The researchers introduce ActReview, a novel post-training framework designed to transform LLM-generated peer reviews from simple diagnostics into actionable revision guidance. To address the limitations of current models that identify weaknesses without providing solutions, the authors decompose the task into two components: diagnostic claim generation and revision suggestion generation. They developed ActReview-40K, a specialized dataset constructed by aligning OpenReview reviewer critiques with corresponding author rebuttals to extract latent supervision for grounded revision plans. For the training methodology, they utilized multi-task supervised fine-tuning followed by Group Relative Policy Optimization (GRPO) on a Qwen3-8B-Base model, employing candidate-aware and weakness-specific rubric rewards to enhance feedback quality. To evaluate these advancements, the team established ActReview-Bench, a human-curated benchmark containing 1,000 instances to measure diagnostic accuracy and revision utility. Experimental results demonstrate that ActReview surpasses previous specialized models in both actionability and grounding, performing competitively against high-parameter prompt-based LLMs. While human evaluations confirm significant improvements in the usefulness of revision suggestions, the study notes a persisting gap regarding technical accuracy. Further analyses indicate that the framework demonstrates robust generalization to unseen papers and maintains consistency across various independent evaluators.