acceptodds
Under review as a conference paper at ICLR 2027

RILO: Feeding Implicit Process Rewards Back into Visual Reasoning for Industrial Anomaly Understanding

Abstract

Industrial anomaly understanding requires reasoning over subtle, localized, and category-dependent visual evidence whose relevance may change as the response evolves. However, standard multimodal large language models (MLLMs) typically encode the input image only once and keep visual-token representations fixed throughout decoding. This creates a mismatch between the evolving visual demands of anomaly reasoning and the static visual representation used for generation. To bridge this gap, we propose RILO (Reward-guided Inspection Loop with Online Visual Modulation), a parameter-efficient post-training framework that uses implicit process rewards not only for policy learning but also as online feedback for visual adaptation. RILO learns state-aligned token-level rewards from answer correctness, evidence compatibility, and structured-output consistency, and periodically uses them to adapt query-image representations during response generation. This generate–evaluate–revise loop redirects visual inspection toward evidence relevant to the current reasoning state, without requiring annotated reasoning traces or additional spatial supervision. Extensive experiments demonstrate that RILO achieves state-of-the-art accuracy on the MMAD benchmark under both one-shot and zero-shot settings, reaching 81.26 and 79.52, respectively. The code will be publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Industrial Anomaly Detection, Large Multimodal Models, Post-Training, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_HySrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RILO targets industrial anomaly question answering on MMAD. After a structured four-field SFT, it trains a policy and an implicit process-reward branch (both LoRA adapters) against a frozen SFT reference. The reward branch learns from a lexical, rule-based process score, and its token-level log-ratio serves as a dense PRIME-style reward for PPO. The distinctive component is visual modulation: every 64 generated tokens, the prefix is scored, the scores are turned into a global magnitude and 20 group weights, a scalar residual is added to contiguous groups of query-image embeddings, and the context is re-prefilled . RILO-4B reports 81.26 (one-shot) and 79.52 (zero-shot) on MMAD.

Critical strengths
  1. Reusing one learned process signal for both credit assignment and visual adaptation is intuitive and clearly presented.

  2. The ablations are careful for the scale of the study: three seeds, a 2×2 decomposition of training-time modulation versus inference-time control, and an explicit cost analysis.

Critical weaknesses

1, Weak motivation. The premise that "static" visual representations cannot follow changing evidence needs overlooks that the decoder already re-reads the image at every step through attention; only the input embeddings are fixed. No pilot study or analysis shows that fixed embeddings are an actual bottleneck.

2., The modulation is a small, coarse perturbation rather than evidence-guided re-inspection. Each of 20 contiguous token groups receives one scalar on all dimensions (α = 0.01), with weights interpolated from answer-token order onto image-token order, and the authors acknowledge that the profile is not tied to image regions . The loop also rarely fires: responses are capped at 160–192 tokens, and the 4B model's mean response length falls to 41 tokens, below the first 64-token boundary.

  1. The gains are not attributed to the claimed mechanism. Most of the improvement comes from the structured verifier plus PRIME-style implicit rewards (68.00 → 73.32 on 3B) and from applying modulation during training (+1.24 on 4B), while inference-time control adds only +0.16 on 4B at 1.88× evaluation cost. All specificity controls are inference-time. Without a training-time control that replaces the reward-guided profile with a magnitude-matched random perturbation, the training-time gain may simply be a regularization or input-noise effect.

  2. Small margins and reporting issues. The margins over the strongest baselines are small (+0.07 over Gemini-2.5-Pro and +0.43 over DCR2-AD (+DK)-7B, one-shot), baseline numbers come from mixed sources, Table 1 duplicates two rows, and β_RM and β_vis are not reported.

Questions for authors

1, What happens if training-time modulation uses a magnitude-matched random or uniform perturbation instead of the reward-guided profile?

2, Can you provide evidence that fixed input embeddings, as opposed to attention, are a bottleneck for anomaly reasoning?

5 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer HySr (1/4)

Q1. On the Motivation for RILO

A1: We thank the reviewer for pointing out that our presentation of the research motivation was not sufficiently precise. We agree that although visual-token representations remain fixed during decoding, the decoder can still dynamically adjust its use of visual information according to the generation context through attention. At the same time, we would like to clarify that RILO does not claim that fixed visual embeddings prevent the model from dynamically attending to the image, and our main objective is not to improve performance through inference-time visual modulation.

RILO focuses on how process-evaluation feedback can be used to adjust the visual state within the current rollout during reinforcement-learning training, thereby improving the final policy learning. Conventional implicit process rewards are mainly used for policy optimization after trajectory generation. RILO further uses this feedback to update the visual features on which subsequent generation within the current rollout depends, allowing the Actor to learn under process-reward-conditioned dynamic visual states. This does not conflict with the decoder dynamically allocating attention according to the text context; rather, it adds a process-evaluation-driven visual feedback path on top of that capability.

This motivation is supported by the experimental results. Table 3 of the original paper shows that, even with the Controller completely disabled at inference time, enabling VM during training increases the 4B model's MMAD Avg. from 79.86% to 81.10% (+1.24 pp). This indicates that the main gains are reflected in the Actor obtained after training, rather than depending on continuously modifying visual features during inference.

To directly examine whether the training-time visual-modulation gains arise merely from general feature perturbations, we added a two-seed RL training ablation for RILO-4B, using two independent training seeds. All models in this training ablation are evaluated with the inference-time Controller disabled, so that the effects of the training-time visual-modulation strategies can be compared independently.

Table 1. Two-seed reproduction of the training-time VM ablation (MMAD Avg., %).

Training VMSeed 1Seed 2MeanLearned Gain (pp)
Learned VM80.6881.2680.97—
Reward-Random79.5280.2479.881.09
Norm-matched Uniform79.2379.6779.451.52

Reward-Random uses modulation weights and strengths sampled from a fixed Learned calibration distribution, independently of the current sample's PRM feedback, with perturbation magnitude calibrated to the actual number of visual tokens. Norm-matched Uniform retains the current sample's global PRM modulation signal and matches the perturbation magnitude using the visual-token-weighted RMS, while setting all visual-group weights to one.

As shown in Table 1, Learned VM achieves the highest accuracy under both independent seeds, with a two-seed mean MMAD Avg. of 80.97%. This is 1.09 and 1.52 percentage points higher than Random (79.88%) and Uniform (79.45%), respectively. Specifically, Learned improves over Random by 1.16 and 1.02 percentage points under the two seeds, and over Uniform by 1.45 and 1.59 percentage points, respectively, showing consistent performance advantages. These results indicate that process-reward-conditioned visual-state updates are more beneficial to final policy learning than the tested random perturbation and uniform offset, rather than merely increasing visual-feature perturbations.

