acceptodds
Under review as a conference paper at ICLR 2027

GroundedVCoT: Grounding Interleaved Text-Image Reasoning with Reinforcement Learning

14 views

Abstract

Visual intelligence requires reasoning about how a scene changes and using the resulting state to guide later decisions. Unified multimodal models can make this explicit by interleaving textual reasoning with generated images. Yet generated images also enter the context for later steps. If any generated image misrepresents an action or contradicts the text, the error can compound across the trajectory. We introduce GroundedVCoT, a reinforcement-learning (RL) framework for grounding interleaved text–image chain-of-thought. At each step, task-specific verifiers assess textual claims and generated visual states against the state implied by the model's own decisions. Step-wise grounding rewards and task success yield a shared advantage that updates both text and image policies. Verification follows the model's trajectory, allowing alternative valid solutions while preserving its self-generated visual history. We evaluate three distinct demands on visual reasoning: planning moves in Rush Hour, constructing a semantic figure in Tangram, and tracking rotations in Rolling Dice. Relative to supervised fine-tuning, GroundedVCoT raises Rush Hour solve rate from 48.0% to 71.3% (+23.3%), Tangram accuracy from 25.6% to 50.5% (+24.9%), and Rolling Dice accuracy from 25.6% to 60.0% (+34.4%), also consistently exceeding terminal reward RL, supporting the value of grounding intermediate steps rather than rewarding task completion alone.

Primary area:
generative models
Keywords:
Unified Models, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_tXAhOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The intermediate visual states generated by the model are fed back into subsequent reasoning steps. Therefore, once an intermediate image becomes inconsistent with the textual decision or the underlying task state, the resulting error may accumulate and propagate throughout the multi-step reasoning process, a phenomenon the authors refer to as compounding drift. To address this issue, the authors propose GroundedVCoT. Its core idea is to verify the textual claims and visual states at each step along the model’s self-generated reasoning trajectory, and to combine these intermediate grounding rewards with the final task reward. A shared cross-modal advantage is then used to jointly optimize the text policy and the image policy.

Critical strengths
  1. The paper points out that once intermediate generated images are fed into subsequent reasoning steps, errors can evolve from isolated generation mistakes into multi-step state drift. This issue is highly relevant to the current development of interleaved multimodal reasoning.
  2. The method is well designed. The paper not only reports overall performance improvements, but also validates its core design through reward ablations and credit assignment ablations. On Rush Hour, terminal-reward RL achieves 63.3%, while the full method improves performance to 71.3%. In addition, the shared advantage outperforms separate advantages.
Critical weaknesses
  1. The current tasks are highly structured and rely on explicit simulators or geometric rules, while GroundedVCoT depends on manually designed verifiers for supervision. Therefore, it remains unclear whether the method can effectively generalize to open-world settings or tasks where state transitions cannot be explicitly verified. The authors also acknowledge this limitation.
  2. GroundedVCoT uses relatively rich step-wise supervision, but the paper lacks a baseline with a comparable level of supervision density that does not explicitly enforce text–visual grounding. As a result, it is still difficult to determine whether the performance gains mainly come from improved grounding consistency or from dense reward shaping more generally.
Questions for authors

Could the authors include a step-wise reward baseline with a similar reward density but without explicit grounding consistency? This would help disentangle the contribution of grounding itself from the more general benefit of dense supervision.

3 comments

Thank you for the positive assessment and the suggestion to distinguish explicit grounding from the general benefit of dense supervision.

W2,Q. Dense-reward comparison We agree that Table 3 alone does not disentangle explicit grounding from the general benefit of dense supervision. To identify whether the performance gains mainly come from improved grounding consistency or from dense reward shaping, we decided to test our reward design against the dense reward design of BRAID [1] which is one of the closest prior work as stated by other reviewers. We contacted the BRAID authors, who confirmed that their code and checkpoints cannot currently be shared pending release approval. We therefore adapted BRAID's published intermediate visual-reward design within our Rush Hour training pipeline, using Qwen3-VL-32B-Instruct instead of its original GPT-5.2 judge. We keep the BAGEL SFT initialization, training data, rollout budget, policy optimizer, and evaluation protocol fixed across all variants.

