acceptodds
Under review as a conference paper at ICLR 2027

Viscot-steer: Disentangling and Transferring Visual Chain-of-Thought via Activation Steering

Abstract

Visual chain-of-thought (visual-CoT), also known as “thinking with images”, equips LVLMs with the ability to locate and use key visual evidence for reasoning. However, acquiring such visual-CoT capabilities requires costly post-training and tightly couples them to a specific backbone, making transfer to other LVLMs difficult. Existing cross-model transfer methods typically treat visual-CoT as a monolithic capability, using a unified representation that ignores the functional distinction between grounding and reasoning and introduces stage-irrelevant interventions. To address this problem, we propose Viscot-steer, a training-free activation steering framework for transferring visual-CoT capabilities across heterogeneous LVLMs. Specifically, Viscot-steer first disentangles visual-CoT into grounding and reasoning vectors and extracts them from the source-model layers associated with the corresponding capabilities. It then aligns the vectors with the activation space of the target model and steers the grounding and reasoning layers during bounding-box prediction and reasoning-trace generation, respectively. Across benchmarks covering high-resolution visual understanding, general visual perception and reasoning, and mathematical visual reasoning, Viscot-steer transfers capabilities from different visual-CoT models to architecturally diverse LVLMs and outperforms the strongest capability-transfer baseline by an average of 1.86 percentage points on the primary evaluation metrics.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
activation steering, visual chain-of-thought, large vision-language models, visual reasoning, visual grounding, cross-model capability transfer

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_2WbJOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Viscot-steer, a training-free approach for transferring visual chain-of-thought (Visual-CoT) capabilities between large vision-language models. Building on activation steering, the authors propose to disentangle Visual-CoT into two capability-specific directions corresponding to visual grounding and reasoning. These directions are extracted through contrastive prompting, aligned with the target model's hidden space, and injected into different transformer layers during the corresponding generation phases.

The method is evaluated on several LVLM backbones and visual reasoning benchmarks. The results show consistent improvements over the base models and generally outperform existing capability-transfer baselines. The authors also provide ablation studies on steering directions, routing strategies, injection layers, and steering strengths.

Critical strengths

1. Clear motivation and a practical framework. The paper addresses a meaningful problem: transferring visual reasoning capabilities across models without repeating expensive post-training. The proposed method is straightforward, computationally lightweight, and does not require updating the target model's parameters. Separating grounding and reasoning interventions is also a reasonable design choice for Visual-CoT, where these two functions operate at different stages.

2. Strong empirical evaluation and presentation. The paper is clearly written and easy to follow. The experimental evaluation is fairly comprehensive, including several target models from different families, multiple benchmarks, comparisons with relevant baselines, and useful ablation studies. The results demonstrate that the proposed framework can improve downstream performance, sometimes by a considerable margin. Overall, the paper is well executed from an empirical perspective.

Critical weaknesses

1. Limited methodological novelty. The framework builds heavily on existing activation steering methods, particularly L2V-CoT. Its main contribution is to separate the original steering direction into grounding and reasoning directions and apply them at different layers and generation phases. While reasonable, this appears to be a relatively incremental extension of prior work.

2. The benefit of disentanglement is not fully convincing. In Table 4, reasoning-only steering already achieves 92.15% on V*, compared with 92.67% for the full method. This raises questions about the additional contribution of the grounding direction in this setting. Moreover, the prompt-based activation differences do not necessarily establish that the extracted vectors represent independent capabilities. The improvements might partly reflect eliciting existing capabilities of the target model rather than transferring newly learned capabilities from the source.

Questions for authors

1. Contribution of disentanglement. Could the authors compare the proposed two-vector approach against a unified steering vector under comparable phase-aware routing and layer-selection settings? This would help clarify whether the improvements come primarily from disentangling the capability vectors or from applying more appropriate intervention strategies.

2. Nature of capability transfer. Have the authors considered extracting steering vectors from the original backbone of DeepEyes, before Visual-CoT post-training, or directly from the target model using the same contrastive prompts? Such comparisons would help determine whether the method transfers source-specific learned capabilities or primarily elicits behaviors already present in the target model.

1 comment

W1: "Limited methodological novelty ... a relatively incremental extension of prior work (L2V-CoT)."

Reply: Thank you for raising this point. While both methods build on activation steering, they address different transfer problems. L2V-CoT transfers reasoning from a text-only reasoning model to a vision-language model. In contrast, our goal is to transfer visual-CoT capabilities from a post-trained visual-CoT model to other LVLMs without additional training.

