Research map / Reinforcement learning / RL from feedback
Learning to Generalize from Sparse and Underspecified Rewards
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Add IML-collected diverse successful trajectories to initialize the MAPO memory buffer (ablation) | MAPO (MAPOX) · WikiTableQuestions and WikiSQL | test accuracy: 42.9 ± 0.5 -> 43.3 ± 0.4; 72.4 ± 0.3 -> 74.2 ± 0.4 | Helped |
| Add meta-learned auxiliary reward to MAPOX (ablation) | MAPOX · WikiTableQuestions and WikiSQL | test accuracy: 43.3 ± 0.4 -> 44.1 ± 0.2; 74.2 ± 0.4 -> 74.8 ± 0.2 | Helped |
| Add Bayesian-optimized auxiliary reward to MAPOX (ablation) | MAPOX · WikiTableQuestions | test accuracy: 43.3 ± 0.4 -> 43.8 ± 0.2 | Helped |
| Add Bayesian-optimized auxiliary reward to MAPOX (ablation) | MAPOX · WikiSQL | test accuracy: 74.2 ± 0.4 -> 74.2 ± 0.2 | No clear effect |
| Use meta-learned auxiliary reward instead of underspecified reward alone (ablation) | MAPO · Instruction following, test | accuracy: 69.8 ± 2.5 -> 74.5 ± 2.5 | Helped |
| Use Bayesian-optimized auxiliary reward instead of underspecified reward alone (ablation) | MAPO · Instruction following, test | accuracy: 69.8 ± 2.5 -> 72.3 ± 2.2 | Helped |
| Replace end-to-end MeRL with post-hoc reranking using trajectory features (ablation) | MAPOX on WikiTableQuestions · WikiTableQuestions | test accuracy: -3.0% | Hurt |
| Replace end-to-end MeRL with post-hoc reranking using policy probability and reward features (ablation) | MAPOX on WikiTableQuestions · WikiTableQuestions | test accuracy: +0.2% | No clear effect |
| Use end-to-end MeRL instead of post-hoc reranking (ablation) | MAPOX on WikiTableQuestions · WikiTableQuestions | test accuracy: +0.8% | Helped |
| Use MeRL instead of MAPO (ablation) | WikiSQL models ensembled over five runs · WikiSQL | test accuracy: 74.2 -> 76.9 | Helped |
Closest papers by meaning
- DISCOVER: Automated Curricula for Sparse-Reward Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- Dense Reward for Free in Reinforcement Learning from Human Feedback (ICML 2024 · similarity 0.88)
- Learning Goal-Conditioned Representations for Language Reward Models (NeurIPS 2024 · similarity 0.88)
- Learning Process Rewards via Visitation Matching for Efficient RL (NeurIPS 2026 · similarity 0.88)
- Regularizing Hidden States Enables Learning Generalizable Reward Model for LLMs (NeurIPS 2024 · similarity 0.87)
- GRAM: A Generative Foundation Reward Model for Reward Generalization (ICML 2025 · similarity 0.87)
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.