MethodOverallL1L2L3L4L5
Terminal-reward RL63.3%100.0%100.0%66.7%40.0%10.0%
BRAID-style visual reward62.7%100.0%93.3%76.7%33.3%10.0%
GroundedVCoT71.3%100.0%100.0%83.3%56.7%16.7%

The BRAID-style dense reward achieves 62.7%, close to terminal-reward RL (63.3%), whereas GroundedVCoT reaches 71.3%. Thus, under the same training framework, this BRAID-style dense visual reward does not reproduce the gains of GroundedVCoT. Our method differs by verifying both textual claims and generated visual states against the evolving task state induced by the model's own trajectory. In this comparison, both GroundedVCoT and the BRAID-style adaptation use the same shared-advantage mechanism to update text and image policies, allowing us to compare intermediate reward designs under matched credit assignment. Table 4 separately examines credit assignment within GroundedVCoT: holding the grounding rewards fixed, sharing their advantage across both policies improves solve rate from 64.7% to 71.3% over modality-specific advantages. Together, these results support our intermediate grounding design and the benefit of shared cross-modal credit assignment. We have added the BRAID-style comparison to the revised manuscript.

[1] Bridging Interleaved Multi-Modal Reasoning as a Unified Decision Process

W1. Generalization beyond structured tasks We agree that our reliance on task-specific verifiers limits the immediate applicability of GroundedVCoT to unstructured reasoning tasks. Our current contribution targets verifiable interleaved visual reasoning, where state transitions or geometric constraints can be checked during training.

To assess transfer beyond structured puzzle solving, we are additionally evaluating GroundedVCoT on broader visual understanding and reasoning benchmarks following ThinkMorph's evaluation protocol [2]. We will share the results during the rebuttal period and incorporate them into the revised manuscript as soon as the evaluations are complete.

[2] ThinkMorph: Emergent Properties in Multimodal Interleaved Chain-of-Thought Reasoning. ICLR 2026.

Thank you for the detailed response and additional experiments.

Regarding W2, the newly added BRAID-style dense reward comparison addresses my main concern. Under the same training setup and shared advantage mechanism, GroundedVCoT still achieves better results, which provides further support for the effectiveness of the grounding reward design. Although this experiment uses an adapted version of BRAID's reward design and cannot fully disentangle the effects of different reward mechanisms, I believe this additional comparison substantially strengthens the evidence supporting the authors' conclusions. Regarding W1, I understand that the authors primarily focus on verifiable, structured visual reasoning tasks, and I appreciate their clarification of the scope of the method.

Overall, the additional experiments have largely addressed my concern about the reward design. I still have some reservations regarding generalization, but I acknowledge the authors' explanation of the current scope and limitations of the method.

Thank you for the thoughtful follow-up and for recognizing that the additional experiments address your main concern about the grounding reward design. Regarding generalization, we have obtained new results that we believe help address your remaining reservation. New experiments: Transfer to broader visual reasoning. Starting from the ThinkMorph SFT checkpoint, we additionally trained GroundedVCoT using a subset of StructCoT (Wang et al., 2026) and evaluated it on six visual reasoning benchmarks following ThinkMorph's protocol.

BenchmarkThinkMorph SFTGroundedVCoT
VStar (out-of-domain)64.40%67.02%
MMVP (out-of-domain)80.67%81.00%
SAT (out-of-domain)51.33%53.33%
VSP (in-domain)73.83%78.17%
VisPuzzle (in-domain)77.25%77.50%
ChartQA (in-domain)79.26%80.12%
GroundedVCoT achieves numerical improvements over its SFT initialization on all six benchmarks, including VStar, MMVP, and SAT, which fall outside the training task families. These results provide encouraging evidence that grounding-based RL can benefit broader visual reasoning capabilities, rather than only improving performance on the tasks used for training.
We agree that extending the training framework to arbitrary open-world settings remains an important challenge. Nevertheless, these findings suggest that the benefits of grounding-based supervision need not be confined to the tasks on which the supervision is provided. No task-specific verifiers are required at inference time.
We hope these additional results further strengthen the evidence for the broader relevance of GroundedVCoT.
Reviewer_zAikOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper aims to solve a problem "compounding drift" that wrong image state generated in the reasoning process of UMMs is propagated to the subsequent reasoning trajectory. It proposed a reinforcement learning framework called GroundedVCoT, which applies step-wise textual and visual grounding rewards and assigns their shared advantage to both policies. Experimental results show the effectiveness of this method.

Critical strengths
  1. The interleaved text-image reasoning problem is interesting and popular. This paper provide a new perspective for this problem from the UMMs settings.

