acceptodds
Under review as a conference paper at ICLR 2027

From Symbolic Mapping Failure to Numerosity Specialization in Vision-Language Models

Abstract

Visual number sense supports counting objects, comparing group sizes, and determining the size of selected or combined groups. Vision-language models exhibit sensitivity to quantity, but their ability to answer numerical questions remains uneven across visual scenes and tasks. We introduce a two-phase post-training framework to develop more accurate and reusable visual number sense in pretrained models. Configuration Coverage first associates shared count labels with complementary object layouts and appearances. A joint update then combines Relational Numerosity Supervision, which links counts through local changes and part–whole relations, with multi-interface supervision for comparison and query-conditioned counting. We also construct a controlled synthetic dataset for training and evaluation across visual configurations and count ranges. Experiments on two VLM families show substantial improvements in synthetic counting and task-dependent gains on external numerical benchmarks. Matched-budget ablations establish the benefit of complementary configuration coverage, while supervision-allocation studies demonstrate the contributions of relational and task-specific training. Representation analyses reveal more accurate intermediate-layer count readout and stronger cross-configuration readout transfer. We characterize the resulting profile as numerosity specialization: high-precision counting develops within targeted visual and numerical regimes, with varying transfer to larger cardinalities, other visual distributions, and numerical tasks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
vision-language models, visual counting, numerosity, post-training, representation analysis

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 3 reviews, above 85% of the others.
Reviewer_a9ocOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Currently, Vision-language models exhibit sensitivity to quantity, but their ability to answer numerical questions remains uneven across visual scenes and tasks. To solve this issue, in this paper, the author introduces a two-phase post-training framework to develop more accurate and reusable visual number sense in pretrained models. In addition, they also construct a controlled synthetic dataset for training and evaluation across visual configurations and count ranges. In experiment, besides the general comparison studies to demostrate the effectiveness of the model, in ablation studies, the author find an interesting fact: Representation analyses reveal more accurate intermediate-layer count readout and stronger cross-configuration readout transfer.

Critical strengths
  1. author construct a controlled synthetic dataset for training and evaluation across visual configurations and count ranges for further use\
  2. the findings-representation analyses reveal more accurate intermediate-layer count readout and stronger cross-configuration readout transfer-might be insteresting and enlightening
Critical weaknesses
  1. The gains on synthetic images are much larger than on natural images. In Table 3, Qwen improves by 19.42 percentage points on Synthetic OOD and 5.21 on PixMo-Count. The corresponding gains for InternVL are 39.88 and 0.57 percentage points. These differences do not by themselves establish overfitting. However, the paper should explain why improvements are large on the authors’ generated images but more limited on natural images, and why transfer differs so much between the two models.
  2. I think the internal diagnostics are narrow. Because there is only one model, one layer, and one probe type in the experiment
Questions for authors

Could the authors provide actual dataset examples, side-by-side comparisons of training and OOD images, and failure cases, together with an analysis of the synthetic-to-real performance gap? Appendix A.2 provides LoRA settings, but are the vision encoder and multimodal projector frozen during both training phases? Which modules are actually trained?

6 comments
AuthorsOpenReview ↗

## Response — Reviewer a9oc

Thank you for recognizing the controlled dataset and the representation findings. We address the transfer gap, differences between backbones, diagnostic scope, concrete examples, and trainable modules below.

A1. Why synthetic gains exceed natural-image gains

The target-highlighting experiment shows that natural-image errors partly depend on target presentation and visual context.

Our synthetic training controls object configurations and count labels, while natural-image questions additionally require selecting and separating the queried objects. On fixed PixMo-Count cases, we preserve pixels within target masks, object positions, and image dimensions, then either darken the background to 20% brightness or replace it with gray. Table A1 reports the gray-background results for Base and the selected Full checkpoint on 20 historical shared-failure cases and nine jointly correct cases (Appendix A.4.5, Table 13).

