acceptodds
Under review as a conference paper at ICLR 2027

LenGuard-GPC: Length Guarding with Guided-Prompt Consistency for Spatial Reasoning Reinforcement Learning

Abstract

Multi-view spatial reasoning requires vision-language models to compare visual evidence across images, align object correspondences, and infer spatial relations over long visual contexts, a setting where chain-of-thought reasoning tends to grow verbose without becoming more accurate. Reinforcement learning with verifiable rewards is a natural fit for this task, but standard GRPO reward relies on sparse outcome-level feedback and gives no signal about where a reasoning trajectory goes wrong, nor any control over its length. We propose LenGuard-GPC, a dense reward framework that addresses both problems together. For each sampled trajectory, it compares the token-wise predictive distributions under a standard prompt and a guided prompt, and uses the resulting token-sum KL divergence as a dense reward signal. Since this KL penalty accumulates over tokens and would otherwise reward shorter responses regardless of their quality, we introduce a staged length bonus that keeps reasoning length within a controlled range without simply encouraging brevity.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Vision-Language Models, Reinforcement Learning, Reasoning Efficiency

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_R85bOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes LenGuard-GPC for multi-view spatial reasoning with GRPO. It uses guided-prompt consistency to provide a denser trajectory-level signal than answer correctness alone and introduces a staged length guard to balance reasoning sufficiency and conciseness. Comparisons with the base model and vanilla GRPO across in-domain and out-of-domain benchmarks, together with fine-grained analyses, generation-length statistics, and ablations, indicate improved spatial reasoning and reduced redundant generation.

Critical strengths
  1. The method directly addresses the stated failure modes. The reward design jointly tackles sparse supervision, verbose reasoning, and collapse to overly short responses. It is simple to integrate and incurs no additional inference-time cost.

  2. The empirical trends are broadly consistent. LenGuard-GPC improves over vanilla GRPO across multiple benchmarks while reducing generation length. The fine-grained results are also broadly aligned with the intended improvement in cross-view spatial reasoning.

Critical weaknesses
  1. Guided-prompt consistency is not directly validated as a measure of reasoning quality. The method assumes that a trajectory should receive a higher reward when the predictive distributions under the standard and guided prompts are more similar. However, both distributions are produced by the same policy, and a low KL may instead reflect weak guidance, model insensitivity to the guidance, or shared prompt biases. The paper demonstrates the utility of this reward mainly through downstream task performance but does not analyze how the KL score relates to correctness, reasoning quality, or grounding in visual evidence.
  2. The experiments do not sufficiently isolate the contributing factors. The inputs, template, filtering procedure, and possible use of answer information in guided-prompt generation are not fully disclosed. Important controls are also missing, including a length-only variant, a flat length bonus, mean per-token KL, shuffled or generic guidance, hyperparameter sensitivity, and multiple random seeds. It therefore remains unclear whether the gains arise from informative guidance, length regularization, or their specific combination.
Questions for authors
  1. What information is available when generating each guided prompt? Can the generating model access the ground-truth answer or a complete rationale? Please provide the generation template, representative examples, and quality-control procedure, and explain how answer leakage or privileged teacher information is excluded.

  2. Could the authors report multi-seed results and add controls for a length-only objective, a flat length bonus, mean per-token KL, and random or generic guidance? These results would materially affect my assessment of the mechanism and robustness of the reported gains.

  3. Couldn the authors analyze the correlation between guided-prompt KL and answer correctness, human-rated reasoning quality, or grounding in visual evidence? Does the reward remain effective with irrelevant, generic, or randomized guidance? Such analyses would help establish that the gains arise from meaningful guided consistency rather than generic prompt regularization.

4 comments

Q1: Effectiveness of Guided-Prompt Consistency as a Reward Signal

R1: We thank the reviewer for this insightful comment regarding the effectiveness of guided-prompt consistency. Guided-prompt consistency serves as an auxiliary training signal that complements the verifiable accuracy reward in GRPO. Its contribution is directly evaluated through the following ablations.

Importantly, our ablation results provide empirical evidence of the effectiveness of this reward component. As reported in Section 4.3, we compare vanilla GRPO, guided-prompt consistency without length regulation, and our full method. We additionally evaluate a variant without the KL reward to isolate the contribution of staged length regulation.