  2. Empirical studies show the effectiveness of GroundedVCoT and provide great mechanisms analysis of why it works.

Critical weaknesses
  1. The methodological novelty is somewhat limited. Prior works on VLMs, such as [1, 2], have already explored perception-aware process rewards for improving visual reasoning through RL. While extending such supervision to joint text-image generation is meaningful, the paper does not sufficiently clarify the novel contributions beyond existing process-level reward learning.

  2. The comparison with closely related UMM methods is insufficient. In particular, BRAID has already explored joint optimization of text and image policies with intermediate visual feedback, but it is not included as a baseline. Moreover, GroundedVCoT relies on task-specific verifiers, which may limit its applicability to less structured reasoning tasks. The benefits of its particular grounding design over more general process-level supervision remain insufficiently established

[1] Yingqian Min, et al. Improving Vision-language Models with Perception-centric Process Reward Models. CVPR 2026.

[2] Tong Xiao, et al. Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward. ICLR 2026.

Questions for authors
  1. In Sec 3.3, the paper stated that "follow the required structure and refer to valid task entities that can be processed by the verifier", which suggests that additional task-specific verification information is required. Is this process computationally heavy? What are the additional annotation and verification costs? Could you provide the comparison of your methods and vanilla RL or baselines under the controlled computation budget setting?

  2. Table 3 demonstrates the complementary benefits of textual and visual grounding rewards. But, there is no evidence in this paper that shows cross-modal supervision is fundamentally necessary for the dense process reward. How about using other similar dense supervision like BRAID?

  3. In table 5, the authors intervene on the visual history while keeping the textual trajectory fixed and regenerating only the final answer step. However, the fixed textual context may already contain information from the original generated images, potentially confounding the ablation. Moreover, the sentence in Line 461-462 "Because the replacements may conflict with claims in the fixed textual trace, their lower solve rate cannot be attributed solely to the quality of the images". This explanation seems somewhat puzzling, as it raises concerns about whether the ablation can isolate the contribution of generated images. Could the authors clarify?

5 comments

Thank you for the detailed feedback. We address your concerns with new controlled experiments, intermediate-state interventions, and measurements of verification overhead.

W1, Q2. Contribution beyond perception-aware process rewards and BRAID We thank the reviewer for pointing out Perceval and Perception-R1. We have added both works to the related-work section and clarified how GroundedVCoT differs from them.

Both methods establish the value of perception-aware rewards for improving textual reasoning over visual observations. GroundedVCoT addresses a different setting in which the model does not only reason about visual evidence, but also generates the visual state that becomes evidence for subsequent reasoning.

Our contribution is to ground both modalities against the evolving task state. Rather than checking only perceptual claims in text, we verify both textual claims and generated images against the state transition induced by the model's own action, or the corresponding geometric constraints. This allows alternative valid trajectories to receive full grounding credit without following a fixed demonstration. The resulting step-level advantage updates both policies, so errors in either modality provide feedback to both. BRAID is particularly relevant because it jointly optimizes text and image policies, but applies its intermediate VLM-based visual-process advantage only to image turns. In contrast, GroundedVCoT combines textual and visual grounding against the evolving task state into a shared step-level advantage that updates both policies. New controlled comparison. We agree that BRAID is an important comparison. We contacted its authors to request the pretrained checkpoint for direct evaluation, but they confirmed that their code, data, and checkpoints are pending release approval and cannot currently be shared. To address this limitation, we adapted BRAID's published intermediate visual-reward design within our Rush Hour pipeline, using Qwen3-VL-32B-Instruct instead of its original GPT-5.2 judge. All variants share the same initialization, training data, rollout budget, policy optimizer, and evaluation protocol, allowing us to compare alternative intermediate supervision designs under the same training framework.

MethodOverallL1L2L3L4L5
Terminal-reward RL63.3%100.0%100.0%66.7%40.0%10.0%
BRAID-style visual reward62.7%100.0%93.3%76.7%33.3%10.0%
GroundedVCoT71.3%100.0%100.0%83.3%56.7%16.7%

The BRAID-style dense reward achieves 62.7%, close to terminal-reward RL (63.3%), whereas GroundedVCoT reaches 71.3%. Under the same training framework, this BRAID-style dense visual reward does not reproduce the gains of GroundedVCoT. In this comparison, both GroundedVCoT and the BRAID-style baseline use shared advantages to update text and image policies, isolating the effect of intermediate reward design. Table 4 separately examines credit assignment within GroundedVCoT: holding the grounding rewards fixed, sharing their advantage across both policies improves solve rate from 64.7% to 71.3% over modality-specific advantages. Together, these results support both our intermediate grounding design and the benefit of shared cross-modal credit assignment. We have added this comparison and clarified the methodological distinction from prior process-reward approaches in the revised version of the paper.

