acceptodds
Withdrawn from ICLR 2027

SPLIT-RL: Staged Perception–Language Reasoning with Claim-Level Advantages for Vision–Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision–Language (VL) reasoning requires a model to both extract relevant and accurate information from an image (visual reasoning, VR), and to infer the answer from it (language reasoning, LR). Reinforcement learning with verifiable rewards typically trains both through a single chain-of-thought with a final-answer reward. This gives every CoT token the same sequence-level advantage failing to distinguish capability specific errors. We propose **SPLIT-RL**, a staged post-training approach that trains VR and LR in disjoint phases. Because a group’s rollouts differ along one capability at a time, the group-relative advantage isolates it, and each phase is optimized using specific reward. We further introduce Claim Level Advantage (CLA-GRPO), which decomposes a VR phase rollouts into atomic visual claims and provides a fine-grained advantage at claim level based on visual-type group formation. Although trained in two phases, trained policy is evaluated like GRPO model, with a single CoT call at inference time. Under this protocol, SPLIT-RL improves average accuracy over GRPO by 1.5–6.1 points across Qwen3-VL models from 2B to 30B-A3B and InternVL3.5-8B. Evaluating each capability using oracle based diagnostic shows that answer-only GRPO leaves perception unchanged, whereas SPLIT-RL improves both VR and LR.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
vision language model, post-training RL

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_p11qOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SPLIT-RL separates visual perception and language reasoning during reinforcement learning to address the coarse credit assignment of final-answer rewards. It trains a single model first to produce useful image descriptions, then to reason from those descriptions without seeing the image. It also proposed CLA-GRPO, which assigns claim-level advantages by comparing the correctness of visual statements within groups of the same visual type. Across five backbones and six visual mathematics and logic benchmarks, SPLIT-RL improves average accuracy over GRPO by 1.4–6.1 points. Ablations validate the effectiveness of training perception before reasoning and combining completeness and correctness signals.

Critical strengths
  1. SPLIT-RL separates perception from reasoning during training, making it easier to target the source of errors, and introduces claim-level advantages to distinguish correct and incorrect visual statements within the same description.

  2. The experiments provide substantial evidence for both effectiveness and the proposed design choices. Improvements over GRPO are consistent across five backbones and six benchmarks, spanning different model sizes and architectures. Beyond aggregate accuracy, the ablations examine reward components and training order, while capability-specific diagnostics provide evidence that both perception and reasoning improve.

Critical weaknesses
  1. The performance gains may not justify the additional training cost. The paper does not provide a quantitative analysis of training efficiency. Based on the method description, SPLIT-RL appears substantially more expensive than vanilla GRPO: it involves two separate training stages, and CLA-GRPO additionally requires a visual judge. These components could increase training time and, depending on how the judge is deployed, memory requirements. The reported gains seem modest relative to this potential overhead.
  2. The evaluation is restricted to multimodal mathematical reasoning benchmarks and does not include perception-centric benchmarks. Given the method’s emphasis on improving perception, demonstrating its effectiveness on such benchmarks is essential.
  3. Minor problems: (1) Presentation: The figures could be improved, particularly Figure 2, where the small font makes the training pipeline and annotations difficult to read. (2) Missing relevant references: The related-work discussion omits several closely relevant studies. PTS: Unleashing Perception-Time Scaling to Multimodal Reasoning Models, which explores explicit perception decomposition and reinforcement learning for improving visual perception; CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning trains a captioning model using the accuracy of a text-only model answering questions from its captions, closely related to the proposed description bottleneck; ClaimDiff-RL: Fine-Grained Caption Reinforcement Learning through Visual Claim Comparison, which uses verified atomic claim differences to construct captioning rewards. Discussing these connections would help clarify SPLIT-RL’s novelty, particularly its staged optimization and claim-level advantage assignment.
Questions for authors
  1. How is the training efficiency of the SPLIT-RL?
  2. How is the SPLIT-RL's performance on more perception-centric multimodal benchmarks?
3 comments

We thank the reviewer for their valuable feedback and recognizing SPLIT-RL’s strengths of introducing fine-grained, capability-specific optimization, supported by consistent gains and comprehensive experimental validation. Below, we provide our response to address the concerns and questions.

weakness 1:

We provide the quantitative analysis of training efficiency below. SPLIT-RL improves average accuracy by 2.7 % while generating 54% fewer training tokens, at 1.75× the total GPU-hours. The breakdown identifies CLA judge waiting as the main source of the measured overhead. However, it is to be noted that our current implementation does not yet exploit large-scale batching of CLA judge requests. Larger batches and greater judge-side parallelism could reduce waiting time, and lowering idle GPU-hours.

