sxiong/DeepControl-Qwen2.5-3B
DeepControl-Qwen2.5-3B is a 3.1 billion parameter Qwen2.5-3B-Instruct checkpoint developed by Siheng Xiong, optimized as a deep search agent. This model is specifically designed for search-augmented large language model reasoning, as detailed in the EMNLP 2026 paper "Adaptive Information Control for Search-Augmented LLM Reasoning." It excels at tasks requiring adaptive information control and retrieval, leveraging a 32768 token context length for enhanced search capabilities.
Loading preview...
DeepControl-Qwen2.5-3B Overview
DeepControl-Qwen2.5-3B is a specialized 3.1 billion parameter model derived from Qwen2.5-3B-Instruct, developed by Siheng Xiong. Its core innovation lies in its function as a deep search agent, designed to enhance reasoning in large language models through adaptive information control. This model is the checkpoint from the research presented in the paper "Adaptive Information Control for Search-Augmented LLM Reasoning" (EMNLP 2026).
Key Capabilities
- Search-Augmented Reasoning: Optimized for tasks that require dynamic and adaptive retrieval of information to support complex reasoning processes.
- Adaptive Information Control: Implements mechanisms for intelligent control over information retrieval, making it suitable for scenarios where precise data access is crucial.
- High Context Length: Supports a context length of 32768 tokens, enabling the processing of extensive search results and complex queries.
- Research-Backed Design: Developed and validated within a research framework, focusing on improving LLM performance in search-intensive applications.
Recommended Usage
This model is particularly well-suited for applications requiring a sophisticated search agent, such as advanced question-answering systems, research assistants, or any scenario where an LLM needs to intelligently interact with and leverage external knowledge bases. It is designed to be integrated with a retrieval system, specifically utilizing a wiki18 corpus and requiring a custom DeepControl system prompt and action parser for full functionality. Recommended inference settings include BF16 dtype, a context length of 8,192 for inference, and specific parameters for generation temperature, maximum new tokens per action, and retrieval top-k.