misalignment
Misalignment refers to the situation where the objectives of an AI system do not correspond to the goals or preferences of its stakeholders, potentially leading to unintended consequences or undesirable outcomes.
- AdvEDM: Fine-grained Adversarial Attack against VLM-based Embodied Agents
- CTRL-ALT-DECEIT Sabotage Evaluations for Automated AI R&D
- Emergent Risk Awareness in Rational Agents under Resource Constraints
- Enhancing CLIP Robustness via Cross-Modality Alignment
- Gradient-Weight Alignment as a Train-Time Proxy for Generalization in Classification Tasks
- Image Stitching in Adverse Condition: A Bidirectional-Consistency Learning Framework and Benchmark
- One Prompt Fits All: Universal Graph Adaptation for Pretrained Models
- Preference Learning with Lie Detectors can Induce Honesty or Evasion
- Rethinking Fine-Tuning when Scaling Test-Time Compute: Limiting Confidence Improves Mathematical Reasoning
- Strategyproof Reinforcement Learning from Human Feedback