After gray-background extraction, Full answers 6/20 shared-failure cases correctly, compared with 4/20 for Base, and retains 8/9 correct answers in the jointly correct group, compared with 7/9 for Base. Full performs better under the same intervention in both selected groups. This supports improved use of clarified target information for numerical answers. The remaining errors and disrupted correct answers also show that target selection and visual context continue to affect natural-image counting.

Table A1. Qwen target-highlighting diagnostics, using the same Base/Full summary as Appendix A.4.5, Table 13. Full denotes the selected seed13 checkpoint. Original-image columns show the historical results used for group selection; gray-background columns show intervention results. Entries are numbers correct.

ModelShared-failure originalShared-failure gray backgroundShared-correct originalShared-correct gray background
Base0/204/209/97/9
Full0/206/209/98/9

A2. Why the backbones show different transfer gains

We added Gemma as a third backbone, with Synthetic OOD performance between Qwen and InternVL.

The selected Gemma Full setting achieves 51.75% on Synthetic OOD 9–16, between Qwen’s 49.42% and InternVL’s 54.13%. Table A2 reports Gemma’s Base, Coverage-only, Interface-only, RNS+Interface, and Full results across synthetic and external tasks. All three backbones improve over their pretrained baselines on Synthetic OOD, with different gain magnitudes. The revised main Table 3 and Appendix D.6, Table 22 include the Gemma results.

Table A2. Gemma results across synthetic count ranges and external tasks. Accuracy (%).

ConfigurationOOD9–16OOD9–12OOD13–16PixMoQuantity
Base42.7567.2518.2567.3060.00
C-only6.6312.500.7563.8160.67
I-only41.2574.757.7565.3360.50
R+I20.2540.500.0064.7660.17
Full51.7580.5023.0069.2061.33

A3. Scope of the representation and behavioral diagnostics

The diagnostic study combines two representation readouts with three numerical-response interfaces.

The main Qwen representation analysis uses the last prompt-token representation at language block 17 across Base, Coverage, RNS+Interface, and Full. Ridge regression measures count readout, while prototype classification tests transfer from Grid to Scattered and Clustered layouts without refitting. The added behavioral study uses a separate set of 384 images for free counting and candidate-number scoring, together with 192 image pairs evaluated in both orders for quantity comparison. The original analysis is detailed in Appendix A.4, and the added diagnostic settings in Appendix A.4.4.

Coverage achieves 100% on comparison and 65.63% on both counting interfaces; Full reaches 86.20% on free counting and 86.98% on candidate scoring, compared with 82.81% for Base (Table A3). Full improves exact-count responses while comparison remains near ceiling. The readout–generation dissociation is consistent with the distinction between available quantity information and symbolic output discussed by Pang et al. [1]. The new behavioral results are included in Appendix A.4.4, Table 12.

Table A3. Qwen behavioral diagnostic. Free counting and candidate scoring use 384 images; quantity comparison uses 192 pairs in both orders. Accuracy (%).

CheckpointFree countingCandidate-number scoringQuantity comparison
Base82.8182.8198.18
Coverage65.6365.63100.00
Full86.2086.9898.70

A4. Training/test examples and natural-image cases

Yes. We provide actual training and evaluation examples, together with matched original and highlighted natural-image cases.

Figure 2 in the main text already illustrates the visual configurations and numerical tasks. Appendix C, Figure 5 adds actual enumerative and factorized Coverage images, an RNS difference-counting pair, local counting, and Synthetic OOD examples at 10 and 14 objects, with their queries and reference answers. Appendix A.4.5, Figure 4 provides paired PixMo-Count examples.

A5. Which modules are trained in each phase

Yes. The vision encoder and multimodal projector remain frozen in both phases; training updates only language-decoder LoRA parameters.

For Qwen3-VL-8B, LoRA targets q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, and down_proj. For InternVL3-9B, it targets wqkv, wo, w1, w2, and w3. Both use rank 16, alpha 32, and dropout 0.05. The pretrained decoder weights, token embeddings, output head, and biases also remain frozen. Full initializes the joint RNS+Interface phase from the Coverage adapter and continues optimizing its LoRA matrices. Appendix B.1, Table 14 lists the backbone-specific trainable modules and the freezing policy for both phases.

