acceptodds
Under review as a conference paper at ICLR 2027

Soft Spatial Reasoning

Abstract

Large Vision-Language Models (LVLMs) commonly perform spatial reasoning through chain-of-thought (CoT), encoding intermediate reasoning as autoregressive sequences of discrete language tokens. Such *hard thinking* requires committing to a single token at each step, even when the correct spatial interpretation remains uncertain. This early commitment constitutes *premature discretization*: an incorrect token selection can propagate errors through subsequent reasoning. We propose **Soft Spatial Reasoning**, a post-training framework that introduces *soft thinking* for spatial tasks in LVLMs. At each intermediate reasoning step, the LVLM forms a continuous soft state by mixing token embeddings rather than selecting a single token, allowing multiple candidate continuations to influence the next step. The appropriate degree of softness, however, can vary across reasoning steps: retaining multiple candidates may preserve a useful spatial interpretation, but if those candidates imply conflicting spatial relations, mixing them may interfere with subsequent reasoning. At the core of Soft Spatial Reasoning is **AdaptSoft**, a controller that uses the current hidden state and predictive uncertainty to adapt the degree of softness at each reasoning step. To train AdaptSoft, we introduce a gradient-alignment learning objective that provides a step-specific learning signal for softness control without intermediate reasoning supervision. Across diverse spatial benchmarks, Soft Spatial Reasoning outperforms hard and fixed-soft CoT baselines using the same backbone, as well as a range of existing LVLMs. The source code is available in the supplementary material.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large vision-language models, spatial reasoning, soft thinking, chain-of-thought reasoning, reinforcement learning, adaptive temperature control

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_Ud3HOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Soft Spatial Reasoning, a post-training framework for adaptive soft reasoning in LVLMs. Intermediate states mix candidate token embeddings, while AdaptSoft predicts a step-specific mixture temperature from hidden-state features and predictive entropy. The LVLM is trained with GRPO, and the controller receives an auxiliary gradient-alignment signal without intermediate reasoning annotations. Using Qwen3-VL-8B-Thinking, the method improves over matched hard- and fixed-soft baselines on OmniSpatial and SpatiaLab, with additional zero-shot results on MindCube supporting transfer to spatial mental-modeling tasks.

Critical strengths
  1. AdaptSoft provides a concrete mechanism for varying token-mixture concentration across reasoning steps. Conditioning on both hidden states and predictive entropy allows control to depend on reasoning context rather than uncertainty alone.

  2. Gradient alignment provides a differentiable training signal for the softness controller without requiring annotated reasoning steps or an external process evaluator. The ablation against rollout-level task-advantage training supports its usefulness in the evaluated perspective-taking tasks.

  3. Matched-backbone comparisons distinguish the gains from ordinary GRPO, fixed soft thinking, and adaptive softness. Evaluation on unseen SpatiaLab and MindCube further supports transfer beyond the post-training benchmark.

Critical weaknesses
  1. Results are reported without variation across independent training runs or confidence intervals. Since both GRPO training and soft rollouts are stochastic, the reproducibility of the gains over fixed soft thinking remains uncertain.

  2. The additional training cost of reconstructing soft states and computing gradient alignment is not quantified. Comparisons of wall-clock time, peak memory, and inference latency would clarify the performance–cost trade-off relative to hard and fixed-soft GRPO.

  3. The experiments do not establish how much improvement requires the proposed controller and alignment objective rather than simpler temperature control. Comparisons with a tuned fixed temperature or an entropy-based temperature schedule would provide stronger evidence for the necessity of the learned mechanism.

Questions for authors
  1. How does AdaptSoft compare with the best fixed temperature selected on validation data, rather than only the default temperature? Would its advantage remain against a simple entropy-based temperature schedule?

  2. What are the training-time, peak-memory, and inference-latency overheads relative to fixed-soft GRPO? Do the accuracy gains persist when the methods are compared under matched computational budgets?

3 comments

W1 Evaluation Variability and Confidence Intervals

Results are reported without variation across independent training runs or confidence intervals. Since both GRPO training and soft rollouts are stochastic, the reproducibility of the gains over fixed soft thinking remains uncertain.

A: We have added repeated stochastic evaluations and a paired confidence interval to assess the stability of the reported gain. Three independent mean@8 evaluations of the same AdaptSoft checkpoint yield 49.65 0.10% overall accuracy (mean standard deviation), showing consistent performance across evaluation runs.

