- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes BiVaR, a training-free model merging method that transfers reasoning capabilities from LLMs to VLMs. The method identifies shared reasoning neurons through cross-model activation distribution consistency. It then uses adaptive low-rank merging to improve reasoning while preserving other general multimodal capabilities of VLMs.
Critical strengths
- Merging LLMs with VLMs to improve VLM reasoning without costly training is an interesting topic.
- The main experiments cover multiple model pairs.
Critical weaknesses
W1. Methodological motivation and novelty are not sufficiently clear.
- The novelty of BiVaR mainly lies in using cross-model activation distribution consistency to identify shared reasoning neurons for model merging, compared with prior methods such as SNRF. However, the motivation for reframing shared reasoning neuron identification as a distribution consistency problem is not sufficiently clear. In particular, why should similar activation distributions indicate that neurons have shared reasoning functions? What is the conceptual advantage of this criterion over existing neuron selection methods such as detecting neuron set intersection by SNRF? The method section, especially Sections 3.3 and 3.4 provides detailed math definitions, but is not easy to follow, especially, the motivation for each design choice is unclear.
W2. Important experiments and ablations are missing.
- The main experiments cover multimodal mathematical reasoning, general multimodal understanding, and text-only reasoning, but non-mathematical, common-knowledge multimodal reasoning is not evaluated, the authors may consider potential benchmark datasets: A-OKVQA, VCR, MMMU-Pro, ReasonVQA, GQA. For text-only reasoning, why is MMLU-Pro reported in Figure 1(a) but not included in the main experiments?
- For the ablations, results are mainly only reported on one model Qwen2.5-VL-7B-Instruct. Are these ablation findings generalizable to other models?
- A key ablation appears missing: to demonstrate the benefit of distribution consistency over neuron overlap, a direct comparison between BiVaR's neuron selection and SNRF's intersection-based selection is needed, using the same merging procedure and comparable numbers of selected neurons.
W3. The performance gains over prior methods and the base VLM are modest overall. Moreover, across all methods, improvements over the base VLM are generally small, and some merging methods even harm performance (Figure 1(b) and Table 1).
- Can the authors elaborate on the significance of these gains? Are the improvements consistent across random seeds? Additionally, besides the base VLM performance, it can be very helpful to also report the LLM performances on each benchmark, so that readers can better understand how much reasoning capability is transferred and how much room for improvement remains.
Questions for authors
-
Why should similar activation distributions indicate that neurons have shared reasoning functions? What justifies using distributional consistency rather than individual neuron responses to identify transferable reasoning neurons?
-
The paper states that distribution consistency corresponds to a density ratio close to 1. However, Equations (5)–(7) use only an upper threshold on the discriminator score. How does this ensure that selected neurons are distributionally consistent rather than simply having low density-ratio scores?
-
In lines 216–218, why is mutual consistency necessary when reasoning is transferred only from the LLM to the VLM? Can the authors clarify the definition of mutual consistency and why unidirectional discrimination is insufficient?
-
Since distribution-based neuron selection is the key contribution, can the authors provide a controlled ablation comparing BiVaR's selection with SNRF's intersection-based selection, using the same merging procedure and comparable numbers of selected neurons? Table 6 partially addresses this but does not fully isolate the effect of neuron selection.
-
GSM8K and Geo3K are used to collect neuron activations. How sensitive are the identified neurons and final performance to the choice and number of reasoning samples? Could different datasets or subsets lead to different results?
-
Given the relatively modest performance gains in Table 1, are these improvements consistent across different random seeds? How much of the source LLM's reasoning capability is actually transferred to the merged VLM?
-
Table 5 studies rank sensitivity but does not demonstrate the advantage of adaptive-rank selection over fixed-rank decomposition.
7 comments
W3 & Q6: Significance, Stability, and Extent of Reasoning Transfer.
(1) Significance of Performance Gains. Although the absolute gains are modest, BiVaR improves reasoning performance without additional training while preserving general multimodal capabilities. As shown in Table 6, SNRF-only neurons provide limited benefits and even degrade GPQA performance, whereas BiVaR-only neurons improve both MathVision and GPQA, highlighting the importance of selective transfer in mitigating negative interference.
