Research map / Reinforcement learning / Visual and representation RL
Contrastive Reinforcement Learning of Symbolic Reasoning Domains
Gabriel Poesia, WenXin Dong, Noah Goodman
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 |
|---|---|---|---|
| use InfoNCE contrastive loss instead of value regression (CVI) (ablation) | ConPoLe · CommonCore Equations | success rate (%): 89.0% -> 92.5% | Improved |
| use InfoNCE contrastive loss instead of value regression (CVI) (ablation) | ConPoLe · CommonCore Sorting | success rate (%): 77.0% -> 100.0% | Improved |
| use InfoNCE contrastive loss instead of value regression (CVI) (ablation) | ConPoLe · CommonCore Fractions | success rate (%): 84.5% -> 96.0% | Improved |
| use InfoNCE contrastive loss instead of value regression (CVI) (ablation) | ConPoLe · CommonCore Multiplication | success rate (%): 100.0% -> 99.5% | No clear effect |
| use only local negatives (ConPoLe-local) instead of all beam-search negatives (ablation) | ConPoLe · CommonCore Equations | success rate (%): 92.5% -> 76.5% | Hurt |
| use only local negatives (ConPoLe-local) instead of all beam-search negatives (ablation) | ConPoLe · CommonCore Fractions | success rate (%): 96.0% -> 86.0% | Hurt |
| use only local negatives instead of all negatives (ablation) | ConPoLe · CommonCore Sorting | success rate (%): 100.0% -> 100.0% | No clear effect |
| use only local negatives instead of all negatives (ablation) | CVI · CommonCore | success rate | No clear effect |
| replace 1-step lookahead with bounded BFS (ablation) | ADI and DAVI · CommonCore | success rate | No clear effect |
| train with contrastive examples (ConPoLe) instead of behavioral cloning | ConPoLe · CommonCore Equations | success rate (%): 4.5% -> 92.5% | Improved |
| train with contrastive examples (ConPoLe) instead of DRRN | ConPoLe · CommonCore Equations | success rate (%): 2.5% -> 92.5% | Improved |
| use ConPoLe representations instead of edit distance | 1-NN classifier · Khan Academy equation categories | accuracy: 0.714 -> 0.905 | Improved |
Closest papers by meaning
- NFT: Bridging Supervised Learning and Reinforcement Learning in Math Reasoning (ICLR 2026 · similarity 0.87)
- AceReason-Nemotron: Advancing Math and Code Reasoning through Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- PUZZLES: A Benchmark for Neural Algorithmic Reasoning (NeurIPS 2024 · similarity 0.86)
- BlendRL: A Framework for Merging Symbolic and Neural Policy Learning (ICLR 2025 · similarity 0.86)
- DeepMath-103K: A Large-Scale, Challenging, Decontaminated, and Verifiable Mathematical Dataset for Advancing Reasoning (ICLR 2026 · similarity 0.86)
- RLKD: Distilling LLMs’ Reasoning via Reinforcement Learning (AAAI 2026 · similarity 0.85)
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.