We will further clarify this distinction in the revised manuscript: RILO does not treat fixed visual representations as an inherent limitation on dynamic attention. Instead, it studies how implicit process evaluation can be converted into visual-state feedback within an RL rollout, thereby improving the final learned policy.

AuthorsOpenReview ↗

Rebuttal to Reviewer HySr (2/4)

Q2. On the Spatial Relevance and Actual Activation Frequency of Visual Modulation

A2: We thank the reviewer for the concern about the visual-modulation strategy and its practical effect. We agree that RILO uses lightweight group-level visual-feature modulation and does not establish an explicit correspondence between language tokens and image spatial positions. We do not claim that fixed visual embeddings prevent the decoder from dynamically attending to image evidence. Rather, RILO investigates whether process-reward feedback can provide an additional visual-state update within the current rollout during RL training. Nevertheless, the reviewer's question about whether this interpolated modulation has actual spatial relevance deserves further examination.

To this end, we added a spatial analysis based on ground-truth defect masks. Specifically, we first map the ground-truth defect masks to the image regions represented by the visual tokens after Qwen3-VL's actual image preprocessing. We then apply RILO's original 20-group index mapping to associate each visual token with its corresponding modulation weight and calculate the normalized proportion of modulation weight assigned to defect-region tokens. The masks are used only for offline spatial analysis and are not involved in training, inference, or modulation-weight generation.

Table 2. Differences in normalized modulation-weight allocation to ground-truth defect regions between Learned and the two controls (percentage points).

Analysis SetQuery ImagesLearned − Position-only (pp)Learned − Cross-sample (pp)
Four defect-related QA5500.1960.199
Defect Localization1190.0590.068

Across 550 query images, Learned achieves higher normalized weight allocations to actual defect regions than Position-only and Cross-sample, with differences of +0.196 and +0.199 percentage points, respectively. The Defect Localization subset shows positive differences in the same direction (+0.059 and +0.068 pp). The Defect Localization subset shows results in the same direction. Learned therefore has a positive advantage in weight allocation to actual defect regions over both controls, further supporting the observation of a spatial association between reward-conditioned modulation and actual defect regions.

Regarding the VM activation frequency, we further examined the actual training logs from the first 188 steps of RILO-4B RL training. The average response length of approximately 41 tokens reported in Appendix L of the original paper corresponds to the later stage of training, rather than the entire training process.

Table 3. VM activation across different training stages.

Training PhaseActivated / TotalRate
Pass 1–414,367 / 16,80085.52%
Pass 5–8890 / 16,8005.30%
Pass 9–1321 / 19,0400.11%
Overall15,278 / 52,64029.02%

The activation statistics cover the first 188 RL training steps actually recorded, totaling 52,640 rollouts. Pass 13 includes only the 8 steps that had been completed. Effective Activation means that at least one actual non-zero visual-state update was performed in the corresponding rollout.

The statistics show that the effective activation rate is 85.52% (14,367/16,800) in the first four training passes, decreases to 5.30% in Passes 5–8, and further decreases to 0.11% in Passes 9–13. In total, 15,278 of the 52,640 rollouts have effective VM activation, corresponding to an overall activation rate of 29.02%. Therefore, we agree with the reviewer that VM is triggered less frequently in the later stages of training. However, a large number of effective visual updates occur during early training, so the use of modulation throughout the training process cannot be judged solely from the response length at the later stage.

In addition, the newly added two-seed matched training ablation shows that the two-seed mean MMAD Avg. of Learned VM is 1.09 and 1.52 percentage points higher than Reward-Random and Norm-matched Uniform, respectively. Learned VM also outperforms both controls under each independent seed. This further supports the training gains of process-reward-conditioned visual modulation.

In summary, the new mask analysis provides supplementary evidence of a positive spatial association between reward-conditioned modulation weights and actual defect regions; the training logs clarify the actual use of VM at different training stages; and the training ablation further supports the learning gains from process-reward-driven visual-state updates. Together, these results address the reviewer's questions about spatial relevance, practical effects, and activation frequency, while clarifying that RILO does not rely on an explicit spatial-correspondence assumption.

AuthorsOpenReview ↗

Rebuttal to Reviewer HySr (3/4)

Q3. On the Attribution of Gains to Training-Time Visual Modulation

A3: We thank the reviewer for the in-depth analysis of the sources of RILO's performance gains. We would like to further clarify that RILO's core contribution is to reuse implicit process rewards not only for policy optimization but also to guide visual-state updates within an RL rollout, thereby improving the final Actor. Table 3 of the original paper already distinguishes this contribution through a stage-wise ablation: even with the inference-time Controller disabled, training-time VM increases the MMAD Avg. of RILO-4B from 79.86% to 81.10% (+1.24 pp). To address the reviewer's key question of whether this gain comes from effective process-reward guidance or merely from the regularization effect of ordinary visual perturbations, we added direct training-time controls.

Specifically, under a unified training and evaluation protocol, we conducted a matched two-seed RL training ablation for RILO-4B, using two independent training seeds to compare Learned VM, Reward-Random VM, and Norm-matched Uniform VM. Reward-Random samples modulation strengths and visual-group weights from a fixed calibration distribution derived from Learned VM. It does not depend on the current sample's PRM feedback, and its perturbation magnitude is calibrated according to the actual number of visual tokens. Norm-matched Uniform retains the current sample's global PRM modulation signal and matches the perturbation strength using the visual-token-weighted RMS, but sets the weights of all visual groups to 1. These two controls separately examine whether ordinary random perturbations can produce the same gains and whether the non-uniform weight allocation formed by process rewards has an additional effect.

As shown in Table 1 in our response to Q1, Learned VM achieves the highest accuracy under both independent seeds, with a two-seed mean MMAD Avg. of 80.97%. This is 1.09 and 1.52 percentage points higher than Random and Uniform, respectively. Furthermore, the per-seed gains of Learned over Random are +1.16 and +1.02 pp, while those over Uniform are +1.45 and +1.59 pp, showing consistent performance advantages. These results provide direct evidence for attributing RILO's training gains: under matched training conditions, non-uniform visual modulation conditioned on process rewards is more beneficial to final policy learning than ordinary random perturbations and uniform offsets of the same strength. This further supports RILO's core design of extending implicit process rewards from policy-optimization signals to visual feedback within a rollout, allowing process rewards not only to participate in training credit assignment but also to improve Actor learning through visual-state updates.

Regarding the inference cost raised by the reviewer, we agree that the online Controller increases computational cost, which has been quantified in Appendix K. The main performance gains of RILO are already retained in the trained Actor, and the inference-time Controller is an optional mechanism for further improving accuracy rather than a requirement for deploying RILO. According to Table 3 of the original paper, the 4B model retains approximately 89% of its overall performance gain with the inference-time Controller disabled. Therefore, for latency-sensitive deployment, the trained Controller-Off Actor can be used directly, retaining the main performance gains while avoiding additional online control overhead.