References

[1] Xingzhou Pang, Yifan Hou, Junling Wang, and Mrinmaya Sachan. Unveiling the Visual Counting Bottleneck in Vision-Language Models. Proceedings of the 43rd International Conference on Machine Learning (ICML), PMLR 306:96335–96355, 2026.

Dear Reviewers, as the discussion period comes to a close, please let us know if you have any remaining concerns or questions. We would be happy to address them. Thank you again for your time and consideration!

Reviewer_F2KoOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tests a two-stage post-training method for improving numerical performance in pretrained vision-language models. The first stage trains on synthetic scenes with counts from 1 to 8, using both 4 by 4 lattice patterns and scenes with varied layouts, object properties, and backgrounds. The second stage jointly uses direct-count and relational examples, along with selected CLEVR and TallyQA tasks. Experiments are run on Qwen3-VL-8B-Instruct and InternVL3-9B-Instruct. The largest improvements are on the paper's synthetic counting tests, while the gains on PixMo-Count and CompareBench are much smaller. For Qwen, the authors also analyze intermediate hidden representations. On the cross-layout diagnostic, the Coverage checkpoint has higher prototype accuracy but lower generated-count accuracy than the Base model.

Critical strengths

1- The paper organizes functional numerosity access around visual configuration, count, and task interface. It uses this structure for controlled training and evaluation across visual configurations, count ranges, and numerical queries (Section 3).

2- On Qwen's local synthetic evaluation, Full is nearly perfect for counts 1 to 8 and reaches about 85% for counts 9 to 12 (Table 2). On InternVL, Full improves over the Base VLM on the separate Synthetic OOD test (Table 3). On Qwen, Full retains most of Interface-only's held-out CLEVR and TallyQA performance (Table 4).

3- The checkpoint analysis compares ridge probes, cross-layout prototype classification, and generated-count accuracy across four Qwen checkpoints. Prototype and generation accuracy are measured on the same target images, and the two measures move in different directions across the tested checkpoints (Table 6).

Critical weaknesses

1- It is hard to tell how much of Full's gain over RNS+Interface comes from the Coverage design itself. Full gets 39,202 Coverage examples before the same downstream phase, so the comparison also adds much more training data (Appendix A.2). The same-sized Coverage variants differ in both count frequencies and visual configurations (Appendices A.1 and A.3.1), and Table 3 does not report Coverage alone.

2- It is still unclear what RNS adds beyond direct counting. RNS includes direct-count examples for counts 9 to 12, so the improvement at these counts does not isolate the effect of its relational examples (Table 7). Several RNS tasks also ask for a value that can be obtained by counting a displayed panel. The paper does not report held-out accuracy for the relational RNS task types or a matched direct-count-only baseline.

3- The evidence for broader generalization is still limited. On PixMo-Count and CompareBench, Full improves over Interface-only by only about 0.2 to 1.1 points across the two backbones, and Table 3 gives no uncertainty estimates for these differences. Synthetic OOD reports one score for counts 9 to 16 even though RNS includes direct-count training for counts 9 to 12, and Full's results are not split into counts 9 to 12 and 13 to 16.

4- The stronger internal readout has unclear behavioral value on the cross-layout diagnostic. Coverage raises prototype accuracy but lowers generated-count accuracy by 21.8 points, while Full keeps the prototype gain but restores generation only to the Base score (Table 6). The paper notes this mismatch, but Full still does no better than Base at generating counts on these images, even within the 1 to 8 Coverage range. The image-statistics control also finds strong low-level cues correlated with count in the diagnostic images, but does not test whether the VLM uses them (Appendix A.4.2).

Questions for authors

1- How much of Full's gain over RNS+Interface comes from the Coverage data's configuration diversity, rather than from the extra direct-count training? Could the authors add a control matched for examples, updates, and count frequencies but without the complementary configurations (Table 3, Appendices A.1, A.2, A.3.1)?

2- Do the gains extend to another VLM family, such as a multimodal Gemma model, beyond Qwen3-VL and InternVL3?