Visual-CoT involves distinct capabilities, including grounding and reasoning. This raises a key question: which capability should be steered at which layer and generation phase? This question has received little attention in L2V-CoT and other activation steering methods. Our key contribution is to address it by steering grounding and reasoning at different layers and their corresponding phases.

Empirically, on Qwen3-VL-4B, L2V-CoT decreases accuracy from 49.66% to 48.20% on MME-RealWorld-Lite and from 72.75% to 72.50% on HR-Bench 8K, whereas our method reaches 50.70% and 74.75%, respectively. Overall, our method outperforms L2V-CoT in 22 of 25 model-benchmark settings.

W2: "Reasoning-only steering already achieves 92.15% on V*, compared with 92.67% for the full method"; the extracted vectors may not represent independent capabilities.

Reply: Thank you for this observation. Our ablations support effective capability disentanglement, with the grounding and reasoning vectors serving their respective functions. V* answer accuracy provides only an indirect measure of grounding capability, so we additionally examine bounding-box IoU and response length. These ablations show their complementary effects on localization and reasoning.

Table R1. Accuracy, response length, and bounding-box IoU on V* with Qwen2.5-VL-7B (original paper Table 16, latest Table 18).

MethodAccuracy (%)Response length (tokens)IoU
Vanilla75.9298.90.3102
Grounding-only86.39100.80.3445
Reasoning-only86.39161.10.2984
Ours93.72162.00.3472

Grounding-only increases IoU from 0.3102 to 0.3445 with little change in response length. Reasoning-only improves accuracy and generates longer trajectories, but does not improve IoU. Combining both achieves the highest accuracy and IoU.

Our matched unified-vector and source-model comparisons further address the nature of disentanglement and transfer in Q1 and Q2.

Q1: Could the authors compare the proposed two-vector approach against a unified steering vector under comparable phase-aware routing and layer-selection settings?

Reply: Thank you for this suggestion. In our original unified-vector ablation (original Table 4, latest Table 4), the vector is applied at every decoding step, with its layer selected on the validation set. Following your suggestion, we additionally evaluate unified vectors at the same capability-specific layers and with the same phase-aware routing as our method.

Table R2. Unified-vector comparison on V* with Qwen3-VL-4B. The matched comparison uses the same capability-specific layers as Ours.

MethodRoutingAccuracy (%)
Unified vector, original (original Table 4, latest Table 4)All decoding steps88.48
Unified vectors (answer unsteered)Phase-aware88.48
OursPhase-aware92.67

Under comparable phase-aware routing and layer-selection settings, our two-vector approach improves accuracy by 4.19 pp, supporting the benefit of capability disentanglement beyond the effect of the intervention strategy. We have clarified the original configuration in the captions and added the matched comparison to revised Tables 4 and 15.

Q2: Have the authors considered extracting steering vectors from the original backbone of DeepEyes ... or directly from the target model using the same contrastive prompts?

Reply: Thank you for this suggestion. We extracted vectors from Qwen2.5-VL-7B, which is both the target model and the original DeepEyes backbone, using the same extraction data and contrastive prompts. We compare this target-self setting with DeepEyes-derived vectors under the same steering settings.

Table R3. Effect of vector source on V* with Qwen2.5-VL-7B.

MethodVector sourceAccuracy (%)
Vanilla—75.92
Target-selfQwen2.5-VL-7B83.77
OursDeepEyes-7B93.72

DeepEyes-derived vectors outperform target-self by 9.95 pp, supporting the contribution of source-specific learned capabilities from visual-CoT post-training, rather than solely eliciting existing capabilities of the target model. We have added this comparison to revised Table 13 and described its settings in Appendix B, under “Effect of vector source.”

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

VisCoT-Steer proposes a training-free approach to transferring visual chain-of-thought behavior across vision-language models. It extracts grounding and reasoning directions using contrastive prompts, maps them into the target model’s hidden space, and applies them at different layers and generation phases. Experiments cover five main target models and five main benchmarks, with additional evaluations using an alternative source model.