The observed improvement over fixed-soft GRPO is 2.73 percentage points (49.65% versus 46.92%). We quantify uncertainty in this difference using a paired bootstrap over test questions with 10,000 resamples, scoring each question by its mean accuracy over eight generations. The resulting 95% confidence interval is [2.44, 3.25] percentage points.

The repeated evaluations show stable performance, and the paired confidence interval supports a positive gain over fixed soft thinking. These analyses use the evaluated checkpoints, as additional independent training runs were infeasible within the rebuttal window.

Q1/ W3 Fixed-Temperature Tuning and Entropy-Based Control

How does AdaptSoft compare with the best fixed temperature selected on validation data, rather than only the default temperature? Would its advantage remain against a simple entropy-based temperature schedule?

A: "Soft Thinking + GRPO" already uses a fixed softness temperature selected on validation data. In preliminary experiments, we performed a grid search over training softness temperatures {0.1,0.3,0.5,1.0} and inference softness temperatures {0.1,0.5}. The best validation setting uses =0.1 during training and =0.5 at inference, compared with the original recipe's default of =0.1 for both (SofT-GRPO, Zheng et al.). Soft Spatial Reasoning exceeds this baseline by 2.75 percentage points on OmniSpatial, 1.85 points on SpatiaLab, and 4.11 points on MindCube. We have added the temperature grid search and selected settings to Appendix B.2 of the revised manuscript.

For entropy-based temperature control, the submitted paper includes an ablation without the hidden-state input. In this ablation, AdaptSoft adjusts softness using only predictive uncertainty (entropy), learning how to do so during post-training. With both variants trained on Perspective Taking tasks, adding the hidden-state input increases accuracy from 39.22% to 42.75%, a gain of 3.53 percentage points. This result supports the value of reasoning context even when the uncertainty-to-temperature mapping is learned.

Q2/ W2 Training, Memory, and Inference Costs

What are the training-time, peak-memory, and inference-latency overheads relative to fixed-soft GRPO?

A: Compared with "Soft Thinking + GRPO," Soft Spatial Reasoning increases training time per step by 32% and peak allocated GPU memory by 3.2 GB. Full-test mean@8 inference takes 142 minutes versus 62 minutes.

MetricSoft Thinking + GRPOSoft Spatial Reasoning
Training time per step319 s420 s
Peak allocated GPU memory37.8 GB41.0 GB
Full-test inference time (mean@8)62 min142 min
Generated tokens per response at inference (mean)249382
Decoding throughput (tokens/s, 4H100)821550

Training response lengths are approximately equal (288 versus 292 tokens). Inference timings cover the complete test set with eight generations per question for mean@8 evaluation. Soft Spatial Reasoning generates longer responses on average and has lower measured decoding throughput, both of which contribute to the end-to-end inference time.

Do the accuracy gains persist when the methods are compared under matched computational budgets?

A: Yes. Soft Spatial Reasoning retains a 4.7-percentage-point accuracy advantage at comparable training budgets of approximately 18 hours on the same GPU configuration. We evaluate both methods on OmniSpatial using an image pixel budget of 128282 during both training and evaluation, preserving aspect ratio and flooring image dimensions to multiples of 28. Both methods use the same evaluation procedure.

Training hoursMethodTraining stepsAccuracy (%)
18 hSoft Thinking + GRPO20044.2
18 hSoft Spatial Reasoning14048.9

Matching training time accounts for the additional computational cost of adaptive softness: "Soft Thinking + GRPO" completes 200 steps while Soft Spatial Reasoning completes 140. We have added this analysis to the Results section of the revised manuscript.

Thanks to the authors for their detailed response, I appreciate the authors' efforts, my problem is basically solved, and I will keep my score

Thank you for your time and constructive feedback, which has helped us strengthen the paper.

Reviewer_W8z1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LVLMs commonly perform spatial reasoning through CoT, encoding intermediate reasoning as autoregressive sequences of discrete language tokens, where this premature discretization leads propagatation of errors.

This paper proposes the Soft Spatial Reasoning, a post-traininig framework with soft thinking for spatial tasks in LVLMs in solving this problem. The soft thinking is achieved through the mixing token embeddings.

It proposes that if those candidates imply conflicting spatial relations, mixing them may interfere with subsequent reasoning. Thus, this paper proposes and trains a AdaptSoft, which is a controller which uses the current hidden state and predictive uncertainty to adapt the degree of softness at each reasoning step.