5 comments

Response to Reviewer F2Ko

Thank you for identifying the attribution, generalization, and diagnostic questions. We address each with the corresponding training control, task-level results, or representation analysis.

F1–F2. Extra training, count frequencies, and configuration coverage

We added a random-resampling control that matches the training budget and count frequencies.

The control uses 39,202 first-stage exposures and the same 4,096-example RNS+Interface phase, matching optimizer updates and the count histogram in Table F1a. Its first stage resamples from the same source pool, producing 25,074 distinct source entries with repeated exposures. This changes systematic coverage and repetition while retaining the source families and numerical distribution. The random control performs better at OOD 9–12, but shows markedly weaker generalization to counts 13–16 and lower accuracy on PixMo-Count. The matched protocol and results are reported in Appendix D.2, Tables 16–17.

Table F1a. Training exposures by cardinality in systematic and random Coverage.

CountSystematic Coverage exposuresRandom Coverage exposures
12,4602,460
22,5112,511
32,7272,727
43,3633,363
54,6134,613
66,4506,450
78,2118,211
88,8678,867
Total39,20239,202

Table F1b. Systematic Coverage and the random-resampling control. Accuracy (%).

ConfigurationOOD9–12OOD13–16PixMoQuantity
Systematic Coverage Full57.2541.2570.7361.67
Random Coverage→RI73.508.0068.4461.67

F3. Coverage-only performance

The Coverage-only checkpoint shows strong within-range counting and limited transfer beyond that training regime.

Table P1 reports 97.96% on local counts 5–8 for Coverage-only, while Tables P1 and F3 report 0.00% on both Synthetic OOD count ranges. We include this intermediate checkpoint alongside RNS+Interface and Full, together with its external-task results. This makes the role of the two optimization phases explicit: Coverage develops counting within the targeted visual and numerical regime, and Full improves larger-count and external-task scores over Coverage-only. The component results and count-range splits are provided in Appendix D.1, Table 15, and Appendix D.4, Table 20.

Table F3. Coverage-only and reference configurations. Accuracy (%).

ConfigurationOOD9–12OOD13–16PixMoQuantity
Base32.2527.7565.5261.00
Coverage-only0.000.0057.0360.00
RNS+Interface48.2530.0069.3359.33
Full57.2541.2570.7361.67

F4. What RNS adds beyond direct counting

Full scores higher on OOD 13–16 and delta counting in the reported comparison.

We replace the RNS allocation with direct-count supervision and report local counting, five relational task types, both OOD ranges, PixMo-Count, and Quantity in Table F4. The selected Full setting reaches 49.25% on Synthetic OOD 9–16 and 41.25% at 13–16; the replacement reaches 39.13% and 12.75%, respectively. The direct-count replacement scores higher on OOD 9–12 and PixMo-Count. The complete replacement comparison is included in Appendix D.3, Table 18.

Table F4. Full and the direct-count replacement: local, relational, OOD, and external tasks. Accuracy (%).

Evaluation / tasknFullRNS→direct-count replacement
Local direct count 1–494100.00100.00
Local direct count 5–89899.66100.00
Local direct count 9–129685.4279.17
Anchor-to-total9661.1175.00
Delta counting9653.820.00
Partition sum6445.3143.75
Removal6467.1957.81
Successor32100.00100.00
Synthetic OOD 9–1680049.2539.13
Synthetic OOD 9–1240057.2565.50
Synthetic OOD 13–1640041.2512.75
PixMo-Count52570.7371.81
CompareBench Quantity60061.6760.67

F5. Relational-task accuracy and panel-counting shortcuts

The task-level evaluation shows improved performance on numerical differences and part–whole queries.

In Table F5, Full reaches 53.82% on delta counting and 67.19% on partition sum, compared with 13.54% and 6.25% for Base, and 32.29% and 21.35% for Interface-only. There are 96 delta questions and 64 partition questions; trained-model results are means over three downstream seeds. We also report anchor-to-total and removal. Delta questions ask for the difference between two panels, while some other tasks can be answered by counting a displayed set. The gains extend to the supervised change and part–whole queries. These task-level results are included in Appendix D.3, Table 19.

