✨Steer2Adapt: Dynamically Composing Steering Vectors Elicits Efficient Adaptation of LLMs

📝 Summary:
STEER2ADAPT adapts LLMs by composing steering vectors from reusable semantic prior subspaces. This lightweight framework dynamically combines basis vectors, offering efficient and flexible adaptation for complex tasks without learning new vectors. It achieves an average performance improvement of...

🔹 Publication Date: Published on Feb 7

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.07276
• PDF: https://arxiv.org/pdf/2602.07276

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#LLM #AI #MachineLearning #ModelAdaptation #SteeringVectors

✨BidirLM: From Text to Omnimodal Bidirectional Encoders by Adapting and Composing Causal LLMs

📝 Summary:
BidirLM adapts causal LLMs into bidirectional encoders, overcoming catastrophic forgetting and integrating specialized models. It employs a prior masking phase, weight merging, and data mixture, outperforming alternatives on text, vision, and audio benchmarks.

🔹 Publication Date: Published on Apr 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.02045
• PDF: https://arxiv.org/pdf/2604.02045

🔹 Models citing this paper:
• https://huggingface.co/BidirLM/BidirLM-Omni-2.5B-Embedding
• https://huggingface.co/BidirLM/BidirLM-0.6B-Embedding
• https://huggingface.co/BidirLM/BidirLM-1.7B-Embedding

✨ Datasets citing this paper:
• https://huggingface.co/datasets/BidirLM/BidirLM-Contrastive

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#LLM #MultimodalAI #DeepLearning #AIResearch #ModelAdaptation