Critical strengths
  1. This paper introduces the Soft Spatial Reasoning, to our knowledge the first framework dedicated to soft thinking for spatial reasoning in LVLMs and thie paper introduces AdaptSoft to control softness at each reasoning step using the current reasoning state and predictive uncertaint.

  2. The Overview of Soft Spatial Reasoning is clear and easy to follow, the architecture design is neat and easy to follow. Extensive experiments covering 19 spatial reasoning categories demonstrate overall improvements over strong spatial reasoning baselines.

Critical weaknesses
  1. The ablation study is not suifficient where there is a lack of baselines in the MindCubebenchmarks.

  2. The soft spatial reasoning is implemented and achieved with the mixed embeddings which indicates that it is possible to introduce a traininig-free/zero-shot adaptation of pre-trained LVLM with fixed soft inference. However, there is no such baselines evaluated. Neither do the baseline experiments conducted by pre-traiened backbones with entropy heuristic.

Questions for authors

See above

4 comments
AuthorsOpenReview ↗

W1: Additional Baselines on MindCube

The ablation study is not suifficient where there is a lack of baselines in the MindCubebenchmarks.

Zero-shot MindCube accuracy (%). Overall is sample-weighted.

MethodOverallRotationAmongAround
Qwen3-VL-8B-Thinking (Base)33.6245.1433.2630.44
Hard Thinking + GRPO36.3246.5336.1432.37
Soft Thinking + GRPO34.0245.8833.6530.81
Soft Spatial Reasoning (Ours)38.1347.3638.1832.37

A: We have added the same Hard Thinking + GRPO and Soft Thinking + GRPO baselines reported in Tables 1–2 to the MindCube evaluation. This extends the controlled comparison to zero-shot spatial mental modeling, where models must combine limited views to infer unseen spatial relations and reason about hypothetical movements. Soft Spatial Reasoning achieves 38.13% overall accuracy, exceeding Hard Thinking + GRPO by 1.81 percentage points and Soft Thinking + GRPO by 4.11 points. Fixed soft thinking falls below hard thinking here, while adaptive softness improves over both. This contrast suggests that spatial mental modeling benefits from adjusting softness as evidence from different views is combined, beyond simply preserving alternative continuations. We have added these baselines to Table 3 of the revised manuscript.

AuthorsOpenReview ↗

W2: Training-Free Soft Inference and Entropy Baselines

The soft spatial reasoning is implemented and achieved with the mixed embeddings which indicates that it is possible to introduce a traininig-free/zero-shot adaptation of pre-trained LVLM with fixed soft inference. However, there is no such baselines evaluated. Neither do the baseline experiments conducted by pre-traiened backbones with entropy heuristic.

OmniSpatial accuracy (%). Average is sample-weighted. † Training-free, applied at inference to the base model.

MethodAverageDynamic Reasoning: ManipulationDynamic Reasoning: Motion Anal.Spatial Interaction: Traffic Anal.Spatial Interaction: LocalizationSpatial Interaction: Geospa. StrategyComplex Logic: Pattern Rec.Complex Logic: Geometric ReasoningPerspective Taking: EgocentricPerspective Taking: AllocentricPerspective Taking: Hypothetical
Qwen3-VL-8B-Thinking (Base)43.9057.1454.2836.0454.9548.6724.7723.1271.5631.1138.82
Soft Thinking†43.9759.4655.7841.1858.1047.2725.7725.1672.5531.1240.96
Soft Thinking + Entropy†43.7158.1153.1844.7156.1950.0023.7124.5272.5532.9838.55
Hard Thinking + GRPO45.9261.3254.3055.8858.3346.1825.9025.0072.9236.5742.53
Soft Thinking + GRPO46.9361.8255.8552.2162.6249.7725.2625.0076.3536.2345.88
Soft Spatial Reasoning (Ours)49.6863.5162.1457.6562.8650.0029.9034.8475.4934.9346.34

Zero-shot SpatiaLab accuracy (%). Average is sample-weighted. † Training-free, applied at inference to the base model.

