acceptodds
Under review as a conference paper at ICLR 2027

MENTOR-ISP: Stage-Wise Credit Mentoring for Single-Agent ISP Tuning

Abstract

An image signal processor (ISP) is a sequential pipeline in which upstream decisions affect downstream modules. Single-agent reinforcement learning efficiently optimizes ISP parameters jointly, but typically relies on a global reward that provides coarse credit across coupled controls. Serialized multi-agent approaches provide finer stage-wise supervision, but require repeated ISP execution at deployment. We present Mentor-ISP, a modular single-agent RL framework that decouples structured credit assignment and inter-stage dependency modeling from serialized ISP execution. Mentor-ISP factorizes the joint action into pipeline-ordered groups, uses sequential action handoffs to condition downstream decisions on upstream predictions, and applies stage-wise credit during training. At inference, all decisions are assembled into a single joint action, requiring only one ISP rendering per optimization step. Across two real-camera image-quality datasets, Mentor-ISP substantially improves peak signal-to-noise ratio (PSNR) over a matched single-agent baseline while approaching serialized multi-agent performance at near-single-agent inference cost. Ablations show that stage-wise credit provides the primary gain, with action handoff offering complementary benefit.

Primary area:
reinforcement learning
Keywords:
Image Signal Processing, Reinforcement Learning, ISP Parameter Tuning, Sequential Decision Making, Stage-Wise Credit Assignment, Efficient Inference

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_hnRhOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper's thesis is that the structure of decision-making, of credit assignment, and of environment interaction need not coincide. Sequential ISP evaluation can be used only to compute training-time credit, while inference stays a single joint action. This is implemented as an autoregressive, factorized PPO actor plus pipeline-ordered stage-wise rewards.

Critical strengths

The principle is clean and could transfer. Using expensive, serialized intermediate evaluations only as a training-time signal, then deploying a single-pass policy, applies to many modular pipelines (rendering, compilers, robotics stacks), not just ISPs. Stated in general terms, it would interest the ICLR audience.

Critical weaknesses

The increment over prior work is narrow. Most components come from Sun et al. (2026), SARL-ISP and MARL-ISP: the stage-wise intermediate-render reward, mixing it with a global reward via ω, the DPOM transition, and the horizon T = 4. Autoregressive action heads are standard (Metz et al. 2017; action branching, Tavakoli et al. 2018). The contribution therefore reduces to applying MARL-ISP's reward inside a factorized single agent. Related ideas the paper does not discuss include privileged or asymmetric training signals (asymmetric actor-critic, teacher–student distillation of a serialized MARL policy into a single pass) and reward decomposition. Distilling MARL into a single-pass student is an obvious baseline that is missing.

Image comparisons are too few for a CV work. Besides, from Fig 3 and 4 (which present only 2 cases, both the same cityscape scene), the improvement over existing methods is barely noticable. The results are not convincing.

Questions for authors
  1. Can you state precisely which components are new relative to RL-SeqISP, SARL-ISP, and MARL-ISP?
  2. Is the global critic used by the actor at all in Mentor-ISP? If not, what is it for?
Reviewer_urVYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Mentor-ISP, a modular single-agent reinforcement learning framework designed to optimize Image Signal Processing (ISP) pipelines. Employing an Actor-Critic architecture that combines a frozen ResNet-50 feature extractor with the PPO algorithm, the work decomposes ISP parameters into five stage-specific policy heads. Through an "action handover" mechanism, downstream decisions are conditioned on upstream predictions, utilizing stage-wise credit assignment as a structured learning signal. Experimental results demonstrate that Mentor-ISP has improvements for certain atasets, However, details regarding certain hyperparameter settings, specific dataset sources, and camera models are omitted, and hence, further verification is required to show the generalizability and reproducibility of the work.

Critical strengths

The paper shows improvement over single agent design, that is, with same light-weight scales as single agent design, but comparable performance as multiagent design

Critical weaknesses

The work focuses on a niche subject of ISP tuning, and the baseline comparisons are limited to only two works. Some presentations are not clear. Consider highlighting Table 3 first. Figure 3 and 4 does not present a clear difference between results, consider enlarging or use tools to show the differences.

Questions for authors

How does the proposed method compare with other baselines, such as the works listed at the end of paragraph 1 in the introduction section.

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

This paper proposes Mentor-ISP, which decomposes joint actions into a sequence of groups arranged in a pipeline. By passing actions sequentially, downstream decision-making can be adjusted based on the predictions from upstream stages. Additionally, a phased contribution assignment mechanism is incorporated during the training process. During inference, a unified joint action is generated to reduce the number of required ISP rendering passes.

Critical strengths

This paper decouples the parameter modules from the prediction process, achieving a level of predictive performance comparable to multi-agent models while maintaining an inference cost roughly equivalent to that of SARL.

Critical weaknesses
  1. The methodology in this paper offers limited innovation. Techniques rely heavily on the existing MARL-ISP approach, such as parameter grouping, sequence prediction, and staged rewards.

  2. In Table 1, comparisons are made only against SARL and MARL, without other mainstream methods of ISP parameter optimization.

  3. Although the authors claim to have improved inference efficiency compared to existing methods, their performance on image quality metrics for the S24 and S7 datasets has also declined. Does this imply that it is not possible to achieve both higher efficiency and better experimental results?

Questions for authors

Please refer to the weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.