Table 1. Ablation of reward components on MMSI.

MethodGuided-Prompt ConsistencyStaged Length GuardMMSI Acc. (%)
Vanilla GRPONoNo32.2
w/o Staged Length Reward (Constant Bonus)YesNo33.5
w/o KL (Length bonus Only)NoYes32.5
LenGuard-GPC (Ours)YesYes36.1

As shown above, introducing guided-prompt consistency improves MMSI accuracy from 32.2% to 33.5%, whereas using only the staged length reward yields a marginal improvement to 32.5%. This suggests that the performance gain cannot be explained by length regulation alone. Combining both components further improves accuracy to 36.1%, outperforming the KL-only and length-only variants by 2.6 and 3.6 percentage points, respectively.

These findings indicate that guided-prompt consistency provides an effective additional training signal, while the staged length guard complements it by regulating reasoning length. Although KL divergence is not intended as a standalone measure of reasoning correctness, the ablation results demonstrate the complementary roles of the two reward components in improving downstream spatial reasoning performance.

We thank the reviewer for this valuable suggestion, which helps clarify the individual contributions of our reward components and strengthen the empirical support for LenGuard-GPC.

Q2: Further Evaluation of Reward Design Choices

R2: We thank the reviewer for the constructive suggestions regarding the contributions of different reward components and alternative design choices. We would like to clarify the existing ablation results and provide additional controlled comparisons.

We clarify that the variant labeled “w/o staged length reward” in Section 4.3 of the original submission retains a constant length bonus while removing the staged bonus design. It therefore corresponds to “KL + Constant Bonus” in Table 2, achieving 33.5% accuracy on MMSI with an average response length of 163.9 tokens. Replacing the constant bonus with our staged length guard improves accuracy to 36.1%, demonstrating the benefit of the staged design. We have revised this point in our paper.

To further investigate the design choices, we additionally evaluate mean KL per token and a constant length bonus. We also report average response lengths to examine the effect of different reward configurations on reasoning efficiency.

Table 2. Ablation of KL aggregation and length regulation on MMSI.

MethodKL AggregationLength BonusMMSI Acc. (%)Avg. Tokens
Qwen3-VL-8B-InstructNoneNone31.1359.0
Vanilla GRPONoneNone32.2307.9
KL onlyToken-sumNone31.613.1
KL + Constant BonusToken-sumConstant33.5163.9
Mean KL onlyToken-meanNone32.6588.5
Mean KL + Staged BonusToken-meanStaged31.8517.7
LenGuard-GPC (Ours)Token-sumStaged36.1261.9

Furthermore, the KL-only variant provides direct evidence of short-response collapse. Without length regulation, its average response length drops dramatically to 13.1 tokens, compared with 307.9 tokens for vanilla GRPO. This behavior is consistent with the length-dependent bias of accumulated KL penalties, which can incentivize the policy to terminate reasoning prematurely.

In contrast, our staged length guard maintains an average response length of 261.9 tokens while improving accuracy to 36.1%, demonstrating that length regulation effectively mitigates short-response collapse while preserving the benefits of guided-prompt consistency.

Regarding guided-prompt construction, we provide the complete generation instructions and guided-prompt template in our response to Q3, clarifying the role of training-set ground truth and external supervision.

We thank the reviewer for these valuable suggestions, which help clarify the motivation and empirical effectiveness of our reward design.

Q3: Further Details on Guided-Prompt Construction and Its Role in Training

R3: We thank the reviewer for raising these important questions. We would like to clarify the role of ground-truth (GT) information in our guided-prompt construction, the distinction between training-time supervision and evaluation-time information leakage, and the reproducibility of our approach.

(1) Clarification on GT usage and token-level supervision.

First, we emphasize that the GT used in our guided-prompt template comes exclusively from the training set, where GT is already available for computing the accuracy reward in standard GRPO. Incorporating GT into the guided prompt provides an additional form of training-time supervision, which should be distinguished from information leakage during evaluation.

Importantly, our method does not generate any GT-conditioned reasoning trajectory. During each RL rollout, the policy samples a single trajectory under the standard prompt, without access to either the hint or GT. The guided prompt, which incorporates task-specific guidance and GT, is used only to evaluate the token-level predictive distributions along this fixed trajectory through teacher forcing. We then compute the KL divergence between the standard and guided distributions at each token position and aggregate these scores into a trajectory-level RL reward.

