclear, limiting the interpretation of why the method works.
- The main comparison leaves some uncertainty about attribution and robustness. H-OPD and LEGO-OPD use different distillation supports, so their comparison does not fully isolate teacher construction from the update procedure. Moreover, the reported standard deviations measure evaluation sampling variability rather than variation across independent training runs.
Questions for authors
- For , what is the distribution of achieved versus target teacher KL, and how often does the calibration coefficient reach its cap?
- Are matched-support comparisons with H-OPD and independent-training-seed results available? These would materially strengthen the evidence for the source and reliability of the gains.
4 comments
Response to Reviewer d9WG [2/2]
Q2 & W2 Robustness across independent training seeds
Computational constraints prevented us from repeating every baseline across three training seeds. We selected H-OPD with student-top-16 support as a representative multi-teacher comparison with aligned support selection. We evaluated this control and LEGO-OPD at and .
The results below use three training seeds (42–44), adding two seeds to the training seed used in the main paper. Each model is trained for 200 steps and evaluated with five inference seeds. Within each training seed, we first average the results over the five inference seeds and compute the category averages. We then report the mean and sample standard deviation across the three training seeds.
Table R8. Multimodal reasoning averaged across three training seeds (%).
| Method | MATH-Vision | MathVerse | MathVista | MMMU-Pro | RBench-M | Olympiad-MM | MM Avg. |
|---|---|---|---|---|---|---|---|
| H-OPD, student top-16 | 31.73 ± 0.49 | 56.60 ± 0.30 | 68.08 ± 0.18 | 39.82 ± 0.31 | 25.68 ± 1.07 | 40.45 ± 0.86 | 43.73 ± 0.10 |
| LEGO-OPD, | 34.83 ± 1.85 | 57.68 ± 0.67 | 69.92 ± 0.39 | 41.12 ± 0.25 | 28.77 ± 0.44 | 44.71 ± 0.63 | 46.17 ± 0.38 |
| LEGO-OPD, | 35.16 ± 0.76 | 58.21 ± 0.89 | 70.11 ± 0.51 | 41.41 ± 0.38 | 28.86 ± 1.06 | 44.00 ± 0.92 | 46.29 ± 0.25 |
Table R9. Text-only reasoning averaged across three training seeds (%).
| Method | MATH-500 | MathVerse-TO | MMLU-Pro | RBench-T | Olympiad-Text | Text Avg. |
|---|---|---|---|---|---|---|
| H-OPD, student top-16 | 81.77 ± 0.26 | 60.03 ± 0.13 | 46.79 ± 0.05 | 37.23 ± 0.56 | 52.29 ± 0.15 | 55.62 ± 0.16 |
| LEGO-OPD, | 82.28 ± 0.11 | 60.07 ± 0.56 | 47.30 ± 0.21 | 38.85 ± 0.35 | 52.51 ± 0.30 | 56.20 ± 0.22 |
| LEGO-OPD, | 82.23 ± 0.45 | 60.30 ± 0.08 | 47.46 ± 0.07 | 39.12 ± 1.01 | 52.56 ± 0.40 | 56.33 ± 0.30 |
Both LEGO-OPD settings achieve higher multimodal and text-only averages than the H-OPD control at each of the three seeds. Across the three seeds, the gains are +2.44/+0.58 pp at and +2.56/+0.71 pp at for multimodal/text-only reasoning. This consistency supports repeatability beyond inference-seed variability. Detailed benchmark results for each training seed are provided in Appendix Tables 26 and 27 in Section E.5 of the revised manuscript.
Thank you for the detailed response and the additional experiments. The extended calibration diagnostics, matched-support comparison with H-OPD, and results across independent training seeds largely address my main concerns and strengthen the evidence for the method’s robustness.
We thank the reviewer for taking the time to examine our additional analyses and experiments. We appreciate their constructive feedback, which helped us clarify the calibration behavior and strengthen the evidence for the method’s robustness.
Response to Reviewer d9WG [1/2]
Thank you for highlighting the distinction between the intended KL target and its practical approximation, and between evaluation variability and training variability. We agree that these distinctions are essential for interpreting the results.
Q1 & W1 Calibration diagnostics at ρ=2
We extended the expert-composition diagnostic in Mainpaper Figure 2 to ρ=2 using exactly the same sampled sequences. The coefficient comparisons and achieved-versus-target teacher KL curves are provided in Appendix Figure 8, and the cap-hit rates in Appendix Table 16, both in Section D.6 of the revised manuscript. Figure 8 reports source-KL-binned means. In both settings, achieved KL tracks the target for small source KL and shows upward deviations over an intermediate-to-large source-KL range.
To complement the KL curves, Table R5 reports how often the practical approximate coefficient reaches its cap.
Table R5. Overall cap-hit rate of the practical approximate coefficient (αmax=6, 24,576 diagnostic prefixes).
The overall cap-hit rate increases modestly from 2.14% at ρ=1 to 3.14% at ρ=2, remaining low in both settings.
Under Mainpaper Eq. (11), increasing ρ from 1 to 2 doubles the target KL budget. With our approximate calibration rule, this globally scales the visual evidence coefficient α by 2 before capping, consistent with the modest increase in the cap-hit rate. These diagnostics show that a larger budget strengthens the contribution of visual evidence as intended. This stronger visual emphasis is reflected in modest gains in multimodal and text-only reasoning averages at ρ=2 in Mainpaper Table 1.
Q2 & W2 H-OPD comparison with aligned support selection
We added an H-OPD control that selects the student's top-16 tokens, following LEGO-OPD's support-selection rule. The original H-OPD uses the union of the two teachers' top-8 sets. We compare both H-OPD settings with Single VLM OPD and LEGO-OPD at ρ=1 and ρ=2 using training seed 42. All models use the 2B student and are evaluated at step 200. Single VLM OPD uses the 4B VLM teacher, while H-OPD and LEGO-OPD use the same 4B language and grounding experts.
Tables R6 and R7 report mean accuracy ± SD across inference seeds for individual benchmarks. Category averages are reported as means only.
Table R6. Multimodal reasoning comparison at training seed 42 (%).
Table R7. Text-only reasoning comparison at training seed 42 (%).
Changing H-OPD's support raises its multimodal/text-only averages modestly from 43.47/55.57 to 43.83/55.60. Single VLM OPD achieves 40.37/51.72, while LEGO-OPD reaches 45.84/56.01 at ρ=1 and 46.28/56.68 at ρ=2. LEGO-OPD's gains persist after aligning support selection. This comparison is included in Appendix Tables 28 and 29 in Section E.6 of the revised manuscript. The Single VLM OPD, original H-OPD, and LEGO-OPD results match Mainpaper Table 1. The student-top-16 H-OPD checkpoint is an additional control.