MethodAvg. AccuracyGenerated tokensGPU-hours (total)GPU-hours (excluding CLA judge waiting)
GRPO72.601267M150150
SPLIT-RL75.3586M262141

Regarding the two-stage design, SPLIT-RL allocates the same combined maximum output-token budget as GRPO: 2,048 tokens per stage versus 4,096 tokens for GRPO. Its actual generated-token count is substantially lower compared to GRPO, decreasing from 1,267M to 586M.

CLA supervision does introduce additional overhead, which we include in the reported GPU-hours (total). Specifically, waiting for the CLA judge accounts for 121 of SPLIT-RL’s 262 GPU-hours. Excluding this waiting time, SPLIT-RL uses 141 GPU-hours, compared with 150 for GRPO, a 6% reduction. Additionally, judge calls are deployed through API calls avoiding any additional GPU memory requirements.

In summary, SPLIT-RL additional training cost is a trade-off for fine-grained supervision that achieves higher accuracy with substantially fewer generated tokens. We will update our manuscript with these results.

Weakness 2:

Our original evaluation followed the benchmark suite used in PEARL [1] (CVPR 2026). However, we agree that direct evaluation of perception is important to show effectiveness of our method. To address this concern, we extend the evaluation to PerceptionBench [2] and MMStar [3]. PerceptionBench [1] evaluates atomic visual perception capabilities through questions designed to isolate perceptual ability from reasoning and external knowledge. MMStar [2] evaluates multimodal understanding through questions selected to require visual input, spanning coarse and fine-grained perception capabilities.

The results below use Qwen3-VL-8B as the backbone, with accuracy reported in percentages.

ModelPerceptionBenchMMStar
Qwen3-VL-8B27.0765.13
GRPO25.6065.00
SPLIT-RL29.7372.67
Δ vs GRPO4.137.67

SPLIT-RL outperforms both GRPO and the base model on both benchmarks. Beyond its gains of 4.13 and 7.67 percentage points over GRPO, it improves over the base model by 2.66 and 7.54 points, respectively. GRPO decreases PerceptionBench accuracy relative to the base model and leaves MMStar performance essentially unchanged. This observation is consistent with our analysis in Fig. 1 that final-answer supervision alone does not necessarily improve visual perception. SPLIT-RL’s improvement on these two benchmarks shows that its benefits extend to atomic as well broader visual understanding.

[1] Chi Zhang et al. Perceptual-Evidence Anchored Reinforced Learning for Multimodal Reasoning. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 41111–41120, 2026

[2] Zichao Lin et al. PerceptionBench: Evaluating Atomic Visual Perception in Multimodal Large Language Models. arXiv preprint arXiv:2607.24957, 2026.

[3] Lin Chen et al. Are We on the Right Way for Evaluating Large Vision-Language Models? Advances in Neural Information Processing Systems (NeurIPS), 2024

Weakness 3

We thank the reviewer for highlighting these presentation issues and relevant studies. We will revise Figure 2 by enlarging labels and annotations, allocating more space to the two training phases, and making the flow between visual description and language reasoning clearer. We will also check all figures at the final publication size to ensure readability.

We will add PTS [4], CapRL [5], and ClaimDiff-RL [6] to the related-work discussion and explicitly explain their connections to SPLIT-RL.PTS uses supervised training followed by RL to teach step-by-step visual measurement, such as estimating a target length through repeated comparisons with a reference segment. SPLIT-RL instead separates the optimization of visual description and language reasoning into dedicated training phases within a single policy. CapRL shares the description bottleneck, rewarding a captioner through a separate text-only model’s answer accuracy. SPLIT-RL additionally optimizes the answering capability of the same policy in its language reasoning phase. ClaimDiff-RL verifies atomic differences between generated and reference captions and aggregates them into a caption-level reward. SPLIT-RL uses claim verification to assign credit within each description: it normalizes each relevant claim’s correctness against claims of the same visual type across rollouts for the same question and applies the resulting advantage to that claim’s tokens.

[4] Yifan Li et al. Unleashing Perception-Time Scaling to Multimodal Reasoning Models. International Conference on Learning Representations (ICLR), 2026.

[5] Long Xing et al. CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning. International Conference on Learning Representations (ICLR), 2026.