Scope and generalization. We are currently evaluating GroundedVCoT on broader visual understanding and reasoning benchmarks, following ThinkMorph’s [1] evaluation protocol. These evaluations assess transfer beyond structured puzzle solving and require no simulator, oracle intermediate images, or task-specific verification feedback during inference.

We will share the results during the rebuttal period and incorporate them into the revised manuscript as soon as the evaluations are complete.

[1] ThinkMorph: Emergent Properties in Multimodal Interleaved Chain-of-Thought Reasoning. ICLR 2026.

Q1. Verification requirements and computational cost GroundedVCoT requires task structure but no additional human annotation of generated trajectories. Rush Hour and Rolling Dice use metadata and deterministic transition rules available from their generators, while Tangram reuses KiloGram's existing polygon geometry and semantic part annotations. The main additional setup cost is implementing the task-specific verifier.

We measured the verification overhead on Rush Hour. A complete 10-epoch GroundedVCoT run on 8 NVIDIA H200 GPUs took 68.6 hours, of which only 25.5 minutes (0.62%) were spent computing all grounding and task rewards. Image generation and policy optimization dominate the training cost.

Our comparison with terminal-reward RL uses the same BAGEL SFT initialization, training data, epochs, number of rollouts per question, and optimization procedure. Under this matched training budget, GroundedVCoT improves solve rate from 63.3% to 71.3%. Thus, the gain does not require additional rollouts or optimization steps, and the computational overhead of verification is small. We have added these cost and annotation details to the revision.

Q3. Visual history intervention and downstream error propagation We agree that the fixed textual trajectory in the original Table 5 may retain information from the original images and conflict with their replacements. Consequently, that experiment measures the contribution of visual history conditional on the retained text, but cannot isolate downstream error propagation.

We therefore conducted a new intervention experiment on 50 Tangram test puzzles. At steps 1–5 of each policy's own trajectory, we translate one recognized piece by approximately 16 pixels so that it partially leaves the target silhouette, preserving the preceding history, action text, and all other pieces. We then freely regenerate all subsequent text and images.

The intervention changes subsequent piece selection in 17.1% of SFT trajectories and 20.0% of GroundedVCoT trajectories, confirming that intermediate visual states influence later decisions. The intervention also reduces final construction accuracy by 0.2497 for SFT and 0.3704 for GroundedVCoT, although this metric directly penalizes the displaced piece and therefore cannot alone establish downstream propagation.

To distinguish downstream propagation from persistence of the displaced piece, we measure the grounding accuracy of subsequently introduced pieces, excluding all pieces recognized before or at the intervention and keeping the remaining inventory as a fixed denominator.

PolicyOriginalCorruptedChange [95% CI]
SFT0.31370.2723−0.0414 [−0.0757, −0.0078]
GroundedVCoT0.47670.4557−0.0209 [−0.0609, +0.0164]
Confidence intervals are obtained by resampling puzzles after averaging eligible intervention points within each puzzle.

Under SFT, corrupting one intermediate image reduces the grounding accuracy of subsequently introduced, unperturbed pieces by 4.14 percentage points. This provides direct evidence of downstream visual-error propagation without fixing subsequent reasoning. GroundedVCoT retains higher downstream grounding accuracy after corruption, although the difference in sensitivity between policies is not statistically conclusive. We have incorporated the intervention results into Table 5 of the revised manuscript, with additional step-wise analysis and a qualitative example in Appendix I.

I really appreciate the authors' substantial effort to address my concerns.

Most of my concerns have been addressed, however, W2 is still not fully resolved. As the authors stated, "GroundedVCoT requires task structure," which means the scope of this method is very limited. Structured and clean tasks today can often be solved through scaling, and for tasks where data is limited, the annotation process can also be expensive.

So, although GroundedVCoT may outperform the BRAID-style baseline on these tasks, I do not think it has a significant impact on the community. I will keep my original score.