Thus, GT serves as privileged training-time information for constructing a stronger reference condition, rather than as a target chain of thought for imitation. The policy is optimized through the standard GRPO objective, without generating or directly imitating teacher-produced reasoning trajectories.

Crucially, neither GT, guided prompts, nor the external hint generator is used during inference. The trained policy receives only the standard prompt and independently generates its reasoning and final answer. Furthermore, all evaluation benchmarks are held out from training, and their GT answers are never used for guided-prompt construction. Therefore, our method does not involve evaluation-set or inference-time GT leakage.

(2) Guided-prompt construction and reproducibility.

As described in Section 3.2, we use GPT-5.3-Chat to generate one-sentence, task-specific hints offline for each training instance. To improve reproducibility, we provide the complete prompt templates below.

Hint-generation instructions (GPT-5.3-Chat):

You are given a set of images, a question, and the ground-truth answer. Generate one caption that helps solve the question without revealing the answer.
Rules:
- Do not state, paraphrase, or imply the ground-truth answer.
- Do not copy words from the ground-truth answer.
- Focus on the key cross-image clues, such as matching objects, relative positions, viewpoint changes, or spatial relations.
Question: {question}
Ground Truth: {answer}
Output: One caption only.

Guided-prompt template:

Question: {}

Please reason step-by-step according to the following instructions:
Provide your detailed reasoning for the raw question within the <thinking> </thinking> tags.
Hint: {}

Ground Truth: {}
Use the multiple perspectives to discover objects or details not visible in a single image. Ensure your logic synthesizes the visual evidence into a coherent conclusion.
Final Output: Provide only the single option letter (e.g., A, B, C, D) within the <answer> </answer> tags.

As shown above, the hint-generation instructions explicitly discourage revealing or implying the GT answer. While the guided-prompt template includes GT to provide a stronger reference condition, it is used exclusively for token-level reward evaluation, rather than for generating additional reasoning trajectories. This distinction is central to our reward design.

We thank the reviewer for this constructive feedback. We have clarified the role of GT-conditioned training supervision, provided the complete prompt templates, and clarified the training-only role of guidance to improve the transparency and reproducibility of our framework.

Q4: Robustness Across Random Seeds

R4: We thank the reviewer for the valuable suggestion regarding experimental robustness. We would like to emphasize that the contributions of different reward components and alternative design choices have been extensively examined in Tables 1 and 2, including the effectiveness of guided-prompt consistency, staged length regulation, and alternative KL aggregation and length bonus strategies.

To further assess training stability, we additionally evaluate our method across multiple random seeds using the same Qwen3-VL-8B-Instruct backbone and identical training configurations.

Table 3. Multi-seed evaluation on MMSI (accuracy, %).

MethodSeed 1Seed 2Seed 3
LenGuard-GPC (Ours)36.135.936.4

Together with the comprehensive ablation studies presented in Tables 1 and 2, these multi-seed experiments provide additional evidence regarding the effectiveness and robustness of our framework.

We thank the reviewer for this constructive suggestion, which helps strengthen the empirical evaluation of our method.

Q5: Effectiveness of Task-Specific Guidance

R5: We thank the reviewer for this insightful suggestion regarding whether guided-prompt consistency benefits from meaningful task-specific guidance rather than generic prompt regularization.

First, as demonstrated in Tables 1 and 2, our ablation studies provide evidence for the effectiveness of guided-prompt consistency and its interaction with length regulation. In particular, guided-prompt consistency with a constant length bonus achieves 33.5% accuracy, while the staged length reward without guided-prompt consistency achieves 32.5%. Combining guided-prompt consistency with the staged length guard further improves accuracy to 36.1%, supporting the complementary roles of the two components.

To further investigate whether the reward depends on meaningful guidance, we introduce an additional Shuffled Guidance control. Specifically, we randomly permute the generated hints across different training instances, breaking the correspondence between each hint and its original question. Importantly, the ground-truth answer remains unchanged for each instance, and the guided-prompt template, KL reward, length guard, and all training hyperparameters are kept identical to our full method.

Table 4. Effect of task-specific versus shuffled guidance on MMSI.

