Research map / Optimization / Optimizers
Stochastic Structured Prediction under Bandit Feedback
Artem Sokolov, Julia Kreutzer, Stefan Riezler, Christopher Lo
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 pairwise preference learning with binary feedback instead of expected-loss or cross-entropy learning | Bandit structured prediction on SMT · SMT test | BLEU: 0.267 -> 0.273; 0.271 -> 0.273 | Improved |
| Use pairwise preference learning with binary feedback rather than expected-loss or cross-entropy learning | Bandit structured prediction on SMT · SMT | iterations to best development score: 370k -> 115k; 281k -> 115k | Improved |
| Use expected-loss minimization rather than pairwise preference or cross-entropy learning | Bandit structured prediction on chunking · Chunking test | F1: 0.914 -> 0.923; 0.891 -> 0.923 | Improved |
| Use pairwise preference learning with continuous feedback rather than expected-loss or cross-entropy learning | Bandit structured prediction on chunking · Chunking | iterations to best development score: 7.5M -> 4.7M; 5.9M -> 4.7M | Improved |
| Use pairwise preference learning instead of cross-entropy minimization | Bandit structured prediction · SMT | squared stochastic gradient norm: 3.04$_{\pm 0.02}$ -> 2.88e-4$_{\pm 3.40e-6}$ (PR(bin)); 3.04$_{\pm 0.02}$ -> 1.03e-8$_{\pm 2.91e-10}$ (PR(cont)) | Improved |
| Use pairwise preference learning instead of cross-entropy minimization | Bandit structured prediction · Chunking | squared stochastic gradient norm: 4.20$_{\pm 0.71}$ -> 7.71e-4$_{\pm 2.53e-4}$ (PR(bin)); 4.20$_{\pm 0.71}$ -> 5.99e-3$_{\pm 7.24e-4}$ (PR(cont)) | Improved |
| Use continuous rather than binary pairwise feedback (ablation) | Pairwise preference learning · SMT | squared stochastic gradient norm: 2.88e-4$_{\pm 3.40e-6}$ -> 1.03e-8$_{\pm 2.91e-10}$ | Improved |
| Use continuous rather than binary pairwise feedback (ablation) | Pairwise preference learning · Chunking | squared stochastic gradient norm: 7.71e-4$_{\pm 2.53e-4}$ -> 5.99e-3$_{\pm 7.24e-4}$ | Hurt |
Closest papers by meaning
- Bandit and Delayed Feedback in Online Structured Prediction (NeurIPS 2025 · similarity 0.90)
- Small steps no more: Global convergence of stochastic gradient bandits for arbitrary learning rates (NeurIPS 2024 · similarity 0.86)
- Bandits with Preference Feedback: A Stackelberg Game Perspective (NeurIPS 2024 · similarity 0.86)
- Randomized Confidence Bounds for Stochastic Partial Monitoring (ICML 2024 · similarity 0.85)
- Non-stationary Bandit Convex Optimization: A Comprehensive Study (NeurIPS 2025 · similarity 0.85)
- Performative Prediction with Bandit Feedback: Learning through Reparameterization (ICML 2024 · 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.