MethodAverage3D Geom.Dep. & Occu.OrientationRelat. Posit.Size & ScaleSpati. Navig.
Qwen3-VL-8B-Thinking (Base)44.7842.0147.1047.0251.8840.4741.35
Soft Thinking†44.5042.8645.5648.0253.3038.4940.51
Soft Thinking + Entropy†45.0042.4448.2645.5453.7739.2941.77
Hard Thinking + GRPO46.2143.7047.8847.5254.2541.2743.88
Soft Thinking + GRPO46.8644.1248.2648.0255.6641.6744.73
Soft Spatial Reasoning (Ours)48.7146.2249.4248.5158.9642.4648.10

Zero-shot MindCube accuracy (%). Overall is sample-weighted. † Training-free, applied at inference to the base model.

MethodOverallRotationAmongAround
Qwen3-VL-8B-Thinking (Base)33.6245.1433.2630.44
Soft Thinking†32.9746.9932.5329.11
Soft Thinking + Entropy†32.2144.4031.5831.35
Hard Thinking + GRPO36.3246.5336.1432.37
Soft Thinking + GRPO34.0245.8833.6530.81
Soft Spatial Reasoning (Ours)38.1347.3638.1832.37

A: We have added both requested training-free baselines: Soft Thinking†, which applies fixed soft inference with mixed embeddings to Qwen3-VL-8B-Thinking (Base), and Soft Thinking + Entropy†, which additionally uses an entropy heuristic. Both operate directly at inference, without post-training, on all three benchmarks.

Both training-free baselines remain close to Base across all three benchmarks. Fixed soft inference produces little overall change, and adding the entropy heuristic does not yield a consistent improvement across benchmarks. Soft Spatial Reasoning, which incorporates soft states into spatial post-training and learns to adjust their softness during reasoning, improves over Base by 5.78, 3.93, and 4.51 percentage points on OmniSpatial, SpatiaLab, and MindCube, respectively. We have added both training-free baselines to Tables 1–3 of the revised manuscript.

The author provides a comprehensive answer to my questions and I will keep my score.

Thank you for your time and thoughtful consideration of our responses.

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

The paper studies adaptive soft reasoning for visual spatial tasks. Building on probability-weighted mixtures of token embeddings, it introduces AdaptSoft to predict a step-specific mixture temperature from the hidden state and predictive entropy. The backbone is post-trained with GRPO, while the controller is trained using gradient alignment. Experiments with Qwen3-VL-8B-Thinking show higher average accuracy than hard and fixed-soft post-training baselines on OmniSpatial and SpatiaLab, with additional transfer results on MindCube.

Critical strengths
  1. Learning when to concentrate or retain candidate information is a useful extension of soft thinking. Soft reasoning need not benefit equally from the same mixture temperature at every step. AdaptSoft makes this choice dependent on the current state and uncertainty, and the gradient-alignment objective provides a concrete training mechanism. Although embedding mixtures themselves are established prior work, learning their step-specific concentration is a meaningful direction to investigate.

  2. The matched-backbone results provide positive evidence for the method. The authors state that the hard and fixed-soft controls use the same backbone and post-training setup. Relative to fixed soft thinking, average accuracy increases from 46.93% to 49.68% on OmniSpatial and from 46.86% to 48.71% on SpatiaLab (Tables 1–2). The latter result shows transfer beyond the training benchmark, while Table 4 provides additional controller ablations. I regard these gains as substantive evidence rather than discounting them simply because the method adds computation.

Critical weaknesses
  1. The added value over established soft-thinking methods needs a stronger comparison. Soft Thinking (Zhang et al., arXiv:2505.15778) already uses probability-weighted embedding mixtures and entropy-based Cold Stop; the submission cites this work and also builds on SofT-GRPO. The central increment is therefore learned softness control and its training, rather than soft reasoning itself. It is not sufficiently clear how the fixed-soft baseline relates to the prior recipe, including temperature tuning and stopping rules. The stated matching of post-training settings is useful, but matched steps do not necessarily match computation. Additional computation is justified if it obtains gains that a reasonably tuned prior method cannot obtain at a comparable budget; the current comparisons do not fully establish that distinction. This limits the evidence for the incremental contribution, not the validity of the reported gains.

  2. The benefits and limitations across tasks are not sufficiently characterized. The average improvements coexist with lower point estimates than fixed soft thinking in OmniSpatial Egocentric (76.35% to 75.49%) and Allocentric (36.23% to 34.93%) categories, while geometry and motion improve substantially (Table 1). Without uncertainty estimates, these differences do not establish a systematic trade-off, and all six SpatiaLab categories improve. Nevertheless, understanding where adaptive mixing helps or interferes is directly relevant to the motivation for step-specific softness. All three evaluated benchmarks concern spatial reasoning, so the results also leave non-spatial capability retention after full fine-tuning unresolved. I would like a clearer account of the method’s useful operating conditions, rather than treating higher aggregate accuracy as a complete characterization.