In summary, the newly added training-time controls show that the gains from RILO are not reproduced by the tested generic visual perturbations; rather, they depend on how process rewards guide visual-state updates. This clarifies the core contribution we aim to emphasize: using closed-loop process feedback during training to improve the Actor's visual reasoning capability, without requiring continuous online visual modulation at inference time.

Reviewer_u3vpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The proposes RILO, a framework for industrial anomaly understanding that uses implicit process rewards to guide visual reasoning. Specifically, it learns token-level rewards from answer correctness, evidence compatibility, and output consistency. These rewards are then used to adapt visual token representations during generation, allowing the model to adjust its visual information for subsequent reasoning. Experiments on MMAD demonstrate competitive performance under both zero-shot and one-shot settings.

Critical strengths
  1. The idea of feeding implicit process rewards back into visual reasoning is interesting. It extends the use of process rewards beyond policy optimization to dynamic visual feature adaptation.
  2. The paper provides extensive experiments, including comparisons with existing methods, component ablations, and reward validation analyses. The empirical results on MMAD are competitive.
Critical weaknesses
  1. Insufficient justification for visual modulation. This appears to be the main novelty, but the underlying mechanism remains unclear. Can the authors provide stronger evidence that (1) token-level rewards reflect actual reasoning errors and (2) negative textual rewards correspond to relevant visual tokens? The current position-based mapping lacks clear semantic or spatial justification.

  2. Limited generalization evaluation. The method does not sufficiently explore the generalization potential of MLLMs. Can the authors provide cross-dataset or unseen-category results, as in AnomalyR1, rather than relying only on MMAD-related data?

  3. Unclear SFT data construction. How are the structured reasoning targets for SFT constructed? Are CoT annotations generated or manually provided, and how is their quality verified? More details are needed to assess reproducibility.

Questions for authors

See weaknesses

5 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer u3vp (1/5)

Q1. On the Rationale for the Visual-Modulation Mechanism

A1: We thank the reviewer for the in-depth questions about RILO's visual-modulation mechanism. We understand that the reviewer's concerns mainly involve two aspects: (1) whether token-level process rewards reflect actual reasoning errors; and (2) whether negative token-level rewards can meaningfully guide the modulation of visual-token groups, and whether position-based interpolation has a practical effect.For the first question, Appendix H of the original paper provides error-sensitivity and visual-condition counterfactual experiments. For the second, we further added a matched RL training ablation for RILO-4B and a spatial analysis based on ground-truth defect masks to examine the effectiveness of reward-conditioned visual modulation and its relationship with defect regions.

(1) Whether token-level rewards are sensitive to reasoning errors. In Appendices H.1 and H.2 of the original paper, we designed field-level error counterfactual experiments and visual-condition counterfactual experiments, respectively, to examine the sensitivity of implicit process rewards to reasoning errors and image–answer consistency. Specifically, H.1 compares reward changes in modified and unmodified fields across 2,800 pairs of counterfactual samples. H.2 fixes the answer text and replaces only the query image to examine whether the reward model can distinguish visual conditions.

Table 1. Existing process-reward validation results.

ExperimentMetricResult
Appendix H.1Modified-field reward drop0.598
Appendix H.1Unmodified-field reward drop0.095
Appendix H.1Field-detection AUROC0.838
Appendix H.2Pairwise discrimination accuracy70.50%

As shown in Table 1, the reward decrease in the modified fields is substantially larger than that in the unmodified fields, and the AUROC for field-error detection reaches 0.838. In the image-swap experiment, where the text remains unchanged, the reward model also achieves a pairwise discrimination accuracy of 70.5%. Together, these two experiments demonstrate the sensitivity of implicit process rewards to both field-level reasoning errors and visual-condition changes, supporting their use as feedback associated with the current reasoning state in RILO.

AuthorsOpenReview ↗

Rebuttal to Reviewer u3vp (2/5)

Q1. On the Rationale for the Visual-Modulation Mechanism (continued)

A1 (continued): (2) The connection between negative textual rewards and visual tokens, and the rationale for position-based interpolation.

We thank the reviewer for raising this question. We would like to further clarify RILO's design: RILO does not assume an explicit one-to-one correspondence between language tokens and image spatial positions. Linear interpolation converts the token-level reward profile from the current reasoning process into modulation coefficients for visual-token groups, allowing visual representations to be dynamically adjusted according to process feedback. Therefore, we focus on testing whether this reward-conditioned modulation can lead to effective policy-learning gains and whether the resulting modulation weights exhibit a spatial association with actual defect regions.

First, we added a two-seed RL training ablation for RILO-4B. Under a unified training and evaluation protocol, it directly compares three visual-modulation strategies: Learned VM, Reward-Random, and Norm-matched Uniform. All models are evaluated with the inference-time Controller disabled.

Table 2. Two-seed training-time VM ablation results (MMAD Avg., %).

Training VMSeed 1Seed 2MeanLearned Gain (pp)
Learned VM80.6881.2680.97—
Reward-Random79.5280.2479.881.09
Norm-matched Uniform79.2379.6779.451.52

Here, Random samples modulation strengths and visual-group weights from a fixed calibration distribution obtained from Learned VM. It does not depend on the current sample's PRM feedback, and its perturbation magnitude is calibrated according to the actual number of visual tokens. Uniform retains the current sample's global PRM modulation signal and matches the perturbation strength using the visual-token-weighted RMS, while removing the non-uniform allocation across visual groups.

As shown in Table 2, Learned VM achieves the highest accuracy under both independent seeds, with a two-seed mean MMAD Avg. of 80.97%, outperforming Random and Uniform by 1.09 and 1.52 percentage points, respectively. Learned VM consistently outperforms both controls across the two seeds, indicating that reward-conditioned non-uniform visual modulation produces better final policy performance than the tested random perturbation and uniform offset. These results provide direct training evidence supporting RILO's use of process feedback to guide visual-state updates.

AuthorsOpenReview ↗

Rebuttal to Reviewer u3vp (3/5)

Q1. On the Rationale for the Visual-Modulation Mechanism (continued)

A1 (continued): Second, to further analyze the relationship between this modulation strategy and actual visual defects, we added a spatial diagnostic experiment based on ground-truth masks.We extract the per-chunk visual-modulation weights recorded during actual RILO-4B RL training and map the ground-truth defect masks to the image regions corresponding to the actual visual tokens processed by Qwen3-VL. Specifically, we first identify the image region represented by each visual token using the same image-preprocessing geometry and the resulting visual-token layout. We then apply RILO's original 20-group index mapping to associate each visual token with its corresponding Learned modulation weight. Finally, we aggregate the modulation weights assigned to defect-region visual tokens and normalize their sum by the total modulation weight across all visual tokens, obtaining the proportion of modulation weight allocated to actual defect regions. The same spatial mapping and normalization procedure is applied to all modulation strategies. Position-only examines the effect of fixed positional allocation, while Cross-sample uses modulation weights from other samples to examine the additional information provided by the current sample's process rewards.