Reviewer_QLV1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2(+1)
Presentation
3
Soundness
3(+2)
Read review
Paper summary

GroundedVCoT studies error accumulation in interleaved text–image reasoning, where an incorrect generated image becomes part of the context and affects subsequent decisions. The proposed framework uses task-specific verifiers to evaluate textual claims and generated visual states along the model’s own trajectory. These intermediate rewards are combined with task rewards to construct a shared step-level advantage for jointly optimizing the text and image policies

Critical strengths
  1. The paper addresses a meaningful failure mode of unified multimodal models: errors in generated intermediate states can propagate through subsequent reasoning.
  2. The results show improvements over both supervised fine-tuning and terminal-reward RL. On Rush Hour, the full method improves solve rate from 63.3% with terminal-reward RL to 71.3%. The reward and credit-assignment ablations provide useful supporting evidence.
Critical weaknesses
  1. The incremental contribution is unclear. BRAID [1] already combines joint text–image RL, shared credit assignment, and intermediate visual feedback.
  2. Evidence for mitigating compounding visual drift remains indirect. In Table 5, removing image history reduces solve rate from 71.3% to 67.3%, while random images yield 68.0%. Since these interventions fix the textual trajectory and regenerate only the final answer, they do not establish how visual errors affect subsequent decisions. Intermediate-step interventions that rerun the remaining trajectory and measure error propagation would provide stronger evidence.

[1] Bridging Interleaved Multi-Modal Reasoning as a Unified Decision Process

Questions for authors
  1. What is the key methodological contribution beyond BRAID? Can controlled comparisons isolate the benefits of task-verifiable grounding from those of the underlying RL procedure?

  2. Can you intervene on images at intermediate steps and rerun the remaining trajectory, allowing subsequent text to change, to establish how visual errors propagate and affect final task performance?

4 comments

Thank you for recognizing the importance of the problem and our improvements over terminal-reward RL. We address both concerns with new experiments in the revised manuscript, summarized below:

W1,Q1. Contribution beyond BRAID and controlled comparison We agree that BRAID is the closest prior work. Our contribution lies in jointly grounding textual claims and generated visual states against the evolving task state, and combining their grounding signals into a shared step-level advantage that updates both policies.

BRAID uses a VLM judge to assess intermediate images for visual correctness, fidelity, reasoning utility, and trustworthiness. Its visual-process advantage is applied to image turns, while text turns receive the shared terminal-reward advantage (BRAID Eq. 10). GroundedVCoT instead (i) verifies both textual claims and generated visual states against the same evolving task state, (ii) follows the state induced by the model's own actions rather than a fixed demonstration, allowing alternative valid trajectories, and (iii) assigns the resulting grounding advantage to both policies at each step. Thus, visual grounding errors also provide feedback to the text policy, and textual grounding errors to the image policy.

This distinction targets a failure identified by MentisOculi [1] that models may generate incorrect intermediate visual states and may also fail to reason reliably from visual states even when those states are correct. GroundedVCoT therefore supervises both the generated state and the reasoning policy that consumes it, rather than treating intermediate visual feedback as an image-only signal.

New controlled comparison. To assess the benefit of our grounding design beyond the underlying RL procedure, we contacted the BRAID authors to request their model checkpoint for direct evaluation. They confirmed that their code, data, and checkpoints are still pending release approval and cannot currently be shared. We therefore adapted BRAID's published intermediate visual-reward design within our Rush Hour training pipeline, using Qwen3-VL-32B-Instruct instead of its original GPT-5.2 judge. All variants share the BAGEL SFT initialization, training data, rollout budget, policy optimizer, and evaluation protocol. This comparison is thus a controlled adaptation of BRAID's reward design rather than an exact reproduction of the full method.

Our ablations isolate these choices while keeping the backbone and RL procedure fixed. On Rush Hour, terminal-reward RL achieves 63.3%, while adding textual and visual grounding reaches 71.3% (Table 3). Holding the grounding rewards fixed, sharing their advantage across both policies improves 64.7% to 71.3% over modality-specific advantages (Table 4), supporting cross-modal assignment of the dense grounding signal.

MethodOverallL1L2L3L4L5
Terminal-reward RL63.3%100.0%100.0%66.7%40.0%10.0%
BRAID-style visual process reward62.7%100.0%93.3%76.7%33.3%10.0%
GroundedVCoT71.3%100.0%100.0%83.3%56.7%16.7%

The BRAID-style reward achieves 62.7%, close to terminal-reward RL (63.3%), whereas GroundedVCoT reaches 71.3%. In this comparison, both GroundedVCoT and the BRAID-style adaptation use shared advantages to update text and image policies, isolating the effect of intermediate reward design. Table 4 separately examines credit assignment within GroundedVCoT: holding the grounding rewards fixed, sharing their advantage across both policies improves solve rate from 64.7% to 71.3% over modality-specific advantages. Together, these results support our intermediate grounding design and the benefit of shared cross-modal credit assignment. In the revised paper, we have clarified these distinctions from BRAID.