(2) Stability Across Random Seeds. We conduct three independent cross-machine runs and evaluate randomness in data sampling, random projections, and discriminator initialization using three seeds (41–43).
| Experimental Setting | MathVista | MathVerse | MathVision | MME | MMMU | ScienceQA | GPQA |
|---|---|---|---|---|---|---|---|
| Qwen2.5-VL-7B-Instruct | 69.07 | 42.43 | 26.32 | 1680.69 | 50.56 | 66.70 | 29.68 |
| Cross-Machine Runs | 70.57±0.09 | 42.71±0.13 | 29.39±0.19 | 1721.83±0.63 | 50.59±0.17 | 67.90±0.10 | 30.89±0.06 |
| Data Sampling | 70.52±0.54 | 42.91±0.24 | 28.95±0.32 | 1720.83±1.29 | 50.63±0.13 | 68.10±0.21 | 30.67±0.13 |
| Random Projection | 70.46±0.32 | 42.72±0.45 | 29.17±0.50 | 1721.77±0.85 | 50.63±0.45 | 67.87±0.15 | 30.78±0.17 |
| Discriminator Initialization | 70.49±0.17 | 42.70±0.66 | 29.28±0.33 | 1721.83±1.15 | 50.59±0.28 | 67.89±0.32 | 31.11±0.41 |
| Experimental Setting | MathVista | MathVerse | MathVision | MME | MMMU | ScienceQA | GPQA |
|---|---|---|---|---|---|---|---|
| Qwen3-VL-8B-Instruct | 72.77 | 45.45 | 17.76 | 1726.94 | 53.00 | 94.65 | 31.69 |
| Cross-Machine Runs | 73.77±0.07 | 46.50±0.06 | 19.08±0.00 | 1750.16±0.43 | 53.15±0.06 | 95.34±0.05 | 32.68±0.18 |
| Data Sampling | 73.73±0.28 | 46.56±0.32 | 19.19±0.19 | 1750.03±0.98 | 53.15±0.28 | 95.34±0.25 | 32.46±0.19 |
| Random Projection | 73.80±0.25 | 46.46±0.30 | 19.08±0.33 | 1750.32±0.84 | 53.15±0.28 | 95.31±0.25 | 32.46±0.50 |
| Discriminator Initialization | 73.73±0.25 | 46.49±0.30 | 18.97±0.19 | 1750.05±0.93 | 53.00±0.29 | 95.39±0.23 | 32.35±0.20 |
BiVaR exhibits small performance variations across both models, supporting the stability of its improvements.
(3) Extent of Reasoning Transfer. We additionally report source LLM performance and introduce Reasoning Gap Closed (RGC) to quantify the proportion of the performance gap closed by BiVaR:
| Target VLM | Benchmark | Source LLM | Base VLM | BiVaR | Improvement (pp) | Reasoning Gap (pp) | RGC (%) |
|---|---|---|---|---|---|---|---|
| Qwen2.5-VL-7B-Instruct | GPQA | 34.15 | 29.68 | 30.82 | +1.14 | 4.47 | 25.50 |
| Qwen2.5-VL-7B-Instruct | MMLU-Pro Math | 45.16 | 40.99 | 42.89 | +1.90 | 4.17 | 45.56 |
| Qwen2.5-VL-7B-Instruct | Math 500 | 68.00 | 63.80 | 65.60 | +1.80 | 4.20 | 42.86 |
| Qwen3-VL-8B-Instruct | GPQA | 36.83 | 31.69 | 32.58 | +0.89 | 5.14 | 17.32 |
| Qwen3-VL-8B-Instruct | MMLU-Pro Math | 78.13 | 70.06 | 74.78 | +4.72 | 8.07 | 58.49 |
| Qwen3-VL-8B-Instruct | Math 500 | 80.6 | 75.60 | 79.20 | +3.60 | 5.00 | 72.00 |
BiVaR achieves higher RGC on mathematical reasoning tasks, consistent with our objective of transferring mathematical reasoning capabilities from the source LLMs. Meanwhile, improvements on GPQA demonstrate that the transferred capabilities also generalize to scientific reasoning in biology, chemistry, and physics, albeit to a lesser extent.