Table 3. Differences in normalized modulation-weight allocation to ground-truth defect regions between Learned and the two controls (percentage points).

Analysis SetQuery ImagesLearned − Position-only (pp)Learned − Cross-sample (pp)
Four defect-related QA types550+0.196+0.199
Defect Localization119+0.059+0.068

As shown in Table 3, Learned exhibits positive differences in weight allocation to actual defect regions relative to both Position-only and Cross-sample, indicating that reward-conditioned visual-group weights exhibit a degree of sample-dependent spatial allocation. This is consistent with RILO's design of using process feedback for dynamic visual-feature modulation. We note that the ground-truth masks are used only for offline spatial diagnosis and label construction for the external test set; they are not used to generate RILO's visual-modulation weights or provided as model inputs during training or inference.

In addition, Appendix J (Table 14) of the original paper further examines the effect of visual-weight allocation on prediction behavior. On selected difficult diagnostic samples, Learned achieves a +3.67-percentage-point accuracy advantage over the Non-defect control, which allocates modulation weights to non-defect regions.

In summary, combining the process-reward validation in Appendix H, the visual-modulation intervention in Appendix J, and the newly added matched RL training ablation and ground-truth-mask spatial analysis, we can more fully address the reviewer's two questions. Implicit process rewards provide feedback related to reasoning errors and visual conditions. Visual modulation constructed from this feedback performs better than random and uniform modulation under matched training conditions. The newly added mask analysis further observes a positive spatial association between Learned modulation weights and actual defect regions. We hope these additional results clarify the rationale for dynamically adjusting visual states based on process rewards in RILO and the practical value of this mechanism for subsequent policy learning.

Reviewer_zdaoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  1. RILO combines implicit process rewards with dynamic visual modulation, allowing reward signals to guide both policy learning and visual adaptation during generation.
  2. The paper provides relatively thorough ablations, multi-seed evaluations, and useful image-swap counterfactual experiments.
  3. The paper is generally well organized, and the methodology and experimental results are clearly presented.
Critical strengths
  1. RILO combines implicit process rewards with dynamic visual modulation, allowing reward signals to guide both policy learning and visual adaptation during generation.
  2. The paper provides relatively thorough ablations, multi-seed evaluations, and useful image-swap counterfactual experiments.
  3. The paper is generally well organized, and the methodology and experimental results are clearly presented.
Critical weaknesses
  1. Section 3.4 requires rescoring the generated prefix under the current visual state. However, the supplementary code's build_prefix_records() does not construct the previous_visual_state field, as noted in the README. This raises questions about whether the reward model receives the complete visual-state information required by the proposed formulation in this evaluation path.
  2. RILO linearly interpolates language-token risk scores onto ordered visual-token groups. However, the paper does not establish a clear alignment between language-level errors and spatial image regions. It remains unclear whether this modulation genuinely redirects attention toward relevant defect regions.
  3. The incremental improvement from inference-time control is small (only +0.16 points for RILO-4B), while Appendix K reports that full control increases the 3B batched evaluation time from 260.2 to 490.4 seconds (1.88×). This raises concerns about the practical accuracy-efficiency trade-off.
Questions for authors
  1. Could the authors clarify the discrepancy between the visual-state rescoring protocol in Section 3.4 and the supplementary evaluation implementation?
  2. What evidence supports the spatial relevance of the interpolated language-token reward profiles?
  3. Given the limited inference-time gains, what are the practical advantages of enabling the full controller rather than using the controller-off model?
3 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer zdao (1/3)

Q1. On Visual-State Consistency in the Evaluation Code

A1: We thank the reviewer for carefully examining our code and pointing out the missing propagation of the previous_visual_state field in the submitted evaluation script. Upon investigation, we confirmed that the complete Strict Visual-State Replay evaluation implementation was inadvertently omitted when the supplementary materials were uploaded. This omission does not mean that the RILO online training mechanism itself lacks the corresponding state handling.

In the complete implementation used for the paper's experiments, whenever a new chunk boundary is reached, the process reward model re-scores the full current text prefix using the visual state actually used for generation in the preceding round, consistent with the online training protocol described in Appendix E. In addition, each visual modulation is reconstructed from the original visual features to avoid unintended cumulative shifts across successive modulation steps.

We have now restored the complete evaluation implementation and verified the correctness of visual-state handling through three rounds of integration testing. These tests cover multi-chunk state replay, synchronized handling of the main visual features and DeepStack, and the correspondence between samples and states when the batch dynamically shrinks. The strict evaluation mode also checks whether the specified visual state has been correctly replayed, preventing evaluation from silently continuing with an incorrect visual state.

To further improve reproducibility, the updated supplementary materials include the complete training and evaluation code, corresponding configurations, and reproduction instructions. We also provide the two-seed RILO-4B models reproduced through training with the original implementation, together with the retained original RILO-3B weights, so that reviewers can directly evaluate the models and verify the results.

AuthorsOpenReview ↗

Rebuttal to Reviewer zdao (2/3)

Q2. On the Spatial Relevance of the Interpolated Reward Profiles

A2: We thank the reviewer for raising this question. We would like to clarify that RILO does not assume an explicit one-to-one correspondence between language-token positions and image spatial coordinates. Linear interpolation maps the reward profile of the current reasoning process onto modulation coefficients for visual-token groups, allowing visual representations to be dynamically adjusted based on process-level feedback.

To address whether this modulation exhibits spatial relevance, we added a spatial analysis based on ground-truth defect masks. Specifically, we examine whether the Learned visual modulation weights generated by the current process rewards exhibit a positive allocation tendency toward actual defect regions, without assuming an explicit language-to-image spatial correspondence. We extract the per-chunk visual modulation weights recorded during actual RILO-4B RL training and combine them with the ground-truth defect masks to calculate the normalized proportion of weights assigned to defect regions. We compare this proportion against two controls: Position-only, which uses a fixed positional weight allocation, and Cross-sample, which uses modulation weights from other samples. These controls allow us to examine whether the observed spatial allocation bias is associated with reward feedback from the current sample. Importantly, the ground-truth masks are used only for offline spatial analysis and do not participate in RILO's training, inference, or modulation-weight computation.

Table 1. Advantage of Learned over the control methods in modulation-weight allocation to actual defect regions (percentage points).