Critical strengths
  1. Transfer without backbone parameter updates. VisCoT-Steer applies activation interventions rather than repeating backbone-specific post-training. It uses separate grounding and reasoning vectors at different layers and generation phases, while leaving backbone parameters unchanged.
  2. The evaluation includes several model families and substantial gains. The experiments include three transfer baselines and GRIT as an alternative source. InternVL2.5-8B improves from 36.25 to 50.38 on HR-Bench 4K, and Qwen2.5-VL-7B from 75.92 to 93.72 on V* Bench. Vanilla models already use the DeepEyes prompting setting, providing a useful prompt-matched comparison.
  3. Additional analyses examine grounding, components, and alignment. Table 16 reports improved bounding-box IoU with grounding steering. Tables 13–14 extend component and routing ablations to InternVL2.5-8B, while Table 15 shows that paired projection-based alignment matches LMN on Qwen3-VL-4B on V* Bench.
Critical weaknesses
  1. The contribution of source post-training is not isolated. The prompt-matched vanilla baseline and grounding-IoU improvements support an effect beyond prompting alone. However, Equation (2) contrasts instructions within the same source model, leaving unresolved whether the transferred directions depend on visual-CoT post-training or primarily elicit behavior already present in the base model; this distinction is central to the paper’s capability-transfer motivation. Question: Please compare DeepEyes-derived and base-checkpoint-derived directions in a representative setting, using matched extraction data and tuning budgets.
  2. The ablations do not sufficiently isolate the claimed functional disentanglement. Tables 4 and 13 support the two-vector design, but do not specify whether the unified-vector comparator uses matched layer placement and routing. Tables 5 and 14 additionally change intervention frequency and answer-phase steering, so their results establish the benefit of selective intervention more directly than correct capability–phase matching. Question: Please clarify the unified-vector configuration and compare correct- and swapped-phase routing at fixed capability-specific layers, with matched intervention rates and strengths and no answer-phase steering.
  3. Calibration and total inference cost are insufficiently documented. Appendix A.2 reports benchmark-specific tuning for the method and baselines, but does not identify the calibration split construction or search budgets. Table 17’s 1.09× ratio measures per-token time rather than end-to-end latency, so it does not establish the claimed negligible query-level overhead when response length and visual interactions can also change. Question: Please specify the calibration splits and search budgets, and report end-to-end latency, generated tokens, and crop calls on the same target model.
Questions for authors

see weaknesses

2 comments

W1: Please compare DeepEyes-derived and base-checkpoint-derived directions in a representative setting, using matched extraction data and tuning budgets.

Reply: Thank you for raising this question. We compared DeepEyes-derived and base-checkpoint-derived directions under the same extraction, steering settings and tuning budgets. The stronger DeepEyes-derived directions suggest that the transferred vectors capture capabilities learned through visual-CoT post-training, rather than solely eliciting behavior already present in the base model.

Table R1. DeepEyes-derived and base-checkpoint-derived directions on V* with Qwen3-VL-4B. The base checkpoint is Qwen2.5-VL-7B-Instruct.

DirectionsOverall (%)Direct (%)Relative (%)
Vanilla86.3989.5781.58
Base-checkpoint-derived89.0191.3085.53
DeepEyes-derived92.6793.9190.79

The DeepEyes-derived directions outperform the base-checkpoint-derived directions by 3.66 pp. This supports our goal of training-free visual-CoT transfer from a post-trained source to another LVLM, without repeating post-training. We have added this comparison to revised Table 13 and described its settings in Appendix B, under “Effect of vector source.”

W2: Please clarify the unified-vector configuration and compare correct- and swapped-phase routing at fixed capability-specific layers, with matched intervention rates and strengths and no answer-phase steering.

Reply: Thank you for this suggestion. In original Tables 4 and 13 (latest Tables 4 and 15), the unified vector is applied at every decoding step, with its layer selected on the same validation set. Following your suggestion, we additionally evaluate unified vectors at the same capability-specific layers and with the same phase-aware routing as our method. We also add answer-unsteered and swapped-routing controls to the phase-aware routing ablations (original Tables 5 and 14, latest Tables 5 and 16).

Table R2. Unified-vector and routing ablations on V* with Qwen3-VL-4B. U denotes unified directions; G and R denote grounding and reasoning vectors.

MethodReasoningBounding boxAnswerAccuracy (%)
Unified vector, original (original Table 4, latest Table 4)UUU88.48
Unified vectors, matched layers and routingUU—88.48
w/o routing (original Table 5, latest Table 5)G + RG + RG + R85.34
w/o routing, answer unsteeredG + RG + R—86.39
Swapped routingGR—85.34
OursRG—92.67

With matched layers and routing, our disentangled vectors improve over the unified-vector control by 4.19 pp. Our routing also outperforms the answer-unsteered and swapped controls by 6.28 pp and 7.33 pp, respectively, supporting the benefit of selecting the corresponding capability vector for each phase.

