acceptodds
Under review as a conference paper at ICLR 2027

OptClimber: Curriculum Experience Learning for LLM Optimization Modeling

Abstract

Large language models have shown strong potential for automated optimization modeling, yet existing methods mainly focus on how modeling knowledge is provided through prompting or fine-tuning, while paying less attention to how such knowledge should be accumulated throughout learning. We introduce OptClimber, a curriculum experience learning framework that organizes optimization problems by domain and difficulty and progressively builds reusable modeling knowledge from solver-verified problem-solving trajectories. OptClimber couples curriculum scheduling with experience accumulation, allowing reliable modeling principles to be acquired and refined as problem complexity increases. The accumulated experiences are further used to construct solver-verified trajectories for supervised fine-tuning, while the same curriculum structure organizes subsequent reinforcement learning with solver-grounded feedback. In this way, OptClimber connects data organization, experience accumulation, supervised learning, and reinforcement learning within a unified framework. Experiments on six optimization modeling benchmarks show that the prompt-based OptClimber improves its backbone by 18.3% on average, with larger gains on challenging tasks. After fine-tuning and curriculum-guided reinforcement learning, OptClimber-14B achieves 72.7% average accuracy and surpasses the current state-of-the-art fine-tuned model.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Optimization Modeling

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_xekeOpenReview ↗
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
  1. The intuition is reasonable, which aligns with the curriculum learning observations.
  2. Extensive experiments are conducted across different benchmarks and model scales.
Critical weaknesses
  1. 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.
  2. The methodological novelty is incremental, primarily combining existing paradigms (experience libraries, context distillation for SFT, and solver-grounded GRPO) into a unified pipeline.
  3. 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.
  4. 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.
  5. 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
  1. What are the exact domains in the predefined domain set (Appendix E.1), and what is the performance without domain partitioning (w/o Domain)?
  2. How does Experience-Guided SFT compare against vanilla Rejection Sampling SFT (teacher generating solver-verified trajectories without retrieved experiences)?
  3. Why does RL degrade OptMATH performance by ~18% on 4B/8B models in Table 13 (or were the OptiBench and OptMATH columns swapped)?
  4. 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?
Reviewer_KzWUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The primary contribution of OptClimber is the integration of difficulty-aware curriculum scheduling with accumulated modeling experience and subsequent model training. Rather than treating problem-solving trajectories as independent examples, the framework organizes them into a reusable knowledge structure that can inform future problems. This is a meaningful extension of experience-based LLM optimization because it considers both the acquisition and deployment of modeling knowledge. The use of solver verification also provides a stronger feedback mechanism than unconstrained text-based reflection. However, the core ingredients, curriculum learning, experience memory, supervised fine-tuning, and reinforcement learning with verifiable rewards, are individually well established. The contribution therefore depends on demonstrating that their integration produces effects beyond the sum of their components. The paper does not sufficiently establish that its curriculum organization is superior to strong adaptive data selection, competence-based sampling, or matched-data training strategies under equivalent computational budgets. The generalization of learned experiences across unseen optimization domains is also less convincing than improvements on benchmarks related to training.

Critical strengths

The paper connects problem ordering, experience accumulation, SFT, and RL through a shared curriculum, offering a practical framework for developing specialized optimization-modeling LLMs.

Critical weaknesses
  • The framework combines established techniques, and its distinct technical novelty over recent curriculum-learning and experience-memory approaches is not sufficiently isolated.
  • Comparisons involving different model families, training datasets, and optimization procedures do not fully isolate the contribution of curriculum-based experience accumulation. For example, OptClimber-Qwen3-14B is compared against SIRL-Qwen2.5-32B, but differences in model family and training conditions complicate interpretation.
  • Strong benchmark accuracy does not necessarily establish that accumulated experiences represent transferable optimization principles rather than recurring templates, memorized solutions, or benchmark-specific modeling patterns.
Questions for authors

The central technical weakness is that the effectiveness of OptClimber depends on how modeling difficulty is defined, how reusable experiences are extracted, and how experience quality is maintained during progressive accumulation. These operations may introduce noise or reinforce incorrect modeling patterns, particularly when the solver verifies execution outcomes without fully verifying the semantic equivalence between a generated formulation and the original natural-language problem. Additionally, the relationship between curriculum ordering and observed improvements may be confounded by problem selection, sample exposure frequency, training volume, and the quality of collected reasoning trajectories. Although the paper includes comparisons against random and hard-first curricula, stronger factorial ablations are required to establish the independent contributions of experience memory, difficulty adaptation, SFT, and RL. The absence of sufficiently detailed cost-normalized comparisons also limits conclusions about training efficiency. More convincing out-of-domain evaluations are needed to establish whether the curriculum produces broadly reusable modeling knowledge.

Here are the key questions to be addressed.

  • What is fundamentally new compared with AlphaOPT, OptiTree, experience-memory systems, competence-based curriculum learning, and existing curriculum-guided RL?
  • What are the performance gains from curriculum scheduling alone, experience accumulation alone, SFT alone, RL alone, and their full combination?
  • Can the authors provide matched-backbone, matched-training-data, matched-token, and matched-compute comparisons with competitive fine-tuning and curriculum baselines?
  • How is modeling difficulty determined, and how sensitive are results to inaccurate difficulty rankings or alternative difficulty metrics?
  • Can experiences learned from one optimization domain improve performance on entirely unseen domains without retrieving closely related training examples?
  • How does the system identify and remove incorrect, conflicting, redundant, or outdated experiences? Does solver verification guarantee semantic correctness?
  • What are the full costs of experience construction, retrieval, solver verification, supervised training, reinforcement learning, and inference relative to the observed improvements?
Reviewer_1T1AOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

OptClimber introduces a curriculum-based framework for training LLMs in optimization modeling. It organizes problems by domain and difficulty, progressively solves them from easy to hard, validates each formulation using a solver, and accumulates lessons from failed attempts. At inference time, relevant lessons are retrieved to guide problem solving. The same experience repository is further leveraged by a teacher model to generate solver-verified trajectories for supervised fine-tuning (SFT).

Critical strengths

1.The question is concrete: modeling knowledge should be accumulated by difficulty, not only injected by a prompt or a fine-tune. 2.OptiTree is compared under the same 200-problem OR-Instruct construction budget, and OptClimber is ahead on the harder sets.

Critical weaknesses

1.The fine-tuned “state of the art” claim compares Qwen3 models trained here with starred numbers from Llama-3 and Qwen2.5 papers. Those numbers were not rerun on the corrected targets or with the variable-type retry used for OptClimber. Qwen3-14B after SFT alone already averages 66.9, close to SIRL-Qwen2.5-32B at about 68, before the curriculum RL step. 2.Curriculum RL is not uniformly helpful. In Table 13, OptMATH falls from 59.7 to 40.4 at 4B and from 62.6 to 45.6 at 8B. The average still rises mainly because OptiBench jumps from 17.1 to 64.5 at 4B and from 21.3 to 65.6 at 8B. The paper does not discuss these drops. 3.Figure 1’s prompt curriculum scores 41.0 on IndustryOR and 31.9 on OptMATH after 200 OR-Instruct problems. The full system in Table 1 scores 61.0 and 55.4 under the same stated budget. The gap between the motivation setup and the full method is unexplained. 4.Kendall’s W near 0.91 shows that five experts agree with each other. Those labels are not used to build the curriculum, and agreement between the automatic rubric and the experts is not reported. 5. Miss Citation: OPT-BENCH: Evaluating LLM Agent on Large-Scale Search Spaces Optimization Problems. https://arxiv.org/abs/2506.10764

Questions for authors

See weakness.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.