MethodGuidance ConfigurationMMSI Acc. (%)
Vanilla GRPONo Guidance32.2
LenGuard-GPCShuffled Hints32.1
LenGuard-GPC (Ours)Original Hints36.1

This controlled comparison isolates the contribution of task-specific hint information while preserving the same GT-conditioned reference setting. Original hints achieve 36.1% accuracy, compared with 32.1% for shuffled hints, supporting the contribution of task-specific guidance beyond an additional prompting condition alone.

We emphasize that guided-prompt KL is not intended as a standalone measure of reasoning correctness. Instead, it serves as an auxiliary reward signal combined with verifiable task rewards. The ablations in Tables 1 and 2, together with the shuffled-guidance control, provide complementary evidence regarding the effectiveness of this reward design.

We thank the reviewer for this valuable suggestion, which helps clarify the role of task-specific guidance in our framework.

Dear Reviewer,

As the discussion period is nearing its end, we would appreciate your feedback on our rebuttal and the additional results in the revised manuscript. Please let us know if any concerns remain so that we can respond promptly.

Best regards,

The Authors

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

It proposes two reward terms for GRPO in multi-view spatial reasoning. The first uses the negative sum of token-level KL divergences between predictions under standard prompts and guided prompts containing hints generated by GPT-5.3-Chat. The second adds a length bonus to discourage overly short responses. The method trains Qwen3-VL-8B on about 13K samples from VST-500K and improves the average score across six benchmarks by 2.0 points over GRPO while generating shorter responses.

Critical strengths
  1. Consistent improvements with shorter responses. Compared with a matched GRPO baseline, the method improves MMSI from 32.2 to 36.1, MindCube from 44.7 to 48.4, and ViewSpatial from 44.4 to 46.9. It also reduces generated tokens by 15% to 38% without adding inference cost.
Critical weaknesses
  1. The core reward mechanism lacks essential implementation details. Neither the guided prompt template nor the GPT-5.3 hint generation instructions are provided. If the hints reveal or strongly suggest the answers, the KL reward may effectively distill privileged information rather than encourage consistency. Generating hints with a closed model also introduces additional supervision that the paper does not acknowledge. Without these details, the main results are difficult to interpret or reproduce.

  2. The experimental evidence is limited, and key design choices remain untested. The experiments use only one backbone and one seed, with a single ablation evaluated on MMSI. Important alternatives are missing, including mean KL per token, KL without a length bonus, and a constant length bonus. The claimed collapse toward short responses is also not demonstrated. Moreover, the base model scores below random chance on MindCube (29.4 vs. 33.0), raising concerns about answer parsing and potentially inflated gains.

Questions for authors

Clarifications That Could Change My Assessment:

  1. Could the authors provide the guided prompt template, hint generation instructions, and example hints? Do these hints reveal or imply the correct answers?

  2. Why does the base Qwen3-VL-8B score below random chance on MindCube? What percentage of responses fail answer extraction?

  3. How does using mean KL per token without the length bonus affect performance?

  4. On ViewSpatial, the average response length is 121.6 tokens, below the bonus threshold of 160. What proportion of responses actually receives the bonus?

Suggestions:

  1. Since the token-level KL values are summed into a single trajectory reward, describing this as "dense process-level feedback" may be misleading. The wording should be clarified.

  2. The paper should discuss related work on learning from hinted prompts (OPSD and PTD-PO) and length control in RL (DAPO, Kimi k1.5, and L1). Key numerical results should also be included in the abstract.

7 comments

Q1: Further Details on Guided-Prompt Construction and Its Role in Training

R1: We thank the reviewer for raising these important questions. We would like to clarify the role of ground-truth (GT) information in our guided-prompt construction, the distinction between training-time supervision and evaluation-time information leakage, and the reproducibility of our approach.

(1) Clarification on GT usage and token-level supervision.

First, we emphasize that the GT used in our guided-prompt template comes exclusively from the training set, where GT is already available for computing the accuracy reward in standard GRPO. Incorporating GT into the guided prompt provides an additional form of training-time supervision, which should be distinguished from information leakage during evaluation.