We have clarified the original settings in the captions and added these controls to revised Tables 4, 5, 15, and 16. Thank you again for helping us clarify these comparisons.

W3: Please specify the calibration splits and search budgets, and report end-to-end latency, generated tokens, and crop calls on the same target model.

Reply: Thank you for this important suggestion. We clarify the calibration protocol and report query-level inference cost below. Original Table 17 (latest Table 19) was intended to show that steering (Eq.5) introduces negligible per-token latency overhead, not that the end-to-end cost of a complete reasoning trajectory is negligible.

1. Calibration. For each benchmark, all methods use the same calibration split, randomly sampled from its training data and disjoint from the test set: 300 examples for V*, 400 for HR-Bench, 1,000 for MME-RealWorld-Lite, and 500 each for MathVista and MathVerse. We tune the steering strength over {0.5, 1, 2, 4}, followed by the applicable steering layers, and select configurations based on calibration accuracy. We tune key parameters to ensure fair comparisons with all baselines, not because our method's gains depend on exhaustive search. Figures 4 and 6 (latest) show robust improvements across a range of evaluated steering strengths and layer settings. We have added details of the calibration splits to Appendix A.2, “Hyperparameter Settings.”

2. End-to-end inference cost. We report accuracy, end-to-end latency, generated tokens, and crop calls for the same target model.

Table R3. End-to-end cost on V* with Qwen3-VL-4B, one H100 80GB GPU and batch size 1.

MethodAccuracy (%)Time (s/query)Tokens/queryCrops/query
Vanilla86.393.64152.41.59
L2V-CoT88.485.22234.11.60
Command-V89.015.18232.91.58
Ours92.675.26210.01.42

Compared with Vanilla, our method generates longer trajectories and achieves higher accuracy with increased end-to-end latency. Compared with L2V-CoT and Command-V, our method achieves higher accuracy at comparable end-to-end latency. We have added the end-to-end cost results to revised Table 20 and clarified the scope of the per-token comparison in the discussion accompanying revised Table 19.

Reviewer_m5CFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Viscot-steer, which transfers visual chain-of-thought behavior through activation steering without updating model weights. It extracts separate grounding and reasoning vectors from a source model, maps them to the target hidden dimension, and applies them at different layers and generation phases. Experiments cover five target LVLMs and five benchmarks, with additional transfer experiments using GRIT. The paper reports an average improvement of 1.86 percentage points over the strongest capability-transfer baseline.

Critical strengths
  1. The method addresses a clear limitation of uniform steering. Separating bounding-box prediction from textual reasoning is a sensible design choice. Phase-aware injection is straightforward to implement and avoids modifying model parameters.

  2. The evaluation covers several meaningful transfer settings. The experiments include different model families and achieve substantial gains on some tasks.

  3. The ablations demonstrate the effectiveness of the proposed method. Separate vectors outperform the unified-vector variant on both Qwen3-VL-4B and InternVL2.5-8B. Table 16 also goes beyond answer accuracy by reporting bounding-box IoU and response length, helping distinguish the effects of grounding and reasoning steering.

Critical weaknesses
  1. The routing ablation does not isolate cross-capability interference. In Table 5, the no-routing variant applies both vectors at every decoding step, whereas the full method also leaves the answer phase unsteered. Its performance drop could therefore reflect interference during reasoning and grounding, disruption of final-answer generation, or both.

  2. Some improvements are too small to interpret confidently without uncertainty estimates. For Qwen3-VL-4B, MathVerse accuracy is 50.69%, compared with 50.63% for Command-V. On MathVista, the full method exceeds the unified-vector ablation by only 0.1 percentage points. These results do not undermine the larger gains, but they provide limited evidence for a consistent advantage across tasks.

  3. The efficiency claim is stronger than the measurements support. Table 17 reports a 9% increase in latency per token, not end-to-end latency. Table 16, using a different target model, shows response length increasing from 98.9 to 162.0 tokens. These results cannot be combined into a runtime estimate, but they show why per-token overhead alone is insufficient to establish negligible inference cost.

Questions for authors
  1. Could you add a routing control that applies both vectors during reasoning and bounding-box prediction but leaves answer generation unsteered? This would separate phase-specific routing from the benefit of avoiding answer-stage intervention.