Questions for authors
  1. [Clarification] How is “Soft Thinking + GRPO” constructed relative to Zhang et al. (arXiv:2505.15778)? Please specify the fixed-temperature tuning and stopping rules, including whether Cold Stop is retained. A common stopping rule is reasonable for isolating temperature control, but should be distinguished from a comparison with the complete prior recipe.

  2. [Clarification and decision-relevant evidence] Please confirm whether hard, fixed-soft, and adaptive-soft runs all use the reported 200 training steps, prompt batches, rollouts, and backbone optimizer updates. What evidence is available that the adaptive gain persists when the prior baselines receive a comparable computational budget, rather than only the same number of steps? I do not regard additional computation as a drawback by itself; the issue is whether it yields an advantage over spending that budget on the baselines.

  3. [Decision-relevant evidence] Are the Egocentric/Allocentric decreases relative to fixed soft thinking reproducible, and what explains the contrast with the geometry/motion gains? If capability-retention evaluations are available, do they distinguish effects of spatial post-training shared by all variants from effects specific to adaptive softness? These results would clarify the scope and trade-offs of the claimed benefit.

3 comments
AuthorsOpenReview ↗

Q1/ W1 [Clarification]

Q1/ W1. Baseline construction, temperature tuning, and stopping rules

How is “Soft Thinking + GRPO” constructed relative to Zhang et al. (arXiv:2505.15778)?

A: "Soft Thinking + GRPO" is a controlled LVLM baseline for comparing fixed and adaptive softness in spatial reasoning. Zhang et al. introduce soft thinking at inference for large language models, while Zheng et al. develop GRPO post-training for soft thinking in large language models. We adopt Zheng et al.'s SofT-GRPO formulation to construct this baseline for Qwen3-VL-8B-Thinking on OmniSpatial, where spatial reasoning requires interpreting images alongside language.

Our LVLM adaptation adds image information to the policy update, which is absent from the original text-only formulation. During rollout generation, Qwen3-VL receives the image features while generating its soft states. During the policy update, we reconstruct the same soft states and restore the image features, ensuring that the update uses the visual information available during generation. This enables GRPO post-training of soft thinking on spatial questions requiring both images and language.

This baseline and Soft Spatial Reasoning share the same backbone, spatial training data, and GRPO setup, including these adaptations for images. The baseline keeps softness fixed across reasoning steps, while Soft Spatial Reasoning learns to adjust it at each step through AdaptSoft.

Please specify the fixed-temperature tuning

A: For "Soft Thinking + GRPO," we use a softness temperature =0.1 during training and =0.5 at inference. This setting achieved the highest validation accuracy in a preliminary grid search over training softness temperatures {0.1,0.3,0.5,1.0} and inference softness temperatures {0.1,0.5}. We have added the baseline definition, evaluated softness temperatures, and selected setting to Appendix B.2 of the revised manuscript.

stopping rules, including whether Cold Stop is retained. A common stopping rule is reasonable for isolating temperature control, but should be distinguished from a comparison with the complete prior recipe.

A: We do not use Cold Stop; both "Soft Thinking + GRPO" and Soft Spatial Reasoning use the </think>-based stopping rule described in Appendix B.2 of the submitted manuscript. At each reasoning step, the framework checks the highest-weight token in the current soft state. When this token is </think>, it stops soft reasoning and generates the final answer.

To assess whether responses reliably follow the required reasoning and final-answer format without Cold Stop, we evaluated a preliminary Soft Spatial Reasoning run before and during GRPO post-training. The format reward encourages this structure during training. All evaluations use soft reasoning, the fixed system prompt in Figure 9 of the revised manuscript, and the stopping rule described above. Correctly formatted responses increase from 88.7% for Qwen3-VL-8B-Thinking (Base) to 99.5% for Soft Spatial Reasoning after three epochs of GRPO post-training.

Model and training stageCorrect format (%)
Qwen3-VL-8B-Thinking (Base)88.7
Soft Spatial Reasoning after one epoch91.3
Soft Spatial Reasoning after three epochs99.5

These preliminary results show that the model reliably maintains the required reasoning and final-answer format without Cold Stop. Appendix B.3 of the revised manuscript describes the stopping rule and reports the preliminary format-reliability results.