[6] Tianle Li et al. ClaimDiff-RL: Fine-Grained Caption Reinforcement Learning through Visual Claim Comparison. arXiv preprint arXiv:2605.20278, 2026.

Questions to Authors

Q1: answered in weakness 1

Q2: answered in weakness 2

We hope this clarifies the reviewer’s concerns. We would be happy to answer any further questions or provide additional clarifications if needed.

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

SPLIT-RL aims to improve credit assignment in reinforcement learning for vision-language reasoning. It trains one model in two stages. In the first stage, the model generates a visual description, which is scored using answer correctness and a completeness reward. CLA-GRPO adds local advantages based on judgments of visual claims, grouped by visual type. In the second stage, the model learns to answer questions from the description without seeing the image. At test time, it uses a single image-conditioned chain-of-thought response. Experiments on five backbones and six benchmarks report average gains of 1.4–6.1 percentage points over GRPO. The main contribution is the combination of staged training and claim-level credit assignment, with additional experiments intended to measure changes in perception and reasoning separately.

Critical strengths
  1. The training design addresses a clear problem. Separating visual descriptions from later reasoning makes it easier to provide feedback on missing or incorrect visual information. The completeness reward and claim-level judgments have clear roles. The ablations on training order and reward components are useful for understanding the overall recipe.
  2. The gains are reported across several backbones with a standard inference setup. Table 1 shows higher average accuracy than GRPO for all five backbones, covering different model sizes, two model families, and both dense and MoE architectures. Evaluation uses the same single-call CoT format as the baseline, which makes the method easier to use in practice.
Critical weaknesses
  1. The contribution of the new advantage rule is not clear from the current comparisons. ViRL [1] already uses visual process supervision and local advantage adjustment. SPLIT-RL differs by working with textual claims and normalizing their scores within visual-type groups, but the value of these changes needs clearer evidence. In Table 5, staged training alone scores 73.0%, compared with 72.6% for GRPO, while the full method scores 75.3%. Adding CLA introduces claim-level correctness labels, local advantages, and type-based normalization together. The ablations do not separate their effects. Comparisons using the same kind of claim supervision as a sequence-level reward, or using local advantages without type grouping, would help show what CLA-GRPO itself adds. Training steps are matched, but total generated tokens and judge costs are not reported.
  2. The evidence for separating improvements in perception and reasoning is limited. The two stages separate the training roles, but changes to their shared parameters can still affect both roles. In Stage 1, the answer reward also depends on the reasoner’s performance, as Section 3.3 acknowledges. The diagnostic in Table 6 is useful, but it measures perception indirectly through Kimi-K3’s accuracy when answering from the model’s descriptions. This score depends on both the visual information extracted and how well Kimi-K3 can use the resulting description. The reported 1.7-percentage-point improvement has no confidence interval. Direct checks of visual facts, together with uncertainty estimates, are needed to support the stronger claim that perception itself improves.
  3. The implementation does not ensure that each claim contains only one fact. Section 3.3.2 describes atomic visual claims, but Appendix 9.2 states that descriptions are split into sentences using PySBD. The judge prompt in Appendix 9.7 also allows a component to contain several claims and assigns its visual type according to the primary claim. Correct and incorrect facts in the same sentence may therefore receive one correctness score and share an advantage. This provides more local feedback than a rollout-level reward, but it does not fully address the problem of assigning credit to individual facts. The paper should report how often the resulting segments contain multiple facts and whether this affects the claimed benefit.

[1] Wang et al. From Illusion to Intention: Visual Rationale Learning for Vision-Language Reasoning. arXiv:2511.23031.

Questions for authors
  1. With the same training data, judge setup, and budget, how does CLA-GRPO compare with simpler ways of using claim-level feedback? Useful controls would include a sequence-level reward built from claim scores, quality-based multiplicative advantage adjustment, and local advantages without visual-type grouping. Please report results across multiple seeds, together with generated tokens and judge costs.
  2. Can you provide direct evidence for the quality of the claim decomposition and the improvement in perception? Please report how often a sentence-based segment contains multiple facts or a mixture of correct and incorrect facts. An independent evaluation of visual-fact correctness and missing information, along with confidence intervals for the Kimi-K3 diagnostic, would help assess both claims.
3 comments
AuthorsOpenReview ↗

Author's response 1/3

We thank the reviewer for their valuable feedback and recognizing the clear motivation and effectiveness of our staged training design, supported by the reward and training-order ablations, as well as the consistent improvements across five diverse backbones without requiring any changes to the standard inference setup. Below, we provide our response to address the concerns and questions.