  2. Could you report paired confidence intervals and sensitivity to the sampled extraction data, especially for the small gains on mathematical reasoning benchmarks?

  3. What are the end-to-end latency, generated-token count, and tool-call count for each method on the same target model? Do the accuracy gains remain under matched token and interaction budgets?

10 comments

W1 & Q1: The routing ablation does not isolate cross-capability interference. ... Could you add a routing control that applies both vectors during reasoning and bounding-box prediction but leaves answer generation unsteered?

Reply: Thank you for this suggestion. Following your proposal, we evaluated an additional routing variant that applies both vectors during reasoning and grounding, while leaving the answer phase unsteered.

Table R1. Routing ablation on V* with Qwen3-VL-4B.

MethodReasoningGroundingAnswerAccuracy (%)
Base———86.39
w/o RoutingG + RG + RG + R85.34
w/o Routing, answer unsteeredG + RG + R—86.39
Ours (phase-aware routing)RG—92.67

Our method improves over w/o Routing, answer unsteered by 6.28 pp, showing the benefit of phase-aware routing even when answer generation is unsteered. We have added this comparison to revised Table 5.

W2 & Q2: Some improvements are too small to interpret confidently without uncertainty estimates. ... Could you report paired confidence intervals and sensitivity to the sampled extraction data, especially for the small gains on mathematical reasoning benchmarks?

Reply: Thank you for highlighting this point. Our primary transfer setting is high-resolution visual reasoning, a major focus of DeepEyes and related visual-CoT models. Our method yields substantial gains on V* and HR-Bench and also improves performance in general and mathematical visual reasoning. On Qwen2.5-VL-7B, our method improves V* accuracy from 75.92% to 93.72%, surpassing DeepEyes at 88.48%. On MathVista and MathVerse, our method achieves 65.6% (base 64.9%, +0.70 pp) and 42.89% (base 38.93%, +3.96 pp), close to DeepEyes at 66.5% and 43.35%, respectively. This suggests effective transfer in mathematical visual reasoning and that the gains may depend on the source model's task-specific capabilities.

Following your suggestion, we report paired confidence intervals below.

Table R2. Accuracy and absolute accuracy differences with paired 95% bootstrap confidence intervals (CIs) for Qwen3-VL-4B.

BenchmarkComparisonAccuracy (%): Ours / comparatorDifference (pp; 95% CI)
V*Command-V92.67 / 89.01+3.66 [0.00, 7.85]
MathVistaCommand-V66.90 / 66.10+0.80 [−2.10, 3.70]
MathVistaUnified vector66.90 / 66.80+0.10 [−2.80, 3.00]
MathVerseCommand-V50.69 / 50.63+0.06 [−2.00, 2.10]

We assess extraction-data sensitivity in terms of data size, sampling seed, and math-source proportion. For data size, Figure 3 shows that even 100 extraction examples yield a strong gain on V* (90.05% vs. 86.39% for the base model). For sampling, we resampled 1,000 extraction examples with three seeds:

Table R3. Accuracy with different extraction samples.

Target modelBenchmarkPaper result (%)Three seeds (mean ± SD, %)
Qwen3-VL-4BV*92.6791.62 ± 0.52
Qwen2.5-VL-7BV*93.7292.84 ± 0.80
Qwen3-VL-4BMathVista66.9066.70 ± 0.50

For data composition, we further varied the proportion of math-source extraction examples.

Table R4. Effect of math-source proportion on MathVista accuracy (%) with Qwen3-VL-4B (mean ± SD over three seeds).

Math-source proportionMathVista OverallGeneralMath
23.5%66.70 ± 0.565.30 ± 0.667.89 ± 0.7
50%66.90 ± 0.565.20 ± 0.668.35 ± 0.6
75%67.00 ± 0.565.00 ± 0.768.70 ± 0.6
100%66.90 ± 0.464.50 ± 0.668.94 ± 0.5

Increasing the math-source proportion improves the Math subset's mean accuracy in these experiments, suggesting that transfer performance also depends on the composition of the extraction data. We have added the data-composition results to revised Table 14.

W3 & Q3: The efficiency claim is stronger than the measurements support. ... What are the end-to-end latency, generated-token count, and tool-call count for each method on the same target model? Do the accuracy gains remain under matched token and interaction budgets?

Reply: Thank you for this feedback. Original Table 17 (latest Table 19) was intended to show that steering (Eq.5) introduces negligible per-token latency overhead, not that the end-to-end cost of a complete visual reasoning trajectory is negligible. We have clarified the scope of the per-token comparison in the discussion accompanying revised Table 19.