AuthorsOpenReview ↗

Q2/W1 [Clarification and decision-relevant evidence]

Q2/W1. Matched training settings and comparable computational budgets

Please confirm whether hard, fixed-soft, and adaptive-soft runs all use the reported 200 training steps, prompt batches, rollouts, and backbone optimizer updates.

A: Yes. All three methods use the same 200 training steps, prompt batches, rollouts, and number of backbone optimizer updates in the reported experiments. We have clarified these shared training settings in the Implementation Details section of the revised manuscript.

What evidence is available that the adaptive gain persists when the prior baselines receive a comparable computational budget, rather than only the same number of steps?

A: At comparable training budgets of approximately 18 hours, Soft Spatial Reasoning achieves 48.9% accuracy, compared with 44.2% for "Soft Thinking + GRPO." We evaluate both methods on OmniSpatial using an image pixel budget of 128282 during both training and evaluation, preserving aspect ratio and flooring image dimensions to multiples of 28. Both methods use the same evaluation procedure and GPU configuration in this comparison.

Training hoursMethodTraining stepsAccuracy (%)
18 hSoft Thinking + GRPO20044.2
18 hSoft Spatial Reasoning14048.9

Within these comparable training budgets, "Soft Thinking + GRPO" completes 200 steps while Soft Spatial Reasoning completes 140. The baseline therefore receives 60 additional training steps, yet Soft Spatial Reasoning retains a 4.7-percentage-point accuracy advantage. We have added this analysis to the Results section of the revised manuscript.

AuthorsOpenReview ↗

Q3/W2 [Decision-relevant evidence]

Q3/W2. Understanding the differences across tasks

Are the Egocentric/Allocentric decreases relative to fixed soft thinking reproducible, and what explains the contrast with the geometry/motion gains?

A: Our interpretation is that some tasks with longer reasoning, such as Geometric Reasoning and Motion Analysis, may benefit more from adaptive softness. With shorter reasoning, its advantage over fixed soft thinking may be smaller. Longer traces provide more opportunities for AdaptSoft's step-specific adjustments to help, while "Soft Thinking + GRPO" uses a fixed softness temperature throughout. To examine this interpretation, we measured how many reasoning steps Soft Spatial Reasoning generated during evaluation. The table reports the median lengths alongside the accuracy differences relative to "Soft Thinking + GRPO" from Table 1:

CategoryMedian reasoning stepsAccuracy change (percentage points)
Egocentric104-0.86
Allocentric218-1.30
Motion Analysis290+6.29
Geometric Reasoning486+9.84

Motion Analysis and Geometric Reasoning have median reasoning lengths approximately three and five times that of Egocentric, respectively, and show the largest accuracy gains. These observations are consistent with our interpretation that longer reasoning may provide more opportunities for adaptive softness to help.

We also examine how the softness temperature varies within reasoning traces. AdaptSoft shows 7.2% less variation in softness temperature during Egocentric reasoning than overall. We measure variation as the difference between the 90th and 10th percentile temperatures within each reasoning trace. This observation is consistent with our interpretation that shorter reasoning may offer fewer opportunities for adaptive softness to help. We have added this discussion to the Results section and the reasoning-length and softness-variation analysis to Appendix A of the revised manuscript.

We also express the accuracy differences from "Soft Thinking + GRPO" as changes in the average number of correct answers under mean@8 evaluation. Egocentric and Allocentric decrease by approximately one and five correct answers on average, while Geometric Reasoning and Motion Analysis gain approximately 15 and 22 correct answers, respectively. The categories contain 102, 375, 155, and 346 test questions, respectively.

Finally, our ablations show that the proposed components also benefit perspective taking in a separate Perspective Taking training setting. Removing the hidden-state input from AdaptSoft reduces Egocentric and Allocentric accuracy by 3.92 and 3.01 percentage points; replacing step-specific gradient alignment with the rollout-level task advantage reduces them by 4.90 and 3.27 points. These results suggest that perspective taking benefits from matching softness to the current reasoning state, beyond simply allowing the softness temperature to vary.

If capability-retention evaluations are available, do they distinguish effects of spatial post-training shared by all variants from effects specific to adaptive softness? These results would clarify the scope and trade-offs of the claimed benefit.