[1] MentisOculi: Revealing the Limits of Reasoning with Mental Imagery. ICML 2026

W2,Q2. Intermediate-state intervention and error propagation We agree that the fixed textual trajectory in the original Table 5 prevents a direct assessment of downstream error propagation. We therefore conducted a new intervention experiment in which all subsequent text and images are regenerated.

On 50 Tangram test puzzles, each policy generates its own trajectory. At steps 1–5, we translate one recognized piece by approximately 16 pixels so that it partially leaves the target silhouette, preserving the preceding history, action text, and all other pieces. We then regenerate the remaining trajectory without constraining subsequent decisions.

The intervention changes subsequent piece selection in 17.1% of SFT trajectories and 20.0% of GroundedVCoT trajectories, showing that altered visual states influence later decisions. The intervention also reduces final construction accuracy by 0.2497 for SFT and 0.3704 for GroundedVCoT. However, because this score directly penalizes the displaced piece, it cannot alone establish downstream error propagation.

To distinguish downstream error propagation from persistence of the displaced piece, we measure new-piece grounding accuracy: the fraction of the remaining piece inventory that is subsequently grounded correctly, excluding all pieces recognized before or at the intervention. Missing or malformed pieces receive no credit.

PolicyOriginalCorruptedChange [95% CI]
SFT0.31370.2723−0.0414 [−0.0757, −0.0078]
GroundedVCoT0.47670.4557−0.0209 [−0.0609, +0.0164]
Confidence intervals are obtained by resampling puzzles after averaging eligible intervention points within each puzzle.

Under SFT, corrupting one intermediate image reduces the grounding accuracy of subsequently introduced, unperturbed pieces by 4.14 percentage points. This provides direct evidence of downstream error propagation rather than merely persistence of the injected error. Moreover, for step-2 interventions, the cumulative grounding deficit grows from 0.0057 at step 3 to 0.1144 at step 7 across 44 paired interventions (Appendix I), illustrating how errors can accumulate.

GroundedVCoT retains higher downstream grounding accuracy after corruption (0.456 vs. 0.272). Although its observed accuracy drop is smaller, the between-policy difference in sensitivity is not statistically conclusive (+0.0153, 95% CI [−0.0453, +0.0725]). We therefore do not claim robustness to externally injected errors. Rather, the intervention provides causal evidence of downstream visual-error propagation in Tangram, while our main results demonstrate the benefit of grounding the model's own generated states during training.

We have added this experiment, further step-wise analysis, and a qualitative example to the revised manuscript (Appendix I).

Reviewer_QLV1OpenReview ↗

responds to authos

Thank you for your response and for running additional experiments to address Question 2. I found those results convincing and will raise my ratings for Informativeness and Soundness. I still see limited differentiation between this work and BRAID, so I will keep my overall score.

Thank you for reconsidering the Informativeness and Soundness assessments. We appreciate your acknowledgment that the new intervention experiments address the concern about downstream visual-error propagation. Regarding BRAID, rather than restating the methodological distinctions discussed in our earlier response, we would like to emphasize what the new controlled comparison shows. If GroundedVCoT's gains primarily came from the components it shares with BRAID, namely joint text-image RL, intermediate feedback, and shared credit assignment, then replacing our grounding signal with BRAID-style visual-process feedback while keeping those components fixed should preserve most of the gain. In our controlled comparison, both variants use the same backbone, training data, rollout budget, optimizer, and shared-advantage update. The main change is the intermediate supervision signal. Under these matched conditions, the BRAID-style variant achieves 62.7%, whereas GroundedVCoT achieves 71.3% on Rush Hour. The 8.6-point gap under matched credit assignment therefore provides direct evidence that the form of intermediate supervision introduced by GroundedVCoT materially contributes beyond the shared RL framework. Table 4 provides the complementary control. When the grounding rewards are held fixed and only their assignment is changed, performance decreases from 71.3% to 64.7%. Together, these experiments separate the contribution of the grounding signal from the contribution of how that signal is propagated across the text and image policies. We have revised the paper to make this empirical decomposition, rather than a broader claim of novelty in joint text-image RL, the basis of our contribution. We hope this addresses the remaining concern about differentiation from BRAID and may lead you to reconsider the overall rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.