Revisions: Neuron-subset results are extended in Table 6 (L487–493), and repeated-run statistics in Appendix C.3.2 (L1436–1446, Tables 25–26). Reasoning-gap closure is added in Appendix C.1.5 (L1069–1079, Table 13).
Q1: Rationale for Distributional Consistency.
(1) Conceptual Motivation. Unlike individual activation responses, which primarily reflect response strength, our method evaluates cross-model compatibility through ID–OOD distribution discrimination. If a discriminator cannot reliably distinguish a neuron's activation distribution from the other model's reference distribution, this suggests shared response behavior under reasoning inputs. We therefore use distributional consistency as a proxy for identifying transferable neurons rather than requiring strict functional equivalence.
(2) Empirical Validation. We compare BiVaR with Hungarian Fusion (HF) and K-means Fusion (KF), following Model Fusion via Retrofitting [1].
| Method | MathVista | MathVerse | MathVision | MME | MMMU | ScienceQA | GPQA |
|---|---|---|---|---|---|---|---|
| Qwen2.5-VL-7B-Instruct | 69.07 | 42.43 | 26.32 | 1680.69 | 50.56 | 66.70 | 29.68 |
| BiVaR | 70.53 | 42.67 | 29.28 | 1721.75 | 50.56 | 67.89 | 30.82 |
| HF | 69.17 | 41.16 | 27.63 | 1711.72 | 50.56 | 66.73 | 29.91 |
| KF | 69.73 | 42.43 | 29.61 | 1720.13 | 50.56 | 66.93 | 30.58 |
| Method | MathVista | MathVerse | MathVision | MME | MMMU | ScienceQA | GPQA |
|---|---|---|---|---|---|---|---|
| Qwen3-VL-8B-Instruct | 72.77 | 45.45 | 17.76 | 1726.94 | 53.00 | 94.65 | 31.69 |
| BiVaR | 73.77 | 46.47 | 19.13 | 1750.08 | 53.11 | 95.33 | 32.58 |
| HF | 73.33 | 45.73 | 18.09 | 1727.20 | 52.56 | 94.74 | 31.70 |
| KF | 73.60 | 46.11 | 18.42 | 1729.22 | 53.11 | 95.04 | 31.92 |
BiVaR achieves better overall performance on both models. Furthermore, our neuron subset ablation shows that B-only neurons outperform S-only neurons selected by activation strength, providing additional evidence for distribution-based selection.
[1] Model Fusion via Retrofitting, ICML 2026.
Revisions: The distributional rationale is clarified in Section 1 (L46–81), with matching comparisons in Appendix C.2.1 (L1126–1133, Table 14) and neuron-subset evidence in Table 6 (L487–493).
Q2–Q3: One-Sided Thresholding and Mutual Consistency.
(1) One-Sided Thresholding. We acknowledge that one-sided thresholding does not strictly guarantee a density ratio close to 1, as the discriminator score differs from the log-density ratio by an additive constant. Instead, we use reference-calibrated thresholds to filter neurons with substantial distributional deviations. As shown below, our approach outperforms two-sided thresholding and its bidirectional intersection, suggesting that stricter constraints do not necessarily improve reasoning transfer.
