Research map / Large language models / Alignment and preferences
Personalized Reward Learning with Interaction-Grounded Learning (IGL)
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 |
|---|---|---|---|
| Use three latent reward states instead of two (ablation) | IGL-P · Production image recommendation | Dislikes: [1.141, 1.337, 1.557] -> [0.751, 1.072, 1.274] | Helped |
| Use three latent reward states instead of two (ablation) | IGL-P · Production image recommendation | Likes: [0.914, 0.949, 0.988] -> [0.985, 1.029, 1.054] | Helped |
| Use three latent reward states instead of two (ablation) | IGL-P · Three-state recommendation simulator | user satisfaction | Helped |
| Use IGL-P(2) rather than IGL-P(3) (ablation) | IGL-P · Two-state recommendation simulator | user satisfaction | No clear effect |
| Add a definitely-negative feedback oracle to disambiguate negative reward events (ablation) | IGL-P(3) · Three-state recommendation simulator | user satisfaction | Helped |
| Use IGL-P(3) instead of CB-Click | Production image recommendation · Production image recommendation | Dislikes: [1.234, 1.311, 1.348] -> [0.890, 0.918, 0.955] | Helped |
| Use IGL-P(3) instead of CB-Click | Production image recommendation · Production image recommendation | Clicks: [0.968, 0.979, 0.990] -> [1.000, 1.010, 1.020] | Helped |
| Optimize comments as the reward instead of personalized latent satisfaction | CB-comment · Facebook news recommendation | user dissatisfaction | Mixed |
| Personalize reward learning across reader communication styles | IGL-P(3) · Facebook news recommendation | recommendation performance across reader groups | Helped |
| Optimize emoji reactions as a fixed reward | CB-emoji · Facebook news recommendation | negative reward | Hurt |
| Reduce the frequency of definitely-negative dislike feedback (ablation) | IGL-P(3) · Three-state recommendation simulator | convergence speed | Hurt |
| Use IGL-P(2) when the underlying environment has three latent reward states (ablation) | IGL-P · Three-state recommendation simulator | user satisfaction | Hurt |
Closest papers by meaning
- Provably Efficient Interactive-Grounded Learning with Personalized Reward (NeurIPS 2024 · similarity 0.90)
- Capturing Individual Human Preferences with Reward Features (NeurIPS 2025 · similarity 0.85)
- MTRec: Learning to Align with User Preferences via Mental Reward Models (NeurIPS 2025 · similarity 0.85)
- PersonalLLM: Tailoring LLMs to Individual Preferences (ICLR 2025 · similarity 0.85)
- An Information Theoretic Approach to Interaction-Grounded Learning (ICML 2024 · similarity 0.85)
- Enhancing Personalized Multi-Turn Dialogue with Curiosity Reward (NeurIPS 2025 · similarity 0.84)
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.