- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper introduces OptClimber, a curriculum experience learning framework for LLM-based optimization modeling. A three-stage framework is designed, including (1) Curriculum Experience Accumulation, which samples problems from easy to hard and distills reusable experiences from solver-verified failures; (2) Experience-Guided SFT, which conditions a teacher LLM on the retrieved experiences to generate solver-verified trajectories and distills this knowledge into a student LLM (trained without experience inputs); and (3) Curriculum-Guided RL, which retains the easy-to-hard curriculum scheduling and optimizes the policy via GRPO with solver-grounded feedback. Experiments across six optimization modeling benchmarks demonstrate strong improvements in both prompt-based (DeepSeek-V3) and fine-tuned (Qwen3-4B/8B/14B) settings.
Critical strengths
- The intuition is reasonable, which aligns with the curriculum learning observations.
- Extensive experiments are conducted across different benchmarks and model scales.
Critical weaknesses
- The claimed "Domain" dimension is under-specified and unvalidated. The paper never lists the predefined domain set (Appendix E.1), ignores domains in the motivation study (Section 3), and lacks a w/o Domain ablation.
- The methodological novelty is incremental, primarily combining existing paradigms (experience libraries, context distillation for SFT, and solver-grounded GRPO) into a unified pipeline.
- Key ablations are missing, notably a vanilla Rejection Sampling SFT baseline (teacher generating solver-verified trajectories without retrieved experiences) to prove the necessity of experience-guided SFT.
- Table 13 shows an unexplained ~18% accuracy drop on OptMATH after RL for the 4B/8B models, and the evaluation uses up to 3 solver-feedback repair rounds (Appendix A.3), raising fairness concerns against single-turn baselines in Table 1.
- The presentation and writing need substantial improvement. (1) Section 3 (Motivation) is largely disconnected from the specific pipeline proposed, merely re-validating well-known observations that curriculum scheduling outperforms random or hard-first ordering; (2) key concepts (e.g., domain isolation, dynamic refinement, experience distillation, and RL reward shaping) are heavily coupled throughout the narrative, making it difficult to disentangle the true core contribution from composite engineering tricks.
Questions for authors
- What are the exact domains in the predefined domain set (Appendix E.1), and what is the performance without domain partitioning (w/o Domain)?
- How does Experience-Guided SFT compare against vanilla Rejection Sampling SFT (teacher generating solver-verified trajectories without retrieved experiences)?
- Why does RL degrade OptMATH performance by ~18% on 4B/8B models in Table 13 (or were the OptiBench and OptMATH columns swapped)?
- What is the single-turn (0 repair rounds) accuracy of OptClimber in Table 1, and did the cited baselines () also use 3 rounds of solver repair?