Importantly, our method does not generate any GT-conditioned reasoning trajectory. During each RL rollout, the policy samples a single trajectory under the standard prompt, without access to either the hint or GT. The guided prompt, which incorporates task-specific guidance and GT, is used only to evaluate the token-level predictive distributions along this fixed trajectory through teacher forcing. We then compute the KL divergence between the standard and guided distributions at each token position and aggregate these scores into a trajectory-level RL reward.

Thus, GT serves as privileged training-time information for constructing a stronger reference condition, rather than as a target chain of thought for imitation. The policy is optimized through the standard GRPO objective, without generating or directly imitating teacher-produced reasoning trajectories.

Crucially, neither GT, guided prompts, nor the external hint generator is used during inference. The trained policy receives only the standard prompt and independently generates its reasoning and final answer. Furthermore, all evaluation benchmarks are held out from training, and their GT answers are never used for guided-prompt construction. Therefore, our method does not involve evaluation-set or inference-time GT leakage.

(2) Guided-prompt construction and reproducibility.

As described in Section 3.2, we use GPT-5.3-Chat to generate one-sentence, task-specific hints offline for each training instance. To improve reproducibility, we provide the complete prompt templates below.

Hint-generation instructions (GPT-5.3-Chat):

You are given a set of images, a question, and the ground-truth answer. Generate one caption that helps solve the question without revealing the answer.
Rules:
- Do not state, paraphrase, or imply the ground-truth answer.
- Do not copy words from the ground-truth answer.
- Focus on the key cross-image clues, such as matching objects, relative positions, viewpoint changes, or spatial relations.
Question: {question}
Ground Truth: {answer}
Output: One caption only.

Guided-prompt template:

Question: {}

Please reason step-by-step according to the following instructions:
Provide your detailed reasoning for the raw question within the <thinking> </thinking> tags.
Hint: {}

Ground Truth: {}
Use the multiple perspectives to discover objects or details not visible in a single image. Ensure your logic synthesizes the visual evidence into a coherent conclusion.
Final Output: Provide only the single option letter (e.g., A, B, C, D) within the <answer> </answer> tags.

As shown above, the hint-generation instructions explicitly discourage revealing or implying the GT answer. While the guided-prompt template includes GT to provide a stronger reference condition, it is used exclusively for token-level reward evaluation, rather than for generating additional reasoning trajectories. This distinction is central to our reward design.

(3) Role of external supervision and alternative guidance generators.

GPT-generated hints and GT-conditioned guided prompts provide training-time supervision beyond the standard outcome reward. Our contribution is to transform guidance-conditioned token-level distributional differences into a GRPO-compatible reward, while keeping trajectory generation entirely on-policy under the standard prompt and without requiring teacher-produced reasoning traces.

Furthermore, the proposed mechanism does not inherently depend on GPT-5.3-Chat. To validate this, we conduct an additional experiment using Qwen3-VL-8B-Instruct as an open-source hint generator, while keeping the policy initialization, training data, reward design, and other training configurations unchanged.

Table 1. Comparison of guidance generators.

RL SettingGuidance GeneratorMMSIViewSpatialMindCube
Vanilla GRPONone32.244.444.7
Our methodGPT-5.3-Chat36.146.948.4
Our methodQwen3-VL-8B-Instruct35.346.648.6

As shown above, replacing GPT-5.3-Chat with Qwen3-VL-8B-Instruct yields comparable performance, with only minor differences across the three benchmarks. Notably, Qwen3-VL-8B-Instruct guidance achieves 35.3% on MMSI, 46.6% on ViewSpatial, and 48.6% on MindCube, consistently outperforming vanilla GRPO. These results demonstrate that the effectiveness of guided-prompt consistency is not tied to a particular proprietary guidance generator and can be retained with a smaller open-source model.

We thank the reviewer for this constructive feedback. We have clarified the role of GT-conditioned training supervision, provided the prompt templates, and added experiments with an alternative guidance generator to improve the transparency, reproducibility, and generality of our framework.

Q2: Further Experimental Validation of Reward Design and Evaluation Reliability

R2: We thank the reviewer for the constructive suggestions regarding reward design, reasoning-length behavior, experimental robustness, and benchmark reliability. We would like to clarify the existing evidence and provide additional analyses addressing these concerns.

(1) Ablation of KL aggregation and length regulation.

