- 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
- Can you state precisely which components are new relative to RL-SeqISP, SARL-ISP, and MARL-ISP?
- Is the global critic used by the actor at all in Mentor-ISP? If not, what is it for?