Similar Items: Beyond Negative Rollouts: Positive-Only Policy Optimization with Implicit Negative Gradients
- Unintended Negative Impacts of Promotional Language in Patent Evaluation
- Implicit Representations of Grammaticality in Language Models
- Mitigating Misalignment Contagion by Steering with Implicit Traits
- DGPO: Beyond Pairwise Preferences with Directional Consistent Groupwise Optimization
- Latent-GRPO: Group Relative Policy Optimization for Latent Reasoning
- Beyond Decodability: Reconstructing Language Model Representations with an Encoding Probe