Table F5. Relational-task accuracy (%). Base is a fixed checkpoint; trained configurations report three-seed means.

Relational tasknBaseInterfaceRNS+IFull
Anchor→total9635.4225.0059.7261.11
Delta counting9613.5432.2949.6553.82
Partition sum646.2521.3557.2967.19
Removal6415.6337.5054.1755.73

F6. External gains and uncertainty

Yes. We now report three-seed variability alongside the external-task gains.

On the complete external evaluations, Full achieves 70.73±2.92% on PixMo-Count and 61.67±0.76% on Quantity, compared with 69.59±1.64% and 60.78±1.08% for Interface-only (Table F6). Base scores 65.52% and 61.00%. Full has the highest mean on both tasks, with a larger gain over Base on PixMo. The ± values are sample standard deviations across three downstream training seeds, describing training-seed variability. Base is a fixed pretrained checkpoint. The three-seed external results are included in Appendix D.5, Table 21.

Table F6. Complete external evaluations: Qwen three-seed mean and sample standard deviation (%); Base is fixed.

ConfigurationPixMo525Quantity600
Base65.5261.00
Interface-only69.59±1.6460.78±1.08
RNS+Interface69.33±3.3459.78±0.67
Full70.73±2.9261.67±0.76

F7. Separating OOD 9–12 from 13–16

We now separate visual-distribution transfer at counts 9–12 from performance at the larger 13–16 counts.

The Synthetic OOD evaluation contains 800 images, with 400 in each range. Full scores 57.25% at 9–12 and 41.25% at 13–16 (Table F7; Appendix D.4, Table 20). RNS includes direct-count supervision at 9–12; counts 13–16 are beyond that direct-count answer range. We also describe the larger sets in RNS panels and larger numerical answers in Interface data, so the two ranges can be interpreted against the actual training exposure.

Table F7. OOD 9–12 and 13–16 alongside external-task results. Accuracy (%).

ConfigurationOOD9–12OOD13–16PixMoQuantity
Base32.2527.7565.5261.00
Coverage-only0.000.0057.0360.00
RNS+Interface48.2530.0069.3359.33
Full57.2541.2570.7361.67

F8–F9. Internal readout, numerical behavior, and visual cues

Count readout improves after Coverage, and generation recovers after joint RNS+Interface training.

Pang et al. [1] distinguish visual individuation, magnitude awareness, and symbolic mapping, and show that a model can retain quantity information while failing to generate the corresponding number. Prototypes formed on Grid images transfer to Scattered and Clustered images without refitting. Coverage improves this readout while generated-count accuracy drops by 21.8 points; Full retains much of the readout advantage and restores generation to the Base level. The original checkpoint analysis is reported in Appendix A.4.

The task-interface experiment shows improved exact-count responses after Full training, with quantity comparison remaining near ceiling. We evaluate Base, Coverage, and Full on a separate 384-image set using free counting, candidate-number scoring, and 192 comparison pairs presented in both orders. Coverage achieves 100% on comparison and 65.63% on both counting interfaces. Full reaches 86.20% on free counting and 86.98% on candidate scoring, compared with 82.81% for Base (Table F8). The contrast between strong comparison and weaker enumeration is consistent with the symbolic-mapping bottleneck studied by Pang et al. [1]. The added behavioral diagnostic is included in Appendix A.4.4, Table 12.

Table F8. Separate behavioral diagnostic: 384 images for counting and candidate scoring; 192 comparison pairs tested in both orders. Accuracy (%).

CheckpointFree countingCandidate-number scoringQuantity comparison
Base82.8182.8198.18
Coverage65.6365.63100.00
Full86.2086.9898.70

The background intervention tests how visual context affects counting while preserving pixels within target masks and their positions. We darken the background or replace it with gray. After gray-background extraction, the reported Full checkpoint answers 6/20 shared-failure cases correctly, compared with 4/20 for Base, and retains 8/9 correct answers in the jointly correct group, compared with 7/9 for Base (Table F8b). Full performs better under the same gray-background intervention in both selected groups. The paired responses also show that changing visual context can recover errors and disrupt previously correct answers. These results are reported in Appendix A.4.5, Table 13; the original area controls and image-statistics analysis are detailed in Appendix A.4.2.