Analysis SetQuery ImagesLearned − Position-only (pp)Learned − Cross-sample (pp)
Four defect-related QA550+0.196+0.199
Defect Localization119+0.059+0.068

Across 550 query images, Learned modulation yields a higher normalized weight proportion assigned to actual defect regions, with differences of +0.196 and +0.199 percentage points over Position-only and Cross-sample, respectively. The point estimates for the defect-localization subset are also positive. This indicates that Learned modulation weights generated from the current sample's process rewards exhibit a higher tendency to allocate weights to actual defect regions, providing additional empirical support for a spatial association between reward-conditioned visual modulation and defect-related regions.

This finding complements the prediction-behavior intervention experiment in Appendix J (Table 14) of the original paper. In that experiment, on selected diagnostic samples, Learned achieved a +3.67-percentage-point accuracy advantage over the Non-defect allocation, further indicating that visual-weight allocation can affect the model's subsequent predictions.

Together, the mask-based spatial analysis and the existing visual-modulation intervention results provide complementary evidence: the learned modulation exhibits a positive allocation tendency toward actual defect regions, while the intervention experiment demonstrates that different visual-weight allocation strategies can affect subsequent predictions. These findings provide additional support for the spatial relevance and practical effect of RILO's reward-conditioned visual modulation.

AuthorsOpenReview ↗

Rebuttal to Reviewer zdao (3/3)

Q3. On the Computational Cost and Practical Deployment of the Inference-Time Controller

A3: We acknowledge the reviewer's concern about the computational cost of practical deployment. This is also an important consideration in the design and evaluation of RILO. We would like to emphasize that RILO does not rely on the online Controller at inference time to obtain its main performance gains.

As shown in Table 3 of the paper, RILO-4B achieves an average MMAD accuracy of 81.10% even with the inference-time Controller disabled, improving by 1.24 percentage points over iPRM Actor-only (79.86%), which is trained without visual modulation. With the Controller enabled, the accuracy further increases to 81.26%. This indicates that RILO retains approximately 89% of its overall performance gain even when no additional online visual modulation is performed during deployment.

For inference-time computational cost, we have reported absolute wall-clock measurements in Appendix K (Table 15).

Table 2. Computational cost at inference time.

ConfigurationTotal Time (s)Throughput (QA/s)Time per QA (s)Relative Time
Controller-Off260.22.6900.3721.00×
Controller-On490.41.4270.7011.88×
Note: Accuracy results are from RILO-4B, whereas inference-time measurements are from RILO-3B (Appendix K).

On a single A800, RILO-3B achieves throughputs of 2.690 and 1.427 QA/s under Controller-Off and Controller-On, respectively. The corresponding amortized wall-clock time per question increases from 0.372 to 0.701 seconds, representing an additional cost of approximately 0.329 seconds per question under batched evaluation. Further cost decomposition shows that the main overhead comes from repeated multimodal prefill rather than the visual-modulation operation itself.

Therefore, RILO supports flexible inference configurations according to deployment requirements. For latency-sensitive scenarios, the Controller-Off Actor can be deployed directly, retaining the main performance gains while avoiding additional online control overhead. For scenarios with sufficient computational resources that prioritize further accuracy improvements, the Controller can be enabled. In essence, the main performance gains of RILO are already retained in the trained Actor, while the inference-time Controller serves as an optional performance-enhancement mechanism rather than a requirement for deploying RILO.

Reviewer_8S7WOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper targets industrial anomaly question answering (MMAD) and proposes RILO. The premise is that anomaly reasoning requires visual evidence at different granularities as the response unfolds — first judging whether an anomaly is present, then its type, location, and appearance — whereas mainstream MLLMs encode the image only once and keep visual-token representations fixed throughout decoding. This creates a mismatch between evolving evidence demands and a static visual representation.

The method has three components. First, the model is required to emit a four-field structured response, <locate><describe><compare><answer>, and a deterministic verifier produces a response-level soft target s_proc = v_fmt·(0.6·answer correctness + 0.2·evidence token overlap + 0.2·template consistency). Second, following PRIME's implicit PRM formulation, token-level rewards are defined as q_t = log π_φ(y_t|c_t) − log π_SFT(y_t|c_t) and the reward branch is trained with BCE(σ(β Σ q_t), s_proc); the distinguishing feature is that the context c_t is bound to the visual state under which that token was generated. Third — and this is the paper's main addition — visual modulation (VM): at every C=64-token boundary during generation, the reward branch and the frozen SFT reference rescore the accumulated prefix to yield a global magnitude P and G=20 group weights w, which modify the query image's visual-token representation via h^(k+1) = h^(0) + α·P·w (α=0.01, adding the same scalar across all feature dimensions of every token in a group). The KV cache is then invalidated, the multimodal prefix is re-prefilled, and generation continues, forming a generate–evaluate–revise–continue loop. Policy optimization uses a PPO-clip objective with RLOO advantages (no critic), combining three reward sources: token-level implicit rewards, the terminal verifier reward, and KL toward the frozen SFT policy. Because different segments of a single response are generated under different visual states, computing the importance ratio requires replaying each chunk under its own visual state.

The implementation builds on Qwen2.5-VL-3B and Qwen3-VL-4B, with format SFT followed by online RL (1,050 QA instances each, LoRA rank 8, 2×A800). Results: 81.26 on MMAD one-shot and 79.52 zero-shot, both the best reported, with the 4B model surpassing several 7B/8B/72B task-specific methods. Gains concentrate on defect-oriented tasks (+16.83 defect mean at 3B; +13.89 zero-shot defect mean at 4B). Appendices A–N additionally provide leakage control, verifier details, replay parity checks, multi-seed ablations, reward calibration, controller sensitivity, cost decomposition, and training-stability analyses.

Critical strengths

S1. The problem is correctly identified and the research direction is valuable. The large gap between object-oriented and defect-oriented tasks on MMAD does indicate that the bottleneck lies in fine-grained evidence acquisition rather than object recognition, and the observation that visual representations remain fixed throughout decoding points to a genuine architectural constraint in current MLLMs. More valuable still, the paper moves implicit process rewards beyond their usual role as post-rollout training returns and feeds them back to the visual input during rollout. Even though the concrete realization is open to question (see W1), the problem formulation itself is meaningful, and to our knowledge no comparable attempt exists in industrial anomaly reasoning.

S2. The ablation design is sensibly structured. Separating attribution between reward learning (answer reward → process target → iPRM) and visual modulation (a training × inference 2×2 decomposition), and equipping the inference-time controller with Identity / Uniform / Shuffled / Random controls to distinguish "the learned modulation" from "repeated re-prefill per se", is the right direction and more careful than most comparable work — although the coverage of these controls still has a gap (see W2).

