Research map / Large language models / Agents, code and math
ORLoopBench: Solver-in-the-Loop Benchmarks for Self-Correction and Behavioral Rationality in Operations Research
Ruicheng Ao, David Simchi-Levi, Xinshang Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train Qwen3-8B with solver-verified GRPO on top of SFT/curriculum | Qwen3-8B · OR-DEBUG-BENCH LP repair | RR@5: 95.3% vs 92.4% (best API) | Improved |
| add SFT on teacher trajectories to base model (ablation) | Qwen3-8B · OR-DEBUG-BENCH pilot validation | RR@5: 51.2% -> 93.1% | Improved |
| add GRPO to SFT baseline (ablation) | Qwen3-8B SFT · OR-DEBUG-BENCH validation (200) | RR@5 / DA: RR@5 91.5% -> 92.0%; DA 68.0% -> 66.0% | Mixed |
| add curriculum stage before GRPO/SFT (ablation) | Qwen3-8B SFT · OR-DEBUG-BENCH validation (200) | RR@5: 91.5% -> 95.0% | Improved |
| add process reward model (PRM) supervision (ablation) | Qwen3-8B SFT · OR-DEBUG-BENCH validation (200) | DA: 68.0% -> 72.7% | Improved |
| combine curriculum with GRPO (ablation) | Qwen3-8B SFT · OR-DEBUG-BENCH validation (200) | RR@5 / DA: RR@5 91.5% -> 95.3%; DA 68.0% -> 62.4% | Mixed |
| remove faithfulness penalty from GRPO reward (ablation) | Qwen3-8B GRPO · OR-DEBUG-BENCH | diagnostic quality | Hurt |
| constraint-level repair instead of whole-model code regeneration | MILP repair · OR-DEBUG-BENCH MILP semantic drift | correct objective preserved: GPT-5.4 regeneration: 90% OPTIMAL but 28.2% correct objective | Improved |
| zero-shot transfer of LP-trained model to MILP, then add MILP-specific training | Qwen3-8B · OR-DEBUG-BENCH MILP repair | RR@5: 78.8% -> 87.1% | Improved |
| add three-stage curriculum (Extreme, Boundary, Full) for bias | Qwen3-8B · OR-BIAS-BENCH OOD | OOD bias: 20.0% -> 10.4% | Improved |
| prompting only (zero-shot CoT or three-shot ICL) instead of training (ablation) | Qwen3-8B · OR-DEBUG-BENCH | RR@5: 23.0% (CoT), 38.7% (3-shot ICL) | Hurt |
| train on OR formulation (ORLM/SIRL/LLMOPT) instead of repair | OR-domain models · OR-DEBUG-BENCH LP repair | RR@5: 71.0%/73.8%/78.4% vs 95.3% | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- LLMOPT — LLMOPT: Learning to Define and Solve General Optimization Problems from Scratch (ICLR 2025) · better on 1 of 1 shared result
- OptiMUS — OptiMUS: Scalable Optimization Modeling with (MI)LP Solvers and Large Language Models (ICML 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- PEARL: Solver-in-the-Loop Interactive Optimization Modeling from Natural Language (NeurIPS 2026 · similarity 0.85)
- OpenClawBench: Benchmarking Process-side Anomalies in Real-world Agent Execution Trajectories (NeurIPS 2026 · similarity 0.84)
- Extrapolative Weight Averaging Reveals Correctness–Efficiency Frontiers in Code RL (NeurIPS 2026 · similarity 0.84)
- MM-OptBench: A Solver-Grounded Benchmark for Multimodal Optimization Modeling (NeurIPS 2026 · similarity 0.84)
- Lean Refactor: Multi-Objective Controllable Proof Optimization via Agentic Strategy Search (NeurIPS 2026 · similarity 0.84)
- DiagnosticIQ: A Benchmark for LLM-Based Industrial Maintenance Action Recommendation from Symbolic Rules (NeurIPS 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