(2) Mutual Consistency. Although reasoning transfer is unidirectional (LLM→VLM), KL divergence is inherently asymmetric, so unidirectional discrimination may introduce uncertainty in identifying distributionally compatible neurons. We therefore evaluate compatibility in both directions (V→L and L→V) and retain their intersection to improve selection robustness.
| Method | MathVista | MathVerse | MathVision | MME | MMMU | ScienceQA | GPQA |
|---|---|---|---|---|---|---|---|
| Qwen2.5-VL-7B-Instruct | 69.07 | 42.43 | 26.32 | 1680.69 | 50.56 | 66.70 | 29.68 |
| BiVaR (Ours) | 70.53 | 42.67 | 29.28 | 1721.75 | 50.56 | 67.89 | 30.82 |
| Two-Sided V→L | 68.60 | 42.55 | 24.01 | 1703.84 | 48.78 | 67.67 | 30.26 |
| Two-Sided L→V | 68.93 | 41.96 | 28.29 | 1696.12 | 51.33 | 67.77 | 29.93 |
| Two-Sided Intersection | 68.10 | 42.61 | 25.99 | 1719.09 | 48.44 | 67.82 | 29.61 |
| Method | MathVista | MathVerse | MathVision | MME | MMMU | ScienceQA | GPQA |
|---|---|---|---|---|---|---|---|
| Qwen3-VL-8B-Instruct | 72.77 | 45.45 | 17.76 | 1726.94 | 53.00 | 94.65 | 31.69 |
| BiVaR (Ours) | 73.77 | 46.47 | 19.13 | 1750.08 | 53.11 | 95.33 | 32.58 |
| Two-Sided V→L | 68.77 | 42.89 | 18.36 | 1730.73 | 53.00 | 94.45 | 31.91 |
| Two-Sided L→V | 68.97 | 42.72 | 18.36 | 1736.10 | 52.89 | 94.70 | 31.91 |
| Two-Sided Intersection | 68.63 | 42.89 | 18.36 | 1744.87 | 53.33 | 95.19 | 32.57 |
As shown in the table and our additional neuron-count-controlled experiments, bidirectional selection achieves better overall performance, supporting the effectiveness of mutual consistency.
Revisions: Thresholding is clarified in Appendix C.2.5 (L1273–1281, Table 20), and bidirectional selection is evaluated in Appendix C.2.3 (L1219–1229, Tables 16–17).
We appreciate the reviewer’s recognition of the importance of training-free reasoning transfer and the comprehensive evaluation across multiple model pairs. We address your concerns below.
W1: Methodological Motivation and Novelty.
(1) Limitations of Neuron Intersection. SNRF identifies shared neurons by intersecting highly activated channels at corresponding positions, implicitly assuming index-wise functional alignment. However, fine-tuning may alter neuronal representations, such that functionally related neurons no longer occupy corresponding positions. Consequently, SNRF may overlook transferable neurons due to this strict positional constraint.
(2) Motivation for Distributional Consistency. We instead formulate neuron selection as an ID–OOD distribution discrimination problem. If a discriminator struggles to distinguish a neuron's activation distribution from the other model's reference distribution, this suggests statistically compatible cross-model response behavior, making it a promising candidate for reasoning transfer without requiring strict functional equivalence.
Revisions: The motivation for distribution-based selection is clarified in Section 1 (L46-81).
W2: Additional Experiments and Ablation Studies.
(1) Non-Mathematical Reasoning Evaluation. We extend our evaluation to A-OKVQA, VCR, MMMU-Pro, ReasonVQA, GQA, and MMLU-Pro. As shown below, BiVaR achieves better overall performance on both models, supporting its generalizability beyond mathematical reasoning.
(2) Generalizability of Ablation Findings. We extend the key ablations to Qwen3-VL-8B-Instruct. As shown below, the results exhibit consistent trends with Qwen2.5-VL, further validating the effectiveness of our key components across different models.
(3) Effectiveness of Neuron Selection (W2-3 & Q4). We compare BiVaR's distribution-based selection with SNRF's activation-based selection using the same merging procedure adopted from SNRF.
BiVaR consistently outperforms SNRF across both models. Furthermore, Table 6 of the original paper shows that BiVaR-only neurons yield greater reasoning improvements than SNRF-only neurons under the same transfer update. Together, these results support the effectiveness of our distribution-based neuron selection.
Revisions: Additional benchmarks are reported in Appendix C.1.2 (L963–971, Table 10). Qwen3 ablations and selection comparisons are added in Appendix C.2.3–C.2.4 (L1219–1241, Tables 16, 18–19) and Table 6 (L487–493).