Critical weaknesses

W1. A clear disconnect between the stated motivation and the actual mechanism, and the most closely related line of work is bypassed

The paper repeatedly describes the method as "redirecting visual inspection toward evidence relevant to the current reasoning state" (abstract, introduction, Section 3.4) — a spatial claim. Yet from the implementation in Appendix D, where the modulation lands is determined entirely by textual position: the risk trajectory is defined over positions of already-generated tokens and linearly interpolated in token order to G values (Eq. 26), while visual tokens are partitioned into G contiguous groups by index alone (Eq. 27), and the risk of the g-th text segment is added to the g-th visual-token group (Eq. 28). In other words, because the describe field is written in the middle of the response, the tokens that get modified are those with middling indices — which has no relation to where the defect actually is. The paper never justifies why a proportional position in the text should correspond to a proportional position among visual tokens. Appendix J is moreover inconsistent with the main text on this point: it states that the profile is defined over "ordered visual-token groups rather than supervised image regions" and explicitly notes that "this experiment is not a spatial-grounding evaluation."

More importantly, the motivation the paper invokes — noticing mid-response that the evidence is insufficient and looking again more closely — is most directly realized by the line of visual-CoT work based on explicit cropping, zooming, and tool invocation. Yet Section 2.2 and the mechanism comparison in Appendix M discuss only four latent-update methods (Mirage, H-GIVR, SIEVE, Reason-IAD) and make no mention of or comparison with that line at all, for instance:

  • DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning (ICLR 2026) — uses RL to learn directly when to invoke crop-and-zoom, without relying on spatial annotation;
  • Zooming without Zooming: Region-to-Image Distillation for Fine-Grained Multimodal Perception (ICML 2026) — obtains fine-grained perception via region-to-image distillation, without actual magnification at inference time.

Both were formally published well before this submission deadline and are not concurrent work; they should be discussed. The second is especially germane to the paper's design trade-off: the question it addresses is precisely whether one can obtain the benefit of magnification without performing it explicitly, which is the implicit premise behind this paper's choice of a latent-space bias. Note also that SIEVE, which the paper does cite, opens its abstract by observing that existing methods rely on external image operations such as zooming or cropping — indicating that the authors are not unfamiliar with this line. Given that the stated motivation points directly at such methods while this paper instead applies a scalar offset on the order of 10⁻² in latent space, the absence of any argument or empirical comparison for this trade-off is the paper's principal shortcoming.

W2. Most of the gain comes from existing methods; the paper's own mechanism contributes marginally, and its cost is not fully reported

Decomposing the improvement stage by stage from Table 3 (3B):

StageAvgΔAttribution
F-SFT67.49——
answer-only RL68.00+0.51standard RL
terminal reward replaced by s_proc70.62+2.62custom verifier
switch to implicit PRM73.32+2.70PRIME (Cui et al.), existing method
add training-time VM75.36+2.04this paper's mechanism
add inference-time control76.09+0.73this paper's mechanism

Of the total +8.60 from 67.49 to 76.09, +5.83 comes from "a custom reward function plus an off-the-shelf PRIME", while the closed-loop visual modulation that this paper actually contributes accounts for +2.77. The contrast is sharper at 4B: iPRM, actor-only already reaches 79.86, a configuration that by itself surpasses every baseline in Table 1 except DCR2-AD (+DK)-7B, with VM adding a further +1.40 to reach 81.26. That is, the claim that "RILO achieves SOTA on MMAD" rests primarily not on the paper's core contribution but on the combination of the verifier design and PRIME; the new mechanism is a modest increment on top of an already strong configuration. The main text does not draw this distinction.

On computational cost, only the inference side is reported (Table 15, 700 examples, one A800):

ModeTimeRel.Peak memory
Off260.2 s1.00×47,793 MiB
Identity (repeated re-prefill only)457.5 s1.76×47,793 MiB
Full490.4 s1.88×65,637 MiB

That is, 88% slower inference and 17.8 GB more peak memory, for +0.16 Avg at 4B. The authors' candor here is commendable (Appendix K explicitly declines to claim low overhead and recommends disabling the controller when latency matters), but the training-side cost is never reported: every chunk of every rollout requires a re-prefill, replay requires segmented forward passes per chunk, and both π_φ and π_SFT must each be run. Given that training-time VM is precisely the paper's main contribution (+1.24 at 4B, +2.04 at 3B), the absence of a wall-clock and memory comparison for RL with and without VM leaves the reader unable to judge the cost–benefit.

More critically, there is a gap in the controls. The four controls in Table 9 (Identity / Uniform / Shuffled / Random) are all at inference time and all share the same VM-trained actor, whereas the dominant gain (+2.04 at 3B) comes from training. The existing experiments therefore cannot distinguish "the gain comes from the learned modulation" from "the gain comes from a regularization effect of small perturbations to the visual input during training." Appendix L makes the latter explanation all the more necessary to rule out: late in 4B training, response length contracts to roughly 41 tokens and rarely crosses the C=64 chunk boundary, so the VM activation rate drops markedly — yet +1.24 is still recorded under conditions where VM is largely inactive. This combination requires explanation.

W3. The multi-seed results exhibit a regularity hard to attribute to randomness, and contradict the main text in one place

The multi-seed reporting in Appendix G exhibits a systematic pattern: across all 18 configurations and paired quantities in Tables 7–10, the three seed values satisfy seed1 < seed2 < seed3 without a single exception.

  • Table 7: F-SFT (67.18 / 67.53 / 67.76), answer-only RL, process-terminal RL, iPRM actor-only — all 4 monotone;
  • Table 8: all 4 configurations of the VM × inference-control 2×2 — monotone;
  • Table 9: all 6 control arms (Off / Identity / Uniform / Shuffled / Random / Learned) — monotone;
  • Table 10: all 4 paired task-level gains (AD / Def. Avg / Obj. Cls. / Obj. Anal.) — monotone.

Some correlation across configurations under a shared seed is expected (identical data ordering), but this ordering holds just as strictly for the Shuffled, Random, and Uniform arms, which are designed to destroy the modulation structure, and several paired differences have standard deviations as low as 0.02 (Table 8, "VM effect, control on": +2.53 ± 0.02; Table 9, "Learned − Identity": +0.66 ± 0.02). For LoRA-based online RL training, this degree of stability exceeds our expectations.

There is in addition a direct contradiction: Table 3 reports ±σ for every 4B configuration (79.86 ± 0.27, 79.91 ± 0.29, 81.10 ± 0.31, 81.26 ± 0.25), yet Appendix G opens by stating that the multi-seed runs span "every matched 3B configuration," and Appendix B.2 describes the 4B run as a single strict execution with 3407 as the request-local rollout seed base. The provenance of the 4B standard deviations is therefore unclear. Likewise, the three weight settings in Appendix C.4 (76.05 / 76.09 / 76.08) are not stated to be retrained runs.