Weakness 1:

ViRL adjusts trajectory-level advantages for visual crop actions based on their overlap with reference bounding boxes. In contrast, SPLIT-RL operates on individual textual visual claims, verifying their correctness against the image without requiring ground-truth bounding boxes or visual annotations. CLA-GRPO further computes claim-specific advantages by normalizing correctness within groups of claims sharing the same visual type. We will add ViRL to our related work and clarify these distinctions.

To investigate whether the improvements arise simply from access to stronger VLM supervision or from how that supervision is incorporated into optimization, we trained an additional GRPO baseline with sequence-level VLM supervision. This baseline uses the same types of recall and claim-correctness feedback, but aggregates individual claim-correctness scores into a single scalar reward, rather than assigning localized claim-level advantages.

MethodMVistaMVisionMVerseWeMathLogicVistaDynaMathperceptionbenchMMstarAverage
Base77.263.263.879.960.474.427.465.563.9
GRPO77.964.567.884.665.675.325.965.165.8
GRPO (recall + CLA)77.969.770.587.067.674.525.164.367.0
SPLIT-RL79.869.469.887.466.079.129.2772.269.1

Adding VLM supervision improves standard GRPO from 65.8% to 67.0% average accuracy. However, SPLIT-RL further improves accuracy to 69.1%, outperforming the judge-augmented GRPO baseline by +2.1 points. These results demonstrate that the performance improvement cannot be attributed solely to stronger VLM supervision. Instead, they provide evidence for the benefit of combining staged capability optimization with fine-grained claim-level credit assignment, rather than aggregating the same types of supervision into a sequence-level reward.

We agree that matching training steps alone does not fully capture training cost. We therefore additionally report total generated tokens and judge calls for the different training approaches. All counts below are in millions.

MethodGenerated tokensJudge calls ( answer/recall/ CLA)Total judge calls
GRPO1267.442/-/-0.442
GRPO (recall+CLA)655.442 / .221 / 1.372.033
SPLIT-RL586.442 / .221 / .6561.319

Compared with standard GRPO, SPLIT-RL generates 54% fewer training tokens, while improving average accuracy on the original six benchmarks from 72.6% to 75.3% (+2.7 points). The trade-off is additional VLM supervision, increasing the number of judge calls by approximately 3×.

More importantly, compared with GRPO augmented with the same types of VLM supervision, SPLIT-RL generates 11% fewer tokens and requires 35% fewer judge calls, while achieving higher accuracy. This indicates that staged capability optimization and claim-level supervision can provide a more effective use of both generated tokens and VLM-judge feedback than incorporating these signals into standard GRPO.

AuthorsOpenReview ↗

Author's response 2/3

Weakness 2:

We agree that, since both stages update the same policy parameters, capability-specific training does not guarantee complete isolation of perception and reasoning improvements. To strengthen our evidence, we conducted an additional direct evaluation of visual descriptions and computed 95% confidence intervals for our existing capability-specific diagnostic.

Direct evaluation of perception quality: To evaluate perception improvement independently of downstream answer accuracy and reasoning style, we conducted a preference-based evaluation of visual descriptions. We randomly selected 600 examples across the six benchmarks and compared descriptions generated by GRPO and SPLIT-RL along two axes: (i) completeness, measuring whether the description provides correct and sufficient visual information, and (ii) conciseness, measuring whether it focuses on query-relevant information without unnecessary details.

We used two VLM judges, Claude Opus 4.8 and Claude Sonnet 4.6, to compare description quality. The following results report SPLIT-RL's win/tie/loss percentages against GRPO and net win rates (win minus loss), with 95% confidence intervals.

MetricOpus 4.8 Win/Tie/Loss (%)Net Win [95% CI]Sonnet 4.6 Win/Tie/Loss (%)Net Win [95% CI]
Completeness17.4 / 70.6 / 12.0+5.3 [+1.0, +9.5]10.6 / 82.9 / 6.5+4.0 [+0.5, +7.4]
Conciseness32.2 / 38.7 / 29.0+3.2 [-3.0, +9.3]28.2 / 48.9 / 22.9+5.3 [-0.3, +10.9]

Both judges consistently favor SPLIT-RL on completeness, with 95% confidence intervals excluding zero. Conciseness also shows positive trends. Unlike the Kimi-K3 diagnostic, this evaluation directly compares the quality of generated visual descriptions without relying on a downstream reasoner's answer accuracy.