Table F8b. Natural-image intervention cases: 20 historical shared-failure images and nine jointly correct original images. Full denotes the selected seed13 checkpoint. The original-image columns give the historical results used to select the groups; the gray-background columns give intervention results (Appendix A.4.5, Table 13).

ModelShared-failure originalShared-failure gray backgroundShared-correct originalShared-correct gray background
Base0/204/209/97/9
Full0/206/209/98/9

F10. Multimodal Gemma

Yes. We added Gemma and report its intermediate and final checkpoints alongside Base.

Table F10 includes Coverage-only, Interface-only, RNS+Interface, Full, both OOD count ranges, and the external tasks. The selected Full setting improves OOD 9–16 from 42.75% to 51.75% and OOD 13–16 from 18.25% to 23.00%. This supplies a third-family test of the framework. These results appear in the revised main Table 3 and Appendix D.6, Table 22.

Table F10. Gemma checkpoints and transfer results. Accuracy (%).

ConfigurationOOD9–16OOD9–12OOD13–16PixMoQuantity
Base42.7567.2518.2567.3060.00
C-only6.6312.500.7563.8160.67
I-only41.2574.757.7565.3360.50
R+I20.2540.500.0064.7660.17
Full51.7580.5023.0069.2061.33

References

[1] Xingzhou Pang, Yifan Hou, Junling Wang, and Mrinmaya Sachan. Unveiling the Visual Counting Bottleneck in Vision-Language Models. Proceedings of the 43rd International Conference on Machine Learning (ICML), PMLR 306:96335–96355, 2026.

Reviewer_PYL6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a systematic post-training (fine-tuning) framework for improving the visual number sense of vision-language models (VLMs). The key contributions are (1) the identification of visual configuration, cardinality, and task as the relevant dimensions along which numerical competence varies, and (2) a systematic post-training test of each dimension's role in VLMs' number-related performance, using Configuration Coverage, Relational Numerosity Supervision, and multi-interface (comparison and query-conditioned counting) supervision. Validation consists of experiments on two VLM families with a controlled synthetic dataset, external numerical benchmarks, matched-budget ablations, and representation (layer-wise count readout) analyses.

Critical strengths
  1. Systematic approach. The paper identifies key dimensions of numerical competence (configuration, cardinality, task) and builds a corresponding post-training framework to test them. A novel finding is that each dimension contributes to a different aspect of numerical performance; for example, configuration coverage improves robustness, while relational training promotes acuity at higher numbers. This dissociation is the paper's most consequential contribution.

  2. Grounding in the cognitive science literature. The paper provides a thorough review of the related psychology and neuroscience literature on number sense, which motivates the chosen dimensions and makes the framework more principled than a purely engineering-driven fine-tuning recipe.

Critical weaknesses
  1. Incomplete factorial design of the post-training conditions. In the nested post-training build-up, only A, A+B, and A+B+C were examined, and it is not clear why. The effect of C estimated from the contrast (A+B+C) vs. (A+B) can differ from the effect estimated by (B+C) vs. (B) or (A+C) vs. (A); i.e., the contributions attributed to each component may depend on what precedes them (interactions or order effects). A full factorization of the post-training conditions would make the claims about each dimension's distinct role much stronger.

  2. Limited generality and novelty of the evaluation. Only two VLM families are tested, so generalization of the findings to other VLMs is limited. In addition, the dataset and evaluation suites do not seem entirely new relative to existing counting/numerosity benchmarks.

Questions for authors
  1. (Clarification) What was the rationale for examining only the nested sequence A, A+B, A+B+C rather than a full factorial set of post-training conditions? Could you report additional conditions (e.g., B, C, A+C, B+C) to test whether the effect of each component is independent of the others, or whether there are interaction/order effects? This could materially affect my assessment of the claim that each dimension contributes to a distinct aspect of numerical performance.