Two presentation issues also merit mention. We appreciate that such slips are hard to avoid entirely under submission pressure and they do not affect our technical assessment, but they do add to the cost of reading and cross-checking: in Section 4.1, "with particularly clear gains on MVTec-LOCO and GoodsAD 3)" has a truncated figure reference (presumably Figure 3); and two rows in Table 1, Qwen2.5-VL-3B and AnomalyR1-3B, are duplicated verbatim. Separately, nearly every appendix section closes with a highly uniform self-limiting paragraph — for example, Appendix J's "We therefore use this analysis only to characterize the behavioral sensitivity of the frozen system to profile allocation, rather than to claim that the learned profile recovers the true defect region." Such caution is in itself appropriate, but recurring in near-identical phrasing across a dozen sections, it reads more as pre-emptive deflection of anticipated objections than as a conclusion arising naturally from each analysis.

Taking these observations together, and considering that the paper's AI use statement already covers "preliminary mathematical reasoning and methodological discussion" as well as "reviewing experimental design and ablation settings for rigor and potential confounding factors," we are unable to rule out that some of the audits and sensitivity analyses in Appendices G–L were written up from pre-existing conclusions rather than all corresponding to actually executed repeated training runs and measurements. We would welcome the authors' explanation and clarification on this point.

Questions for authors

Q1 (re W1). Please justify the index alignment from "textual token position" to "visual-token group index" in Eqs. 26–27. Given that the g-th segment of the risk profile invariably acts on the g-th visual-token group, that this group's image extent varies with input resolution, and that the textual positions mapping to a given group keep drifting across successive chunks, how can this mapping support the main text's characterization of "redirecting visual inspection toward relevant evidence"? Did you consider using the attention from low-scoring tokens to visual tokens to determine the modulated region, and comparing that against the current index alignment?

Q2 (re W1). Please include the explicit crop-and-zoom line in the related work (at minimum DeepEyes and Zooming without Zooming, above) and explain the advantage of a scalar offset on the order of 10⁻² in latent space over explicitly re-encoding a magnified region. If a full reproduction is infeasible under your compute budget, could you at least report one such method on MMAD, or explain concretely why it would fail in this setting of small defects without spatial annotation?

Q3 (re W2). On the cost and attribution of VM, please add: (a) a wall-clock and peak-memory comparison of the online RL stage with and without VM; (b) training-time controls using Random / Uniform / Identity offsets — the four existing controls in Table 9 are all at inference time and cannot separate "the learned modulation" from "a regularization effect of small training-time perturbations," which is precisely where the dominant gain (+2.04 at 3B) originates; and (c) an explanation of where the +1.24 training-time VM gain comes from, given that Appendix L shows 4B responses contracting to roughly 41 tokens and rarely crossing the C=64 boundary, i.e. a markedly reduced VM activation rate.

Q4 (re W3). Please clarify: (a) what exactly differs between the three runs in Appendix G, whether they are mutually independent full retrainings, and why all 18 configurations in Tables 7–10 — including the Shuffled / Random / Uniform control arms — exhibit strictly monotone seed1 < seed2 < seed3 without exception; (b) where the ±σ values for the 4B configurations in Table 3 come from, given that Appendix G states the multi-seed runs cover only 3B configurations while Appendix B.2 describes 4B as a single strict execution; and (c) whether the three weight settings in Appendix C.4 (76.05 / 76.09 / 76.08) correspond to three independent retrainings.

9 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer 8S7W (1/8)

Q1. On the Purpose and Effectiveness of Index Mapping

A1: We thank the reviewer for raising this question about Index Mapping in visual modulation. We would first like to clarify the purpose of this mechanism to avoid misunderstandings about RILO's core contribution.

RILO does not assume a natural or explicit correspondence between the relative positions of text tokens and image spatial regions. Index Mapping converts the token-level process-reward profile of the currently generated prefix into a set of visual-feature modulation coefficients, allowing process-evaluation feedback to participate in updating the visual state used for subsequent generation within the current rollout. This index mapping is a deterministic feedback-allocation rule, rather than an explicit text-to-image spatial localization mechanism. Therefore, the effectiveness of RILO does not rely on these positional indices having predefined spatial semantics.

In response to the reviewer's suggestion of Attention-based Mapping, we further implemented modulation weights constructed using the attention from low-reward text tokens to visual tokens, and compared this approach with the original Index Mapping. The results are shown in Table 1.

Table 1. Final accuracy on 494 eligible QA questions (%).

Mapping StrategyCandidate QAQA Reaching the First Update BoundaryFinal Accuracy (%)
Index Mapping70049478.54
Risk-weighted Attention70049478.54

On the frozen RILO-4B Step-45 Actor, we fixed 700 candidate QA questions in advance, of which 494 naturally reached the first visual-update boundary. Starting from the same generation prefixes, the two mapping methods produced different subsequent continuations on some samples, while their final answer selections remained identical. Both achieved an accuracy of 78.54% on the 494 eligible questions. These results show that, although the mapping strategy can affect intermediate generation, the more complex Risk-weighted Attention Mapping did not provide additional final-answer accuracy gains over Index Mapping in this controlled comparison.

AuthorsOpenReview ↗

Rebuttal to Reviewer 8S7W (2/8)

Q1. On the Purpose and Effectiveness of Index Mapping (continued)

A1 (continued): RILO's primary benefit comes from training-time visual-state updates that improve the final Actor, rather than requiring individual inference-time modulation steps to change the final answer. To further verify whether the gains from training-time visual modulation result from effective process-reward guidance, we added a two-random-seed RL ablation under a unified training and evaluation protocol, comparing three training strategies: Learned VM, Reward-Random, and Norm-matched Uniform. All models in this training ablation are evaluated with the inference-time Controller disabled, so that the effects of the training-time visual-modulation strategies can be compared independently.

Random samples modulation strengths and visual-group weights from a pre-fixed Learned calibration distribution. It does not depend on the current sample's PRM feedback and matches the actual visual perturbation magnitude. Uniform retains the current sample's global PRM modulation signal and matches the perturbation strength using the visual-token-weighted RMS, but removes the non-uniform allocation across visual groups. These two controls examine whether ordinary random perturbations can produce the same gains and whether the non-uniform weight allocation formed by process rewards has an additional effect.

Table 2. Two-seed reproduction of the training-time VM ablation (MMAD Avg., %).