Confidence intervals for the Kimi-K3 diagnostic. We additionally computed 95% confidence intervals for the perception improvements reported in Table 6, using the same macro-averaged accuracy metric across the six benchmarks.

Comparison (Kimi-K3 as fixed reasoner)Δ Accuracy (pp)95% CI
SPLIT-RL vs. Base+1.72[+0.37, +3.15]
SPLIT-RL vs. GRPO+1.66[+0.41, +2.96]

Although the absolute gains are modest, both confidence intervals exclude zero, providing additional statistical support for the observed perception improvement.

Together, these analyses provide two complementary pieces of evidence: (i) improvements in visual description completeness measured directly through VLM-based preference judgments, and (ii) statistically supported improvements in downstream accuracy when the reasoner is held fixed. These evidences show that SPLIT-RL improves perception beyond the gains attributable to language reasoning alone.

AuthorsOpenReview ↗

Author's response 3/3

Weakness 3:

We agree that our current implementation uses sentence-level segmentation rather than strictly atomic fact decomposition, so a single segment may contain multiple visual facts. While CLA-GRPO provides finer-grained credit assignment than sequence-level GRPO, it does not guarantee independent advantages for every individual fact. Our VR-stage prompt instructs the model to present enumerable visual facts in a structured, bullet-point format, encouraging separation of individual visual observations. The judge instruction to select the primary claim type when a component contains multiple claims is included as a fallback to accommodate imperfect segmentation by PySBD, rather than as an intended feature of our claim decomposition.

To quantify this limitation, we conducted an additional analysis measuring the proportion of type-mixed segments, i.e., segments containing visual claims belonging to multiple types, at different stages of VR training.

Training StepSegmentsType-mixed (%) [95% CI]
087156.4 [52.5, 60.5]
541,49343.8 [40.5, 47.2]
1083,48323.6 [21.6, 25.6]

The proportion of type-mixed segments decreases substantially from 56.4% to 23.6% over training, suggesting that the model increasingly produces descriptions with more homogeneous visual claim types within each segment. This makes our type-specific advantage assignment more targeted as training progresses.

Questions: Q1: answered in weakness 1 response Q2: answered in weakness 3 response

We hope this clarifies the reviewer’s concerns. We would be happy to answer any further questions or provide additional clarifications if needed.

Reviewer_8XP6OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SPLIT-RL, a staged RLVR recipe for VLMs that separates perception ("VR") from reasoning ("LR") within a single policy. Stage 1 (VR): the model writes a question-relevant image description with no reasoning or answer. Stage 2 (LR): starting from Stage-1 weights, standard GRPO trains the model to answer from one on-policy description, without the image. Inference is ordinary single-pass CoT with the image.

Critical strengths
  1. Consistent gains across five backbones with a simple recipe that adds no inference cost. SPLIT-RL improves average accuracy over step-matched GRPO on every backbone tested.

  2. A clean, plausible decomposition with some supporting evidence. Using the description as an information bottleneck, so that answer correctness also verifies description sufficiency, is an elegant way to turn (image, question, answer) triples into perception supervision without caption labels. The stage-order ablation (VR→LR 75.3 vs. LR→VR 72.1, same rewards and budget) supports the claim that perception should be trained before reasoning.

Critical weaknesses
  1. Unmatched supervision. The VR stage receives per-claim correctness verdicts from Claude Sonnet 4.6 (which sees the image) plus a Claude recall judgment. The GRPO baseline gets final-answer correctness only.

  2. Staging alone (VR→LR, answer reward only) scores 73.0 on 8B vs. GRPO's 72.6 (Table 5), which is within noise. Most of the gain appears only once the Claude-judged recall and CLA signals are added (75.3).

  3. Without an entangled-CoT GRPO baseline given the same Claude perception signals, the paper cannot attribute the improvement to staging or CLA rather than to distilling judgments from a stronger VLM.

  4. Reasoning-style shaping. The LR prompt (App. 9.6) explicitly forbids "wait"/"actually"/re-interpretation and asks for "one plan, one derivation, one final answer.

  5. The headline qualitative example (Fig. 3) shows GRPO hitting the token limit after 49 "Wait"s versus 3 for SPLIT-RL.

  6. Part of the gain may therefore come from shorter, more decisive outputs and fewer truncations rather than from better perception. No output-length, truncation or no-answer statistics are reported.