We clarify that the variant labeled “w/o staged length reward” in Section 4.3 of the original submission retains a constant length bonus while removing the staged bonus design. It therefore corresponds to “KL + Constant Bonus” in Table 2, achieving 33.5% accuracy on MMSI with an average response length of 163.9 tokens. Replacing the constant bonus with our staged length guard improves accuracy to 36.1%, demonstrating the benefit of the staged design.

To examine the design choices more comprehensively, we additionally consider mean KL per token and a constant length bonus. We also report the average generated tokens to directly analyze the effect of length regulation.

Table 2. Ablation of KL aggregation and length regulation on MMSI.

MethodKL AggregationLength BonusMMSI Acc. (%)Avg. Tokens
Qwen3-VL-8B-InstructNoneNone31.1359.0
Vanilla GRPONoneNone32.2307.9
KL onlyToken-sumNone31.613.1
KL + Constant BonusToken-sumConstant33.5163.9
Mean KL onlyToken-meanNone32.6588.5
Mean KL + Staged BonusToken-meanStaged31.8517.7
LenGuard-GPC (Ours)Token-sumStaged36.1261.9

(2) Evidence of short-response collapse.

We appreciate the reviewer's suggestion to provide direct evidence of short-response collapse. As shown in Table 2, the KL-only variant produces extremely short responses, with an average length of approximately 13.1 tokens, compared with 307.9 tokens for vanilla GRPO. This behavior is consistent with the length bias introduced by token-sum KL: since the cumulative KL penalty increases with the number of generated tokens, the policy can reduce its penalty by prematurely terminating its reasoning.

Our staged length guard increases the average response length to 261.9 tokens while further improving accuracy to 36.1%. This demonstrates that the staged length reward mitigates short-response collapse while preserving the benefits of guided-prompt consistency. Notably, the resulting trajectories remain shorter than those produced by vanilla GRPO, indicating a better balance between reasoning sufficiency and compactness.

(3) Robustness across random seeds.

We agree that evaluating multiple random seeds would strengthen the evidence regarding training stability. We therefore include additional runs using the same Qwen3-VL-8B-Instruct backbone and consistent training configurations.

Table 3. Multi-seed evaluation on MMSI (accuracy, %).

MethodSeed 1Seed 2Seed 3
LenGuard-GPC (Ours)36.135.936.4

Q3: Clarification of the Qwen3-VL-8B Baseline on MindCube

R3: We thank the reviewer for inviting further clarification of the baseline results on MindCube. We would first like to emphasize that the reported 29.4% accuracy is consistent with previously published results. Specifically, [1][2][3] have reported the same 29.4% MindCube accuracy for Qwen3-VL-8B, consistent with our results. Therefore, this relatively low baseline is not unique to our evaluation. Some of the models featured in the MindCube report [4], such as Qwen2.5-VL-7B-Instruct, SpaceMantis[5], and Spatial-MLLM[6] exhibit performance levels that fall below those of random selection.

We additionally examined the model's answer parsing behavior and found that 5.1% of the base Qwen3-VL-8B-Instruct responses could not be successfully parsed into valid answer options. These responses are counted as incorrect under our evaluation protocol, which partially explains the relatively low baseline accuracy. Our evaluation consistently applies the same answer-parsing criteria to all methods. Additionally, the below-chance performance may partly reflect MindCube's structured data distribution, as its construction relies on canonical viewpoints (e.g., front, left, right, and back) and predefined spatial configurations, which may induce systematic prediction biases in models that struggle with cross-view spatial reasoning.

Furthermore, our fine-grained analysis reveals two particularly weak MindCube categories: “behind the observer” (3/77, 3.90%) and “left/right of an object from a specified viewpoint” (31/175, 17.71%), both substantially below the 25% random-choice baseline for four-option questions. Metadata-based analysis, supported by image spot checks, reveals systematic spatial errors: in 39/77 (50.65%) “behind” questions, the model selected an object in front of the camera, while in 113/175 (64.57%) left/right questions, it selected the front-facing scene background rather than the requested side. These patterns suggest that the model may confuse visible backgrounds with observer-relative spatial relations and fail to integrate evidence across views, helping explain its performance in these challenging categories.

Importantly, our approach is evaluated against vanilla GRPO under the same evaluation protocol, with MindCube accuracy improving from 44.7% to 48.4%. Moreover, our method also demonstrates consistent gains on MMSI (32.2% to 36.1%) and ViewSpatial (44.4% to 46.9%), indicating that the observed improvements are not limited to MindCube.