A: Soft Spatial Reasoning maintains general multimodal performance close to Base on MMMU and MMBench, with higher accuracy than both post-trained baselines. We conducted additional capability-retention experiments, comparing all three post-trained variants against Qwen3-VL-8B-Thinking (Base) to distinguish the effects of shared spatial post-training from those specific to adaptive softness. The experiments used the same post-trained checkpoints evaluated in Tables 1–3, with all four models evaluated zero-shot under the same settings and without additional training on either benchmark.

ModelMMMU (%)(pp)MMBench (%)(pp)
Qwen3-VL-8B-Thinking (Base)72.44—90.89—
Hard Thinking + GRPO71.56-0.8891.49+0.60
Soft Thinking + GRPO70.22-2.2290.37-0.52
Soft Spatial Reasoning72.22-0.2291.66+0.77

: change relative to the pretrained backbone, in percentage points.

Soft Spatial Reasoning's 2.75-point OmniSpatial gain over "Soft Thinking + GRPO" comes with better capability retention: the MMMU decrease relative to Base shrinks from 2.22 to 0.22 percentage points. On MMBench, the 0.52-point decrease with fixed soft thinking becomes a 0.77-point improvement over Base with adaptive softness. Hard Thinking + GRPO also remains close to Base on MMMU (-0.88 points) and MMBench (+0.60 points). The larger MMMU loss with fixed soft thinking is therefore not a loss shared by all three post-trained variants. We have added these results to the Results section of the revised manuscript.

Comments on the paper

3 comments

We thank all reviewers for their time and constructive feedback, which have helped us strengthen the paper. The revised manuscript includes additional experiments and clarifications, with additions and revisions highlighted in blue\color{blue}{\text{blue}}. We welcome further questions or suggestions and would be happy to clarify our responses or any aspect of the revised manuscript during the discussion.

Dear Reviewers, with the discussion period closing tomorrow, we would be happy to provide any further clarification on our responses or the revised manuscript during the remaining time. Thank you again for your time and consideration.

AuthorsOpenReview ↗

Summary of Responses and Revisions

Dear ACs, SACs, and Reviewers,

Thank you for your time and constructive feedback. We appreciate the reviewers’ recognition of learning how much to mix candidate tokens at each reasoning step (Reviewers mqii, W8z1, and Ud3H), training this control without annotated reasoning steps (Reviewer Ud3H), and the gains over baselines using the same backbone, including on unseen benchmarks (Reviewers mqii and Ud3H).

  1. Baseline tuning and computational cost (Reviewers mqii and Ud3H). We clarified how the fixed softness temperature was selected on validation data and how both methods use the same stopping rule. In the new OmniSpatial comparison, both methods trained for approximately 18 hours on the same GPU configuration. Soft Spatial Reasoning reached 48.9% versus 44.2% for Soft Thinking + GRPO, a 4.7-point gain with 140 versus 200 training steps. We also report training time per step, peak memory, and inference time.

  2. Training-free baselines and MindCube transfer (Reviewer W8z1). Applying fixed soft thinking to the pretrained backbone, with or without an entropy heuristic, kept overall accuracy close to the backbone across all three benchmarks. With the two post-trained baselines added to MindCube, Soft Spatial Reasoning reached 38.13% zero-shot accuracy, compared with 36.32% for Hard Thinking + GRPO and 34.02% for Soft Thinking + GRPO.

  3. Evaluation stability on OmniSpatial (Reviewer Ud3H). Three evaluations of the same Soft Spatial Reasoning checkpoint yielded 49.65% ± 0.10 percentage points (mean ± standard deviation). The 2.73-point gain over fixed soft thinking had a paired 95% confidence interval of [2.44, 3.25] points for the evaluated checkpoints.

  4. Task-level gains and reasoning-state ablation (Reviewers mqii and Ud3H). Geometric Reasoning and Motion Analysis showed the largest gains and had longer reasoning traces than Egocentric, while Egocentric and Allocentric showed small decreases. In the separate Perspective Taking ablation, adding the current reasoning state to a controller that already learned to adjust softness from predictive uncertainty increased accuracy from 39.22% to 42.75%.

  5. General multimodal capability retention (Reviewer mqii). After spatial post-training, Soft Spatial Reasoning scored 0.22 points below the backbone on MMMU and 0.77 points above it on MMBench. It scored higher than both post-trained baselines on both benchmarks.

The revised manuscript includes the additional baseline comparisons, analyses, and clarifications, with changes highlighted in blue. Thank you again for helping us strengthen the paper.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.