  7. Format-confounded probe. Each probe block uses exactly the format SPLIT-RL was trained on (describer prompt; text-only reasoning from a description), while GRPO never saw either. This favors SPLIT-RL by construction; the same applies to Table 7.

  8. Even so, the measured VR gain is small: +1.7 for the full method, only +0.6 for the pure bottleneck reward (VR→LR), and −0.2 for VR+Recall. So by the paper's own numbers, the core claim that the description bottleneck "pushes the model toward more complete and accurate perception" is barely supported. The VR gain appears only with Claude-judged CLA.

  9. No variance estimates. All results are single runs with greedy decoding: no seeds, no confidence intervals.

  10. Novelty positioning. The bottleneck idea is close to Vision-SR1 and MoCA, and the perception/reasoning decoupling is close to VLM-CapCurr and PAPO. Vision-SR1 and MoCA are discussed but not compared experimentally.

Additional issues

  • Text contradicts Table 1 for 30B-A3B. Sec. 4.2.1 claims the largest gains are on WeMath/LogicVista/DynaMath (+3.2 to +4.2) and that SPLIT-RL trails GRPO only on MathVista and MathVision. Table 1 shows WeMath +3.5, LogicVista +0.1, DynaMath +5.8, and SPLIT-RL trailing on MathVista (−0.8) and MathVerse (−1.7); MathVision is +1.5.
  • Inconsistent headline range. The gain range is "+1.4 to +6.1" (abstract, intro, conclusion) but "+1.5 to +6.1" in Sec. 4.2.1. The conclusion also reports the gains in "%" rather than points.
  • Fig. 6 contradicts Fig. 1. In Fig. 1 the gold answer 2.20 with r=10.5r=10.5 implies ∠BOC=12∘\angle BOC = 12^\circ, and reading 168° as ∠BOC\angle BOC is labeled the error. Yet Fig. 6's recall-judge example calls 168° "the central angle … the arc-length formula requires." The worked example of the text-only recall judge thus endorses the very misreading the paper criticizes, which also illustrates that a text-only judge cannot know which facts are actually needed.
  • CLA design.
    • Same-type sibling groups can pool unrelated facts (e.g., a count of points vs. a count of bars), so the baseline is a type-level average correctness rate rather than a comparison of alternative statements of the same fact.
    • Setting Acla=0A_{cla}=0 beyond 60 claims is described as a penalty on long generations, but a zero advantage is not a penalty.
    • Judge accuracy is not validated against human labels.
  • Isolation claim. The paper says Stage-1 rollouts "differ only in what was described," but each description is scored by a single stochastic reasoner sample at temperature 1, so reasoning noise enters the VR advantage.
  • Terminology. "Visual Reasoning" names a phase whose prompt forbids reasoning; "perception" would be clearer.
  • Cross-references and typos. "Appendix 7" should be App. 9.5 / Table 7. "Section (3.3) CLA-GRPO and recall reward" is garbled. Eq. 11 is referenced before it is defined. Typos: "from from", "diagonistic", "addition quanlitative". Figure 5 is hard to read.
Questions for authors
  1. What are the mean output lengths and the truncation and no-answer rates at inference for Base, GRPO and SPLIT-RL? How does GRPO perform with a larger training token cap, or with the same "linear, no-wait" reasoning instruction used in your LR prompt? This would separate style and length effects from capability gains.

  2. The probe evaluates VR and LR in formats SPLIT-RL was trained on but GRPO was not. Can you measure perception in the single-pass setting instead, e.g., have the same CLA judge score the visual claims inside single-pass CoTs from each model? Given that VR→LR gains only +0.6 VR, does the bottleneck reward itself improve perception?

  3. For CLA-GRPO, how does type-based sibling grouping compare with simpler alternatives, such as a global claim pool or a fixed baseline (y−0.5y-0.5), at the same λ\lambda? What is the judge's claim-level accuracy against a small human-annotated sample?

6 comments
AuthorsOpenReview ↗

Author's response 1/n

We thank the reviewer for their valuable feedback and recognizing SPLIT-RL’s strengths of consistent improvements across five backbones without additional inference cost, as well as the effectiveness of our perception–reasoning decomposition and the supporting evidence from the stage-order ablation. Below, we provide our response to address the concerns and questions.

Weakness 1,2,3:

To address this, we conducted an additional experiment with an entangled-CoT GRPO baseline augmented with VLM-judge supervision, and extended the evaluation to two additional perception-focused benchmarks, PerceptionBench and MMStar.