Table R5. End-to-end cost on V* with Qwen3-VL-4B, one H100 80GB GPU and batch size 1.

MethodAccuracy (%)Time (s/query)Generated tokens/queryCrops/query
Vanilla86.393.64152.41.59
L2V-CoT88.485.22234.11.60
Command-V89.015.18232.91.58
Ours92.675.26210.01.42

Compared with Vanilla, our method generates longer trajectories and achieves higher accuracy with increased end-to-end latency. Compared with L2V-CoT and Command-V, our method achieves higher accuracy at comparable end-to-end latency.

Following your proposal, we further evaluate all methods under matched token and interaction budgets.

Table R6. V* accuracy (%) under matched token and crop budgets (budget-exhausted trajectories without a complete answer are counted as incorrect).

Token / crop limitsVanillaL2V-CoTCommand-VOurs
128 / 142.4137.1741.3657.07
256 / 384.8285.8685.8691.62

Under the 256-token/3-crop limits, our method improves over Vanilla by 6.81 pp. We have added the end-to-end cost and matched-budget results to revised Tables 20 and 21.

Thank you for your detailed response! After reading your response, my concerns on the ablation of the phase-aware routing (W1&Q1) and inference cost (W3&Q3) are addressed.

However, I'm still concerned about the limited empirical gains of the proposed method. Table R2 shows that the improvements on the mathematical benchmarks fall within the confidence intervals. Moreover, increasing the math-source proportion from 23.5% to 100% only brings limited performance gain (66.70 to 66.90 on MathVista, while the performance of the base model is 66.0). This may indicate that the proposed method may not effectively transfer visual-mathematical reasoning capabilities for LVLMs, which limits the applicability of the method.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear AC,

Thank you for your time and effort throughout the review process. Below is a concise summary of the reviews and our rebuttal.


Before rebuttal

All reviewers recognized the paper’s clear presentation and strong empirical evaluation. Their main concerns were:

  • Novelty and the source of transferred capabilities (bNHr, 2WbJ).
  • Controlled comparisons for disentanglement and phase-aware routing (m5CF, bNHr, 2WbJ).
  • Small mathematical gains, uncertainty, and extraction-data sensitivity (primarily m5CF).
  • Calibration details and end-to-end inference cost (m5CF, bNHr).

During rebuttal

We supplemented the evidence:

  • Disentanglement and routing (m5CF, bNHr, 2WbJ): added matched unified-vector, answer-unsteered, and swapped-routing comparisons.
  • Source of transferred capabilities (bNHr, 2WbJ): compared DeepEyes-derived vectors with base-checkpoint-derived and target-self vectors.
  • Robustness (m5CF): reported paired confidence intervals and sensitivity to extraction-data size, sampling seed, and composition.
  • Efficiency (m5CF, bNHr): measured end-to-end latency, generated tokens, crop calls, and accuracy under matched budgets.
  • Calibration (bNHr): specified data sources, split sizes, and checks for overlap with evaluation sets.

These additions are documented in the revised paper and our responses. We thank all reviewers for their prompt engagement and constructive feedback.

Reviewer 2WbJ: main concerns addressed; rating raised from 2 to 3.

The reviewer confirmed that the new experiments addressed their main concerns and raised their rating from 2 to 3. Regarding novelty, we clarified our contribution: transferring visual-CoT between LVLMs by matching each capability to its steering layer and generation phase, beyond the single-capability transfer focus of activation-steering methods such as L2V-CoT. We appreciate their positive reassessment.

Reviewer m5CF: most concerns addressed; follow-up provided on mathematical gains, awaiting response.

We clarified that our primary objective is strong high-resolution visual reasoning gains while preserving general and mathematical reasoning performance. Transfer gains also depend on the source: DeepEyes primarily targets high-resolution reasoning and obtains smaller mathematical gains. We believe our results and scope clarification address this concern, but have not yet received a reply to our follow-up.

Reviewer bNHr: most concerns addressed; follow-up provided on calibration, awaiting response.

We specified the calibration datasets, splits, and checks ensuring no overlap with evaluation sets, and added these details to Appendix A.2. We believe this addresses the remaining question, but have not yet received a reply to our follow-up and await the reviewer’s response.

We have responded to all raised concerns and remain happy to provide further clarification. Thank you again to the AC and all reviewers for their time and constructive feedback.

Best regards,
Authors #20791

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.