Training VMSeed 1Seed 2MeanLearned Gain (pp)
Learned VM80.6881.2680.97—
Reward-Random79.5280.2479.88+1.09
Norm-matched Uniform79.2379.6779.45+1.52

As shown in Table 2, Learned VM achieves the highest accuracy under both independent seeds, with a two-seed mean MMAD Avg. of 80.97%. This is 1.09 and 1.52 percentage points higher than Random (79.88%) and Uniform (79.45%), respectively. Specifically, the per-seed gains of Learned over Random are +1.16 and +1.02 pp, while those over Uniform are +1.45 and +1.59 pp, showing consistent performance advantages. These results further support RILO's training-time mechanism: non-uniform visual-state updates formed using process rewards are more beneficial to final policy learning than reward-independent random perturbations and uniform modulation of the same strength. They also indicate that RILO's training gains do not come only from introducing visual-feature perturbations; the reward-conditioned modulation strategy and its weight allocation also play an important role.

In addition, we added a spatial analysis based on ground-truth defect masks, as shown in Table 3. Specifically, we first map the ground-truth defect masks to the image regions represented by the visual tokens after Qwen3-VL's actual image preprocessing. We then apply RILO's original 20-group index mapping to assign the recorded per-chunk Learned modulation weights to the corresponding visual tokens. Finally, we calculate the normalized proportion of modulation weights assigned to visual tokens corresponding to actual defect regions. The same spatial mapping and aggregation procedure is applied to both Position-only and Cross-sample controls.

Table 3. Difference in modulation-weight proportion within actual defect regions between Learned and the two controls (percentage points).

Analysis SetQuery ImagesLearned − Position-only (pp)Learned − Cross-sample (pp)
Four defect-related QA types550+0.196+0.199
Defect Localization119+0.059+0.068

Across 550 query images, Learned achieves higher normalized weight proportions assigned to actual defect regions than Position-only and Cross-sample, with differences of +0.196 and +0.199 percentage points, respectively. This result provides supplementary evidence of a positive spatial association between the reward-conditioned visual modulation produced by Index Mapping and actual defect regions.

In summary, Index Mapping is a lightweight implementation through which RILO introduces process rewards into visual-state updates, rather than an assumption of explicit alignment between text and image spatial positions. The matched training ablation supports the learning benefits of reward-conditioned non-uniform visual modulation, while the mask-based analysis provides supplementary evidence of its positive spatial association with actual defect regions. Meanwhile, the additional attention-based comparison shows no final-accuracy advantage for the more complex mapping method under the current frozen-model intervention setting. We will further clarify the role of Index Mapping and its distinction from explicit spatial localization in the revised manuscript to avoid misunderstandings about the method design.

AuthorsOpenReview ↗

Rebuttal to Reviewer 8S7W (3/8)

Q2. On Visual Zooming and Region-to-Image Distillation Methods

A2: We thank the reviewer for pointing out that our discussion of explicit visual cropping and zooming methods in the related work was insufficient. We agree that DeepEyes (ICLR 2026) and Zooming without Zooming (ICML 2026) are closely related to the motivation of this paper, and we will add a discussion of these works in the revised manuscript.

Specifically, DeepEyes uses reinforcement learning to enable active visual perception, allowing the model to select and inspect relevant image regions through zoomed crops during inference. Zooming without Zooming adopts Region-to-Image Distillation, using micro-cropped regions to obtain region-grounded VQA supervision and distilling this supervision back to full-image inputs, enabling fine-grained perception without inference-time zooming. These methods provide effective approaches to fine-grained visual perception, but their mechanisms differ from RILO.

Unlike crop-based approaches, RILO directly modulates visual features extracted by the frozen vision encoder using process-reward feedback within an RL rollout. It does not require explicit crop-region prediction, additional spatial supervision during training, or repeated visual encoding of enlarged image crops. Instead, it updates precomputed visual features and reconstructs the multimodal prefix for subsequent generation. The associated prefix re-prefill overhead is quantified in Appendix K.

More specifically, RILO is not designed as a substitute for high-resolution image zooming. Rather, its objective is to investigate how implicit process-reward feedback can guide visual-state updates within an RL rollout, thereby improving the final learned policy. In RILO, α=0.01\alpha = 0.01 is the scaling coefficient for visual updates; the actual update magnitude is jointly determined by the reward strength PP of the current prefix and the group weight ww. This mechanism performs process-feedback-conditioned feature adjustment on existing visual representations, without explicitly predicting crop regions or re-encoding enlarged images.

Following the reviewer's suggestion, we further evaluated the publicly available ZwZ-4B model in the one-shot setting on all 19,080 QA questions in MMAD. Moreover, since RILO's task-specific post-training uses MMAD data, we also compared the two methods on a newly constructed independent external dataset based on MVTec 3D-AD. This dataset contains 250 test images and 355 QA questions covering 10 object categories that were not used in RILO training, with no overlap between the test images and RILO's training images. This allows us to examine whether the above performance advantage is limited to MMAD. We have updated the supplementary materials with the new dataset and detailed usage instructions so that reviewers can directly conduct the evaluation.

Table 4. Comparison between RILO-4B (Controller-Off) and ZwZ-4B.

Dataset / MetricZwZ-4B (%)RILO-4B (%)RILO Gain (pp)
MMAD Avg.71.2281.109.88
MMAD Defect Avg.66.4979.0412.55
MVTec3D-ExternalQA Overall62.2563.941.69

The results show that both ZwZ-4B and RILO-4B are based on the Qwen3-VL-4B model family. Under a unified evaluation protocol, ZwZ-4B achieves an MMAD Avg. of 71.22% and a Defect Avg. of 66.49%. In comparison, the RILO-trained 4B model achieves 81.10% (from Table 3 of the original paper) and 79.04%, respectively, under the Controller-Off inference configuration, yielding gains of +9.88 and +12.55 percentage points. On the newly constructed small-scale dataset, RILO-4B achieves an Overall Accuracy of 63.94%, higher than ZwZ-4B's 62.25% (+1.69 pp). This shows that RILO maintains competitive performance and obtains a higher accuracy point estimate on an independent dataset and under unseen-category conditions.

Taken together, RILO-4B achieves higher accuracy than ZwZ-4B on the full MMAD benchmark while remaining competitive on the independent external test set. Since the two methods use different post-training data and optimization objectives, these results provide a task-level comparison rather than a controlled attribution of gains to a specific visual-refinement mechanism.

We will add the discussion of DeepEyes and Zooming without Zooming to the related work, together with the full MMAD evaluation results of ZwZ-4B, in the revised manuscript. We will further clarify that RILO's main contribution is to use implicit process rewards during training to form visual-state feedback and improve the final policy.

Comments on the paper

1 comment
Area_Chair_xiveOpenReview ↗

Please engage with the authors rebuttal by Nov 18

Dear Reviewers,

The authors have posted their rebuttal. Please engage with it by November 18, 2026, noting which of your concerns are resolved or which remain. Even brief engagement helps greatly.

Best regards,
Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.