More is Less: The Pitfalls of Multi-Model Synthetic Preference Data in DPO Safety Alignment
Paper • 2504.02193 • Published • 1
Bolian Li
lblaoke
AI & ML interests
None yet
Recent Activity
authored a paper about 3 hours ago
More is Less: The Pitfalls of Multi-Model Synthetic Preference Data in
DPO Safety Alignment authored a paper about 3 hours ago
DRIFT: Learning from Abundant User Dissatisfaction in Real-World
Preference Learning authored a paper about 3 hours ago
Learning Self-Correction in Vision-Language Models via Rollout AugmentationOrganizations
authored 5
papers about 3 hours ago
DRIFT: Learning from Abundant User Dissatisfaction in Real-World Preference Learning
Paper • 2510.02341 • Published • 4
Learning Self-Correction in Vision-Language Models via Rollout Augmentation
Paper • 2602.08503 • Published • 3
Why Reasoning Fails to Plan: A Planning-Centric Analysis of Long-Horizon Decision Making in LLM Agents
Paper • 2601.22311 • Published
Addressing Performance Saturation for LLM RL via Precise Entropy Curve Control
Paper • 2604.26326 • Published • 11
submitted a
paper to Daily Papers about 14 hours ago
authored 8
papers 12 months ago
ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time
Paper • 2410.06625 • Published • 1
Bayesian Computation in Deep Learning
Paper • 2502.18300 • Published
Cascade Reward Sampling for Efficient Decoding-Time Alignment
Paper • 2406.16306 • Published • 1
Entropy-MCMC: Sampling from Flat Basins with Ease
Paper • 2310.05401 • Published
Long-tailed Classification from a Bayesian-decision-theory Perspective
Paper • 2303.06075 • Published
Trustworthy Long-Tailed Classification
Paper • 2111.09030 • Published
Graph Communal Contrastive Learning
Paper • 2110.14863 • Published
Identifying Incorrect Classifications with Balanced Uncertainty
Paper • 2110.08030 • Published