We thank the reviewer for these valuable suggestions. The additional ablations, reasoning-length analysis, and multi-seed evaluation help clarify our design choices and strengthen the empirical evaluation of LenGuard-GPC.

[1] Scaling Spatial Intelligence with Multimodal Foundation Models (CVPR 2026)

[2] Holi-Spatial: Evolving Video Streams into Holistic 3D Spatial Intelligence (ICML 2026)

[3] SpatialOPSD: Self-Distilling Spatial Intelligence from Verified Coding Agent Traces (arxiv 2610.11366)

[4] Spatial Mental Modeling from Limited Views(ICLR 2026)

[5] SpatialVLM: Endowing Vision-Language Models with Spatial Reasoning Capabilities (CVPR 2024)

[6] Spatial-mllm: Boosting mllm capabilities in visual-based spatial intelligence.

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

This paper proposes a dense reward RL framework named LenGuard-GPC. By calculating the token-wise KL divergence between standard and guided prompts as a dense reward signal and introducing an innovative staged length bonus, the method effectively prevents short-answer collapse while maintaining reasoning compactness. Experiments show that this approach significantly outperforms baseline models across multiple in-domain and out-of-domain multi-view spatial reasoning benchmarks.

Critical strengths

Dense Process-Level Reward Guidance

By computing KL divergence under standard and guided prompts, it provides fine-grained dense feedback to precisely pinpoint errors in cross-view object correspondence and spatial relation inference.

Ingenious Staged Length Balancing Mechanism

Introducing staged length boundaries leverages KL penalties to encourage conciseness while preventing the model from collapsing into overly short, content-poor reasoning chains, achieving a dynamic balance between compactness and reasoning sufficiency.

Critical weaknesses

Lack of Training/Inference Platform Details, Compute-Benefit Trade-offs, and Insufficient Detail in Token Efficiency Analysis

The paper fails to specify the underlying hardware platforms for training and inference (such as GPU types or cluster scales), making the claimed training overhead difficult to reproduce or benchmark across different environments. Furthermore, while Section 4.4 touches upon token efficiency, the analysis lacks sufficient depth: it fails to precisely quantify the Pareto frontier between the extra token consumption introduced by guided prompts and the performance gains/savings from shortened reasoning trajectories. Consequently, readers cannot thoroughly evaluate the overall cost-benefit ratio of this dense reward design for practical deployment.

Over-reliance on GPT-5.3-Chat for Guided Prompts and Lack of Validation on Other LLMs

The innovation of this approach—the guided prompts required for dense reward computation—is entirely generated offline by GPT-5.3-Chat. This raises critical methodological concerns regarding its persuasiveness: does this heavy reliance on a state-of-the-art proprietary model undermine the framework's self-contained nature? If weaker open-source models (e.g., Qwen or Llama series) or alternative commercial LLMs were used to generate the guidance, how would the effectiveness of the supervision signal and final RL performance degrade?

Questions for authors

None

4 comments

Q1: Training/Inference Platforms and Compute-Efficiency Trade-offs

R1: We thank the reviewer for this helpful comment. We agree that hardware specifications and a more comprehensive cost-benefit analysis are important for reproducibility and practical evaluation. We would like to clarify the computational overhead of our method and provide additional quantitative comparisons.

(1) Hardware and training overhead.

Our experiments are conducted using 8 NVIDIA H100 GPUs. The training implementation employs full-shard FSDP and gradient checkpointing, with the detailed configurations provided in Appendix B.1.

As already described in Section 4.2, although each trajectory is evaluated under both standard and guided prompts, our method does not require two independent autoregressive generations. Specifically, each trajectory is sampled only once from the standard prompt, while the guided-prompt distribution is evaluated along the fixed trajectory using teacher forcing. This allows token-level predictions to be computed through parallel forward evaluation rather than additional autoregressive decoding. Consequently, the additional evaluation increases the overall training time by approximately 1.6×, rather than doubling it. The shorter trajectories induced by our method further offset part of the inference computational cost.

(2) Accuracy–token efficiency trade-off.

We have further clarified the relationship between accuracy improvements and token savings by reporting both metrics together:

Table 1. Accuracy and generation efficiency.