MethodMVistaMVisionMVerseWeMathLogicVistaDynaMathPerceptionBenchMMStarAverage
Base77.263.263.879.960.474.427.465.563.9
GRPO77.964.567.884.665.675.325.965.165.8
GRPO (with VLM judge)77.969.770.587.067.674.525.164.367.0
SPLIT-RL79.869.469.887.466.079.129.2772.269.1

The results show that stronger VLM supervision does benefit standard GRPO, improving its average accuracy from 65.8% to 67.0%. However, SPLIT-RL still outperforms judge-augmented GRPO by +2.1 points, demonstrating that the improvement cannot be explained solely by the availability of stronger VLM supervision.

More importantly, the differences are particularly pronounced on the two perception-focused benchmarks. SPLIT-RL outperforms judge-augmented GRPO by +4.17 points on PerceptionBench and +7.9 points on MMStar. Interestingly, adding VLM-judge supervision to entangled-CoT GRPO slightly reduces performance on both benchmarks relative to standard GRPO, whereas SPLIT-RL improves substantially. This suggests that simply introducing additional perception supervision into an entangled CoT is insufficient; how this supervision is incorporated into training and assigned to individual visual claims also matters.

We agree that staging alone provides a relatively modest improvement (+0.4 on the original six benchmarks). Our central contribution, however, is the combination of capability-specific staged optimization and fine-grained claim-level credit assignment, rather than staging in isolation. The new comparison provides further evidence that this structured optimization is more effective than simply augmenting conventional GRPO with stronger VLM-judge supervision. We will include these results and clarify this distinction in the revised manuscript.

AuthorsOpenReview ↗

Author's response 2/n

Weakness 4,5,6:

Since the LR stage operates without access to the image, we observed that the model sometimes engages in repeated reinterpretation or speculative reasoning when the visual description is incomplete or incorrect. The LR-stage prompt was designed to discourage this behavior. However, we agree that it is important to disentangle improvements in perception from potential gains due to reasoning style, shorter outputs, and fewer truncations. To address this, we conducted two additional experiments and analyzed the output lengths and truncation rates.

1. Controlling for the LR-stage prompting style.

We trained an additional GRPO baseline using the same reasoning-style instructions as our LR stage, isolating the contribution of the prompt itself.

Method (Qwen3-VL-8B)MVistaMVisionMVerseWeMathLogicVistaDynaMathAVGΔ vs. GRPO
GRPO77.964.567.884.665.675.372.60—
GRPO + LR-stage prompt78.969.070.384.859.773.972.76+0.16
SPLIT-RL79.869.469.887.466.079.175.25+2.65

The results show that the LR-stage prompting style alone provides only a +0.16-point improvement over standard GRPO, while SPLIT-RL outperforms this prompt-controlled baseline by +2.49 points. This indicates that the performance gains cannot be explained merely by the specific reasoning instructions.

Importantly, the LR-stage prompt is used only during training. At inference, SPLIT-RL and all baselines use exactly the same single-pass CoT prompt and decoding settings, as described in Section 4.1.

2. Direct evaluation of perception quality.

To provide evidence of perceptual improvement independent of answer accuracy and reasoning style, we conducted an additional preference-based evaluation of the models' visual descriptions. We randomly selected 600 examples across the six benchmarks and compared descriptions produced by GRPO and SPLIT-RL along two axes: (i) completeness, measuring whether the description provides correct and sufficient visual information, and (ii) conciseness, measuring whether it focuses on query-relevant visual information without unnecessary details.

We used two independent VLM judges, Claude Opus 4.8 and Claude Sonnet 4.6, to determine which description was better on each axis. The results below report SPLIT-RL's win/tie/loss percentages against GRPO and net win rates (win minus loss) with 95% confidence intervals.

MetricOpus 4.8 Win/Tie/Loss (%)Net Win [95% CI]Sonnet 4.6 Win/Tie/Loss (%)Net Win [95% CI]
Completeness17.4 / 70.6 / 12.0+5.3 [+1.0, +9.5]10.6 / 82.9 / 6.5+4.0 [+0.5, +7.4]
Conciseness32.2 / 38.7 / 29.0+3.2 [-3.0, +9.3]28.2 / 48.9 / 22.9+5.3 [-0.3, +10.9]

Both judges consistently favor SPLIT-RL on completeness, with positive confidence intervals that exclude zero. Conciseness also shows positive trends, although its confidence intervals include zero. These findings provide direct evidence that SPLIT-RL improves the quality of visual descriptions, beyond its effect on downstream answer accuracy. They also complement our existing Kimi-K3 oracle-based capability analysis (Section 4.3.2), which shows that standard GRPO leaves perception largely unchanged, whereas SPLIT-RL improves it.

