🚨 New Paper Alert! 🚨 Thrilled to share our latest work: GRACE: Generative Representation Learning via Contrastive Policy Optimization 🧠📄✨ GRACE transforms LLMs from opaque encoders into transparent, interpretable representation learners. 🚀
3
1
4
Replies
Contrastive learning has long treated LLMs as black-box encoders — suppressing their reasoning and generation abilities. GRACE flips this paradigm: instead of minimizing contrastive losses, we treat them as rewards that guide a generative policy! 🤯
0
0
0
💡 Key Innovations: 🧩 Contrastive-as-Reward: Turns similarity objectives into policy-gradient rewards for LLMs. 🗣️ Rationale-Generating Policy: The model explains why texts are similar — producing interpretable, human-readable reasoning traces. 🎯 Reinforcement-Driven Embeddings
0
0
0