acceptodds
Under review as a conference paper at ICLR 2027

BeTAR: Belief-Guided Trajectory Learning with Agentic Reinforcement Learning for Multi-Degradation Surgical Image Restoration

Abstract

Surgical images are frequently affected by multiple coupled degradations, including smoke, low illumination, specular highlights, blur, water droplets, and low resolution, making fixed restoration pipelines difficult to generalize across complex clinical scenes. To address this challenge, we propose BeTAR, a degradation-aware agentic framework that integrates structured degradation perception with belief-guided restoration planning for multi-degradation surgical image restoration. First, a vision-language model is trained to identify degradation types and severity levels and organize them into an interpretable belief state, providing structured environmental information for subsequent decision making. Second, conditioned on this belief state, we formulate restoration planning under a Partially Observable Markov Decision Process (POMDP) framework and optimize an Agentic-RL policy to generate complete open-loop restoration trajectories, allowing the agent to jointly determine tool selection, execution order, and termination without relying on a predefined restoration chain. The policy is further optimized using a reference-free trajectory-level reward that combines no-reference image quality assessment with tool-use constraints, enabling training without paired clean images. Experiments on both in-distribution and independent surgical image test sets show that BeTAR achieves an F1 score of 0.908 for degradation recognition and ranks first on five of six restoration metrics on both test sets.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Surgical Image Restoration, Agentic Reinforcement Learning, Vision-Language Models, Multi-Degradation Image Restoration, Tool-Augmented Agents, Trajectory Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 2 reviews, above 0% of the others.
Reviewer_YpsJOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors propose BeTAR, a degradationaware agentic framework that integrates structured degradation perception with belief-guided restoration planning for multi-degradation surgical image restoration. An VLM is first used to determine the degradation types and severity, and different restoration tools are used for the final recovery. BeTAR achieved SOTA performance on self-constructed datasets.

Critical strengths
  1. Agentic recovery for surgical image restoration in mininally invasive surgery is both meaningful and interesting.
  2. The performance improvement seems to be significant.
Critical weaknesses
  1. Figures 1, 2, and 3 require substantial improvements. Figure 1 is overly schematic and lacks specific quantitative values, clearly defined metrics, and sufficient experimental details to support its conclusions. Figures 2 and 3 also suffer from small text, unclear graphical elements, and poor visual organization, making the proposed framework difficult to understand. The authors should redesign these figures to improve both readability and technical clarity.

  2. Figures 2 and 3 lack sufficient technical details to support reproducibility. Figure 2 repeatedly uses generic <think> placeholders without clearly illustrating the five-stage reasoning process or the corresponding reward design. Figure 3 mainly presents a generic perception–planning–execution pipeline without adequately highlighting the technical innovations of Agentic-RL. The authors should clarify how the belief state is incorporated into the policy, how restoration trajectories are generated, and how trajectory-level rewards are used for optimization.

  3. The experimental evaluation is insufficient to demonstrate generalizability. Although the paper discusses public datasets and benchmarks such as DeSmoke-LAP and SurgClean, the experiments are primarily conducted on the authors' own data. While an independent test set is included, its characteristics and relationship to the training data are not sufficiently described. Evaluations on multiple established public benchmarks would provide stronger evidence of generalizability across different surgical scenarios and degradation types.

  4. Comparisons with existing agentic restoration methods are missing. The paper discusses RestoreAgent, AgenticIR, JarvisIR, and Q-Agent, but Table 1 mainly compares BeTAR against conventional restoration networks. Since the primary contribution lies in agentic tool scheduling rather than restoration backbone design, direct comparisons with representative agentic restoration methods are essential. Ideally, these methods should be evaluated using the same restoration tool library to ensure fairness.

  5. The manuscript requires further language polishing. The Introduction, Method Overview, and Conclusion repeatedly use phrases such as "tool selection, execution order, and termination" and "structured belief state." Similar conceptual descriptions appear throughout the paper without adding substantial technical information. The authors should reduce repetition, improve the overall conciseness, and devote more attention to explaining the methodological contributions.

  6. The necessity of the POMDP formulation is insufficiently justified. The framework generates complete restoration trajectories in an open-loop manner, without updating its belief state based on intermediate restoration results. This raises questions about how the proposed formulation differs from conventional belief-conditioned sequence generation. The authors should clarify the practical advantages of introducing a POMDP and provide more targeted comparisons to demonstrate its necessity.