3. Output length and truncation analysis.

We additionally measured generation lengths and truncation rates under the identical inference setup, with a maximum of 16,384 output tokens.

ModelMean TokensMedian TokensTruncated (16K limit)
Qwen3-VL-8B (Base)3,11835415.7%
GRPO7,6842,50943.6%
SPLIT-RL4,62969621.7%

SPLIT-RL produces shorter generations than GRPO and reduces the truncation rate from 43.6% to 21.7%, a reduction of 21.9 percentage points. We acknowledge that fewer truncations may contribute to the overall accuracy improvement however fewer truncation is a consequence of improved perception as previous results. However, the prompt-controlled baseline shows that reasoning-style instructions alone yield negligible aggregate gains, while the independent perception evaluation demonstrates improvements in the visual descriptions themselves. Together, these results support our claim that SPLIT-RL's benefits extend beyond shorter or more decisive reasoning and include improved perception.

We will incorporate these additional experiments and the output-length analysis into the revised manuscript.

AuthorsOpenReview ↗

Author's response 3/n

Weakness 7,8:

We agree that the capability-specific diagnostic probes use the same generation formats as SPLIT-RL's training stages, which may favor SPLIT-RL. We would like to clarify the purpose of these analyses and provide additional evidence of perception improvement.

1. Main results are evaluated in a setting that favors GRPO.

Importantly, all our main results (Table 1) use standard single-pass CoT inference. This is precisely the format GRPO is trained on, whereas SPLIT-RL is never directly optimized in this format. Despite this, SPLIT-RL consistently outperforms GRPO across all five backbones, including a +2.7-point improvement on Qwen3-VL-8B. Table 7 is provided only as a supplementary analysis to examine performance under two-pass inference, rather than as evidence for our main performance claims.

2. Additional direct evaluation of perception quality.

We acknowledge that our Kimi-K3 oracle-based diagnostic (Section 4.3.2) is not completely format-neutral. Its purpose is to approximately isolate VR and LR capabilities by holding the other component fixed, rather than to serve as the primary evidence of overall effectiveness.

To provide complementary evidence, we conducted an additional preference-based evaluation of visual descriptions, independent of downstream text-only reasoning accuracy. As detailed in our response to the previous concern, we evaluated 600 samples across all six benchmarks using two VLM judges, Claude Opus 4.8 and Sonnet 4.6. Both judges favor SPLIT-RL over GRPO on description completeness, with net win rates of +5.3% [95% CI: +1.0, +9.5] and +4.0% [95% CI: +0.5, +7.4], respectively. Conciseness also shows positive trends. The consistent improvement in completeness provides additional evidence that SPLIT-RL learns to produce more complete and accurate visual descriptions.

3. Understanding the modest VR gains and the role of CLA-GRPO.

We agree that the standalone VR improvement from staging is modest (+0.6), and adding the recall reward alone slightly reduces it (−0.2). However, this is consistent with the motivation for introducing complementary correctness and completeness signals.

The recall reward evaluates whether the description contains all information needed to answer the question, but does not verify whether those visual claims are correct, since the recall judge never sees the image. Consequently, a description may be complete but contain incorrect visual claims, potentially explaining why recall alone does not improve VR accuracy. CLA-GRPO directly addresses this limitation by verifying individual claims against the image and assigning fine-grained advantages to their corresponding tokens.

When combined, recall and CLA-GRPO improve VR by +1.7 points, compared to no improvement under standard GRPO. This is particularly meaningful given that our diagnostic shows standard GRPO primarily improves LR (+3.8) while leaving VR unchanged. Furthermore, the complementary benefits extend to downstream performance, where combining recall and CLA improves overall accuracy from 73.0 to 75.3, substantially more than either component alone.

We agree that the description bottleneck alone does not explain the full perception improvement. Rather, our central contribution is that the bottleneck enables capability-specific optimization, while CLA-GRPO and recall provide the fine-grained correctness and completeness signals needed to improve perception effectively. We will clarify this distinction and the limitations of the diagnostic probes in the revised manuscript.

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

Related papers

Discussion (0)

Sign in to comment.