- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes CR², a framework for multimodal latent reasoning that first constructs a visual-semantic latent block and then reuses it as context for answer generation. Separate visual and semantic branches receive supervision from RoI features and answer-related concepts. Training and inference follow the same latent construction procedure. Experiments with Qwen2.5-VL 3B and 7B across six benchmarks demonstrate improvements over standard SFT and several latent reasoning baselines.
Critical strengths
-
The paper identifies two relevant limitations of existing latent visual reasoning methods: conflicting representational objectives and differences between training and inference.
-
Evaluation covers diverse multimodal capabilities, with particularly strong improvements in counting, spatial relations, OCR, and real-world visual understanding at the 7B scale.
-
Latent replacement experiments and representation analyses support the constructed latents contribute to answer prediction.
Critical weaknesses
-
The training–inference gap comparison is not directly matched. The 12.8-point gap for LVR compares ground-truth RoI features with predicted latents, whereas the 2.02-point gap for CR² compares RoI-matched and fixed latent lengths. LVR also uses eight latents, while CR² reports its best fixed-length result with sixteen. These differences make it difficult to attribute the smaller gap to the proposed method. A comparison under a unified evaluation protocol would strengthen this claim.
-
Baseline for 7B is limited. The 7B baselines are limited. Beyond the original backbone and SFT, only LVR is evaluated. Additional recent latent reasoning baselines would strengthen the comparison.
-
The ablations conflate architecture and supervision. Removing each auxiliary loss while retaining both branches would better isolate the contribution of visual and semantic supervision.
-
Inference costs are not sufficiently reported. Latency, memory usage, and FLOPs comparisons would clarify the overhead of the two-pass framework.
Questions for authors
see the above weakness
6 comments
Response to Reviewer kH7n(2/2)
Weakness-III
The ablations conflate architecture and supervision. Removing each auxiliary loss while retaining both branches would better isolate the contribution of visual and semantic supervision.
Thank you for the suggestion. We added ablations that retain both branches while removing the visual loss, the semantic loss, or both. As shown in Table 3, full supervision outperforms all three settings across the six reported metrics.
Table 3: Ablations of latent-head architecture and auxiliary supervision.
| Variant | Hallusion | MME | RWQA | V* | HR-4K | HR-8K |
|---|---|---|---|---|---|---|
| CR² | 61.1 | 2189.3 | 62.2 | 73.8 | 65.1 | 60.3 |
| VisualHeadOnly | 57.7 | 2133.9 | 57.9 | 72.2 | 62.1 | 59.0 |
| SemanticHeadOnly | 56.9 | 2146.4 | 60.7 | 71.7 | 63.6 | 59.6 |
| SharedHead | 55.7 | 2138.6 | 59.5 | 71.2 | 63.2 | 59.3 |
| SharedHead w/o Sup. | 56.6 | 2130.25 | 60.5 | 69.6 | 62.3 | 58.3 |
| VisualSup. | 60.3 | 2110.3 | 57.9 | 70.6 | 62.6 | 58.5 |
| SemanticSup. | 59.7 | 2124.6 | 56.7 | 71.2 | 64.6 | 59.2 |
| NoSup. | 58.1 | 2093.5 | 56.3 | 70.1 | 62.8 | 59.6 |
Action taken: We have revised in Section 5.3.
Weakness-IV
Inference costs are not sufficiently reported. Latency, memory usage, and FLOPs comparisons would clarify the overhead of the two-pass framework.
We evaluate inference efficiency on an NVIDIA A100 80GB GPU, using 100 samples for latency and memory and 20 samples for profiler-supported FLOPs. CR² maintains stable latency as latent steps increase, achieving a 1.28×–2.02× speedup over LVR with a modest increase in memory and profiler-supported FLOPs.
Table 4: Inference latency, peak memory, and profiler-supported FLOPs.
| Setting | End-to-end latency | Peak GPU memory | FLOPs |
|---|---|---|---|
| qwen2.5-vl-3B | 1246.44 ms | 7.05 GiB | 2.63 T |
| LVR-3B (step 4) | 762.58 ms | 7.10 GiB | 2.66 T |
| LVR-3B (step 8) | 986.66 ms | 7.10 GiB | 2.68 T |
| LVR-3B (step 16) | 1240.20 ms | 7.10 GiB | 2.73 T |
| CR_2 (step 4) | 594.59 ms (1.28x) | 7.18 GiB | 2.68 T |
| CR_2 (step 8) | 605.67 ms (1.63x) | 7.18 GiB | 2.72 T |
| CR_2 (step 16) | 612.92 ms (2.02x) | 7.18 GiB | 2.81 T |
Action taken: We have added relevant discussion in Appendix D.
A Friendly Reminder to Review Our Rebuttal
Dear Reviewer kH7n,
Thank you for your thoughtful review and constructive suggestions. We would like to briefly follow up on our rebuttal, in which we have addressed your concerns and updated the manuscript.
To clarify the training–inference gap comparison, we evaluated LVR under multiple latent lengths, including RoI-matched lengths, and distinguished the effects of feature substitution from those of latent length (Weakness-I). We also added recent 7B baselines and vision-centric benchmarks (Weakness-II), as well as controlled ablations that retain both branches while removing the auxiliary losses (Weakness-III). Finally, we reported inference latency, peak memory, and profiler-supported FLOPs to characterize the overhead of our two-pass framework (Weakness-IV).
We hope that these clarifications and additional experiments are helpful. We would be grateful if you could consider our response, and we would be happy to discuss any remaining questions.
Thank you again for your time and consideration.
Best regards,
All Authors
Response to Reviewer kH7n(1/2)
Thank you for raising this point. The LVR comparison can involve two sources of mismatch: the latent features used during training and inference, and the number of latent steps.
Controlling latent length in LVR. We additionally evaluated LVR with 8, 12, and 16 fixed latent steps, as well as RoI-matched steps (Table 1). Even when the number of predicted latents matches the RoI token count, a 10.4% gap remains between using ground-truth RoI features and predicted latents.
Distinguishing the remaining gap in CR². CR² instead uses the same latent construction-and-replay procedure during training and inference. This removes the specific train–test substitution of ground-truth RoI features with predicted latents. A length mismatch can still remain: under our CR² evaluation, it differs by 2.02%.
Table 1: LVR accuracy gaps under fixed and RoI-matched latent lengths.
We have revised the manuscript to distinguish feature-substitution and latent-length effects, rather than presenting the two numbers as a matched reduction of the same gap.
Action taken: We have added relevant discussion in Appendix C.
Thank you for the suggestion. We have added Monet-7B, LIVR-7B, and CoVT-7B alongside the original baselines, and expanded evaluation to V, CV-Bench, HRBench, MMVP*. As shown in Table 2, CR²-7B achieves the highest reported scores among the listed 7B models on MME, OCRBench, V*, CVBench-2D, HRBench-4K/8K, and BLINK Spatial Relation. In particular, it exceeds the strongest listed baselines by 2.5 percentage points on OCRBench and 3.5 points on BLINK Spatial Relation.
Table 2: Benchmark comparison with additional 7B latent-reasoning baselines.
Action taken: We have revised in Section 5.2.