BenchmarkVanilla GRPO AccuracyOurs AccuracyAccuracy GainVanilla TokensOurs TokensToken Reduction
MMSI32.2%36.1%+3.9307.9261.914.9%
ViewSpatial44.4%46.9%+2.5196.5121.638.1%

These results demonstrate that our method achieves simultaneous improvements in accuracy and generation efficiency. In particular, it reduces the average generated tokens by 46.0 on MMSI and 74.9 on ViewSpatial, while improving accuracy by 3.9 and 2.5 percentage points, respectively. Thus, our method achieves a better accuracy–token-cost trade-off than vanilla GRPO on both benchmarks.

(3) Additional prompt cost and inference efficiency.

We also clarify that the guided prompt consists of a single sentence generated by GPT-5.3-Chat for each training instance. Its average length is 35.9 tokens. This additional guidance is used exclusively for training-time reward computation and is not required during inference. Therefore, it introduces no additional prompt tokens or forward passes at inference time. At deployment, the model only receives the standard prompt, and the reduced output length directly translates into lower autoregressive generation costs.

We have added the hardware specifications and clarified the training-time computational overhead and accuracy–token efficiency comparison in the revised manuscript. We thank the reviewer for this valuable suggestion, which helps us present a more complete picture of the practical efficiency of our method.

Q2: Generality of the Guided-Prompt Generation Model

R2: We thank the reviewer for raising this important question regarding the dependence of our method on GPT-5.3-Chat. We would like to clarify that GPT-5.3-Chat is used solely to generate one-sentence guidance for each training instance offline. It is not involved in policy optimization or inference. The core mechanism of LenGuard-GPC is the token-level distributional comparison between standard and guided prompts, which is independent of the particular model used to generate the guidance. Therefore, GPT-5.3-Chat serves as one possible guidance generator rather than an essential component of our framework.

To further validate this, we conduct an additional experiment by replacing GPT-5.3-Chat with Qwen3-VL-8B-Instruct, an open-source model, to generate the guided prompts. We keep the training data, guidance-generation instructions, policy initialization, GRPO hyperparameters, reward coefficients, and staged length guard unchanged, isolating the effect of the guidance generator.

The results are reported below:

Table 2. Comparison of guidance generators.

RL SettingGuidance GeneratorMMSIViewSpatialMindCube
Vanilla GRPONone32.244.444.7
Our methodGPT-5.3-Chat36.146.948.4
Our methodQwen3-VL-8B-Instruct35.346.648.6

As shown in Table 2, Qwen3-VL-8B-Instruct guidance achieves comparable performance across the three benchmarks and consistently outperforms vanilla GRPO. These results support the effectiveness of our reward framework with a smaller open-source guidance generator.

Importantly, since guidance generation is performed offline and only used during training, replacing the guidance generator does not introduce any additional dependency at inference time. The trained policy remains fully independent of the guidance model.

We have included this additional comparison and clarified the role of the guidance generator in the revised manuscript. We thank the reviewer for this valuable suggestion, which helps establish the generality and practical applicability of our framework.

Dear Reviewer,

As the discussion period is nearing its end, we would appreciate your feedback on our rebuttal and the additional results in the revised manuscript. Please let us know if any concerns remain so that we can respond promptly.

Best regards,

The Authors

Dear reviewers,

Thank you again for raising these concerns. In the revised manuscript, we report the hardware setup and clarify the trade-off: training takes approximately 1.6× as long as vanilla GRPO, while inference uses only the standard prompt. On MMSI and ViewSpatial, LenGuard-GPC improves accuracy by 3.9 and 2.5 percentage points while reducing average output length by 14.9% and 38.1%, respectively.

We also repeated training with hints generated by Qwen3-VL-8B-Instruct. This open-source alternative remained within 0.8 percentage points of the GPT-5.3-Chat results on all three benchmarks and consistently outperformed vanilla GRPO. We appreciate your suggestions and hope these additions address your concerns.

Best regards,

The Authors

Comments on the paper

1 comment
Area_Chair_mcdVOpenReview ↗

rebuttal ddl

Dear reviewers, the ICLR deadline is only 4 days away. The authors have already rebuttaled the issues raised. If you still have questions, please participate in the rebuttal.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.