Questions for authors

The current presentation gives the impression of an insufficiently polished submission. Several figures are overly generic and schematic, with limited technical information beyond what is already stated in the text. It would be much better if the authors provide more informative figures and improve the overall manuscript presentation and proofreading.

Reviewer_rKpqOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes BeTAR for multi-degradation surgical image restoration. A vision-language model is trained with supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO) to predict degradation types and severity levels. These predictions condition a policy that generates an open-loop sequence of fixed restoration tools. The trajectory policy is trained with reference-free image-quality and tool-use rewards. The paper reports degradation-perception comparisons, restoration results on two test sets, and component ablations.

Critical strengths

The paper proposes a possible approach to multi-degradation surgical image restoration by linking degradation perception with tool-based restoration planning. This is a worthwhile problem to investigate.

Critical weaknesses
  1. Concerns about the credibility of the experimental results: Table 1 (page 6) reports exactly the same values for five metrics for MambaIRv2 on the independent test set of 1,003 images and SFHformer on the in-distribution test set of 1,500 images. The repeated values are shown below. IS differs, so the duplication concerns five metrics rather than the entire row.
MetricMambaIRv2 - Independent test set, 1,003 imagesSFHformer - In-distribution test set, 1,500 images
MANIQA0.19960.1996
MUSIQ44.090844.0908
PI5.91345.9134
LIQE1.94621.9462
FADE0.52280.5228
IS2.14981.7578

Table 2 (page 8) reports Recall = 0.306 and F1 = 0.477 for Hulu-Med-7B. Under the standard definitions, with the same evaluation units and aggregation weights, precision cannot exceed 1 and the following bound holds, where R denotes recall.

F1≤2R1+R.F_1 \leq \frac{2R}{1+R}.

At R = 0.306, the largest possible F1 is approximately 0.4686. Even allowing for rounding to three decimal places, the upper bound is approximately 0.4692, below the minimum value of 0.4765 that could round to the reported F1. The manuscript does not specify an alternative aggregation convention that explains this discrepancy.

Together, the unexplained duplication across different models and test sets and the incompatible metric values raise serious concerns about possible fabrication of experimental results. I cannot regard the current evaluation as reliable without a satisfactory explanation supported by the original records.

  1. Limited novelty and insufficient evidence for the claimed contribution: The methodological novelty appears largely limited to applying established SFT and reinforcement-learning techniques to surgical images. Table 2 includes several 7B baselines, such as HuatuoGPT-Vision-7B, Lingshu-7B, and Hulu-Med-7B, without describing comparable task-specific fine-tuning for those baselines. It therefore appears to compare the adapted BeTAR model with off-the-shelf models, which does not isolate the benefit of the proposed method from the benefit of task-specific adaptation.

Table 3 mainly confirms the expected benefit of standard SFT and GRPO on the chosen base model. This does not establish a meaningful new contribution. Figure 1 provides schematic trends without numerical axis values, a defined relative-score metric, or uncertainty, and supplies no usable quantitative evidence for the claimed scheduling improvement.

Figure 3, Stage III(C) (page 5), contains an extra closing parenthesis in the numerator of the advantage formula. The denominator also ends with an unmatched closing parenthesis, corresponding to a missing opening parenthesis if grouping was intended.

Questions for authors
  1. Why do MambaIRv2 on the 1,003-image independent test set and SFHformer on the 1,500-image in-distribution test set have identical values for five metrics in Table 1? Please provide the original per-image scores and evaluation outputs for these entries.
  2. How were Recall and F1 computed and aggregated in Table 2? Please explain how Recall = 0.306 and F1 = 0.477 for Hulu-Med-7B can arise, and provide the metric implementation and sufficient underlying prediction/label statistics to reproduce them.
  3. Were the Table 2 baselines given comparable task-specific fine-tuning? Which comparison establishes an improvement beyond the benefit of adapting the base model with standard SFT and GRPO?
  4. What actual measurements underlie Figure 1? Please provide the numerical results and protocol for comparing learned and random execution orders while keeping the tools and invocation counts fixed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.