  2. (Clarification/suggestion) Do the main findings hold for VLMs beyond the two families tested (e.g., different architectures or scales)?

  3. (Clarification) How do the proposed dataset and evaluation suites differ from existing counting and numerosity benchmarks, and what do they enable that prior datasets do not?

3 comments

Response to Reviewer PYL6

Thank you for highlighting the organization around configuration, cardinality, and task. We address your questions through expanded component comparisons, a third model family, and a clearer account of the resource’s experimental capabilities.

P1. Component contributions and interactions

The full combination delivers the broadest performance gains, while individual components show different strengths.

Table P1 includes Coverage-only, RNS-only, Interface-only, Coverage→RNS, Coverage→Interface, RNS+Interface, and Full. Coverage is the first optimization phase; RNS and Interface are combined in the second phase for Full. Full achieves strong local counting and the highest Synthetic OOD 13–16 and Quantity scores among the reported configurations. These results support the complete combination as an effective overall design. The expanded component comparison is included in Appendix D.1, Table 15.

Table P1. Qwen component combinations and transfer results. Accuracy (%).

Training conditionLocal count 1–4Local count 5–8Local count 9–12Synth. OOD 9–12Synth. OOD 13–16PixMo 525Quantity 600
Base98.9460.2028.1232.2527.7565.5261.00
C-only100.0097.960.000.000.0056.7659.83
R-only100.0092.8668.7567.506.0070.8661.33
I-only97.8757.4827.7840.5029.4269.5960.78
C→R100.0098.9892.7188.500.5069.9060.50
C→I100.0096.9426.0419.500.2567.2459.33
R+I100.0093.5474.6548.2530.0069.3359.78
Full100.0099.6685.4257.2541.2570.7361.67

P2. Generalization to a third VLM family

Yes. The added Gemma experiment reproduces counting gains in a third VLM family.

In the single-seed evaluation, the selected Full setting improves Synthetic OOD 9–16 from 42.75% to 51.75%. The two count ranges improve from 67.25% to 80.50% at 9–12 and from 18.25% to 23.00% at 13–16. The Gemma table also includes Coverage-only, Interface-only, RNS+Interface, and the external tasks. This extends the model-family evidence beyond Qwen and InternVL. The new backbone results appear in the revised main Table 3 and Appendix D.6, Table 22.

Table P2. Gemma results across synthetic count ranges and external tasks. Accuracy (%).

ConfigurationOOD9–16OOD9–12OOD13–16PixMoQuantity
Base42.7567.2518.2567.3060.00
C-only6.6312.500.7563.8160.67
I-only41.2574.757.7565.3360.50
R+I20.2540.500.0064.7660.17
Full51.7580.5023.0069.2061.33

P3. What the dataset enables

The resource makes configuration coverage measurable and directly controllable during post-training.

Inspired by configuration-recognition accounts of subitizing, we provide all 39,202 occupied-cell configurations of 1–8 objects on a 4×4 lattice, together with factorized layouts and appearance variations. This provides an explicit reference space for measuring covered, omitted, and repeated configurations. Under the same example budget, combined coverage achieves 96.42% mean accuracy across 12 visual conditions, compared with 91.77% for enumerative coverage and 93.41% for factorized coverage (original Table 9). The resource also supports numerical-change and part–whole supervision, followed by behavioral and representation tests on matched images. These capabilities enable controlled studies of how training coverage and supervision affect counting. The complete enumerative pool is distinguished from the samples actually used in combined training. The resource construction and examples are detailed in Appendix A.1 and Appendix D.7; the coverage-design comparison appears in Table 9.

Dear Reviewers, as the discussion period comes to a close, please let us know if you have any remaining concerns or questions. We would be happy to address them. Thank you again for your time and consideration!

I thank the authors for their detailed responses. My first question, about the factorial design, is fully resolved. On the second, adding only one more model family is a modest improvement in generality. On the third, it remains still unclear how the current manuscript moves beyond existing counting/numerosity benchmarks. I will therefore keep my score at 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.